Meteosat Second Generation (MSG) Cloud Mask, Cloud Property Determination and Rainfall Comparison with In-situ Observations

Transcription

1 Meteosat Second Generation (MSG) Cloud Mask, Cloud Property Determination and Rainfall Comparison with In-situ Observations Peter Silla Masika March, 2007

2

3 Meteosat Second Generation (MSG) Cloud Mask, Cloud Property Determination and Rainfall Comparison with In-situ Observations By Peter Silla Masika Thesis submitted to the International Institute for Geo-information Science and Earth Observation in partial fulfilment of the requirements for the degree of Master of Science in Geo-information Science and Earth Observation in Water Resources and Environmental Management Programme Specialisation: Advanced use of Remote Sensing in Water Resources Management, Irrigation and Drainage Thesis Assessment Board Prof. Dr. Ir. Z. Su Dr. Ir. M. Booij Dr. B. H. P. Maathuis Dr. T. H. M. Rientjes Chairman (ITC, Enschede) External Examiner (Twente University, Enschede) Primary Supervisor (ITC, Enschede) Member (ITC, Enschede) INTERNATIONAL INSTITUTE FOR GEO-INFORMATION SCIENCE AND EARTH OBSERVATION ENSCHEDE, THE NETHERLANDS

4 Disclaimer This document describes work undertaken as part of a programme of study at the International Institute for Geo-information Science and Earth Observation. All views and opinions expressed therein remain the sole responsibility of the author, and do not necessarily represent those of the institute.

5 Dedicated to my wife, daughter and departed soul of my son & To my parents

6

7 Abstract To obtain accurate estimates of surface and cloud parameters from satellite data an algorithm has to be developed which identifies cloud-free and cloud-contaminated pixels. Data from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board Meteosat Second Generation (MSG) satellites have been available since February The data is accessible to National Meteorological and Hydrological Services (NMHSs). Unfortunately for NMHSs from developing countries the data until recently has never been exhaustively exploited for rainfall estimation, one very important variable in the atmosphere. Developed countries through research institutions have to some extent set in place ways and means of exploiting the MSG data that has been made possible through data distribution policy of EUMETSAT (free access to the data for research and education). This study attempts to utilize available MSG data for developing simple cloud mask and height algorithms and thereafter compare and determine the relationship between cloud height and observed rainfall on a ground station. A multispectral threshold technique has been used: the test sequence depends on solar illumination conditions and geographical location whereas most thresholds used here were empirically determined and applied to each individual pixel to determine whether that pixel is cloud-free or cloud-contaminated. The study starts from the premise of an acceptable trade-off between calculation speed and accuracy in the output data. For this reason, only three infrared channels of MSG satellite were used alongside climatological data provided by National Oceanographic and Atmospheric Administration (NOAA) and also land surface climatological data available from the WorldClim website. The accurate measurement of spatial and temporal variation of tropical rainfall around the globe remains one of the critical unresolved problems in the field of meteorology. This study attempted to compare computed cloud height and observed rainfall on ground station (CGIS-Butare, Rwanda) and derived cloud height-total rainfall relationship from storms over the same station. Results from the simple cloud mask algorithm were validated using EUMETSAT cloud mask products for a tropical region ( 11 N - 14 S and 6-51 E) over Africa. Overall accuracy of the simple cloud mask developed here was found to be 87% for four scenes which were during day- and nighttime as well as twilight time as defined by sun elevation angles. Analysis of recorded rainfall at CGIS and comparison of the same with computed cloud height showed that rainfall mainly occurred when cloud heights were greater than 3000m. Further, deriving a relationship between the observed rainfall and the cloud height was found to follow a Gaussian model in which clouds at approximate heights between 4000m and 5000m produced higher amounts of rainfall. Below and above this height range, rainfall amounts were found to be generally low. The derived cloud height-total rainfall relationship was applied to other storms over this station. Initial results show low correlation between estimated and observed rainfall. More synoptic observations have to be used to evaluate the derived relationship. Next to this a better procedure to differentiate nimbostratus and cumulonimbus has to be incorporated. Different relations between height and observed rainfall for the two types of clouds may be derived which may improve the overall results. KEY WORDS: MSG-SEVIRI, cloud mask, cloud height/type, rainfall comparison/estimation. i

8 Acknowledgements First and foremost I am grateful to the government of the Netherlands for providing me the scholarship under the Netherlands Fellowship Programme to pursue the M.Sc. course in Water Resources and Environmental Management in this unique institution, ITC. I am equally grateful to my organization, Kenya Meteorological Department for allowing me to fulfil my long-term dream of attaining M.Sc. in Remote Sensing-related course. My greatest gratitude goes to my supervisor, Dr. Ben H.P. Maathuis for his critical comments and inputs. I wish to say your support was tremendous. Great thanks to all WREM staff especially Prof. Z. Zu, Dr. A. Gieske, Dr. T. Rientjes, and Dr. R. Becht, for their valuable critical comments in this study. This study could not have been accomplished without EUMETSAT s favourable data distribution policy for research and education. Special thanks go to this great organization and equally appreciate their efforts in promoting satellite meteorology. Yet again many thanks go to all ever dedicated staff members of WREM department at ITC for imparting this valuable knowledge during that past 18 months. Equally thanks to the Programme Director, Dr. Arno van Lieshout for his excellent assistance and cooperation. I would like to appreciate MSG laboratory staff, specifically Mr. G. Reinink, for his tireless assistance in retrieving satellite data. I am grateful to all WREM 2006 course mates especially Essayas, Jose, Beyene, Edna, Anoja, Irena, Marie, Musefa, and Mohammad for their continuous support and friendship bestowed on me during that one and half years. Special thanks go to all my friends with whom I shared my days in Enschede for eighteen months. Fellow Kenyans cannot be left out for their encouragement during that period was great. To my dear wife, Nancy and our lovely daughter, Faith thanks for your patience, prayers and daily expectation of seeing me again. I sincerely owe you alot for all these. To my caring parents, my brothers and sisters, and all friends in Kenya, your support and prayers cannot fail to be appreciated too. ii

9 List of Acronyms A/MSU ANN APOLLO A/TOVS AVHRR BSC BTH CCD CCS CGIS CPC CST DEM DMSP EUMETcast EUMETSAT GAC GHCC GIS GOES GRIB HIRLAM HIRS HRPT HRV IAPP IDL ILWIS ITCZ ITPP KLAROS KNMI LAC MPE MSG NOAA Advanced/Microwave Sound Unit Artificial Neural Networks AVHRR Processing Scheme Over clouds, Land and Ocean Advanced/Tiros-N Operational Vertical Sounder Advanced Very High Resolution Radiometer Bi-spectral Spatial Coherence Bi-spectral Threshold and Height Cold Cloud Duration Cloud Classification System Geographic Information Systems and Remote Sensing Regional Outreach Centre in Butare, Rwanda Climate Prediction Centre Convective Stratiform Technique Digital Elevation Model Defence Meteorological Satellite Program European Organisation for the Exploitation of Meteorological Satellites Broadcast System for Environmental Data European Organisation for the Exploitation of Meteorological Satellites Global Area Coverage Global Hydrology and Climate Centre Geographic Information Systems Geostationary Operational Environmental Satellite General Regularly-distributed Information in Binary form (GRIded Binary) High Resolution Limited Area Model High-resolution Interferometer Sounder High Resolution Picture Transmission High Resolution Visible International Advanced/Tiros-N (Television Infrared Observation Satellite-Next generation) Operational Vertical Sounder (A/TVOS) Processing Package Interactive Data Language Integrated Land and Water Information System Inter-tropical Convergence Zone International Tiros-N (Television Infrared Observation Satellite-Next generation) Operational Vertical Sounder (TVOS) Processing Package Royal Netherlands Meteorological Institute (KNMI) Local APOLLO Retrievals in an Operational System Royal Netherlands Meteorological Institute Local Area Coverage Multi-sensor Precipitation Estimate Meteosat Second Generation National Oceanographic and Atmospheric Administration iii

10 NOAA-HL NODC NMHS NWP NWS PERSIAN SAFNWC SCM SCH/T SEVIRI SSM/I SST SYNOP TAMSAT TIP-data TRMM USGS WMO National Oceanographic and Atmospheric Administration (NOAA) High Latitude National Oceanographic Data Centre of the National Oceanographic and Atmospheric Administration (NOAA) National Meteorological and Hydrological Service Numerical Weather Prediction National Weather Service Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks Satellite Application Facility for supporting Nowcasting and very short range forecasting Simple Cloud Mask Simple Cloud Height/Type Spinning Enhanced Visible and Infrared Imager Special Sensor Microwave Imager Sea Surface Temperature World Meteorological Organization synoptic code for weather observations Tropical Applications of Meteorology using SATellite TIROS-N (Television Infrared Observation Satellite-Next generation) Information Processor data Tropical Rainfall Measuring Mission United States Geological Survey World Meteorological Organization iv

11 Table of contents Abstract.. i Acknowledgements ii Table of contents... iii List of figures v List of tables. vii List of Acronyms... viii 1. Introduction Background Significance of the Study Research Objectives Research questions General Specific Research Hypothesis Scope of Study Logical Sequence of Research Approach / Methodology Outline of the Thesis Literature Review Introduction Meteosat Second Generation (MSG) Satellite Cloud Masking Météo-France (SAFNWC) Cloud Mask Météo-France (Ocean and Sea Ice SAF) Cloud Mask Meteosat VIS-IR and NOAA-A/TOVS Image fusion Cloud Mask KNMI Cloud Mask Algorithm KLAROS Cloud Mask Algorithm APOLLO Cloud Mask GHCC Cloud Mask AFWA Cloud Mask Algorithm Precipitation Processes Satellite Rainfall Estimation Cloud-Indexing Methods Bi-spectral Methods Life-history Methods Cloud Model-based Techniques Blending Techniques EUMETSAT Multi-sensor Precipitation Estimate (MPE) Materials and Methods Data Acquisition...27 v

12 MSG Satellite Data Climatological Data Dew Point Temperature Synoptic Data and Field work Cloud Masking Method Rainfall Estimation Method Data Processing and Results MSG Satellite Images Generation of MSG Satellite and Solar Angles Day-time Cloud Mask Night-time Cloud Mask Twilight Cloud Mask Rainfall Estimation (A case of CGIS Weather station) Direct Comparison of Cloud Height and Rainfall Intensity Direct Comparison of Cloud Height and Total Rainfall Discussions of Results Cloud Mask Results Cloud Height/Type Results Rainfall Estimation Results Conclusions and Recommendations Conclusions Recommendations References Appendices Appendix A: ILWIS Script for Simple Cloud Mask and Height Algorithms Appendix B: Samples of Batch Files Appendix C: Sample of CGIS Weather Station Data Appendix D: Storms over CGIS Weather Station used for Developing Cloud height Rainfall Intensity Regression Function Appendix E: Storms over CGIS Weather Station used for Developing Cloud height Total Rainfall Regression Function Appendix F: Sample of Rain gauge (tipping bucket) Rainfall Data (Nairobi- Dagoretti Meteorological Station) vi

13 List of figures Figure 1-1: MSG/SEVIRI image of 5 th July 2006 at 12:00 UTC as a false Colour Composite...4 Figure 1-2: Three major phases and important steps in the methodology...5 Figure 2-1: Cloud types grouped into different families according to height range and form (Source: Strahler, 1965)...8 Figure 2-2: MSG image false colour composite (BGR) of 25 th December 2006 at 12:00 UTC...10 Figure 2-3: MSG cloud mask for 25 th December 2006 at 12:00 UTC (EUMETSAT, 2006)...11 Figure 2-4: GOES-East and MSG satellites SST products (in C) for 25 th Dec 2006 at 13:00 UTC...14 Figure 2-5: An example of pixel array under consideration with T min at (i,j)...23 Figure 2-6: Some factors influencing the differences between space- and time-collocated TMI and SSMI...24 Figure 2-7: The PERSIAN CCS model structure (source:(hong et al., 2004a))...25 Figure 3-1: Flow chart for simple cloud mask (SCM) and cloud height/type (SCH/T) retrieval...27 Figure 3-2: MSG Data Retriever window (Courtesy of ITC)...28 Figure 3-3: False colour composite (bands 1, 2, and 3 in BGR) (left) and Band 9 (10.8µm) (right) on 19/12/2006 at 12:00 UTC...29 Figure 3-4: Climatological Temperature (in K) images; (a) day-time (b) night-time (c) mean, of Africa and part of Atlantic Ocean for the month of May...30 Figure 3-5: Schematic view of temperature lapse rates in an idealized convective cloud...32 Figure 3-6: Locations of the four stations shown on MSG satellite false colour composite image...33 Figure 3-7: (a) Image acquisition by SEVIRI radiometer, and (b) Schematic diagram on MSG satellite and Rain gauge observation time...35 Figure 4-1: Flow chart for generating MSG satellite and Sun angles...39 Figure 4-2: Sun (for 26 th December 2006 at 15:00 UTC) and MSG Satellite (0 N and 0 E)...40 Figure 4-3: Solar illumination conditions on 26 th December 2006 at 15:00 UTC...40 Figure 4-4: Description of the test sequences for Land surface (left) and Sea surface (right)...41 Figure 4-5: Description of test sequence for land surface (left) and sea surface (right)...43 Figure 4-6: Description of test sequence for land surface (left) and sea surface (right)...44 Figure 4-7: Cloud masks for (a) day-time, (b) twilight time, and (c) night-time; for MSG-1 image of 7 th March 2006 at 15:30 UTC Figure 4-8: Solar illumination conditions on 7 th March 2006 at 15:30 UTC...45 Figure 4-9: Cloud mask (a) and false colour composite (b) for MSG image of 07/03/2006 at 15:30 UTC...46 Figure 4-10: Cloud height (in Meters) image (MSG image of 07/03/2006 at 15:30 UTC)...47 Figure 4-11: Classified cloud height image of 07/03/2006 at 15:30 UTC...47 Figure 4-12: Segments (yellow lines) of cloud mask of 23/11/2005 at 13:30 UTC overlaid on False colour composite (VIS006, VIS008, and NIR016) in (BGR)...48 Figure 4-13: Diurnal cloud height and Rainfall intensity changes on (a) 5 th May 2006, and (b) 10 th May Figure 4-14: Rainfall intensities within cloud height classes...50 Figure 4-15: Gaussian model fit, X = Average cloud height (m), Y = Average rainfall intensity (mm/hr)...51 vii

14 Figure 4-16: Observed and estimated rainfall intensity for different storms Figure 4-17: Diurnal cloud height and Total rainfall changes on (a) 5 th May 2006, and (b) 10 th May Figure 4-18: Gaussian model fit, X= Average storm height (m), Y= Total rainfall (mm) Figure 4-19: Observed and Estimated total rainfall plotted with the error bars Figure 4-20: Relationship between the observed and the estimated total rainfall Figure 5-1: Flow chart on segmentation and visualization of EUMETSAT CLM and SCM Figure 5-2: EUMETSAT cloud mask assigned feature classes for 26 th December 2006 at 15:00 UTC Figure 5-3: Cloud mask segments of EUMETSAT CLM (yellow lines) and SCM (red lines) for 26 th December 2006 at 15:00 UTC, on a false colour composite Figure 5-4: EUMETSAT CTH (a), SCH/T (b), and Difference (between CTH and SCH/T) (c) images for 25 th December 2006 at 11:45 UTC (height is in meters) Figure 5-5: Diurnal height and Total rainfall changes on 1 st March 2006 (left) and 28 th October 2006 (right) over Naivasha station viii

15 List of tables Table 2-1: Spectral channel characteristics of SEVIRI in terms of central, minimum and maximum wavelength of the channels and the main application areas of each channel (Source: (EUMETSAT, 2006))...9 Table 2-2: Definition of illumination conditions (SAFNWC); solar elevation is in degrees...12 Table 2-3: Test sequence over land (SAFNWC)...12 Table 2-4: Test sequence over sea (SAFNWC)...12 Table 3-1: Locations of the four stations within Eastern Africa...33 Table 4-1: Observed storms and their total amount of rainfall...53 Table 4-2: Storm heights and estimated total rainfall...54 Table 5-1: Contingency table for MSG image of 25 th December 2006 at 12:00 UTC...60 Table 5-2: Contingency table for MSG image of 26 th December 2006 at 15:00 UTC...60 Table 5-3: Contingency table for MSG image of 4 th January 2007 at 22:00 UTC...61 Table 5-4: Contingency table for MSG image of 10 th January 2007 at 17:00 UTC...61 ix

16

17 1. Introduction 1.1. Background For more than 40 years, meteorological satellites have been the best way to observe the changing weather on a large scale (EUMETSAT, 2006). Typically, operational meteorology utilizes two types of satellites, namely; polar orbiting and geostationary satellites, to provide the required information. Polar orbiting satellites fly at relatively low altitudes of approximately 800km above the earth surface and can provide information based on a high spatial resolution. Geostationary satellites, on the other hand, are in the equatorial plane and at high altitudes of about 36000km above the earth surface. Their revolution time is the same as that of the earth itself and therefore the satellites are always viewing the same area on the earth. They have low spatial resolution due to their altitudes. However, they can perform frequent imaging, in animated mode, which can depict the ever-changing atmospheric processes. The first generation of European meteorological satellites dates back to 1977, with the launch of Meteosat-1. Since then this series have advanced to Meteosat-7 which is currently located around 57 E and manoeuvring to replace Meteosat-5 located at 63 E. These series are followed by Meteosat Second Generation (MSG) satellites of which the first one (MSG-1 now Meteosat-8) was launched on 28 th August 2002 and became operational in early The second of this series (MSG-2 now Meteosat-9) was launched on 20 th December These two satellites are located at 0 N and 0 E. MSG satellites are spin-stabilized and capable of greatly enhanced earth observations (EUMETSAT, 2006). The Spinning Enhanced Visible and Infrared Imager (SEVIRI) sensor on board MSG has a high temporal resolution of 15 minutes and spatial resolution of 3 km (sub-satellite) for all channels except 1km for high resolution visible (HRV) channel. The major improvement for this series of satellites is the enhanced spectral resolution of 12 channels. The presence of the 3.9µm channel in the current sensor has allowed analyses of cloud cover especially at night-time. The primary mission of MSG satellites is the continuous observation of earth s full disk with a multispectral imager. The repeat cycle of 15 minutes for full-disk imaging provides multi-spectral observations of rapidly changing phenomena such as deep convection. They also provide better retrieval of wind fields which are obtained from the tracking of clouds, water vapour and ozone features. In this study, main attention is given to cloud properties, such as cloud height, that may be associated with rainfall amounts observed a ground station. Presence and characteristics of clouds gives information about the state of the atmosphere. For many cloudy situations, the reflected visible radiation and the emitted thermal radiation are not simple to interpret because the cloud is not the only reflecting/radiating source (Dlhopolsky and Feijt, 2001). Of 1

