METEOSAT SECOND GENERATION DATA FOR ASSESSMENT OF SURFACE MOISTURE STATUS Simon Stisen (1), Inge Sandholt (1), Rasmus Fensholt (1) (1) Institute of Geography, University of Copenhagen, Oestervoldgade 10, 1350 Copenhagen K (Denmark) Email: ss@geogr.ku.dk ABSTRACT Two approaches for estimation of remotely sensed land surface temperature, Ts, based on MSG-SEVIRI data have been tested in a semi-arid environment in West Africa. A local algorithm, based on in situ measured Ts and a regional algorithm, based on MODIS LST products. They are both based on the split-window technique and show encouraging results, however applications are limited to the region for which they are derived. The land surface temperature and vegetation index (NDVI) has in combination been used as an indicator of soil moisture and/or evapotranspiration through a dryness index. The index, denoted TVDI, is based on an evaluation of the Ts/NDVI space. The introduction of MSG-SEVERI data in the estimation of Ts and TVDI offers new possibilities in the application of these parameters since the improved temporal resolution allows for analysis of diurnal variations and greater likelihood of finding cloud free pixels. Plots of diurnal variation in TVDI for selected locations show little variation around midday, suggesting that the index is relatively independent of diurnal variations in Ts. 1. INTRODUCTION Remotely sensed surface temperature, Ts, carry important information about the moisture conditions at the land surface, which has been widely acknowledged for decades [1]. Extensive studies on the subject has been carried out and reported in the literature [2] and [3], and it has been shown that the surface temperature and a vegetation index in combination can be used as drought indicator [4], for fire risk assessment [5], for irrigation management, and not the least, for assessment of surface water status and evapotranspiration [6]. The approach is theoretically justified, and straightforward to implement using remotely sensed observations as the only input. A parameterization of the information in the Ts/NDVI space is the Temperature Vegetation Dryness Index (TVDI) suggested by [7]. Many applications rely on information of very dynamic surface moisture conditions in the form of time series, which unfortunately often is hampered by the high sensitivity of the infrared signal to the atmosphere. In this study, the potential of using Meteosat Second Generation SEVIRI data for assessment of surface water status is explored for West Africa, and in particular the diurnal variation in the observed parameters are examined, in order to reach robust indices carrying information about the moisture conditions at the surface. MSG-SEVERI data have the great advantage to data from the polar orbiting satellites, that the frequency of observations is so much higher, thus the chances to avoid cloud cover are much improved. MSG-SEVIRI data are used in this study to derive TVDI maps for the Senegal River basin in West Africa. 2. TEMPERATURE VEGETATION DRYNESS INDEX (TVDI) The partitioning of radiant energy at the land surface is a determining factor for the surface temperature, Ts. Extreme values of Ts will thus occur at surfaces with highest possible evaporative rates, where water supply is unlimited, and for surfaces with no evapotranspiration. The actual values of Ts will depend on the amount of vegetation, the fractional vegetation cover, here represented by a vegetation index, NDVI. It turns out, that the range in Ts is low for high vegetation cover, and high for less dense canopies or bare soil. The resulting Ts/NDVI space, is triangular in shape (Fig. 1), and the boundaries can be estimated using SVAT modeling [3]. Alternatively, the lines of the triangle can be estimated empirically from a scatter plot of observations of Ts and NDVI obtained from satellite sensors, if certain assumptions are met: The atmospheric forcings are uniform over the study area Uniform aerodynamic resistance over the study area Proc. Second MSG RAO Workshop, Salzburg, Austria 9 10 September 2004 (ESA SP-582, November 2004)
