Precipitation process and rainfall intensity differentiation using Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager data

Transcription

1 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 113,, doi: /2008jd010464, 2008 Precipitation process and rainfall intensity differentiation using Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager data Boris Thies, 1 Thomas Nauß, 1 and Jörg Bendix 1 Received 21 May 2008; revised 6 August 2008; accepted 25 August 2008; published 6 December [1] A new day and night technique for precipitation process separation and rainfall intensity differentiation using the Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager is proposed. It relies on the conceptual design that convective clouds with higher rainfall intensities are characterized by a larger vertical extension and a higher cloud top. For advective-stratiform precipitation areas, it is assumed that areas with a higher cloud water path (CWP) and more ice particles in the upper parts are characterized by higher rainfall intensities. First, the rain area is separated into areas of convective and advective-stratiform precipitation processes. Next, both areas are divided into subareas of differing rainfall intensities. The classification of the convective area relies on information about the cloud top height gained from water vapor-ir differences and the IR cloud top temperature. The subdivision of the advective-stratiform area is based on information about the CWP and the particle phase in the upper parts. Suitable combinations of temperature differences (DT , DT , DT , DT ) are incorporated to infer information about the CWP during nighttime, while a visible and a near-ir channel are considered during the daytime. DT and DT are particularly included to supply information about the cloud phase. Intensity differentiation is realized by using pixel-based confidences for each subarea calculated as a function of the respective value combinations of the previously mentioned variables. For the calculation of the confidences, the value combinations are compared with ground-based radar data. The proposed technique is validated against ground-based radar data and shows an encouraging performance (Heidke skill score for 15-min intervals). Citation: Thies, B., T. Nauß, and J. Bendix (2008), Precipitation process and rainfall intensity differentiation using Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager data, J. Geophys. Res., 113,, doi: /2008jd Laboratory for Climatology and Remote Sensing, Faculty of Geography, Philipps-University Marburg, Marburg, Germany. Copyright 2008 by the American Geophysical Union /08/2008JD Introduction [2] Precipitation as a major part of the global water cycle affects all aspects of human life. However, the high spatial and temporal variability of this parameter still hampers its correct spatiotemporal detection and quantification. In this context geostationary weather satellites with their high spatial and temporal resolution offer the potential for areawide rainfall retrievals. [3] Most of the retrieval techniques based on optical data of geostationary satellite systems rely on a relationship between IR cloud top temperature and the rainfall probability and intensity [e.g., Adler and Negri, 1988]. Such IR retrievals are appropriate for convective clouds, that can be easily identified in the IR and/or water vapor channels [e.g., Levizzani et al., 2001; Levizzani, 2003], but show considerable drawbacks concerning midlatitudinal precipitation processes in connection with extratropical cyclones (hereinafter denoted as advective-stratiform precipitation areas) [e.g., Ebert et al., 2007; Früh et al., 2007]. Such precipitating clouds are characterized by relatively warm and spatially homogeneous cloud top temperatures that differ not significantly between raining and nonraining regions. Therefore, retrieval techniques based solely on IR cloud top temperature lead to an underestimation of the detected rain area and to uncertainties concerning the assigned rain rate. [4] To overcome this drawbacks, several authors suggest to use optical and microphysical cloud parameters derived from multispectral data of new generation satellite systems to improve rainfall retrievals [e.g., Lensky and Rosenfeld, 2003a, 2003b; Ba and Gruber, 2001; Nauss and Kokhanovsky, 2006, 2007; Thies et al., 2008a, 2008b]. [5] The new geostationary system Meteosat Second Generation (MSG) with its payload, the Spinning Enhanced Visible and Infrared Instrument (SEVIRI) [Aminou, 2002], provides the enhanced spectral resolution enabling the retrieval of optical and microphysical cloud properties, together with the application of existing IR retrievals. On the basis of the improved information concerning advectivestratiform precipitation, in combination with the information about convective precipitation, the objective of the present 1of19