18 importance is to determine cloud properties by first distinguishing cloud-free pixels from cloudcontaminated pixels. Quantitative data sets obtained from the cloud-contaminated pixels have several potential applications one of which is for water resources and environmental management. In general, effective integrated water resources management requires timely, accurate and comprehensive meteorological, hydrological and other related information. Use of satellites in observing variables such as rainfall, evaporation and soil moisture has enhanced provision of these data in a timely and effective manner for the water resources management sector. These meteorological variables needs to be monitored effectively and since they are associated with atmospheric moisture hence clouds, there is need to identify the clouds first through masking all cloud-contaminated areas in satellite images. Cloud masking allows identifying cloud-free areas where other products such as land or sea surface temperatures may be computed. It also allows identifying cloudy areas where other products (e.g. cloud types and cloud top temperature/height) may be derived. Cloud type on the other hand provides a detailed cloud analysis. It may be used as input to an objective meso-scale analysis which in turn may be used in a simple nowcasting scheme (Météo-France, 2005b). Cloud type product is essential for generation of cloud top temperature and height products and for identification of precipitating clouds which in turn may be used to estimate rainfall intensity/amount Significance of the Study For a considerable long time, series of precipitation amounts are recorded worldwide. Such amounts, mainly expressed in millimetres (mm) and collected during a day or an hour are not only useful for general meteorological and climatological practices, but are of special interest for hydrology and agricultural meteorology. Surface-based observations of precipitation is accomplished primarily by gauges and, where economically viable, by radars. Over the world s oceans these measurements are often done on buoyancies which are few worldwide. On the other hand, over the land areas the coverage from surface observations is not uniform. Worse still, ground measurements from the conventional rain gauges have deteriorated over the last couple of years and thus an alternative is being sought to continue providing precipitation measurements not only on spatial basis but also on higher temporal scale. The field of remote sensing has advanced and through various meteorological satellites precipitation estimation has been made possible to reasonable scales, for instance, 3km (spatial resolution) and 15 minutes (temporal resolution) for MSG satellites. However, satellites measure cloud properties (e.g. brightness temperature) an important product that provides crucial information that can be used to infer rainfall intensity and/or total rainfall. Understanding these properties, and the crucial information that can directly or indirectly be used in water resources management, is important. Thus there is need for determining cloud type/height using readily available MSG satellite data in order to estimate rainfall. (Maathuis et al., 2006) showed that MSG retriever software developed at ITC can be used to retrieve MSG data and to estimate rainfall over the entire MSG field of view which covers the whole of Africa and part of Europe. The results in the study showed bias for low and high intensity rainfall amounts due to different types of clouds. 2

19 Further, potential applications of developing cloud masking and cloud type algorithm are many. Some of the most important ones nowadays are, in operational weather forecasting and in energy and water balance studies. Clouds represent the most significant source of error in the extraction of earth surface energy and water balance parameters out of meteorological satellite data (Valk et al., 1998). Energy and water balance models are used to estimate fluxes in cloudy conditions. In order to develop an accurate energy balance mapping algorithms, cloud masking is essential. Cloud mask and type software modules have been developed by the Centre de Météologie Spatiale of Météo-France and are embedded in the Satellite Application Facility for supporting NoWCasting and very short range forecasting (SAFNWC)/MSG software package that is distributed by EUMETSAT (Derrien and Le Gléau, 2005). These cloud mask and type algorithms uses transfer functions derived from atmospheric models which are not published. Most of National Hydrological Services, especially in Africa, have no access to these transfer functions and even then may not be in a position to derive, on their own, the transfer functions. Besides, due to financial limitations for most of these National Hydrological Services, shareware or freeware (such as ILWIS) can be used for masking clouds and determining the cloud type through semi-automated processing. Thus there is a need for masking out clouds and determining their basic properties in order to improve forecasted rainfall estimates from MSG. It is envisaged that improving rainfall estimation will assist most of National Hydrological Services to provide information on the status of water resources within their area of jurisdiction. It is also expected that it would further improve timely decision making for, areas prone to disasters related to weather such as floods, landslides or areas frequently affected by droughts. This therefore calls for a need to develop simple cloud mask (SCM) and cloud type/height (SCH/T) algorithms which may be embedded in readily available shareware such as ILWIS Research Objectives This study addressed the following two main objectives; Determination of cloud mask and cloud height/type on all daily MSG images; and Relating derived cloud height with rainfall at a ground rainfall station. Specifically the study focused on: Developing simple cloud mask and height algorithms; Analyzing relationship between MSG cloud height images and observed rainfall intensity and/or total rainfall at a ground station; and Using derived relation to estimate total rainfall from storms at various heights Research questions General Can simple cloud mask (SCM) and cloud height/type (SCH/T) algorithms be developed, using ancillary input data from general climatology databases, and applied to MSG images covering the whole of Africa? 3

20 Specific a) Can the cloud height from the masked cloud at every moment be used to determine the cloud type? b) Is there any relationship between the rainfall intensity and/ or total rainfall and the cloud height/type? c) Can the cloud mask and height algorithms developed be able to process automatically MSG satellite data as they are received from EUMETSAT through EUMETCast every 15 minutes? 1.5. Research Hypothesis The study set the following hypotheses: The smaller the number of cloud forms appear in the atmosphere, the easier it becomes to identify them in satellite imagery; The more complex the cloud mask algorithm structure becomes, the better it performs; and Satellite retrieved cloud properties can be related to rainfall observed on the earth surface Scope of Study The study was conducted on the MSG field of view which covers the whole of Africa ( 39 N - 38 S and 34 W - 53 E). MSG images as in figure 1-1 were subjected to the developed cloud mask and cloud height algorithms. However, for validation of the results a small portion ( 11 N - 14 S and 6 E - 51 E) of the view was considered. Rainfall data from Geographic Information Systems and Remote Sensing Regional Outreach Centre (CGIS), Rwanda was used. Also for the same purpose in-situ rainfall data was collected from Nairobi-Dagoretti, Kisumu, and Naivasha in Kenya. Figure 1-1: MSG/SEVIRI image of 5 th July 2006 at 12:00 UTC as a false Colour Composite (BGR) of bands 1, 2, and 3 (EUMETSAT, 2006) 4

21 1.7. Logical Sequence of Research Approach / Methodology The methodology in this study consists of three phases namely; pre-field work, field work campaign, and post field work as shown in figure 1-2. Literature review Retrieving MSG data Pre-field work Phase Retrieving land surface temperature Retrieving sea surface temperature Developing cloud mask and height algorithms Field work Phase Rainfall data collection Post field work Phase Rainfall data analysis MSG image processing with developed algorithm Comparing cloud height and total rainfall and finding statistically significant relation between total storm rainfall and cloud height Validation of the derived relation Rainfall Estimation Figure 1-2: Three major phases and important steps in the methodology 5

22 1.8. Outline of the Thesis The thesis consists of six chapters. Chapter 1 is the current chapter within which this section is contained. As has been noted, this chapter briefly introduces the study, outlining its justification, the objectives, and the scope of the study. General approach to carry out the study has also been shown. Chapter 2 contains literature review on various methodologies adopted for cloud mask algorithm. A few of these algorithms are presented in this chapter. Also presented here is an overview of satellite rainfall estimation methods. Chapter 3 provides general approach to this current study with details of data requirement and acquisition. Sources of various data are pointed out in this chapter. Steps undertaken to process some of the data are explained. Chapter 4 elaborates on the data processing and presents the results on various stages in the study. Chapter 5 provides detailed analysis of various results and discussions attached to these results. Chapter 6 finally presents conclusions and recommendations drawn from the study. 6

23 2. Literature Review 2.1. Introduction Condensation or deposition of water above the earth s surface creates clouds which develop in an air mass that becomes saturated. The air mass may have passed over warm bodies of water, or over wet surfaces and carried upward by turbulence or convection. Saturation occurs by way of atmospheric mechanisms that the temperature of the air mass is cooled to its dew point. The lifting required for cooling and condensing this water vapour results from several processes, and study of these processes provides a key for understanding the distribution of rainfall in various parts of the world. The major mechanisms or processes that occur causing cloud development include: A) Orographic uplift: - Occurs where air is forced to rise because of the physical presence of elevated land. As the air parcel rises, it cools as a result of adiabatic expansion at a rate approximately 1 C per 100m until saturation. Beyond saturation level the parcel rises moist adiabatically at a rate of 0.6 C per 100m (Strahler, 1965). B) Convectional lifting: - Associated with surface heating of the air at the ground surface. Once enough heating occurs, the air mass becomes warmer and lighter than the surrounding environment and then rise expanding and cooling. When sufficient cooling takes place, saturation occurs forming clouds. This is the common phenomena within tropics forming cumulus cloud and or cumulonimbus clouds (thunderstorms). C) Convergence or Frontal lifting: - Takes place when two air masses come together of which in most cases they have different temperature and moisture characteristics. In frontal lifting, one of the air masses is usually warm and moist while the other is cold and dry. The leading edge of the cold dry air acts as an inclined wall or front causing the moist warm air to be lifted thereby cooling and saturation is finally reached. This is common phenomena in mid-latitudes whereas near the equator winds from both northern and southern hemisphere meet at the Intertropical Convergence Zone (ITCZ) and lifting, cooling, saturation of the air mass occur forming clouds. D) Radiative cooling:- Occurs when the sun no longer supplies the earth surface or water body surface and overlying air with energy derived from solar insolation (e.g. at night). Here the surface of the earth now begins to lose energy in the form of longwave radiation which cools the ground and air above it. Clouds that result from this type of cooling take the form of surface fog. Basically clouds consist of extremely tiny droplets of water ( 0.02 to 0.06mm) in diameter, or minute crystals of ice. Clouds appear white when thin or when sun shines on the outer surface. When dense and thick, clouds appear grey or dark underneath. The presence and movement of clouds is often the only clue that indicates a significant meteorological process occurring in the atmosphere. They indicate the presence of moisture and some type of cooling mechanism. 7

24 Cloud types may be classified on the basis of two characteristics: general form and altitude (Strahler, 1965). There are two major forms of clouds namely; stratiform or layered types and cumiliform. Stratiform are further subdivided according to the level of elevation at which they lie. Clouds lie within high (cirrus and its related forms), middle (alto-stratus and alto-cumulus) and low (stratus, nimbostratus, and stratocumulus) levels. Vertically developed clouds mainly brought about by thermal convection or frontal lifting consists of fair weather clouds and cumulonimbus type of clouds. Figure 2-1 below shows different cloud types grouped into different families according to their altitudes. Figure 2-1: Cloud types grouped into different families according to height range and form (Source: Strahler, 1965) Based on cloud altitudes, three major families are classified namely; family A (high level clouds), family B (middle level clouds), and family C (low level clouds). From figure 2-1, the approximate altitudes of these families are shown. In addition and of high importance in precipitation occurrence is a fourth family type D, which includes clouds with vertical developments mainly due to convection. Within these families and particularly family C and D, there exist the multi-layered clouds which are those with higher depths which give precipitation hydrometers a better environment to develop and grow. Some of the clouds which exist as multi-layered are nimbostratus and cumulonimbus. Nimbostrati are considered multi-layered clouds because their vertical extent often goes well into the middle cloud region. These clouds are dark, usually overcast, and are associated with large areas of continuous precipitation. Cumulonimbuses on the hand are clouds that can produce lightning, thunder, heavy rains, hail, and strong winds. They are the tallest of all clouds that can span all cloud layers. They usually have large anvil-shaped tops which form due to strong winds at high levels of the atmosphere. Thus in the atmosphere the most relevant type of clouds in relation to contribution to precipitation are the multi-layered clouds. 8

25 2.2. Meteosat Second Generation (MSG) Satellite Meteosat Second Generation (MSG) satellites provide vital data for meteorology and climatology at frequent intervals and over wide areas. These series of satellites provides information for the entire African continent at much higher spatial and temporal resolutions as compared to the earlier meteorological satellite series. Coupled with these characteristics is the higher spectral resolution (12 bands/channels) of the SEVIRI instrument as provided in table 2-1. Table 2-1: Spectral channel characteristics of SEVIRI in terms of central, minimum and maximum wavelength of the channels and the main application areas of each channel (Source: (EUMETSAT, 2006)) Band Spectral Characteristics of Spatial Main observational No. Band spectral band Resolution application (µm) (µm) (km, Subsatellite) λ cen λ min λ max 1 VIS Surface, clouds, wind fields 2 VIS Surface, clouds, wind fields 3 NIR Surface, Cloud phase 4 IR Surface, clouds, wind fields 5 WV Water vapor, high level clouds, atmospheric instability 6 WV Water vapor, atmospheric instability 7 IR Surface, clouds, atmospheric instability 8 IR Ozone 9 IR Surface, clouds, wind fields, atmospheric instability 10 IR Surface, clouds, atmospheric instability 11 IR Cirrus cloud height, atmospheric instability 12 HRV Broadband ( ) 1 Surface, clouds One of the key objectives of setting the MSG programme by EUMETSAT was the extraction of meteorological and geophysical fields from satellite image data in support of general meteorological, climatological and environmental activities. Of importance among the products of MSG satellite is cloud information which as stated earlier gives information about the state of the atmosphere. In order to study the behaviour and properties of clouds, they must first be identified from the satellite image. 9

26 The following section briefly discusses a number of cloud extraction methods already developed by individuals and institutions in an attempt to identify all cloudy pixels in various satellite(s) images Cloud Masking Cloud detection from remote sensing data is required for many applications. Some of these are such as determination of cloud cover, identification of cloudy pixels for the retrieval of cloud-related parameters, or exclusion of pixels with even minor cloud contamination if further processing would be affected by the presence of clouds (Schröder et al., 2002). Several methods can be used to perform cloud detection. Some of these methods are such as multispectral thresholding techniques that can be applied to individual pixels (Saunders and Kriebel, 1988), (Derrien et al., 1993), (Stowe et al., 1999). Dynamic cloud cluster analysis relying on histogram analysis was suggested by (Desbois et al., 1982) whereas (Bankert, 1994) indicated use of artificial neural networks which needs manual training. Another approach was suggested by (Ebert, 1987) which involve pattern recognition techniques based on large scale texture analysis. Figure 2-2 is a false colour composite (NIR01.6, VIS0.8, and VIS0.6 as Blue, Green, and Red respectively) image of 25 th December 2006 at 12:00 UTC for a small portion of Eastern Africa continent, which indicates presence of various types of clouds and how they appear in MSG satellite image. Uganda Convection Kenya Low level clouds Rwanda Burundi Thin cirrus Tanzania Figure 2-2: MSG image false colour composite (BGR) of 25 th December 2006 at 12:00 UTC 10

27 Usually low level clouds such as stratus are difficult to identify on an infrared image but on a visible bands colour enhancement (and false colour composite) they appear white to grayish. They appear layered and as semi-transparent. Deep convective clouds on a false colour composite appear cyan in colour and with sharp edges i.e. with distinct boundaries with the rest of nearby clouds. On the other hand high level clouds such as thin cirrus appear cyan in colour and feather-like in pattern. These features can be seen on figure 2-2 as indicated for each type of these clouds. Therefore clouds in such an image could be masked for various studies, one of which is for rainfall estimation. Methods as enumerated in the following sections were an attempt to detect and mask all clouds in various satellite images including MSG. Some of these methods proved to have some disadvantages. For instance, (Dlhopolsky and Feijt, 2001), indicated that the histogram analysis method is time consuming and in many cases cannot make accurate threshold without human interaction. Various authors and institutions have developed cloud mask algorithms for detection of clouds. The most relevant algorithms are explained in the proceeding sections Météo-France (SAFNWC) Cloud Mask Satellite Application Facility for supporting NoWCasting (SAFNWC), within Météo-France and run by a consortium of institutions namely; Spanish Meteorological Institute (INM), Météo-France, the Swedish Meteorological Institute, and the Austrian Meteorological Institute, is tasked to develop and maintain a software package allowing the extraction from MSG/SEVIRI imagery a set of 12 products useful for nowcasting purposes on any user defined area in the MSG (Derrien and Le Gléau, 2005). In the software, cloud mask and cloud type software modules are implemented. The cloud mask algorithm is based on multispectral threshold technique applied to each pixel of the image. Figure 2-3 below shows an example of cloud mask as developed by SAFNWC and accessed from EUMETSAT through EUMETCast. Figure 2-3: MSG cloud mask for 25 th December 2006 at 12:00 UTC (EUMETSAT, 2006) 11