A full range of surface conditions (bare soil full canopy, wet-dry) are present in the data set Soil moisture is the main source of variation for Ts The concept is shown in Fig. 1. An index, the Temperature Vegetation Dryness Index, TVDI, identifying the relative location of an observation in the Ts/NDVI space can be defined. Letting observations at the upper edge, denoting the warm, dry edge with no evapotranspiration having a value of 1, and observations at the bottom, cold and wet edge having values of 0, relative moisture conditions can be quantified as the relation between line A and line B in Fig. 1. The resulting equation to derive the dryness index is thus: TVDI T T s s,min = (1) a + b NDVI Ts,min infrared bands themselves contain information on surface temperature, the difference in the band brightness temperatures carries important information for assessment of atmospheric correction. The two thermal infrared channels on the MSG- SEVERI sensor, situated with center bandwidth at 10.8 µm and 12.0 µm are therefore ideal for assessment of land surface temperature [9]. Estimation of land surface temperature is however significantly complicated by the influence of land surface emissivity. Spatial and temporal variation of band specific emissivity is in theory needed to make accurate estimations of land surface temperature, since the emissivity is band specific and thereby affecting the brightness temperature difference, used in the split-window method. This study however uses a simplified split-window approach, assuming a uniform emissivity in both time and space. This approximation yields the following algorithm: TCH9 TCH 10 TCH9 TCH 10 T s = a + b + c (2) 2 2 Where T s is surface temperature in K, T CH9 and T CH10 are brightness temperatures in the two MSG-SEVERI channels and a, b and c are regression constants. Fig. 1. The TVDI index [7] Where a and b are coefficients of the dry edge equation and T s,min is the threshold minimum surface temperature. The parameters a, b and T s,min are estimated empirically from the scatter plot of Ts vs. NDVI. 3. LAND SURFACE TEMPERATURE The split-window approach to land surface temperature estimation was introduced by [8] in the 80 s and has been used in a long range of applications since. The principle behind the split-window approach is the utilization of the atmospheric information within the thermal infrared signals in the 10.8 12.0 µm regions of the electromagnetic spectrum. Whereas the thermal The simplified approach is calibrated against two different data sources, in situ measurements and MODIS LST Products, resulting in a local and a regional algorithm. 3.1 Local algorithm The local algorithm is derived using observations of land surface temperature from a metrological mast located in the very homogeneous savannah landscape of northern Senegal in West Africa. Hourly MSG-SEVERI data for five relatively cloud free days during the rainy season 2003 is compared with corresponding field observations of infrared surface temperature, measured using a Raytech thermometer. Fig. 2 and 3 show the mean brightness temperatures and split-window corrected temperatures as functions of in situ measured Ts respectively. It is evident that the diurnal change in the atmosphere can be corrected for, using the splitwindow algorithm. The regression analysis yields very good results with a R 2 of 0.99 and a RMSE of 1.0 K. The algorithm must
however be regarded as very local and applications are restricted to a specific season and location. Furthermore the comparison of point observations and pixel brightness temperatures is questionable, since the point observation is likely to be considerable higher than the true area averaged LST. This is due to the fact that the savannah landscape is dominated by grasses with patches of trees and bushes. The mast is located in a pure grass area whereas the MSG pixel will encompass some larger vegetation types, which are colder and creates shade and thereby cooling the surface. Keeping that in mind, Fig. 4 shows encouraging results of a remote sensing derived diurnal temperature curve for two days in October 2003. A single split-window algorithm seems able to estimate both daytime and nighttime land surface temperatures. Mean Band9-Band10 MSG [K] 314 312 308 306 304 302 298 296 294 292 Observed [K] Fig. 2. Mean brightness temp. vs. observed LST Split-window corrected MSG LST [K] 335 340 Observed [K] Fig. 3. MSG Split window LST vs. observed LST 00:42 03:42 06:42 07:42 08:42 09:42 10:42 11:42 12:42 13:42 14:42 15:42 16:42 17:42 21:42 00:42 03:42 06:42 07:42 08:42 09:42 10:42 11:42 12:42 13:42 14:42 15:42 16:42 17:42 21:42 01:01 Land Surface Temperature [K] 335 10/10 2003 11/10 2003 Observed LST [K] Estimated MSG LST [K] 00:00 06:00 12:00 18:00 00:00 06:00 12:00 18:00 00:00 06:00 12:00 Fig. 4. Observed and MSG derived LST 3.2 Regional algorithm In order to derive a more regionally justified algorithm a second approach including MODIS LST products as ground truth was carried out. The MODIS 5 km LST product from the Terra and Aqua sensors can produce up to four daily observations of LST at app. 10:30, 13:30, 21:30 and 01:30 local time. Therefore it is believed that these images to some extend can describe the diurnal variation in LST. For the MODIS LST v.4 product temperatures are extracted in Kelvin with a day/night LST algorithm applied to a pair of MODIS daytime and nighttime observations. This method retrieves the land surface temperature and emissivities in bands 20, 22, 23, 29, and 31-32, simultaneously. The MODIS LST algorithm uses MODIS data as input, including geolocation, radiance, cloud masking, atmospheric temperature and water vapor [10]. 