28 The cloud mask algorithm developed here follow series of tests, the first being to identify pixels contaminated by clouds or snow/ice and applied to land or sea which depend on the solar illumination and on the viewing angles as defined in the table 2-2 with the tests presented in tables 2-3 and 2-4. Most of the thresholds are determined from satellite-dependent look-up tables and NWP forecast fields data and also from ancillary data (elevation and climatological data). Some thresholds are computed at a spatial resolution defined by the user and others are empirical constants. The second process is applied to all pixels, even already classified cloud-free or contaminated pixels, to identify dust clouds and volcanic ash clouds. Spatial filtering is applied to reclassify isolated pixels having a class type different from their neighbours. Table 2-2: Definition of illumination conditions (SAFNWC); solar elevation is in degrees Night-time Twilight Day-time Sunglint Solar elevation<-3-3<solar elevation<10 10<Solar elevation Solar elevation>15 Table 2-3: Test sequence over land (SAFNWC) Day-time Twilight Night-time Snow detection Snow detection T10.8 T10.8 T10.8 T10.8-T12.0 R0.6 R0.6 T8.7-T10.8 T10.8-T12.0 T10.8-T12.0 T10.8-T8.7 T8.7-T10.8 T8.7-T10.8 T10.8-T3.9 T10.8-T3.9 T10.8-T8.7 T3.9-T10.8 T3.9-T10.8 T10.8-T3.9 Local spatial texture Local spatial texture T3.9-T10.8 T8.7-T3.9 Local spatial texture T8.7-T3.9 Table 2-4: Test sequence over sea (SAFNWC) Day-time Sunglint Twilight Night-time Ice detection Ice detection Ice detection SST SST SST SST T10.8-T12.0 R0.8(R0.6) T10.8-T12.0 R(0.8)R0.6 T8.7-T10.8 R1.6 T8.7-T10.8 T10.8-T12.0 T10.8-T3.9 T10.8-T12.0 Local spatial texture T8.7-T10.8 T12.0-T3.9 T8.7-T10.8 R0.8(R0.6) T10.8-T8.7 T3.9-T10.8 T10.8-T3.9 T10.8-T3.9 T10.8-T3.9 Local spatial texture T3.9-T10.8 Low clouds in sunglint T3.9-T10.8 Local spatial texture Local spatial texture T8.7-T3.9 12

29 Some of the notable tests are such as those using IR10.8 and IR12.0 (here in the table indicated as T10.8 and T12.0 respectively) in which over the sea a pixel is flagged as cloudy when its estimated sea surface temperature (SST) value is lower than a monthly climatological SST value by 4K. The sea surface temperatures are estimated from T10.8 and T12.0 brightness temperatures using a non-linear split window algorithm (Le Borgne et al., 2003). If this test is not applied, IR10.8 is used in which case the threshold is determined from surface temperatures forecast by using Numerical Weather Prediction (NWP) model. This test is applied over both the land and the sea surfaces. More importantly is the fact that the threshold is derived from a global Pathfinder night-time bulk SST climatology covering a period of 10 years and available at a 1/9 th degree (approximately 12 km) horizontal resolution. Over land, IR8.7 together with IR10.8 is used in which the difference (IR10.8-IR8.7) should be greater than /cos (θ sat ), during night-time or in case of low sun elevation e.g. at twilight, for any pixel to be flagged cloud-contaminated. θ sat is the satellite zenith angle. Usually, low clouds are characterized at night-time by high IR10.8-IR3.9 brightness temperature differences, which allows their identification over land (Derrien and Le Gléau, 2005). This detection may be less efficient at large viewing angles hence the need to use a different channel (here IR8.7). An empirical test ( /cos (θ sat )) has been developed based on the observation that decrease of IR8.7- IR10.8 with the satellite zenith angle is much stronger for low clouds than for vegetated areas. However, during day-time the empirical test threshold is greater ( *(1/cos (θ sat )-1); where 1/cos (θ sat ) is the secant of the satellite zenith angle, for any pixel to be flagged cloudy. Most of the other test thresholds provided in tables 2-3 and 2-4 are computed from simulation of the surface (ocean, land or snow) top of atmosphere reflectance (for the visible and near infrared bands) by adding an offset and a correction factor. Top of atmosphere reflectance is simulated as: R toa ( 1 a R ) = a (2.1) 0 + a1 * Rsurf / 2 * surf ) Where: a 0, a 1, and a 2 are coefficients computed from satellite and solar angles, water vapour and ozone content using look-up tables. R surf is the land, ocean or snow surface reflectance. Offsets of various percentages and correction factors are added to the above expression. However, they are not described further in the literature. Hence this makes this method difficult to apply. The dynamic thresholds applied to thermal bands differences are obtained by interpolation into lookup tables using the satellite zenith angles and the NWP forecast using radiative transfer models. It is no doubt that the tests are many and therefore require special software to handle. In addition, many African National Meteorological and Hydrological Services (NMHSs) have no capability to handle numerical weather prediction (NWP) model forecasts using radiative transfer models Météo-France (Ocean and Sea Ice SAF) Cloud Mask The developments of the Ocean and Sea Ice Satellite Application Facility (O&SI-SAF) Sea surface temperature algorithms of Météo-France for the determination of Atlantic sea surface temperature require cloud masking. The SST products have three components namely; GOES-East, MSG and 13

30 NOAA-HL derived SSTs (Météo-France, 2005a). In all these components cloud mask algorithm applied is the same and is based on multispectral thresholding technique. Specific adaptation to marine conditions was introduced in the development of the algorithm. These conditions mainly consider temporal stability of SST and climatology. An example of combination of cloud-free GOES- East and MSG satellites SST product in MSG georeference for 25 th December 2006 at 13:00 UTC is provided in figure 2-4 below. Figure 2-4: GOES-East and MSG satellites SST products (in C) for 25 th Dec 2006 at 13:00 UTC Temporal stability of SST was suggested by (Wu et al., 1999) and is applied in Ocean and Sea Ice SAF under the following form: for a clear sky pixel, channel 11µm (IR10.8) temperatures at time H (T11H) are compared to the maximum value of the corresponding temperatures at time H-30 minutes (T11H 1 ) and time H+30 minutes (T11H 2 ). If T11H-Max (T11H 1, T11H 2 ) <Threshold, the pixel is considered as cloudy (Météo-France, 2005a). Threshold set here as suggested by (Wu et al., 1999) is - 0.5K. The process tries to address change of temperature over a pixel. A negative change implies setting in of cloud in that pixel which is in most cases colder that the pixel temperature within a time range (1 hour). This process implies that all pixels real time state will not be determined since it requires the future (next 30 minutes) state. Furthermore, the comparison process may take some time. In considering climatology, the climatologic minimum temperature at any time of the year in question is compared with calculated SST value. Too low SST is indicative of cloud contamination and too low threshold depends on the distance of the considered pixel to the pre-calculated cloud mask and the location of the pixel with respect to the coast. Here it is assumed that near a cloud, too cold temperatures are more suspect and the control of the calculated SST against climatology should be more severe. The scheme works as follows: If T smin -T s > t, the pixel is considered as cloudy where: T smin is the climatologic minimum temperature T s is the calculated SST 14

31 t is the threshold SST is calculated from a non-linear algorithm in the form of expression given below: T S = guess ) ( A + A S) T ( B + B S + B T )( T10.8 T C + C S Corr (2.2) where: A 0, A 1, B 0, B 1, B 2, C 0, and C 1 are constants for which in case of MSG are: , 0, 0, , , , and 0 respectively S = sec(θ)-1, with θ: satellite zenith angle T guess is the climatological SST Corr is correction factor and is 0.2 for MSG Climatologic minimum temperature, T smin is derived from the Pathfinder archive (AVHRR data from 1985 to 1995). The climatology has been made on a decadal (10 day) basis and includes minimum and mean values re-mapped over the MSG disk at same resolution as the thermal infrared bands in MSG. Various thresholds used depend on the position of the pixel with respect to the coast hence it applies only over the sea. The method appears less complicated as compared to the previous one and therefore easy for implementation Meteosat VIS-IR and NOAA-A/TOVS Image fusion Cloud Mask (Casanova et al., 2005) developed an automatic method of cloud classification for direct application in civil aviation. Here visible and infrared channels of Meteosat satellite were used alongside data provided by the A/TOVS (Advanced/Tiros-N Operational Vertical Sounder) onboard NOAA polar satellites. Different spectral techniques were used for different purposes. In their study, an automatic method of cloud classification which provided, in real time, the cloud cover over civil airports on the Iberian Peninsula was developed emphasizing on rain clouds. The method consisted of a series of algorithms based on the physical properties of cloud surfaces and thermodynamic state of the atmosphere. TIP-data included in the telemetry of the high resolution picture transmission (HRPT) satellites were taken and processed through International TOVS Processing Package (ITPP) or International ATOVS Processing Package (IAPP) software, depending on whether the datum type was TOVS or A/TOVS, in order to transform the high-resolution interferometer sounder (HIRS) and advanced/microwave sound unit (A/MSU) sensors radiances into atmospheric data. In this method, albedo classification was performed which revealed that most clouds were good reflectors since it depends first on their thickness and to some extent on the nature of the cloud particles. To avoid sunglint contamination, illumination geometric conditions reflectance thresholds were set. From here reflectivity image was obtained which was further reclassified according to various categories based on surface type. Threshold tests were performed by comparing historic data of mean temperatures at ground level to the calculated brightness temperature values. This allowed detection of non-cloudy pixels which were to be removed before further processing in order to obtain the linear relationship between height and cloud top temperature. This was done by applying the temperature value to the equation obtained through the geopotential and temperature images provided by the A/TOVS data calculated at different pressure levels. 15

32 This method can be seen to use thermodynamic sounding of the atmosphere which in many cases is not available at many meteorological weather stations. Many African National Hydrological Services are also not able to handle the atmospheric sounding due to the cost involved KNMI Cloud Mask Algorithm The Royal Netherlands Meteorological Institute (KNMI) has developed an algorithm for cloud detection and characterization called MetClock (METeosat CLOud Characterisation KNMI). According to (Valk et al., 1998), MetClock algorithm comprises two threshold tests to perform cloud detection on the basis of Meteosat IR data, the relative and absolute infrared tests. A cloudy pixel is determined by comparing Meteosat apparent brightness temperatures with the earth surface temperatures. The relative infrared test uses images of different observation times to compare changes in temperature at the earth surface with changes in temperature of a pixel. A pixel is classified as cloudy when the change in pixel temperature exceeds the change of the earth surface temperature with a certain threshold. On the other hand absolute infrared test uses a single image to directly compare pixel temperatures with the earth surface temperatures. Here again a pixel is classified as cloudy when it exceeds a certain threshold. From these two tests it is clear that the results depend strongly on the accuracy of the surface temperature maps. The surface temperatures are provided by Numerical Weather Prediction (NWP) model, the High Resolution Limited Area Model (HIRLAM). The relative test group consists of five tests, thresholding on the IR imagery and VIS imagery between various time difference including previous imageries and also future imageries either in hours or days. This is the shortcoming of this particular cloud classification method since it takes into account forecast environment which may not be always correct. Besides, surface temperature over complex terrain and high mountains may not be accurate and therefore the method may not be very much applicable in such areas. The algorithm was developed for the first series of Meteosat satellites (Meteosat 1-7) and can be as well applied to MSG satellites KLAROS Cloud Mask Algorithm In their study (Dlhopolsky and Feijt, 2001) developed KNMI Local implementation of APOLLO retrievals in an Operational System (KLAROS) algorithm for processing MSG data. AVHRR data was used as prototype data set with which to produce cloud products expected to be derived with MSG. KLAROS is more or less the same as MetClock. However this method was aimed at improving cloud detection by using a radiative transfer model to help link radiances from the different wavelengths and produce physically meaningful variables. The method was designed to work with use of thresholds defined in databases. The temperature database is derived from the HIRLAM NWP model while reflectivity database was created from two years of NOAA AVHRR data clear skies and verified with synoptic observations. 16

33 It is further stated that KLAROS consists of a set of programs for cloud detection and cloud property retrieval. The two modes of operation are one in C tool which does the data processing and the other is user interface written with the Interactive Data Language (IDL) APOLLO Cloud Mask AVHRR Processing scheme Over clouds, Land and Ocean (APOLLO) was designed to make use of all five spectral channels during day-time and to discretize all AVHRR data into four different groups called cloud-free, fully cloudy, partially cloudy, and snow/ice, before deriving physical properties (Saunders and Kriebel, 1988). Within APOLLO, clouds are discretized into three layers according to their top temperature (Kriebel et al., 2003). The layer boundaries are set to 700hPa and 400hPa and the associated temperatures are derived from standard atmospheres. Each cloudy pixel is checked to see whether it is thick or thin cloud, depending on its channel 4 and 5 temperatures and, during day-time, its channel 1 and 2 reflectances. Thin clouds are taken as ice clouds, i.e. cirrus, whereas thick clouds are treated as water clouds. APOLLO is designed to process AVHRR HRPT (High Resolution Picture Transmission) data as well as Local Area Coverage (LAC) and Global Area Coverage (GAC) data. Those pixels in which the solar elevation is more than 5 above the horizon are processed by means of the day-time algorithm whereas all others are processed by the night-time algorithm. APOLLO uses five threshold tests applied to each pixel and this allows establishing the group of cloud-free and contaminated pixels. These tests are based on AVHRR channels 1, 2, 4 and 5 and rely on simple physical principles. Every pixel which is brighter than a threshold in the solar channels or colder than another threshold in the thermal channels is called cloudy. The use of physical parameters and self-adjusting thresholds minimizes the influence of differences between the instruments aboard different satellites. However, since AVHRR is polar-orbiting it is not to establish meaningful time series of cloud products. Thus the thresholds needs to be carefully selected in order to establish whether a pixel is cloudy or is partially cloudy or is cloud-free GHCC Cloud Mask The Global Hydrology and Climate Centre (GHCC) in Huntsville, Albama receives GOES-East and West satellite data in real-time from their ground stations and produces a number of products from the Imager and Sounder in support of research and operational activities. Cloud detection method used here is bi-spectral spatial coherence (BSC) which uses two spatial tests and one spectral threshold to identify clouds in the GOES Imager or Sounder imagery (Jedlovec and Laws, 2003). The performance of the BSC method is adequate during the day; however it performs poorly near sunrise/sunset and at night. Bi-spectral Threshold and Height (BTH) which is built on BSC uses spatially and temporally varying thresholds. This method also provides cloud top pressure information with the cloud mask. The underlying principle in this cloud detection method with GOES imagery is that the emissivity difference of clouds at 10.7µm and 3.9µm varies from that of the surface (land or ocean) and can be detected from channel brightness temperature differences. (Jedlovec and Laws, 2003) noted that while 17

34 emissivity of clouds at 3.9µm is considerably less than at 10.7µm, reflected solar radiation at 3.9µm makes effective brightness temperatures (sum of emission and reflective components) quite large. The key to cloud detection in the BTH technique is the use of multispectral channel differences to contrast clear and cloudy regions. The 10.7µm and 3.9µm channels on the Imager and similar channels on the Sounder are used to produce an hourly difference image (longwave minus shortwave) for this purpose (Jedlovec and Laws, 2003). Two composite images are created for each hour which represent the smallest negative and smallest positive difference image values (values closest to zero) from the preceding 20 day period (for each time). These composite images serve to provide spatially and temporally varying thresholds for the BTH method. An additional 20 day composite image is generated for each hour using the warmest longwave (10.7µm) brightness temperature for each location from the 20 day period. This composite image is assumed to represent a warm cloud-free thermal image for each time period. BTH method uses the images generated as enumerated above in a four step cloud detection procedure. These include testing adjacent pixels, one-dimensional spatial variability (which fills-in between the cloud edges), minimum difference (which compares the current difference image value to the composite images), and Infrared threshold (which uses an hourly 20-day composite of the warmest 10.7µm channel values at each pixel location. It is no doubt that this method requires high memory space for storing the images generated. It also implies that a lot of iterations have to be done every time a comparison and composite images are to be generated AFWA Cloud Mask Algorithm Kidder et al., (2005) developed various Meteosat Second Generation (MSG-1) cloud mask algorithms for implementation at the United States Air Force Weather Agency (AFWA). These algorithms are named; cloud mask, nocturnal cloud mask, daytime cirrus, nocturnal thin cirrus, precipitating clouds, and multi-channel skin temperature. Cloud mask algorithm uses difference-from-background technique by constructing 10-day infrared background for each hour of the day. The process assumes that in 10 days each pixel is observed to be cloud-free at least once. This method exploits the tendency of clouds being colder than the underlying surface. Pixels whose radiance is less than the background radiance by more than a threshold value are flagged as cloudy. Here 8.7µm channel is used in constructing the 10 days background image. Nocturnal cloud mask test uses the brightness temperature at 10.8µm and the albedo at 3.9µm to detect ice clouds, liquid water clouds, and clear scenes. 8.7µm channel data are used to screen desert pixels. Various empirical thresholds are used in the test. Albedo and brightness temperature background database contains the 3.9µm albedo and 10.8µm brightness temperature data observed over MSG-1 pixel each day for the previous 10-day period. Day-time cirrus test for MSG-1 data utilizes three reflective channels, 0.6µm, 0.8µm, and 1.6µm. The measured radiances are converted to albedos (0 to 1) by dividing the radiances by the solar irradiance 18

35 and by the cosine of the solar zenith angle, then multiplying by phi. This test utilizes the fact that liquid water clouds are highly reflective at all three wavelengths and thus will appear white. Ice clouds (and snow on the ground) are highly reflective at 0.6µm and 0.8µm, but poorly reflective at 1.6µm and therefore they appear cyan in colour. When the albedos are represented in the RGB colour cube, cirrus is having a cyan colour due to ice particles. Nocturnal thin cirrus test uses albedo at 3.9µm, which is calculated from measured radiances at 3.9µm and 10.8µm (Kidder et al., 2005). Radiation from below the cloud leaks through thin cirrus, which results in a negative albedo. They indicated that nothing else results in a negative 3.9µm albedo, hence this is a very sensitive test for thin cirrus at night. Precipitating clouds test uses the brightness temperature difference between the 6.2µm water vapour channel and the 10.8µm window infrared channel to detect high, thick clouds, which are likely to be precipitating. As clouds are usually formed within troposphere, the maximum water vapour content is also expected within this level. Casonova et al., (2005) indicated that in an adiabatic atmosphere, the relation between the height and temperature within troposphere is linear. This linear relation can be obtained by use of A/TOVS data for pressure levels between 1000 hpa (approximately sea level) and 300 hpa (approximately tropopause). It is for this reason that WV06.2 is appropriate to use with IR10.8 in order to extract the most probable raining clouds. Kidder et al., (2005) indicated that at 6.2µm the atmosphere is opaque due to water vapour absorption and that low clouds are not sensed at 6.2µm. Only deep clouds penetrate the water vapour to be sensed at both 6.2µm and 10.8µm, and when this happens then the brightness temperature difference at these two wavelengths becomes small (Kidder et al., 2005). Empirically determined threshold temperature difference of 11K is used to flag out cloudy pixels, (i.e. if WV06.2-IR10.8 < 11K, then the pixel is rain cloud). Multi-channel skin temperature procedure employs two channels (10.8µm and 12.0µm). Both 10.8µm and 12.0µm radiances are affected by water vapour in the column between the surface and the satellite, but 10.8µm is less affected than 12.0µm. The difference between the brightness temperatures at 10.8µm and 12.0µm can be used to correct the 10.8µm brightness temperature for water vapour absorption to yield an estimate of the skin temperature, that is, the brightness temperature which would be observed if there was no water vapour in the atmosphere (Jedlovec and Laws, 2003). The temperature retrieved here is that of the surface only and not the air temperature. However, the method applies to clear-sky pixels only. Most of the above cloud mask methods have not revealed the actual threshold values used. However, under the strength of the arguments in the theories used in these methods to develop the mask algorithm, a simple cloud mask (SCM) could be attempted which could further be used in other applications such as rainfall estimation, weather forecasting, water and energy balance studies, among others. 19

36 2.4. Precipitation Processes Precipitation is water in some form, falling out of the air, and settling on the surface of the earth. Precipitation occurs due to condensation in the atmosphere and not condensation that occurs at the surface such as dew. The major two models of precipitation formation are: collision-coalescence and ice crystal models. An important distinction between the two processes is the temperature of the cloud. Warm clouds are the ones whose mass lies above freezing level while cold clouds primarily exist where the temperature is below freezing level. The collision-coalescence model applies to warm clouds that form in the tropics whereas ice crystal model applies to the process of precipitation in the mid and high latitudes. For precipitation to form under collision-coalescence model there needs to be a variety of different size condensation nuclei. Large condensation nuclei will create large water droplets while smaller condensation nuclei create small ones. In the ice crystal model, cloud water exists in liquid form even though the temperatures are cold enough to freeze water. Water has a temperature below freezing but still in liquid state i.e. super-cooled water. The following section discusses various methods of estimating rainfall (one of the major form of precipitation) from space by use of satellites. Some of these methods attempt to address rain formation processes as explained in this section Satellite Rainfall Estimation The measurement of the surface precipitation is very important to studies of the hydrological cycle, water management planning, flash flood identification, input to hydrological and agricultural models, verification of weather modification experiments and the study of convective systems (Kamarianakis et al., 2006). Rainfall affects lives and economies of a majority of the earth s population. Heavy rain systems are crucial to sustaining the livelihood of many countries. Excess rainfall can cause floods, landslides and loss of property. Rainfall is among the atmospheric parameters, one of the most difficult to measure because of its high temporal and spatial variability and discontinuity. Moreover the coverage of precipitation measurements by ground conventional means (rain gauge networks or weather radars) is much less than adequate especially in the African continent. With the advent of meteorological satellites, improved identification and quantification of precipitation at time scales consistent with the nature and development of cloud has been realised (Levizzani et al., 2002). Meteorological satellites expand the coverage and time span of conventional ground-based rainfall data for a number of applications, above all hydrology and weather forecasting. The primary scope of satellite rainfall monitoring is to provide information on rainfall occurrence, amount and distribution over the globe for meteorology at all scales, climatology, hydrology, and environmental sciences (Levizzani et al., 2002). The uneven distribution of rain gauges and weather radars over most part in the world has called for this new technology of satellite rainfall measurement. 20

37 It is also a well know fact that precipitation is one of the most variable quantity in both space and time especially within the tropics. Geostationary weather satellite visible (VIS) and infrared (IR) imagers provide the rapid temporal update cycle needed to capture the growth and decay of precipitating clouds. The swath widths of satellites in tropical orbit such as the Tropical Rainfall Measuring Mission (TRMM) and of sensors in polar orbits like the Special Sensor Microwave Imager (SSM/I) series leave substantial gaps all over the globe (Levizzani et al., 2002). The quantitative rainfall determination from a variety of precipitating systems differ both dynamically and microphysically and this prompts for non-unique solutions based on physics of precipitation formation processes. Based on a number of earlier studies, (Levizzani et al., 2002) reviewed several satellite-based rainfall estimation methods. In this section a few of these satellite-based rainfall estimation methods that use thermal infrared are briefly explained. They revealed the four main categories of cloud classification as: cloud-indexing, bi-spectral, life-history, and cloud model-based. Each of these categories stresses a particular aspect of sensing cloud physics properties using satellite imagery and in final rainfall estimation Cloud-Indexing Methods According to (Kidder and Haar, 1995), cloud-indexing is the oldest precipitation estimation technique which assumes that it is fairly easy to identify cloud types in satellite imagery. This method assigns a rain rate to each cloud type. The rain at a particular location or in a particular area can be written as: R = r f i, (2.2) i i Where r i is the rain rate assigned to cloud type i, and f i is the fraction of time that the point is covered with (or fraction of the area covered by) cloud type i. This method lacks the validity of assigning a constant rain rate to a particular cloud type. Depending on the cloud formation processes, rain rate may vary significantly even for a particular type of cloud Bi-spectral Methods Bi-spectral methods are based on the very simple, although not always true, relationship between cold and bright clouds and high probability of precipitation, which is characteristic of cumulonimbus (Levizzani et al., 2002). Lower probabilities are associated to cold but dull clouds (thin cirrus) or bright but warm (stratus). Generally cirrus clouds are cold but do not produce as much precipitation as some warmer clouds. Techniques in this category are based on cloud classification and using either radar derived rainfall or good network of ground stations as training data. From here rainfall estimation over a given array of pixels can be derived. The underlying assumption here is that all rain bearing clouds are successfully classified, which is not always the case. 21

38 Life-history Methods These are the methods mainly in the family of techniques that specifically require geostationary satellites images. They rely upon a detailed analysis of the clouds life cycle, which is particularly relevant for convective clouds (Levizzani et al., 2002). One of the most famous techniques which use this method is Tropical Applications of Meteorology using SATellite (TAMSAT) of the Reading University, United Kingdom. The assumption inherent in the TAMSAT procedure is that the relationship between the quantity of rainfall and the cold cloud duration (CCD) is linear, provided there is adequate averaging of data either in space and time. The averaging procedure reduces the false alarms that CCD may cause especially over a short time period or over a very small area which is associated with convective clouds. (Dugdale and Milford, 1985) developed the concept of cold cloud duration (CCD), using the thermal channel of Meteosat, to generate time series of cloud temperature for tropical altitudes, where most rainfall comes from convective activities. They suggested that the duration above a certain threshold temperature value is representative of the amount of rain that is generated. However, according to (Grimes et al., 1999) the basic assumptions in this method are: 1) Rainfall is predominantly convective in origin and that the raining clouds can be identified as those with cloud top temperatures below a threshold temperature (T), 2) The number of hours for which a given pixel is colder than T (the CCD) is linearly related to the rainfall over the same time period, that is: R s = a0 + a1d + e (2.3) where: R s is the rainfall over the pixel, D is the CCD over the pixel, e is the error with zero mean, E[e] =0, and homogeneous variance Var[e]=ε 2 3) The threshold temperature, T and the parameters a 0, a 1, can be estimated for a given region and a given time of year by the analysis of historic data for that region and time of year, that is: = a + a D (2.4) R s 0 1 where: R s, a 0, and a 1 are estimated values. R s, a 0, and a 1 are calculated for each month for a number of empirically determined calibration zones. This method has proved to be successful in tropical regions especially for convective clouds occurring in the region of Inter-tropical Convergence Zone (ITCZ) in which case the first two assumptions are reasonable. However, due to inter-annual variability in rainfall-ccd relationship, there could be overor underestimation of rainfall for a particular month locality if fixed calibration is applied. 22

39 Cloud Model-based Techniques Cloud model-based techniques aim at introducing the cloud physics into the retrieval process for quantitative improvement deriving from the overall better physical description of the rain formation processes (Levizzani et al., 2002). (Adler and Mack, 1984) derived a one-dimensional cloud model by first identifying locations of convective cells and then assigning rain parameters. Associated anvil stratiform rainfall area is identified by a threshold brightness temperature, the value of which is calculated from the satellite data. The model calculates maximum rain rate and maximum volume rain rate from a sequence of model runs as a function of maximum cloud height or minimum cloud model temperature. The use of the one-dimensional cloud model to account for ambient temperature, moisture and shear conditions provides a stronger physical (less empirical) basis for the cloud height - rain relationships. (Adler and Negri, 1988) utilized data from Geostationary Operational Environmental Satellite (GOES) infrared ( µm) channel, in 30 minutes, for an area in the southern Florida for a onedimensional cloud model relating cloud top temperature to rain rate and rain area in the Convective Stratiform Technique (CST). The method here followed a few steps, first of which is to identify the candidate thunderstorm or regions of enhanced convection by searching for the minimum in the GOES temperature T b array. The second step involves eliminating local minima temperature that represent thin, non-precipitating cirrus by calculating slope parameter for each minimum temperature T min as: S = T 1 6 Tmin (2.2) where, T 1 6 is the average temperature of the six closest pixels. If the T min is located at (i,j), 1 = i 2, j i 1, j i+ 1, j i+ 2, j i, j+ 1 i, j T 6 ( T + T + T + T + T + T 1) / 6 (2.3) Note that due to pixel offset along the scan line which is approximately half as large as that across the scan, 6 pixels (highlighted in figure 2-5) are taken as the closest pixels to the one under consideration. i, j Figure 2-5: An example of pixel array under consideration with T min at (i,j) After this step empirical discrimination of thin cirrus, from active convection, in the temperature/slope plane using radar and visible imagery data is derived (Adler and Negri, 1988). Once this is done, correction for the field of view between the GOES field of 8km and that of the cloud model (approximately 1km in horizontal dimension) as used by (Adler and Mack, 1984). Rain parameters are then assigned to the feature based on one-dimensional cloud model and thereafter a threshold temperature is used to identify the anvil stratiform region. This threshold is expected to 23

40 coincide with the relatively thick portion of mature anvils hence modal T b in the frequency distribution of anvil T b is used as anvil background temperature. The above four methods/techniques have been widely applied in satellite rainfall estimation with different scientists endeavouring to optimally estimate rainfall over different space and time ranges. The following section briefly explains the blending techniques used to estimate rainfall over a given area Blending Techniques With the advent of passive microwave measurements, several VIS/IR techniques have been reexamined and integration sought that could help adjusting some of the well known problems of the top down approach of these methods, which generally derive precipitation only from cloud top information (Levizzani et al., 2002). These methods have suggested combining the observations delivered by satellite instruments of different type to improve averaged rainfall estimation by using multi-source data. Combining IR data from geosynchronous satellites attempt to take advantage of both IR and MW techniques. They benefit from the fact that there is excellent time and space coverage of IR images and from the direct connection of the MW observations with precipitation. However, these techniques require some precautions as (Turk et al., 2003) indicated that raining and idealized non-raining conditions as observed by low orbiting earth may cause discrepancies as a result of viewing geometries (see figure 2-6). This depends upon the three-dimensional structure of the cloud and the azimuthal direction that the sensor views it. Moreover, the timing and foot print offsets usually cause significant difference between geostationary satellites and polar orbiting satellites. Figure 2-6: Some factors influencing the differences between space- and time-collocated TMI and SSMI observations under idealized precipitating cloud conditions (Source: (Turk et al., 2003)) These methods often use statistical integration of the satellite IR and MW data. The choice of which variables to match in order to provide the final product may rest on one part the extent of accuracy required and on another the processing time. (Marzano et al., 2005) indicated a possibility of direct combination of microwave brightness temperatures and thermal infrared radiances, in which the 24

41 advantage is the exploitation of the observable information without any post-processing and the disadvantage of space-time collocation matching. Maathuis et al., (2006) applied blending technique using TRMM derived rainfall intensity to calibrate MSG image in which a relation between thermal infrared observations and the passive microwave observations was derived. In order to obtain good correlation between these two variables, as well as to solve collocation problems, they used equal temperature classes to average rainfall intensities. The regression function obtained here was applied to MSG images for regionalization of rainfall intensity over eastern part of Africa. Statistical methods in these techniques are significant and are applied to empirically-trained retrieval algorithms in order to estimate rainfall to a reasonable accuracy. However, the best approach would be based on physically-based retrieval algorithms which, on the other hand, would need a climatological and microphysical tuning. This again would resort to approaches whose aim would be cumulative estimates but not instantaneous estimates. Besides, most of the above methods consider rainfall events and not a specific type of cloud. Different clouds and different stages of development (e.g. for convective clouds) may have different cloud height- rainfall intensity/amount relationship. This calls for an approach that may address the problem of comparing rainfall and cloud height of different clouds that are at different stages of their development. Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIAN) Cloud Classification System (CCS) as described by (Hong et al., 2004a) addressed the problem of comparing different type of clouds by extracting cloud features from infrared geostationary satellite imagery. The PERSIAN algorithm fits the pixel brightness temperature and its neighbour temperature textures, in terms of means and standard deviations, to the calculated pixel rain rates based on an Artificial Neural Networks (ANN). The general approach in the algorithm is as provided in figure 2-7 below. Figure 2-7: The PERSIAN CCS model structure (source:(hong et al., 2004a)) 25

42 Firstly images are pre-processed through cloud segmentation procedure as in (a). Several inputs of feature extraction of the cloud patches are applied as in (b). Once cloud patches are identified cloud classification follows clustering them accordingly as in (c). The final step is to develop non-linear temperature and rainfall fitting for all classified cloud clusters as in (c). The parameters of the temperature rainfall curves are calibrated based on rain estimates from sources such as Radar networks over a region under study. This method offers probably the best idea in comparing cloud properties and rainfall intensity or rainfall amount. However, the enormous data requirement in this algorithm, which is unavailable in many regions, renders it inapplicable in many regions. Consequently a simple satellite rainfall estimation method is vital for regions with low passive microwave datasets. Despite low reliability results that may be obtained in using few satellite cloud data, rainfall estimation can be derived as first approximation for users such as hydrologists or general water management resource organisations EUMETSAT Multi-sensor Precipitation Estimate (MPE) The EUMETSAT multi-sensor precipitation estimate (MPE) has been developed in order to derive instantaneous rainfall intensities from MSG. The method is based on the blending of brightness temperatures of the MSG infrared channels with rainfall intensities from Special Sensor Microwave/Imager (SSM/I) on the United States Defence Meteorological Satellite Program (DMSP) satellites (Heinemann, 2003). The basic assumption of the Multi-sensor Precipitation Estimate (MPE) method is that colder clouds are more likely to produce precipitation than warmer clouds. Here Heinemann, (2003) pointed out that the relationship between the cloud top temperature and the surface rainfall intensity is non-linear and that it depends strongly on the current weather situation. Temporally and spatially co-registered SSM/I and MSG measurements are used to derive look-up tables which describe rainfall intensities as a function of the MSG infrared brightness temperature. The look-up tables are applied to MSG images in order to derive rainfall intensities in full spatial and temporal resolution. This method is indicated to efficiently estimate the spatial distribution and strength of convective precipitation over not only large scale tropical convection but also small scale convective processes and cold fronts. It is however not suitable for estimating precipitation from warm fronts and also orographically induced precipitation which is usually detected but miss-located to great distances, sometimes upto 100km (Heinemann, 2003). These products are available in (EUMETSAT, 2007) and can be downloaded in GRIB2 format. MPE products can be imported into any geographic information systems (GIS) packages such as ILWIS by using windows based GRIB2 import package which can be obtained from ftp://ftp.cpc.ncep.noaa.gov/wd51we/wgrib2/windows_xp/ (NOAA-NWS-CPC, 2005). The respective rainfall intensities can then be viewed in ILWIS. GRIB viewer from SatSignal software and available in (David, 2006) can as well be used to view the amounts at a desired location. The following chapter outlines materials and methods used in this study with detailed information on data acquisition required for cloud masking and also for rainfall comparison as well as rainfall estimation. 26

43 3. Materials and Methods 3.1. Data Acquisition The study attempted to use various data whose source was either straightforward to obtain or needed pre-processing. The general approach in developing the cloud mask is as given in figure 3-1. The proceeding sections explain how each of the retrieval and computation processes was done with ILWIS scripts in Appendix A. Retrieve data from MSG online archive Retrieve land Surface Temperature (LST) ( C) Retrieve Sea Surface Temperature (SST) ( C) VIS Range (VIS006, VIS008, VIS016) Daytime only TOA temperature (IR_039, IR_108, IR_120) Merge SST and LST Time series (K) Cloud mask algorithm Dew point temp calculation and set temp threshold(s) Perform Cloud Mask Not ok Visualize the cloud mask on Colour composite Ok Classify into different height classes Figure 3-1: Flow chart for simple cloud mask (SCM) and cloud height/type (SCH/T) retrieval MSG Satellite Data The Spinning Enhanced Visible and Infrared Imager (SEVIRI) sensor onboard MSG satellite provides data to EUMETSAT at Darmstadt (Germany) which is processed and then uplinked to HOTBIRD-6 in wavelet compressed format (Gieske et al., 2004). The images are received and archived at ITC in compressed form on external drives which are linked to the ITC network and hence accessed through ordinary personal computers. The image geocoding and radiometric calibration coefficients are supplied in so called EPI and PRO files. The data is not atmospherically corrected. Therefore direct ground observation(s) can only be related to the satellite observation(s) (at the required resolution) after atmospheric correction of the images. In this study this step is not necessary since the focus is on 27