12 MODIS LST images from August 2004 were selected, three at each acquisition time. The MODIS LST images, covering one MODIS tile (h16v07) of app. 1100 by 1100 km, were masked using the quality flag to contain pixels with an estimated accuracy of ± 3 K only. To produce the equivalent MSG images the exact acquisition time of the MODIS scenes were extracted from the MODIS LST product and used to create mosaics of MSG bands 9 and 10 that match the acquisition time of the MODIS LST product to within ± 7.5 min. Pixel values from the MODIS LST images and the brightness temperatures from bands 9 and 10 from MSG are extracted without regard to the differences in pixel size, since the scaling problems are regarded as
minor when comparing the 3 km MSG and 5 km MODIS resolutions. Finally a regression analysis is performed for each of the four acquisition times separately and then for all data. Fig. 5 and 6 illustrate the effect of the split window correction on the estimation of MSG LST for three days in August 2004. The scatter plot of DOY 212 is shifted as compared to DOY 223 and 235. This must be due to different atmospheric conditions, since a lower MODIS LST corresponds to a higher MSG brightness temperature compared to DOY 223 and 235. Like the local algorithm the regional algorithm seems to be able to correct for the different atmospheres, since the results in Fig. 6. shows a much more uniform scatter plot including data from all three days, and with a RMSE of 2.65 K (Tab.1.). between acquisition times (Tab.1.). However, if all data are grouped and the regression analysis is performed on all available data the regression analysis yields equally satisfying results with a RMSE of 2.9 K. This can be seen from both Fig. 7 and the RMSE s in Tab. 1. MSG LST [K] 340 All data MSG CH9 CH10 Mean [K] DOY 212 DOY 223 DOY 235 285 280 280 285 MODIS Terra LST day [K] Fig. 5. Mean brightness temp. vs. MODIS LST at 10:30 Approximately 10:30 AM 340 280 280 340 MODIS Terra/Aqua LST day [K] Fig. 7. MSG Split window LST vs. MODIS LST, all data The results are in general satisfying, although the regional algorithm has only been tested on rainy season images and for few atmospheric conditions. The MSG images used in this study were only cloud masked with a very simple threshold value of 280 and 285 K for night and day images respectively. On top of this the cloud mask from the MODIS LST product was applied. Due to sensor angle differences this might not be sufficient masking and the RMSE might be improved if the MSG images were masked more rigorously. MSG LST [K] 280 280 340 MODIS Terra LST day [K] Fig. 6. MSG Split window LST vs. MODIS LST at 10:30 on DOY 212, 223 and 235 Although the results in figures 5 and 6 look encouraging the regression coefficients a, b and c vary greatly Tab. 1. Regression analysis results No. Obs. RMSE [K] a b c Terra day 3,565 2.65-43.17 1.158 4.013 Terra night 2,424 1.95 46.815 0.847 3.333 Aqua day 2,779 3.03 38.621 0.888 2.369 Aqua night 3,913 2.12 124.53 0.587 1.164 All data 12,684 2.90-3.584 1.021 4.296 4. TVDI, RESULTS For two relatively cloud free days in October 2003 the TVDI is calculated for hourly sets of MSG-SEVERI LST and NDVI for a region covering the 375.000 km 2 Senegal River Basin in West Africa. NDVI images were estimated using MSG-SEVERI band 1 (red) and 2 (NIR) without any atmospheric correction. The local
split-window algorithm is used to derive the LST. The Ts/NDVI scatter plot falls within a well defined triangle for all acquisition times, and as expected the triangle moves up and down along the Ts-axes as the LST increases in the morning and decreases in the afternoon. This can be seen from Fig. 8, where the scatter plot for the 9:42, 11:42 and 13:42 o clock images are shown. Ts 345 340 335 11:42 13:42 Ts-NDVI Oct. 11. 2003 Ts-NDVI 13:42 Ts-NDVI 11:42 Ts-NDVI 09:42 Fig. 10 is an illustration of a TVDI map of the Senegal River Basin derived from MSG-SEVERI data for the 10 th of October 2003 at 10:42. The map shows a clear distinction between the irrigated river valley in the North-Western part of the basin and the surrounding drier areas. The low TVDI values in the valley are in agreement with the parameterization of the TVDI since this area represents moist conditions with high evapotranspiration rates. Maps like this can hopefully be applied to water management planning and as validation and calibration of distributed hydrological models. TVDI 10-10 2003 10:42 TVDI [-] High : 1.0 9:42 Low : 0.0 cloud mask 0.0 0.1 0.2 0.3 0.4 0.5 0.6 NDVI Fig. 8. Ts-NDVI