44 clouds which are the main atmospheric parameters aimed at removing from satellite images for direct ground observation(s) relations. The retrieval of MSG data is straightforward using import utilities developed. External batch files were created using the MSG data Retriever software available at ITC. For more details about the software refer to (Maathuis et al., 2005). Figure 3-2 shows MSG data Retriever window with the command line indicating all parameters that are to be retrieved from MSG image. Figure 3-2: MSG Data Retriever window (Courtesy of ITC) The executing commands are saved as a text file and therefore any time a different image is required the changes are only done using a text editor and saved as a batch file in ILWIS software. These batch files are provided as Appendix B. The software has been used intensively in this study as the MSG geometric model is implemented. In this study, only bands 4, 9, and 10 (3.90µm, 10.8µm, and 12.0µm respectively) were used in the cloud masking process. Visible bands 1, 2, and 3 (0.06µm, 0.08µm, and 1.60µm respectively) were only for visualization in order to ensure optimal (visual) cloud mask validation especially during the day. Examples of raw images of band 9 (10.8µm) and visible bands 1, 2, and 3 (0.06µm, 0.08µm, and 1.60µm respectively), with the visible bands viewed as false colour composite, are provided in figure

45 Figure 3-3: False colour composite (bands 1, 2, and 3 in BGR) (left) and Band 9 (10.8µm) (right) on 19/12/2006 at 12:00 UTC On the left figure cyan areas are cloud patches and dark areas are water bodies. Right figure with pseudo colour representation shows colder areas in deep blue to green while warmer areas appear orange to red in colour Climatological Data Climatological data required in this study were minimum, maximum, and mean land surface temperature as well as sea surface temperature. Minimum, maximum, and mean land surface temperature were obtained from WorldClim database available for download from (Hijmans et al., 2005b). This dataset contains global climate grids with a spatial resolution of a square kilometre and can be used for mapping and spatial modelling in GIS (Hijmans et al., 2005a). Sea surface temperatures (SST) were derived from climatological data using the NOAA National Oceanographic Data Centre (NODC) and the University of Miami Rosenstiel School of Marine and Atmospheric Science (RSMAS) AVHRR Version 5.0 Pathfinder SST dataset available at ftp://data.nodc.noaa.gov/pub/data.nodc/pathfinder/version5.0_climatologies/ (NOAA- NODC, 2006) for the period 1985 to This averaged data was already resolved to 4km and in 5- day, 7-day, 8-day, monthly, seasonal, and annual periods and each period provided daytime-only, night-time-only, and day-night combined. Here day-night combined monthly mean sea surface temperature HDF file dataset was imported to ERDAS and then into ILWIS. The dataset provided needs to be rescaled and transformed to represent SST in degree Kelvin. The scale and offset provided are and -3 K respectively, so that the expression for calculation of SST appears as given in equation 3.1. SST ( SST * ) (3.1) K = orig where: SST K and SST orig are corrected SST (in K) and original SST (in C), respectively. 29

46 At this stage the land surface minimum, maximum, and mean temperatures given in monthly were merged with the generated mean monthly SST. The final images of climatological monthly day-time, night-time, and mean temperature of the entire globe were generated. These were used in computing the dew point temperature as well as performing cloud mask. Figures 3-5(a - c) shows these climatological images for the month of May for the African continent and part of Atlantic Ocean. The temperature is in K. (a) (b) (c) Figure 3-4: Climatological Temperature (in K) images; (a) day-time (b) night-time (c) mean, of Africa and part of Atlantic Ocean for the month of May Dew Point Temperature Dew point temperature is an important geophysical parameter that indicates the state of moisture content in the air under given conditions (Hubbard et al., 2003). This is the critical temperature at 30

47 which air is fully saturated and below which condensation normally occurs. Figure 3-5 shows schematic diagram of idealized atmosphere in which a convective cloud is developing vertically. In their study, Hubbard et al., (2003) presented a temperature-based (daily maximum, minimum, and mean) daily dew point temperature estimation method for historical studies in the Northern Great Plains in USA. They developed four regression-based methods incorporating daily maximum, minimum, and mean temperatures and also daily precipitation for different locations in the plains. After statistical analysis of the results obtained from the four methods, they concluded that the model that performed satisfactorily was in the form: T = α T ) + β ( T ) + γ ( T T ) + λ (3.2) d ( m n x n where: T d is the daily dew point temperature in C T m is the daily mean air temperature in C T n is the daily minimum air temperature in C T x is the daily maximum air temperature in C α, β, γ, and λ are coefficients of the regression equation. The method was further supported by the fact that the associated data set required are easily available in most typical meteorological weather stations. The model as Hubbard et al., (2003) pointed out can estimate dew point temperature with sufficient accuracy under varied climatic conditions. Moreover, they also indicated that the climatic conditions observed within the Northern Great Plains are representative of many other regions in the world. Based on the above statements, dew point temperature was therefore computed by use of the model as given in equation 3.2 and consequently equation 3.3 with all the associated coefficients was adopted. However, here use of the climatological monthly mean, maximum, and minimum temperatures was made instead of the daily temperatures. Thus monthly climatological dew point temperature was obtained as follows: T ( T ) ( T ) ( T T ) (3.3) d = m n x n where: T d is the calculated monthly climatological dew point temperature in C T m is the mean monthly temperature in C T n is the minimum monthly temperature in C T x is the maximum monthly temperature in C Minimum, maximum and mean monthly temperatures used here were those obtained from the centres mentioned in section above. However, it is expected that some slight differences may occur in the final dew point temperature values obtained since there was no recalibration of the model with local (African region) data which would otherwise provide more suitable coefficients and subsequently more accurate estimation of dew point temperature. In addition to this, differences due to use of monthly instead of daily temperatures are expected since the regression is based on daily temperatures. 31

48 Top of the cloud Lapse rate = 0.6 C/100m Condensation level Dew point temperature Cloud height Lapse rate = 1 C/100m Earth surface Figure 3-5: Schematic view of temperature lapse rates in an idealized convective cloud. From dew-point concept as visualised in figure 3-5, cloud height can be extracted. The reference point is the earth surface. Dry adiabatic lapse rate of 1 C per 100m and saturated/moist adiabatic lapse rate of 0.6 C per 100m were used here as suggested by Strahler, (1965) and widely accepted in many studies. An example of cloud height (considering only one pixel value) calculation is given here below: Supposing that: surface maximum monthly (e.g. for May) climatological temperature is 300.4K; dew point temperature (at the base of the cloud) is 299.3K (as obtained from equation 3.3); and that brightness temperature observed by the satellite at the top of the cloud is 242.7K. Here brightness temperature was taken as the mean of IR10.8 and IR12.0. In addition supposing an ideal situation of unstable atmosphere where change of temperature with height (lapse rate) is approximately 1 C per 100m (for dry adiabatic) and approximately 0.6 C per 100m (for moist adiabatic), THEN: The cloud top height would be calculated as given in the following expression. H = (( Tx Td ) *1*100) + (( Td Tb ) * 0.6*100) (3.4) = (( ) *1*100) + (( ) *0.6 *100) = 3506 where: H is the cloud height in meters T x is the maximum monthly climatological temperature in K T d is the dew point temperature (in K) as calculated from equation 3.3 T b is the brightness temperature (in K) at the top of the cloud Thus in this example the cloud top height would be 3506m 32

49 Synoptic Data and Field work Synoptic data was collected (from Eastern part of Africa) during a fieldwork campaign and also through a request made to CGIS, Rwanda. Fieldwork activity focused on collecting rainfall data which was measured by setting two rain gauges (tipping bucket type with data loggers); one at Dagoretti corner (Nairobi, Kenya) and one at Kisumu Airport (Kenya). Rainfall data was also collected from a rain gauge at the Ministry of Water and Irrigation, Naivasha (Kenya). Locations of the four stations are as shown in table 3-1 (with the time of rain gauge recording interval) as well as in figure 3-7. Table 3-1: Locations of the four stations within Eastern Africa Station Coordinates Time of Recording Interval Latitude Longitude Nairobi (Dagoretti) 1 18 S E not regular (on tipping) Naivasha 0 24 S E not regular (on tipping) Kisumu 0 06 S E not regular (on tipping) CGIS, Butare 2 36 S E 30 minutes Figure 3-6: Locations of the four stations shown on MSG satellite false colour composite image 33

50 The choice of CGIS station was based on availability of long term series of data that is recorded after every 30 minutes from an hour (for instance; 12:00, 12:30, 13:00, 13:30 UTC etc) (see also appendix C). With this temporal scale it would be easy for comparison with the MSG satellite images which are received in 15 minutes. Besides, CGIS is far enough from the other stations in Kenya and sparsely distributed rainfall stations in this current study are important especially in the rainfall estimation for the obvious reason that at different climatological regimes rainfall estimation methods may have different regression functions. The stations from Kenya (Dagoretti Nairobi, and Kisumu) were selected again due to their distance from one another which is about 400km. In addition Kisumu is situated next to Lake Victoria and thus in cloud mask it would be interesting to investigate the influence of water body (e.g. lake or sea) to cloud mask. Setting own rain gauge (tipping bucket) was therefore necessary to accurately observe the rainfall amount over these two stations. The tipping bucket type of rain gauge used here has the capability of showing the date and time of rainfall observation (see Appendix F). It is therefore easy to compare observations with MSG satellite images which are acquired in after every 15 minutes. However, the limitation here was that the time the tipping occurs could be between MSG image acquisition times. This is well explained by figure 3-7 (b) with the shaded area showing time between two tips and in which satellite image is acquired between the two tips. 34

51 Image acquisition by the SEVIRI radiometer (Source: Meteosat Second Generation System Overview, EUM TD 07 Issue 1.1, 25 May 2001) (EUMETSAT, 2001) Archive repeat cycle-time (e.g. 12:12UTC) Appr. 12:06UTC at equator Dissemination repeat cycle-time (e.g.12:00utc) (a) Rain gauge observation times Time MSG image acquisition time (b) Figure 3-7: (a) Image acquisition by SEVIRI radiometer, and (b) Schematic diagram on MSG satellite and Rain gauge observation time In addition figure 3-7 (a) shows the time stamping of MSG satellite image data, which indicates that within the equatorial region the data is acquired between the two times, say for12:00 UTC image, between 12:00 UTC and 12:12 UTC from south to north of the whole disk. The remaining 3 minutes 35

52 are used in calibrating and turning of the radiometric mirror ready for the next image acquisition. Thus from the two diagrams, it is clear that MSG image acquisition time and gauge observations may not necessarily be at exactly the same time. Thus, in order to solve the problem of different time of observations between the rain gauge and the satellite, there was a need to average over a period of time and space. Here observations within one hour, four MSG top of atmosphere brightness temperatures of the three infrared channels (IR03.9, IR10.8, and IR12.0) as well as total amount of rainfall observed at the station, were averaged. In addition to the two stations in Kenya, rainfall data was collected from a rain gauge at the Ministry of Water and Irrigation, Naivasha. The rain gauge (tipping bucket) was installed in the year 2004 and therefore long period of data was available. Sample of data set collected from the field using this type of rain gauge is given in Appendix F for Nairobi (Dagoretti) which is similar to that of Kisumu and Naivasha Cloud Masking Method Cloud mask method chosen in this study was based on multispectral thresholding technique. This included creating ILWIS scripts in order to generate the necessary images required for processing the cloud mask image. In this method, a number of tests that allowed identification of pixels contaminated by clouds were performed. The main characteristic of these tests, applied to sea or land pixels, depended on the solar illumination conditions and on the satellite viewing angle. The definitions of day-time, night-time, and twilight are as given in table 2-2. The quality of the cloud detection process was assessed by visualizing with the visible bands (for day-time) of the same day and time. Night cloud mask were checked by using one thermal band. Here use of non-linear algorithm, as developed by (Météo-France, 2005a), was made in order to compute sea surface temperature using climatological SST. Split window approach was used with IR10.8 and IR12.0 (brightness temperature of bands 9 and 10) averaged and applied in the algorithm which is in the form given below. ( ( sec 1) SST )* ( T ) T s = T θ 10.8 T (3.5) where: Ts is the calculated sea surface temperature (SST) in C T10.8 and T12. 0 are brightness temperature (in C) of bands 9 and 10 respectively SST is the climatological sea surface temperature ( C) Sec θ is the inverse of cosine of the satellite zenith angle. Here climatological night-time temperature was used as SST. A number of tests were performed in an attempt to extract cloud-contaminated pixels. These tests are explained in the next chapter as well as in the ILWIS scripts which are given in Appendix A. The final cloud mask obtained was then ready for further cloud classification. 36

53 Cloud mask results were also validated using available cloud mask products from EUMETSAT, thanks to the ITC Geodata software development section that recently developed the GRIB2 decoder. During initial stages of this study, validation was not possible since there were no available cloud mask products in the same format as MSG products readily available within ITC. An example of cloud mask from EUMETSAT is as given in figure 2-3 showing cloud as white, clear land as green, and clear sea as blue. The following section outlines how the mask was classified according to heights and used to compare and estimate rainfall from clouds (here going to be referred to as storms) Rainfall Estimation Method As earlier pointed out, rainfall is one of the most difficult atmospheric parameter to measure due to its variability in space and time. However, (Jobard, 2001) stated that rainfall can be inferred from infrared satellite observations in which case brightness temperature for thermal bands for e.g from 10.8µm and 12.0µm measured over cloudy area is related to the cloud height. Clouds with very cold top temperature indicate deep convection which is associated with observed surface precipitation especially in the tropics. The relationship between infrared temperature and rainfall intensity or amount is entirely indirect (Jobard, 2001). Generally it is difficult to discriminate the convective part of the system producing heavy rainfall from the stratiform part of the system or cirrus clouds which are also cold at the top and hardly produce rainfall. In this case therefore, it is worthy attempting to relate cloud height with observed rainfall. The height has to be obtained from methods as explained in the previous section 3.2 in which an attempt has been made to extract clouds at various heights. One-dimensional cloud model-based technique was used in this study to find the basic relationship between cloud height and observed rainfall at a ground station. This was based on relating cloud top height to rainfall amount. Regression-based model as given in equation (3.3) was used to obtain cloud height images. Cloud top height was processed from the cloud top temperature as recorded by MSG satellite. Here brightness temperatures for bands 9 and 10 were averaged. Figure 3-5 indicates schematic temperature lapse rate in relation to cloud height. The general assumption here was that temperature lapse rate, as depicted in the schematic diagram holds in an unstable atmosphere in which case deep convective clouds are formed. In this study, rainfall estimation was based on comparison between point observation and satellite estimation using the one-dimensional cloud-model based technique as mentioned above. In this case therefore, and as Maathuis et al., (2006) pointed out; there is a need to incorporate an averaging procedure in order to account for the collocation problems such as spatial and timing offsets. Here spatial average was carried over 5x5 pixels and temporal average was carried for an hour (four MSG images in an hour). Retrieved temperatures of infrared bands 4, 9, and 10 were averaged and used in the simple cloud mask algorithm developed as an ILWIS script given in Appendix A. See also the next chapter for explanations of different thresholds used. Cloud height obtained in this algorithm was used to compare and estimate rainfall from storms. 37

54 The method used in this study follows the idea that clouds produce different amount of rainfall or have different rainfall intensity at different stages of development. This is best represented by (Hong et al., 2004b) in their study on cloud patch-based rainfall estimation using satellite image classification approach. It should be noted here that due to different climatological regimes, empirical relations (either temperature versus rainfall intensity or cloud height versus rainfall intensity) derived may vary significantly. (Adler and Mack, 1984) studied the impact of the regime-to-regime various on empirical rain estimation schemes based on satellite-observed cloud height or cloud temperature information in which curves representing coastal and inland regimes were strikingly different. They pointed out that these differences had obvious implications for the application of an empirical satellite rain estimation derived in one location and applied in other climatological regimes even with a simple local adjustment. Varying synoptic situations may also cause these types of differences. Rainfall data from CGIS station in Rwanda was investigated to identify storms which produced rainfall over a given period (recorded after every 30 minutes). The observations were recorded at e.g. 1000hrs, 1030hrs, 1100hrs, 1130hrs, etc. The data was in local time and was converted to Universal Time Convention (UTC). For the case of Rwanda, 2 hours are subtracted from local time to change to UTC and for Kenya 3 hours are subtracted from local time. Appendix C shows sample of meteorological records including rainfall obtained from CGIS weather station and used to investigate cloud height variation over a given period. This would give an idea on how the two relate over the selected region which may provide water resource management authorities first approximations of rainfall amount expected from a storm at a particular height. Twelve storms of different days and time from CGIS station were used to develop a regression function between the height and observed rainfall. The function was consequently used to estimate rainfall amount from other storms over the same station to validate the performance of the method developed and thresholds selected during the simple cloud mask algorithm development. 38

55 4. Data Processing and Results This chapter attempts to analyse the processed MSG images in detail and the final cloud mask obtained as well as the cloud height classification images. Analysis of field data and rainfall estimates drawn from developed regression function(s) are also presented MSG Satellite Images The first step was retrieval of relevant bands as well as calculation of solar illumination angles. Solar illumination angles were based on conditions as given in table 2-2. MSG satellite and solar zenith angles were also generated since this change due to earth rotation about its axis Generation of MSG Satellite and Solar Angles Generating MSG satellite and sun angles was done by creating a batch file which could be adapted for any date and time in case new angles were required. This particular applet, which can be executed into an active directory, works in a java environment which must be installed in the system. Generated angles were imported into ILWIS for further processing. Figure 4-1 shows the flow chart for generating the satellite and solar angles. The processing was done for mainly MSG field of view covering Africa ( 39 N - 38 S and 34 W - 53 E). Create MSG and Sun zenith angles (In Java environment) Import satellite and sun zenith angles into ILWIS Add MSG georeference Resample the satellite and sun angles into required georeference Calculate secant angle of the satellite and the sun MSG secant angle image Sun secant angle image Apply sun elevation angle & generate threshold maps Figure 4-1: Flow chart for generating MSG satellite and Sun angles 39

56 Figure 4-2 shows examples of the sun and MSG satellite zenith angles. Sun position as from the left image can be seen to be overhead in the south west of the image (brightest part) for 26 December 2006 at 15:00 UTC. MSG satellite is situated at 0 N and 0 E and can be seen to appear at the same position (brightest part) in the right image Figure 4-2: Sun (for 26 th December 2006 at 15:00 UTC) and MSG Satellite (0 N and 0 E) Zenith angles (left and right hand image respectively) After calculating sun elevation angle, solar illumination conditions were generated by use of threshold mentioned in table 2-2 in which the condition is day-time when the sun elevation angle is greater than 10 and night-time when the sun elevation angle is less than -3. The condition is twilight, that is, either before night-time or before day-time when the sun elevation angle is between -3 and 10. An example of such an image of 7 th March 2006 at 15:30 UTC is provided in figure N 0 E Figure 4-3: Solar illumination conditions on 26 th December 2006 at 15:00 UTC As earlier pointed out, the algorithm is based on a multispectral threshold technique applied to each pixel of the image. A number of tests for each solar illumination condition in which an example of 40

57 cloud mask is given in figure 4-7 were applied. The tests applied in the algorithm attempted to address both the land and sea surfaces based on their characteristics. Flow charts for these tests are given in the following respective sections Day-time Cloud Mask Stddev<1 No IR10.8>293.15K Yes Yes Stddev>1 and T smin -SST cal <-1K Yes No IR10.8> (T max - stddev/2) Yes No No Cloud Clear Cloud Clear Figure 4-4: Description of the test sequences for Land surface (left) and Sea surface (right) Figure 4-4 shows the description of test sequence used in cloud mask during day-time. Over the land surface, during the day cloud contaminated pixels were identified by using standard deviation from the climatological surface temperatures (mean, minimum, and maximum) which should be greater than 1K. Minimum surface temperature was taken as the monthly climatological night-time temperature as processed from the WorldClim database whereas maximum surface temperature was taken as day-time temperature and the mean surface temperature was average of the day-time and night-time surface temperatures. To remove false cloud assignment to pixels over desert areas, brightness temperature of band 9 (IR10.8) less than K was applied otherwise the pixels with a higher temperature were considered cloud free. Further, all pixels already defined as cloudy were subjected to tests in order to avoid cool areas or higher elevated areas. These involved using monthly climatological temperature standard deviation (amplitude). Cloudy pixels with brightness temperature (IR10.8) less than maximum (T max ) day-time monthly climatological temperature less half the monthly standard deviation are assigned cloudy else not cloudy. This does not affect the ocean areas since the climatological standard deviation is very small. This test allows us to reduce misclassifications to the minimum except in high elevated areas and desert areas. Over the sea surface, cloud-contaminated pixels were identified by using standard deviation from the climatological surface temperatures (mean, minimum, and maximum) which should be less than 1K. Further, small difference of -1K (and above) between the local sea surface temperature (as calculated 41

58 using equation 3.5), here referred to as SST cal, and minimum monthly climatological sea surface temperature (here referred to T smin ) was also used to mask cloudy pixels over the sea surfaces. As earlier explained, Météo-France have developed cloud mask in which this study adopted some of the basic ideas to develop some of the thresholds used. Estimating SST by using IR10.8 and IR12.0 brightness temperature together with minimum monthly climatological SST was used by Météo- France. Here the same two bands are used, as top of atmosphere (in K) together with minimum monthly SST (here taken as the night time temperature). Météo-France took a small difference of 4K between estimated SST (by using IR10.8 and IR12.0 brightness temperatures) and the monthly climatological minimum SST. Monthly climatological minimum SSTs are derived from a global Pathfinder night-time bulk SST climatology. The bulk night-time SST, as (Derrien and Le Gléau, 2005) pointed out, does not account for the thermal heating at midday observed in infrared satellite measurements. In this study this difference was set at -1K over the sea surface as stated above. Brightness temperature of band 9 (IR10.8) was applied by Météo-France as well as by (Kidder et al., 2005) in which the idea was to estimate the temperature that would be observed if there was no water vapour in the atmosphere. Météo-France computed threshold from surface temperatures forecast by NWP model. In this study threshold of K was set as the maximum temperature for any pixel to be flagged cloud contaminated. Météo-France again used IR10.8 and IR12.0 difference to detect thin cirrus clouds and cloud edges characterized by higher IR10.8-IR12.0 values than cloud-free surfaces. Here use of IR10.8 less than the maximum climatological surface temperature (with half amplitude of the climatological monthly minimum, maximum, and mean temperatures) was to extract thin cirrus clouds as well as to avoid confusion of moist, warm, cloud free areas with clouds. With these few tests day-time cloud mask was obtained of which an example is as given in figure 4-7 (a). Notable features of this cloud mask are such as sharp boundary between the land and the sea that appears along some coastal areas, especially in this particular case to the North West of the continent. This depicts cloudy conditions over the ocean and non-cloudy conditions over the land, which may not be always the case. The sharp boundary is due to the land-sea temperature effects and increases as we move from equatorial regions to higher latitudes where temperatures are generally low over the sea such as the case in the north-western part of the continent (over the Atlantic ocean). This is more pronounced especially when desert (usually with high temperatures) areas lie next to water body. Cloud mask image shows presence of clouds over the northern part of Africa whereas from the false colour composite of the visible and near infrared bands does not show the same situation. Over central Africa and Atlantic Ocean (the specific region of interest in this study) most of the cloudy pixels (as can be seen from the false colour composite image) have been masked out. Also as can be seen from the false colour composite image in figure 4-9 (b), there appears no thick clouds in the northern part of the continent. However, the algorithm has classified the region to be under low level clouds which are semi-transparent in the visible and near infrared bands. 42

59 Night-time Cloud Mask Figure 4-5 shows the description of the test sequences used in cloud masking at night time. Details of these tests are explained below. Stddev<1K and T min -IR10.8<9K and T mean >283.15K Yes Stddev>1K and T mean -SST cal <10K No Yes Yes No [(IR10.8*IR12.0)/IR03.9]- T mean <2K No Cloud Clear Cloud Clear Figure 4-5: Description of test sequence for land surface (left) and sea surface (right) During night-time, the standard deviation of the monthly climatological temperature was set at a minimum of 1K and the mean brightness temperature (T mean ) of IR10.8 and IR12.0 was taken less than K over the land surface for any pixel to be flagged as cloudy. The difference of the monthly minimum climatological temperature and the brightness temperature of IR10.8 is used is set to be greater than 9K for any pixel to be assigned cloudy. This ensured avoiding cooler areas at night which would otherwise be assigned cloud contaminated. Use of mean brightness temperature for IR10.8 and IR12.0 followed Météo-France developed cloud mask idea in which the difference between the two is used to detect thin cirrus clouds and cloud edges characterised by higher difference (IR10.8-IR12.0) values than cloud-free surfaces. However, in this study the mean of the two was expected to simply avoid the confusion of very moist, warm, cloud free areas with clouds. Over the sea, the standard deviation of the monthly climatological temperature is less than 1K. The difference between the local calculated sea surface temperatures (SST cal ) and the monthly mean climatological temperature was taken to be greater than 10K. Low clouds over the sea were screened by use of IR03.9 to scale down aggregated temperatures of IR10.8 and IR12.0 (i.e. IR10.8*IR12.0). The difference between their mean temperatures and the scaled temperature is set at a minimum threshold value of 2K for the cloudy pixels. This test is based on the fact that the water cloud emissivity is lower at IR03.9 than in IR10.8 or IR12.0. The test allows detecting low clouds at night time. The approach is the same as that of Météo-France using the difference between IR03.9 and IR10.8. An example of night-time cloud mask is given in figure 4-7 (c). 43

60 Twilight Cloud Mask Figure 4-6 shows the test sequences used in extracting clouds during twilight time. Details of these tests and dynamic thresholds are explained below the figure. Stddev<1K and T min -IR10.8<9K and T mean >283.15K No Yes Stddev>1K and T mean -SST cal <5K No [(IR10.8*IR12.0)/IR03.9]- T mean >2K Yes Yes No Cloud Clear Cloud Clear Figure 4-6: Description of test sequence for land surface (left) and sea surface (right) At twilight time the difference between climatological minimum temperature and the brightness temperature of band 9 (IR10.8) was set at a threshold of 9K such that any pixel with greater difference than this value and with mean brightness temperature (IR10.8 and IR12.0) less than K were cloud contaminated. This ensured screening cloudy pixels over the land surfaces where also standard deviation of the monthly climatological temperatures was set at a minimum of 1K. Over the sea, the difference of mean monthly climatological SST and the calculated SST was taken to be greater than 5K for the cloudy pixels. Here Météo-France used IR10.8 and IR12.0 brightness temperatures to estimate SST by using a nonlinear split window algorithm. A pixel is flagged cloud contaminated if its estimated SST value is lower than a minimum monthly climatological SST value by 4K. However, Météo-France does not apply this test where climatological SST is lower than K. In this study low clouds were extracted by use of IR3.90 to scale down brightness temperature of bands 9 and 10 (IR10.8 and IR12.0, respectively). Here maximum threshold value of 2K as the difference between the scaled temperature and the mean brightness temperature of IR10.8 and IR 12.0 was used. Threshold for the difference between estimated SST and the climatological SST from Météo-France gives a threshold of 4K which is comparable to the set value in this study. Météo-France uses IR03.9 and IR10.8 difference to extract low clouds for both day-time and twilight time basing the fact that solar reflection at IR03.9 (approximated by the IR03.9-IR10.8 brightness temperature difference) may be rather high for clouds (especially low clouds), which is not the case for cloud free areas. An example of twilight cloud mask from this study is as given in figure 4-7 (b). 44

61 (a) (b) (c) Figure 4-7: Cloud masks for (a) day-time, (b) twilight time, and (c) night-time; for MSG-1 image of 7 th March 2006 at 15:30 UTC. Cloudy pixels are represented as green whereas grey represents non-cloudy pixels. Merging the three images resulted in final cloud mask as given in figure 4-9 (a). Colour composite image of the same day and time is as in figure 4-9 (b). Here solar illumination conditions are as given in figure 4-8 below. Figure 4-8: Solar illumination conditions on 7 th March 2006 at 15:30 UTC 45

62 Legend Cloud Cloud free a b Figure 4-9: Cloud mask (a) and false colour composite (b) for MSG image of 07/03/2006 at 15:30 UTC. In the cloud mask, green are clouds and white are no cloudy pixels. False colour composite of the visible and near infrared bands (VIS006, VIS008, and NIR016) is more visible from central Africa to the Atlantic Ocean. This is the day-time region as can be seen from figure 4-8. To the eastern part, it is not easy to visualise since this area already falls under twilight and night conditions. Sharp boundary between the land and the sea can be seen to appear along some coastal areas, especially in this particular case to the North West. This depicts cloudy conditions over the ocean and non-cloudy conditions over the land, which may not be always the case. The sharp boundary is due to the land-sea temperature effects. This is more pronounced especially over desert (usually with high temperatures) areas next to water body. Cloud mask image shows presence of clouds over the northern part of Africa whereas from a visual check using the false colour composite of the visible and near infrared bands does not show the same scenario. Over central Africa and Atlantic Ocean most of the cloudy pixels (as can be seen from the false colour composite image) have been masked out. The next step was to process heights for the extracted clouds based on the formula for estimating dew point temperature as given in equation 3.4 in which an example is given in figure

63 Figure 4-10: Cloud height (in Meters) image (MSG image of 07/03/2006 at 15:30 UTC) The cloud height images were classified into three different classes namely; low clouds (50m-1500m), middle clouds 1500m-3000m, and high clouds (> 3000m). These classes were chosen based on occurrence of different types of clouds at different levels as is shown in figure 2-1. Based on this classification and as explained in section 2.1, it is possible to show areas where precipitation is likely. However, this is further investigated in the proceeding section of rainfall estimation. An example of classified cloud height image is given in figure 4-11 below. Figure 4-11: Classified cloud height image of 07/03/2006 at 15:30 UTC 47

64 In order to check whether all cloudy pixels have been extracted properly for the area of study, segmentation of the classified cloud image was done and overlaid on to a false colour composite image of visible and near-infrared bands of the same day and time. Figure 4-12 shows an example of a small window (eastern part of Africa) of such an overlay for MSG satellite image of 23 rd November 2005 at 13:30 UTC. This area was under twilight condition on 7 th March 2006 at 15:30 UTC and thus such an overlay is not provided here. Segmentation was performed on the classified cloud height image of 23 rd November 2005 at 13:30 UTC and the segments overlaid on the false colour composite image of the same date and time. Figure 4-12: Segments (yellow lines) of cloud mask of 23/11/2005 at 13:30 UTC overlaid on False colour composite (VIS006, VIS008, and NIR016) in (BGR) Clouds appear as cyan in colour in the false colour composite image. As can be observed visually from figure 4-12, most of the cloudy pixels have been identified. This is more visible over areas where deep cyan colour (mostly deep convective clouds) appears. Some semi-transparent clouds have not been masked out. However, this is not of serious concern in this current study since most of these semi-transparent clouds do not contribute to precipitation, and if they do, very little rainfall is expected from them. Further discussions to the accuracy of the simple cloud mask algorithm developed are provided in the next chapter Rainfall Estimation (A case of CGIS Weather station) As earlier pointed out, CGIS weather station has a rain gauge which is set to measure rainfall among other meteorological parameters in every 30 minutes. The advantage to such a type of record of data is that it is possible to compare with MSG satellite observation(s) of parameters such as top of atmosphere brightness temperatures of infrared bands, cloud height, and cloud type; and develop a relationship that can be used to infer rainfall intensity or amount from observed (non atmospherically 48

65 corrected) satellite parameters. Various methods as explained in section 2.5 can be used in developing the relations between these parameters. In this study rainfall versus processed cloud height were investigated to get an idea on how they relate and also to find a function that can be used to forecast rainfall intensity or amount expected on a ground station (e.g. CGIS). An attempt was made to relate rainfall intensities from clouds of different dates and times with processed cloud heights of the same dates and times. The relationships were generally too low. This was mainly due to the fact that at different dates and times the cloud/storm over the station is not necessarily at the same development stage. It is likely that in such an approach, relationships are being drawn for storms at different stages of their development (and probably of different types) over the station. As explained in section 2.5.5, satellite-based rainfall estimation algorithm, Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) Cloud Classification System (CCS) by Hong et al., (2004a), extracts cloud features from IR10.8 geostationary satellite imagery in estimating fine scale rainfall distribution. The algorithm processes satellite cloud images into pixel rain rates by; separating cloud images into distinctive cloud patches, extracting cloud features, clustering cloud patches into well-organized subgroups, and calibrating cloud-top temperature and rainfall relationships for the classified cloud groups using gauge-corrected radar hourly rainfall data. Based on this method, therefore, relating rainfall intensities and cloud heights of different dates and times was not expected to yield good relationship. Thus storms over the station were treated separately in order to relate their rainfall intensities or total rainfall amounts with their heights. The relationships between cloud height and the rainfall intensity as well as between cloud height and the total rainfall from a storm were then developed. The following sections detail the results of these approaches Direct Comparison of Cloud Height and Rainfall Intensity Rainfall estimation method used here was based on average storm height during its existence over ground station. Firstly, diurnal trend of observed rainfall intensity and processed cloud heights, over the weather station, were investigated. Two days were selected for this purpose and figure 4-13 shows how cloud height and rainfall intensity varied during the selected days. MSG images of 30 minutes interval were processed to obtain cloud heights whereas rainfall intensities observed at the station at the same time of MSG image acquisition were used. In both cases it can be seen that clouds at a height of above 3000m contribute to a large fraction of the rainfall recorded at the station. The highlighted part clearly shows the time rainfall was observed at the station which agrees with the time of high cloud heights as processed by the simple cloud height algorithm developed using the dew point temperature and lapse rate concepts as applied in equation

66 Cloud height Cloud height 6000 R/Intensity R/Intensity 50 Cloud Height (m) Rainfall Intensity (mm/hr) Cloud Height (m) Rainfall Intensity (mm/hr) :30 3:30 6:30 9:30 12:30 15:30 18:30 21:30 Local time (Hrs) Local time (Hrs) (a) (b) Figure 4-13: Diurnal cloud height and Rainfall intensity changes on (a) 5 th May 2006, and (b) 10 th May :30 3:30 6:30 9:30 12:30 15:30 18:30 21:30 Secondly, identification of storms on different days and plotting their rainfall intensities against processed cloud height, in this case in class intervals of 500m, followed. This was meant to check whether the above two day s cases were a mere coincidence or is the true scenario expected from this particular station. Here cloud heights were grouped from 2500m in intervals of 500m. Fifteen storms with their 30 minutes interval processed heights and their respective observed rainfall intensities were plotted as shown in figure For details of these storms refer to Appendix D. 60 Rainfall Intensity (mm/hr) Rain intensity Height (m) Figure 4-14: Rainfall intensities within cloud height classes From this plot it is clear that high rainfall intensities are observed within cloud heights of 4000m to 6000m. This indicates the same situation as in the case of 5 th May 2006 and 10 th May 2006 as shown in figure However, in each cloud height class there are low rainfall intensity observations. This could be associated with early stages of cloud formation or late stages (dissipating stage) of the cloud. Rainfall intensity within each class was averaged and plotted against average cloud height in each class. The best model fit was found to be Gaussian, whose regression function is: (y=a*exp ((-(xb)^2)/(2*c^2)), where: a = 9.8, b = , and c =1200.3; with correlation coefficient of 0.95 and standard error of 1.03). This model agrees with the fact that very low clouds (e.g. stratocumulus, 50

67 cumulus, and stratus) and very high clouds (e.g. cirrus, cirrocumulus, and cirrostratus) produces very low rainfall. Average rainfall intensity (mm/hr) Average cloud height (m) Figure 4-15: Gaussian model fit, X = Average cloud height (m), Y = Average rainfall intensity (mm/hr) An attempt was made to use this function to estimate rainfall intensity for various storms. Figure 4-16 represents plots of all observed and estimated values within each cloud height class. It is clear that the function overestimated the rainfall intensities from these storms except in very few levels (height) where it underestimated. This appears so when cloud height is between 2500m and 3000m. Thus there was a need to adopt a different approach by either using rainfall intensity or total rainfall amount from different storms. Rainfall Intensity (mm/hr) Observed Estimated Cloud height (m) Figure 4-16: Observed and estimated rainfall intensity for different storms Based on the above results, it can be observed that the developed regression function did not perform well. There was general overestimation of rainfall intensity. Further investigation of the relationship between cloud height and total amount of rainfall from a storm was carried out as explained in the following section. 51