scatter plot for hourly MSG data From the TVDI maps for the 11 th of October 2003 three locations are selected to illustrate the diurnal variation in TVDI for three different environments; the irrigated Senegal River valley and a location North and South of the river. These diurnal curves, illustrated in Fig. 9, show relatively small diurnal variations around midday, but a clear separation between the different locations. It therefore seems as if a single midday TVDI value might be representative for a given day, although the robustness of the TVDI-method needs further investigation for several daily observations. TVDI 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 DOY 284 0.1 River valley South of Senegal river North of Senegal river 0 7:12 9:36 12:00 14:24 16:48 19:12 Hour Fig. 9. Diurnal TVDI-curves for three locations around the Senegal River Valley 0 200 400 km Fig. 10. TVDI-map covering the Senegal River Basin 5. CONCLUSIONS AND PERSPECTIVES A local and a regional split window algorithm have been derived using in situ measurements and MODIS Ts data. The performance of the algorithms are satisfactory, both algorithms have RMSE values less than 3 K. The local algorithm has a RMSE value of 1 K, but the use of the algorithms is restricted to the area where it was calibrated. Assuming the MODIS LST data are the best estimate of land surface temperature, our simple split window algorithm which does not take emissivity and angular effects into account, can be used for estimation of Ts from MSG SEVIRI data with an accuracy comparable to MODIS 5 Km LST. The diurnal variation in the Ts/NDVI shape was found to be in agreement with expectations. The triangular shape grows during the day, and collapses towards sundown. The TVDI was estimated for different locations in the test area, and it was found that diurnal variations were less than the variation between the selected locations. The diurnal variation of TVDI was low, however, the values tended to be lower in the
morning and the late evening, when the surface cools off. This is in accordance with theory. MSG-SEVERI data offers the opportunity to investigate spatial variations in surface characteristics in a much improved temporal resolution. In large scale application where the 3 km spatial resolution of MSG-SEVERI is adequate, this will certainly improve the use of remote sensing. One such application could be distributed hydrological modeling of large river basins, where conventional data is usually scarce. Further research is needed on the dynamics of the surface characteristics and the application of MSG- SEVERI data in an energy balance approach to evapotranspiration estimation. 6. REFERENCES 7. Sandholt I., Rasmussen K. and Andersen J., A simple interpretation of the surface temperature/vegetation index space for assessment of surface moisture status, Remote Sensing of Environment, vol. 79, 213-224, 2002. 8. Price C. Land surface temperature measurements from the split window channels of the NOAA 7 Advanced Very High Resolution Radiometer, Journal of Geophysical Research, vol. 89, 7231-7237, 1984. 9. Sobrino J.A. and Romaguera M., Land surface temperature retrieval from MSG1-SEVERI data Remote Sensing of Environment, vol. 92, 247-254, 2004. 10. MODIS/Terra Land Surface Temperature/ Emissivity Daily L3 Global 5km SIN Grid, http://edcdaac.usgs.gov/modis/mod11b1v4.asp, September 7th 2004. 1. Jackson R.D., Reginato R.J. and Idso S.B., Wheat canopy temperature: A practical tool for evaluation of water requirements. Water Resources Research, vol. 13, 651-656, 1977. 2. Moran S., Clarke T.R., Inoue Y. and Vidal A., Estimating crop water deficit using the relation between surface-air temperature and spectral vegetation index, Remote Sensing of Environment, vol. 49, 246-263, 1994. 3. Gillies R.R, Carlson T.N., Cui J., Kustas W.P. and Humes K.S., A verification of the triangle method for obtaining surface soil water content and energy fluxes from remote measurements of the Normalized Difference Vegetation Index (NDVI) and surface radiant temperature, International Journal of Remote Sensing, vol. 18, 3145-3161, 1997. 4. Wan Z., Wang P. and Li X., Using MODIS Land Surface Temperature and Normalized Difference Vegetation Index products for monitoring drought in the southern Great Plains, USA, International Journal of Remote Sensing, vol. 25, 61-72, 2004. 5. Mbow C., Goita K. and Benie G.B., Spectral indices and fire behavior simulation for fire risk assessment in savanna ecosystems, Remote Sensing of Environment, vol. 91, 1-13, 2004. 6. Nishida K., Nemani R.R., Glassy J.M., and Running S.W.,Development of an evapotranspiration index from aqua/modis for monitoring surface moisture status, IEEE Transactions on Geoscience and Remote Sensing, vol. 41, 493-501, 2003.