68 Direct Comparison of Cloud Height and Total Rainfall Diurnal trend of observed total rainfall and processed cloud heights, over the weather station, were investigated. Two days were selected for this purpose and figure 4-17 shows how cloud height and total rainfall varied during the selected days. MSG images of 30 minutes interval were processed to obtain cloud heights whereas total rainfall observed at the station at the same time of MSG image acquisition was used Cloud height Total Rainfall Cloud height Total Rainfall Cloud Height (m) Total Rainfall (mm) Cloud Height (m) Total Rainfall (mm) :30 3:30 6:30 9:30 12:30 15:30 Local time (Hrs) 18:30 21:30 (a) (b) Figure 4-17: Diurnal cloud height and Total rainfall changes on (a) 5 th May 2006, and (b) 10 th May :30 3:30 6:30 9:30 12:30 Local time (Hrs) 15:30 18:30 21:30 It can be observed that rainfall was recorded at the station when the processed cloud height was at high levels (above 3000m). It is clear then that high clouds over this station are the main rain producing rainfall clouds. This indicated that there is a relationship between cloud height and total rainfall produced by the cloud at certain height. The general idea followed in this comparison is that the more the cloud is sustained at a certain height while producing rainfall, the more the rainfall is observed at a ground station. This idea is borrowed from the case of CCD as explained in section in which the relationship drawn from the life-history cycle of the storm was found to be linear provided spatial and temporal average are considered. However, as (Grimes et al., 1999) pointed out, the most important assumption is that rainfall is predominantly convective in origin and that the raining clouds can be identified as those with cloud top temperatures below a certain temperature threshold. Here, cloud height is used to compare the total amount of rainfall observed at CGIS. A regression function can be derived using as many storms from the station as possible. Twelve storms were used to derive a regression function that was later used to estimate total amount of rainfall from other storms for validation purpose. Comparison with the observed station amount over the same period with the estimated rainfall amount showed slight overestimation for some storms and underestimation for others. Table 4-1 shows the date and time of the storms used to develop the relationship between cloud height and the observed total storm event rainfall. Appendix E shows processed details of the twelve storms. 52

69 Table 4-1: Observed storms and their total amount of rainfall Storm Date Time Duration Average Total (UTC) (hrs) Cloud Height Rainfall (m) * (mm) 1 07/03/ /03/ /03/ /03/ /04/ /04/ /05/ /05/ /05/ /07/ /07/ /08/ * Above the terrain Determination of the model fit showed Gaussian fit as the best for CGIS by using the 12 storms that appeared over the station selected for this analysis. The above storms were used to determine a regression function between the two variables. The best fit obtained was again a Gaussian model (y=a*exp ((-(x-b)^2)/(2*c^2)); where: a = 60.6, b = , and c = with correlation coefficient of 0.96 and standard error of 6.56mm. This is presented graphically in figure The model agrees with the fact that very low clouds (e.g. stratocumulus, cumulus, and stratus) and very high clouds (e.g. cirrus, cirrocumulus, and cirrostratus) produces very low rainfall. Total rainfall (mm) Average cloud height (m) Figure 4-18: Gaussian model fit, X= Average storm height (m), Y= Total rainfall (mm) 53

70 The model whose equation is (4.1) was used to estimate the total amount of rainfall from other storms over the same station. 2 ( x ) y = 60.6*exp (4.1) 2 2*583.0 where: y is the estimated total storm rainfall x ix the average cloud height Five storms were taken for estimating the total amount of rainfall expected from them and the results are presented in table 4-2 and graphically shown in figure 4-19, together with the standard error. Table 4-2: Storm heights and estimated total rainfall Avg. Obs. Est. Storm Total Total Height Rainfall Rainfall Difference Storm Date Time (UTC) (m) (mm) (mm) (%) 1 08/03/ /04/ /04/ /05/ /05/ From these results it can be seen that two out of five storms have been estimated to a reasonable accuracy. These were the storms of 21 st April 2006, and 14 th May The 8 th March 2006 storm was overestimated by 198% which is quite high whereas that of 21 st April 2006 was overestimated by very low percentage of 10%. The rest of the storms were underestimated with lowest at 17% (14 th May 2006). Figure 4-19 shows a plot with error bars whose value is 6.56 mm (standard error of the derived function). Total Rainfall (mm) Storms Observed Estimated Figure 4-19: Observed and Estimated total rainfall plotted with the error bars 54

71 Furthermore relating the observed and the estimated for these five storms, a relation in the form of: Observed Rainfall = *Estimated Rainfall , was obtained. 14 y = x R 2 = Observed rainfall (mm) Estimated rainfall (mm) Figure 4-20: Relationship between the observed and the estimated total rainfall Goodness of fit (R 2 ) of approximately 0.36 was obtained in this relationship. This shows that there is low correlation between observed total rainfall and that estimated using the derived regression. However, considering the approach as enumerated above, all the estimates can be said to be nearly the same as the observed total rainfall. 55

72 56

73 5. Discussions of Results This chapter discusses in detail results of the simple cloud mask and rainfall estimation as obtained in the previous chapter. Various results of previous studies are compared to some results of the current study Cloud Mask Results The simple cloud mask (SCM) algorithm has been checked and validated using the EUMETSAT cloud mask products. The choice of this algorithm for validation was based on the fact that EUMETSAT validation procedure of their cloud mask was more realistic since they use database that is automatically built and is collocated with the MSG satellite data and that of surface observations. The surface data used are hourly weather observations, coded by observers into the World Meteorological Organization (WMO) synoptic code (SYNOP). In addition meteorological information extracted from the French NWP model Action de Recherche Petite Echelle Grande Echelle (ARPEGE) forecast fields is used. Based on these facts, it can be concluded that EUMETSAT cloud mask algorithm as developed by Météo-France is more robust as compared to other cloud algorithms for MSG satellite images. The current study was not able to collect such enormous data for validation thus made use of EUMETSAT products. These products are available from EUMETSAT in GRIdded Binary (GRIB2) format. GRIB is a World Meteorological Organization s (WMO s) standard binary format for exchanging gridded data. The raw data from EUMETSAT was imported into ILWIS after gluing the six segments (provided in the original EUMETCast data stream) of MSG satellite field of view that are provided. In ILWIS a procedure to cross check the accuracy of the cloud mask developed in this study is given in the flow chart (figure 5-1) below. 57

74 EUMETSAT CLM (GRIB2) Simple Cloud Mask (SCM) Reclassification Mask clouds only Sub-map study area Sub-map study area Segmentation of cloud mask Segmentation of cloud mask Overlay onto a VIS/NIR false colour composite image VISUALIZATION Figure 5-1: Flow chart on segmentation and visualization of EUMETSAT CLM and SCM EUMETSAT CLM raw data received at ITC was processed to check on the accuracy of the developed simple cloud mask (SCM) in this study. The EUMETSAT CLM of MSG of 26 th December 2006 at 15:00 UTC is given in figure 5-2. GRIB2 import routine developed at ITC was used to convert the data and appropriate classes were assigned manually. These were named as; cloud, clear land, water, and background. In addition GRIB2 import routine assigns the proper geometric model and therefore the cloud mask can be directly integrated with other processed results. 58

75 METEOSAT SECOND GENERATION (MSG) CLOUD MASK, CLOUD PROPERTY DETERMINATION AND RAINFALL COMPARISON WITH IN Figure 5-2: EUMETSAT cloud mask assigned feature classes for 26th December 2006 at 15:00 UTC The clouds were selected and a small window ( 11 N - 14 S and 6-51 E), covering a part of the tropics over the African continent considered in this study in developing the SCM, was extracted. Segmentation was done for both SCM and EUMETSAT CLM. Both were overlaid onto a false colour composite of the same day and time (e.g. 26th Dec 2006 at 15:00 UTC) and figure 5-3 shows the results of these overlays. Figure 5-3: Cloud mask segments of EUMETSAT CLM (yellow lines) and SCM (red lines) for 26th December 2006 at 15:00 UTC, on a false colour composite 59

76 From figure 5-3 it can be seen that the segments of the EUMETSAT CLM and those of SCM match in some areas and mismatch for other areas. Detection of clouds from the developed algorithm will be assumed to be accurate when the two lines exactly overlaid on each other in this visualization procedure. It can be seen that over the Indian Ocean, the SCM did not extract majority of the cloudy pixels. This could indicate that the climatological SST has to be reviewed either to use daily data instead of the monthly used in this study. However, the other technique used to evaluate the accuracy of the SCM was based on creating confusion (contingency) matrix. This method compares all pixels, within a selected window, to find out whether the pixels are assigned as cloudy or non-cloudy in both masks. Here the two cloud masks images were crossed to built a contingency table that indicates the number of pixels in each category. This will show the ability of the SCM to detect cloudy and non-cloudy events based on EUMETSAT CLM. In order to get better results of accuracy of the SCM, there is a need to use a number of images. Since the SCM algorithm was developed based on different solar illumination conditions, it was appropriate to choose MSG images based on these three conditions. Here four days images were used for validation and their contingency matrices are as given in the following tables 5-1 to 5-4 for the specified day and time. Table 5-1: Contingency table for MSG image of 25 th December 2006 at 12:00 UTC Simple cloud mask (SCM) (Number of pixels) EUMETSAT CLM (Number of pixels) Cloudy Not cloudy Total Error of commission (%) Cloudy Not cloudy Total Error of Omission (%) Overall Accuracy: 89.1 % Table 5-2: Contingency table for MSG image of 26 th December 2006 at 15:00 UTC Simple cloud mask (SCM) (Number of pixels) EUMETSAT CLM (Number of pixels) Cloudy Not cloudy Total Error of commission (%) Cloudy Not cloudy Total Error of Omission (%) Overall Accuracy: 88.9 % 60

77 Table 5-3: Contingency table for MSG image of 4 th January 2007 at 22:00 UTC Simple cloud mask (SCM) (Number of pixels) EUMETSAT CLM (Number of pixels) Cloudy Not cloudy Total Error of commission (%) Cloudy Not cloudy Total Error of Omission (%) Overall Accuracy: 88.0 % Table 5-4: Contingency table for MSG image of 10 th January 2007 at 17:00 UTC Simple cloud mask (SCM) (Number of pixels) EUMETSAT CLM (Number of pixels) Cloudy Not cloudy Total Error of commission (%) Cloudy Not cloudy Total Error of Omission (%) Overall Accuracy: 83.3 % From these tables it can be seen that cloud mask of 25 th December 2006 at 12:00 UTC, 26 th December 2006 at 15:00 UTC and that of 4 th January 2007 at 22:00 UTC had the highest accuracies of 89.1%, 88.9% and 88.0% respectively. In the same cloud masks only 10.2%, 5.7% and 7.8% (respectively) of total number of pixels detected by EUMETSAT, were not detected by the SCM. This can be termed as the cloud failure score or underestimation of cloudy events. On 10 th January 2006 at 17:00 UTC, there was relatively higher cloud failure of 15.7%. This could be associated with non-detection of low clouds or thin, semi-transparent broken clouds at night over both the land and the sea. On this day SCM also depicted a slightly lower overall accuracy of 83.3%. On 25 th December 2006 at 12:00 UTC (day-time), 8.0% of total number of cloudy pixels from the SCM was not under cloudy conditions as per the EUMETSAT CLM. This is very low as compared to other days where the total number of pixels assigned cloudy were almost double; 14.2% (26 th December 2006 at 15:00 UTC), 16.7% (4 th January 2007 at 22:00 UTC), and 22.5 % (10 th January 2006 at 17:00 UTC). This implies that the day-time algorithm was able to differentiate the cloudy and non-cloudy pixels to a greater accuracy as compared to the night-time and twilight time algorithms. Besides, only 7.0% and 7.1% of the EUMETSAT cloudy pixels were assigned non-cloudy by the SCM for 26 th December 2006 at 15:00 UTC and 4 th January 2007 at 22:00 UTC, respectively. This was low as compared to the SCM of 25 th December 2006 at 12:00 UTC and 10 th January 2007 at 17:00 UTC, which assigned EUMETSAT CLM cloudy pixels as non-cloudy at 15.1% and 12.0%, respectively. This could be associated to the use of NWP models by EUMETSAT which is likely to model the more variable day atmospheric profile to a greater accuracy. This is not possible with the simple thresholds used in this study and thus the high difference in assigning cloudy pixels to noncloudy despite good results in overall accuracy of 89.1% for the day SCM. 61

78 Generally underestimation could be occurring over the sea areas since climatological SST used was monthly mean SST. Use can be made of 5-, 7-, or daily- mean SST which most likely would improve the accuracies. During day-time (25 th December 2006 at 12:00UTC), SCM overestimated cloudy events by only 11.9%, which is slightly lower than the other times. It is likely that this low failure is due to thresholds used. Use of NWP model by EUMETSAT to compute some thresholds for twilight time and nighttime seems to improve the accuracies during these times and hence the higher differences from the SCM. Better results of the day-time cloud mask could be associated with the fact that convective activities could be present and thus easy to screen the clouds based on the thresholds used in this study. Moreover, the period selected here is when the Inter-tropical Convergence Zone (ITCZ) is generally within the region under consideration and therefore convective activities, with low presence of thin and semi-transparent cirrus clouds, are common. On 26 th December 2006 at 15:00UTC (twilight time), 4 th January 2007 at 22:00UTC (night-time), and 10 th January 2007 at 17:00UTC (twilight time), overestimation score was 17.1%, 15.5%, and 17.4%, respectively. In general terms and considering the four situations, the overall accuracy of the study is 87.3%. This indicates that SCM performed well and thus the simple thresholds used were able to extract cloudy pixels as intended. It is worthy noting that the time mentioned here refers to the defined solar illumination conditions that are occurring over most part of the selected region. The overall accuracies obtained in this study may depict good performance of the SCM algorithm developed. However, on superimposing the EUMETSAT CLM on a false colour composite showed that not all cloudy pixels were correctly screened. Thus a more robust validation method or cloud mask algorithm may be sought Cloud Height/Type Results An attempt was made to validate the simple cloud height/type (SCH/T) algorithm using the EUMETSAT cloud top height (CTH) products also available from EUMETSAT and accessed through EUMETCast. EUMETSAT CTH products are based on sea surface whereas SCH/T products are based on the earth surface. Thus digital elevation model (DEM) products, sourced from ftp://edcftp.cr.usgs.gov/pub/data/gtopo30/global (USGS, 2007) were used to compute the EUMETSAT CTH products based on the earth surface. Examples of the EUMETSAT cloud top height and SCH/T images for 25 th December 2006 at 11:45 UTC, for a part of African tropical region and Atlantic Ocean, are given in figure

79 (a) (b) (c) Figure 5-4: EUMETSAT CTH (a), SCH/T (b), and Difference (between CTH and SCH/T) (c) images for 25 th December 2006 at 11:45 UTC (height is in meters) The results show significant differences in cloud heights for those pixels assigned cloudy by both EUMETSAT cloud mask products and by the SCM algorithm. The difference image (figure 5-4 (c)) is also provided. Various other days EUMETSAT CTH products were investigated and the same high differences were obtained. However, based on figure 2-1, it can be seen that the EUMETSAT CTH products might be too high. Cloud height computed using the SCH/T algorithm appears to provide estimates which may be realistic based on the same figure 2-1. Thus validation with the EUMETSAT CTH products may not provide reliable results. Further validation of simple cloud height/type algorithm was not carried Rainfall Estimation Results Understanding that rainfall estimation from one-dimensional cloud-based model technique lacks a strong physical basis, it was essential to estimate total amount of rainfall from individual storms. This method provided an idea of how much rainfall a cloud at a certain height can produce. Although this approach has limitations given the assumptions used (e.g. wind shear over the station does not 63

80 strongly affect the storm), it is possible to establish cloud height- rainfall relations that may be used as first approximations. Thus the results would be more general than existing methods, so that the technique would not be tied to one storm or one climatological regime or one synoptic situation. As given in figure 4-14, rainfall intensities were plotted against cloud height classes and the clusters indicates that cloud height between 4000m and 5500m produced rainfall of significant intensities. Total rainfall from storms can also be seen from figure 4-18 to depict the same trend where high amounts are within cloud heights of above 4000m and below 5500m. Diurnal trends (see figure 4-13 and 4-17) of cloud heights and rainfall intensity and/or total rainfall are rather interesting. They showed that rainfall occurred at the station when the cloud height was higher than 3000m. This is by no means a coincidence of results of the two days selected for investigating trends which again confirm the results as shown in figures 4-14 and However, despite clear relationship depicted by the direct cloud height rainfall intensity plots, estimated rainfall intensities for other storms using the derived function, was way above the observed intensities at the station as can be seen from figure From statistical analysis, between observed and estimated total rainfall from other five storms, correlation coefficient of 0.96, root mean square error of 3.72mm, and a skill score index of 0.23 were obtained. (Laurent et al., 1998) pointed out that the non-dimensional skill score index, as here applied, is the relative distance between the estimated values and the observed values and it depends on the standard deviation (error) of the observed data. Skill score is equal to one when the estimates are perfect and equal to zero when there is best constant estimates. The skill score obtained here indicates the estimates were not perfect and that is confirmed from the values as can be seen from table 4-2. Results presented here are for only a few storms and do not depict general results of all storm situations that may occur over the station. However, they may give a first approximation of rainfall amount expected from storms that occur at a certain height. As Heinemann, (2003) pointed out, one of the major difficulties in relating precipitation observed at a ground station and measured satellite signals is that the amount of precipitation reaching the ground depends very much on the structure of the atmospheric layer under the precipitating cloud. This can be said to aggravate the error in the estimates since the atmospheric layers below the precipitating clouds are not modelled in this study to incorporate them. It should also be noted that twelve storms were used to derive the regression function. This may not be enough to derive regression that may be representative of all types of storms that may occur over this station. There is a need to use more storms in order to derive a representative regression function. However, given the data available (from February to August 2006) this was not possible. Besides, there were several days with no precipitation occurrence given that during this period there was only one rainfall season over this region. In order to check whether the results of comparing diurnal change of total rainfall and storm height were mere coincidence, rainfall data from and independent station were investigated. Data from Ministry of Water and Irrigation, Naivasha were used. Two days, one with long rainfall records and 64

81 the other without rainfall were chosen. Here the day with rainfall records was 1 st March 2006 and the day without rainfall was 28 th October Simple cloud height (SCH) algorithm was applied to compute the cloud heights for the two days. Results to this are presented graphically in figure Cloud height Total rainfall Total rainfall Cloud height 1 Cloud height (m) Total rainfall (mm) Cloud height (m) Total rainfall (mm) :30 3:30 6:30 9:30 12:30 15:30 18:30 21:30 0:30 3:30 6:30 9:30 12:30 15:30 18:30 21:30 Local time (Hrs) Local time (Hrs) Figure 5-5: Diurnal height and Total rainfall changes on 1 st March 2006 (left) and 28 th October 2006 (right) over Naivasha station As can be observed, rainfall occurred generally when cloud height was slightly higher than 3000m especially in the afternoon (the shaded part in the left graph). During this time the likely clouds over the station are convective type of clouds which predominantly occur in the afternoon over this region. Early in the morning, no rainfall was recorded even though the cloud height was slightly more than 3000m. These are likely to be cirrus clouds which mainly occur after dissipation of convective clouds. Thus the convective clouds that produced rainfall in the afternoon and in the night must have dissipated and cirrus clouds appeared in early morning. The right graph of 28 th October 2006 shows that cloud heights above 3000m occurred in the night and there was no rainfall recorded on this day. This may be attributed to the fact that the rainfall was measured at a point and that it may have rained away from the rain gauge. The situation over CGIS station is slightly different as can be observed that on 5 th May 2006 rainfall occurred in the afternoon whereas on 10 th May 2006 it occurred in the morning. This means that convective clouds (the likely clouds producing this rainfall) over this region may be sustained at various times during the 24 hours period. Nevertheless, the results of Naivasha station and those of CGIS station are similar in that rainfall is observed when cloud height was more than 3000m. 65

82 66

83 6. Conclusions and Recommendations 6.1. Conclusions The main objective is to develop simple cloud mask and height algorithms that can be used for further studies. As enumerated some of the most important studies include; operational weather forecasting and energy and water balance studies. Clouds represent the most significant source of error in the extraction of earth surface energy and water balance parameters out of meteorological satellite data (Valk et al., 1998). Energy and water balance models are used to estimate fluxes in cloudy conditions. Thus the focus is to develop a simple cloud mask algorithm in order to be able to accurately develop other algorithms. Further to developing simple cloud mask and cloud height algorithms, rainfall estimation is a focus in this study. Availability of satellite images based on thermal infrared bands is essential and is the first focal point in this study. In addition to this is the importance of geospatial data on meteorological parameters that are associated with formation of clouds. Various sources were explored in order to obtain the long term climatological meteorological data, specifically temperature (minimum, maximum, and mean). Firstly, climatological data from the identified sources were used to process input data for the simple cloud mask algorithm. Secondly, field work campaign was carried for collection of ground rainfall data that was used for comparison with the processed cloud heights on various satellite images. During simple cloud mask algorithm development, various thresholds (multi-spectral threshold technique) were explored in order to optimise on extracting all clouds present on any particular day and time. Based on the developed simple cloud mask algorithm, the following comments were drawn: Setting thresholds for screening all cloudy pixels in satellite images is the most difficult part in threshold techniques. The main problem is that the thresholds are functions of many variables such as; surface type (land, ocean, ice), surface conditions (vegetation, soil moisture), recent weather (which changes surface temperature and reflectance significantly), atmospheric conditions (temperature inversions, haze, foggy), season, time of day and even satellite-earth-sun geometry (hence bidirectional reflectance and sun glint) (Kidder and Haar, 1995). An automatic simple cloud mask algorithm has been presented ready for use in other applications among them those interested in identification of cloudy pixels for the retrieval of cloud-related parameters (e.g. cloud heights) especially those for clouds which contribute to rainfall (e.g. cumulonimbus and nimbostratus). Additionally, exclusion of cloudy pixels for further processing (if required) would be affected by the presence of pixels e.g. for land surface, ocean colour and aerosol observations. Thus given its aim, a compromise between 67

84 calculation speed and accuracy of the results was necessary and use of only three channels of MSG satellite aimed at developing the simple algorithms as per the study objective. The simplicity in the algorithm and significant accuracies based on EUMETSAT data thirsty cloud mask algorithm, and the possibility of automation into shareware or freeware such as ILWIS, may greatly improve cloud detection for specifically weather forecasting in most of the African National Hydrometeorological Services (NHMS). Thus it was envisioned at the developmental phase that this algorithm would be simple and physically sound and that the MSG satellite imageries and the necessary processing tools (software e.g. ILWIS) would be available in these NHMSs in Africa. Based on the simple cloud height/type (SCH/T) algorithm developed and consequently comparison with the observed rain gauge data, the following conclusions were drawn: That dew point temperature concept can be used to estimate cloud height which can thus be used to infer rainfall observed on the earth surface. Despite empirical formulation in obtaining geospatial dew point temperature and replication from a different region (USA Northern Great Plains), high correlations when comparing rain gauge observations and processed cloud heights have been obtained. That satellite convective rainfall estimation schemes using thermal infrared data depend on empirically-derived relations between satellite-observed clouds and rainfall. Worse still, derived relationship from one specific location or climatological regime is not replicable to another and thus general low correlations between satellite data and rain gauge observations. That deriving a concrete regression function for rainfall estimation may be rather difficult from simple data inputs such as cloud height or even cloud top temperature. This may require complex model of high computational strength in order to be able to estimate rainfall to a reasonable accuracy. Besides, earlier studies have shown that unless for strong convection, there is low correlation between VIS/IR features and precipitation. This is the same reason as to why rainfall estimates from the derived regression function in this study were of low accuracies since not all cases were conclusively discerned as convective activities. That there is always need for spatial and temporal averaging of satellite data in order to get better results while comparing point observations on the earth surface Nevertheless, the author is aware that the small area considered for the validation of the cloud mask algorithm may not entirely reflect the overall accuracy of the algorithm. However, this gives an indication of the expected results for specifically equatorial Africa Recommendations Regarding the research methods and ability to improve in the simple cloud mask (SCM) and cloud height/type (SCH/T) algorithms, further research can be considered as follows: 68

85 Improving on thresholds tests based on different cloud microphysical processes on formation of cloud particles. Improving on threshold tests based on variables such as surface type, surface conditions, recent and prevailing weather conditions, and atmospheric conditions. Recalibration or deriving a relationship between dew point temperature and readily available meteorological data e.g. minimum, maximum, and mean temperatures, as suggested by Hubbard et al., (2003) for any region under consideration. On the part of rainfall estimation/comparison method, further research can be considered as follows: Improving on rainfall estimation scheme by determining the environment of the convection (in cases where rainfall is assumed to emanate from convective activities) in terms of temperature, moisture and wind shear. Developing a model that is characterised by the significant transience, heterogeneity, and variability to associate rainfall with the extremely complex and yet imperfectly understood precipitating processes in order to produce higher quality estimates as suggested by Hong et al., (2004). The SCM and SCH/T algorithms seem to work well, but they will benefit a lot from a more thorough validation method. SYNOP data could be used for this purpose. However, as stated by Casanova et al., (2004) subjectivity of the meteorological observer, which in most cases depends on expertise as well as experience, is an issue while using SYNOP as a validation method. Last but not least, final recommendations for development of cloud mask and cloud height algorithms would be that; having observed that earth surface features dictates setting of threshold tests, there is a need to develop cloud mask algorithm that will consider all these features. In addition, it is recommended that rainfall estimation/comparison method be based on accurately known cloud formation processes in order to integrate with one-dimensional physical cloud models. 69

86 References Adler, R.F. and Mack, R.A., Thunderstorm Cloud height - Rainfall Rate Relations for use with Satellite Rainfall Estimation Techniques. Journal of Climate and Applied Meteorology, 23: Adler, R.F. and Negri, A.J., Satellite Infrared Technique to Estimate Tropical Convective and Stratiform rainfall. Journal of Applied Meteorology, 27: Bankert, R.L., Cloud Classification of Avhrr Imagery in Maritime Regions Using a Probabilistic Neural-Network. Journal of Applied Meteorology, 33(8): Casanova, C., Romo, A., Hernández, E., Casanova, J.L. and Sanz, J., Rapid response for cloud monitoring through Meteosat VIS-IR and NOAA-A/TOVS image fusion: civil aviation application. A first approach to MSG-SEVIRI. International Journal of Remote Sensing, 26(8): David, T.,2006.Grib Viewer, February 2007 Derrien, M., Farki, B., Harang, L., Legleau, H., Noyalet, A., Pochic, D. and Sairouni, A., Automatic Cloud Detection Applied to Noaa-11/Avhrr Imagery. Remote Sensing of Environment, 46(3): Derrien, M. and Le Gléau, H., MSG/SEVIRI cloud mask and type from SAFNWC. International Journal of Remote Sensing, 26(21): Desbois, M., Seze, G. and Szejwach, G., Automatic Classification of Clouds on Meteosat Imagery: Application to High-level Clouds. Journal of Applied Meteorology, 21: Dlhopolsky, R. and Feijt, A.J., Cloud products retrieval for MSG: Final Report. BCRS Report 2000 : USP-2 Report Beleidscommissie Remote Sensing (BCRS), Delft, 88 pp Dugdale, G. and Milford, J.R., Rainfall estimation over the Sahel using METEOSAT thermal infra - red data, Proc. ISLSCP conference, 2-6 December 1985 Rome, Italy, pp (ESA - SP ; 248). Ebert, E., A Pattern Recognition Technique for Distinguishing Surface and Cloud Types in the Polar Regions. Journal of Climate and Applied Meteorology, 26: EUMETSAT, METEOSAT SECOND GENERATION (MSG) System Overview, EUM TD 07, Issue 1.1, 25 May EUMETSAT,2006.MSG CHANNELS Interpretation Guide: Weather, Surface conditions and Atmospheric constituents, January 2007 EUMETSAT, 2007.EUMETSAT - Product Generation Description Gieske, A.S.M., Maathuis, B.H.P., Hendrikse, J.H.M., Retsios, V., Leuwen, B., Romaguera, M., Sobrino, J.A., Timmermans, W.J. and Su, Z., Processing of MSG-1 Seviri data in the Thermal Infrared- Algorithm Development with the use of the SPARC2004 data set, ESA 70

87 Proceedings WPP-250: SPARC Final workshop, 4-5 July, ESA, 2005, Enschede, the Netherlands, pp. 8. Grimes, D.I.F., Pardo-Igúzquiza, E. and Bonifacio, R., Optimal Areal Rainfall Estimation using Rain gauges and Satellite data. Journal of Hydrology, 222: Heinemann, T., EUMETSAT MPE Validation Status Report. Hijmans, R.J., Cameron, S.E., Parra, J.L., Jones, P.G. and Jarvis, A., 2005a. Very High Resolution Interpolated Climate Surfaces for Global Land Areas. International Journal of Climatology, 25: Hijmans, R.J., Cameron, S.E., Parra, J.L., Jones, P.G. and Jarvis, A., 2005b.WORLDCLIM, June 2006 Hong, Y., Hsu, K., Sorooshian, S. and Gao, X., 2004a. Precipitation Estimation from Remotely Sensed Imagery using an Artificial Neural Network Cloud Classification System. Journal of Applied Meteorology, 43(12): Hong, Y., Sorooshian, S. and Hsu, K., 2004b. Cloud Patch-based Rainfall Estimation using a Satellite Image Classification Approach, 2nd Workshop of the International Precipitation Working Group, Monterey, October Hubbard, K.G., Mahmood, R. and Carlson, C., Estimating Daily Dew Point Temperature for the Northern Great Plains Using Maximum and Minimum Temperature. Agron J, 95(2): Jedlovec, G.J. and Laws, K., GOES Cloud detection at the Global Hydrology and Climate Center, 12th Conference of Satellite Meteorology and Oceanography, February 9-13, 2003, Long Beach, CA, USA. Jobard, I., Status of Satellite Retrieval of Rainfall at Different Scales using Multi-source data, MEGHA-TROPIQUES 2nd Scientific Workshop, 2-6 July 2001, Paris, France. Kamarianakis, Y., Chrysoulakis, N., Feidas, H. and Kokolatos, G., Comparing Rainfall Estimates Derived from Rain gauges and Satellite Iimages at the Eastern Mediterranean Region, 9th Conference on Geographic Information Science, Visegrád, Hungary. Kidder, S.Q. and Haar, T.H.V., Satellite Meteorology: An Introduction. Academic Press, 466 pp Kidder, S.Q., Kankiewicz, J.A. and Eis, K.E., Meteosat Second Generation Cloud Algorithms for use at AFWA, BACIMO 2005, October 12-14, 2005, Monterey, CA, USA. Kriebel, K.T., Gesell, G., Kástner, M. and Mannstein, H., The Cloud Analysis tool APOLLO: Improvements and Validations. International Journal of Remote Sensing, 24(12): Laurent, H., Jobard, I. and Toma, A., Validation of satellite and ground-based estimates of precipitation over the Sahel. Atmospheric Research, 47-48: Le Borgne, P., Legendre, G. and Marsouin, A., Ocean and Sea Ice SAF Product from MSG data, Proceedings of the 2003 EUMETSAT Meteorological Satellite Conference, 29 September - 3 October EUM P39 (Darmstadt: Eumetsat), Weimar, Germany, pp

88 Levizzani, V., Amorati, R. and Meneguzzo, F., MUSIC : Multiple sensor precipitation measurements, integration, calibration and flood forecasting : A Review of Satellite-based Rainfall Estimation Methods, Istituto di Scienze dell'atmosfera e del Clima, Bologna, Italy. Maathuis, B.H.P., Gieske, A.S.M., Retsios, B., Hendrikse, J.H.M. and Leeuwen, B., MSG Data Retriever: Tool for converting raw MSG SEVIRI L1.5 files into Raster-GIS or Raster image file format. Maathuis, B.H.P., Gieske, A.S.M., Retsios, V., van Leeuwen, B. and Hendrikse, J.H.M., Meteosat-8: from temperature to rainfall. ISPRS 2006 : ISPRS mid-term symposium 2006 remote sensing : from pixels to processes, 8-11 May 2006, Enschede, the Netherlands Marzano, F.S., Cimini, D., Coppola, E., Verdecchia, M., Levizzani, V., Tapiador, E. and Turk, F.J., Satellite Radiometric Remote Sensing of Rainfall Fields: Multi-sensor Retrieval Techniques at Geostationary Scale. Advances in Geosciences, 2: Météo-France, 2005a. O & SI SAF Project team: Atlantic Sea Surface Temperature; Product Manual, version 1.5; Nov 2005: SAF/OSI/M-F/TEC/MA/121. Météo-France, 2005b. User Manual for the PGE of the SAFNWC/MSG: Scientific part SAF/NWC/IOP/MFL/SCI/SUM/01. (Issue 1, Rev.2. 65). NOAA-NODC, 2006.Satellite Oceanography at NODC, ftp://data.nodc.noaa.gov/pub/data.nodc/pathfinder/version5.0_climatologies, June 2006 NOAA-NWS-CPC, 2005.Monitoring and Data, ftp://ftp.cpc.ncep.noaa.gov/wd51we/wgrib2/windows_xp/. February 2007 Saunders, R.W. and Kriebel, K.T., An Improved Method for Detecting Clear Sky and Cloudy Radiances from Avhrr Data. International Journal of Remote Sensing, 9(1): Schröder, M., Bennartz, R., Schüller, L., Preusker, R., Albert, P. and Fischer, J., Generating cloudmasks in spatial high-resolution observations of clouds using texture and radiance information. International Journal of Remote Sensing, 23(20): Stowe, L.L., Davis, P.A. and McClain, E.P., Scientific basis and initial evaluation of the CLAVR-1 global clear cloud classification algorithm for the advanced very high resolution radiometer. Journal of Atmospheric and Oceanic Technology, 16(6): Strahler, A.N., Introduction to Physical Geography. John Wiley & Sons, Inc., New York.London.Sydney Turk, F.J., Bidwell, S.W., Smith, E.A. and Mugnai, A., Investigating Inter-Satellite Calibration for the GPM ERA, 12th Conference on Satellite Meteorology and Oceanography. USGS, 2007.GTOPO30 Documentation, ftp://edcftp.cr.usgs.gov/pub/data/gtopo30/global, February 2007 Valk, P.d., Feijt, A.J., Roozekrans, H., Roebeling, H. and Rosema, A., Operationalisation of an algorithm for the automatic detection and characterisation of clouds in METEOSAT imagery. BCRS Report 1998: USP-2: NRSP-2: NUSP, Beleidscommissie Remote Sensing (BCRS). Wu, X., Menzel, W.P. and Wande, G.S., Estimation of sea surface temperatures using GOES- 8/9 radiance measurements. Bulletin of the American Meteorological Society, 80(6):

89 Appendices Appendix A: ILWIS Script for Simple Cloud Mask and Height Algorithms // STEP 1: DATA IMPORT // // Edit batch file to select date and time for thermal bands (Band IR_039, Band IR_108 and Band //IR_120, imported in Kelvin)! import_t.bat! import_tt.bat! import_ttt.bat! import_tttt.bat // Edit batch file to select date and time for visual bands (band VIS006, Band VIS008 and Band //VIS016, imported as 8 bit) //Only used for visual inspection of the segments that result from processing cloud heights! import_v.bat // Filling the undefined pixels in thermal band 4 (IR_039) by setting minimum brightness //temperature at 204K T1_bd39:=iff((T1_band_1<204),204,T1_band_1) T1_band_1:=ifundef(T1_bd39,204) 73