2 study is to investigate the potential of MSG SEVIRI to differentiate the rain area into subareas of differing precipitation processes and rainfall intensities. [6] A proper rainfall intensity classification offers the potential for an improved process-oriented rain rate assignment which represents the next step for an enhanced satellite-based rainfall retrieval. The determination of distinct regression functions between cloud properties and the rain rate for each subarea can improve the representativeness of these functions and improve rain rate assignment. For this reason, the focus of the present study lies on the differentiation of subareas of differing rainfall intensities. In this context, the high spatial (3 by 3 km at subsatellite point) and especially temporal (15 min) resolution provided by MSG SEVIRI is of special benefit, as it permits a quasicontinuous observation of the rainfall distribution in nearreal time, and allows to account for short-time precipitation dynamics, such as short-living convective events with high rainfall intensities. [7] The structure of the article is as follows. In section 2 the data and methods are introduced. Section 3 gives a presentation of the theoretical background and the conceptual design of the proposed retrieval scheme, which is introduced in section 4. An appraisal of the developed scheme follows in section 5. Finally, section 6 gives a short summary and some conclusions. 2. Data and Methods [8] For this study, MSG SEVIRI data together with corresponding ground-based radar data are required. The SEVIRI data have been received at the Marburg Satellite Station (MSS) [Reudenbach et al., 2004; Bendix et al., 2003]. The raw data have been processed by the FMet tool [Cermak et al., 2008]. The radar data are provided by the ground-based C band radar network of the German Weather Service [Deutscher Wetterdienst, 2005; Seltmann, 1997] (Deutscher Wetterdienst report available from Content/Oeffentlichkeit/TI/TI2/Downloads/Radarbrosch_ C3_BCre,templateId = raw,property = publicationfile.pdf/ Radarbrosch%C3%BCre.pdf). The SEVIRI data have a nominal spatial resolution of 3 by 3 km at sub satellite point, while that of the radar data is 4 by 4 km. The scan interval for both data sets is 15 min. Because of the differing viewing geometries between both systems the radar data were projected to the viewing geometry of SEVIRI. The projection, which is realized by a backward resampling with next neighbor, assures the pixel matching between satellite and radar data, necessary for the spatial comparison of both data sets. [9] The data sets consist of 2955 scenes from precipitation events between January and August For this study the data are divided into a training and a validation data set. The training data set is used for the development of the technique and consists of 1559 scenes of precipitation events from January to August The validation data set is considered for the appraisal of the proposed technique and contains 1396 scenes of precipitation events from January to August The precipitation events taken for the training data set are independent from those taken for the validation data set. Because of the differing information content about the cloud properties between daytime and nighttime scenes both data sets are divided into daytime and nighttime scenes (refer to section 4.2.2). The training data set consists of 850 daytime and 709 nighttime scenes. The validation data set comprises 720 daytime and 676 nighttime scenes. [10] For the development and the appraisal of the proposed retrieval technique standard verification scores for dichotomous data sets are applied. The selection of the appropriate scores, which are shortly introduced in the following paragraphs, follows the suggestions of the International Precipitation Working Group (IPWG) [Turk and Bauer, 2006]. The respective subarea mentioned in the description represents each of the five rainfall intensity areas classified by the new developed technique (see section 3 and 4). [11] The bias describes the ratio between the number of pixels that have been assigned to the respective subarea by the satellite and the radar techniques. The probability of detection (POD) describes the fraction of pixels that have been correctly identified by the satellite technique, according to the radar product. The probability of false detection (POFD) indicates the fraction of the pixels incorrectly identified by the satellite algorithm. The false alarm ratio (FAR) gives the fraction of pixels classified as the respective subarea by the satellite technique that actually were not classified by the radar. The Critical Success Index (CSI) indicates how well the classified pixels of the respective subarea correspond to the pixels of the respective subarea observed by the radar. All scores range from 0 to 1. The optimum value for the CSI is 1, while it is 0 for the FAR. Since the POD can be increased by just increasing the respective subarea, it has to be analyzed in connection with corresponding values of the FAR and the POFD since both measure the fraction of the satellite pixels, that have been incorrectly assigned to the respective subarea. The equitable threat score (ETS) indicates how well the classified pixels of the respective subarea correspond to the pixels of the respective subarea observed by the radar, also accounting for pixels correctly classified by chance. Its values range from 1/3 to 1 with the optimum value 1. The Heidke skill score (HSS) also considers the probability of correctly classified pixels by chance. In contrast to the ETS the pixels correctly classified as not belonging to the respective subarea are also incorporated for the calculation of the HSS. It ranges from negative infinite to 1. The optimum value is 1. The Hansen-Kuipers discriminant (HKD) is a measure how well the satellite based technique can distinguish between pixels of the respective subarea and pixels of differing subareas. It ranges from 1 to 1 with an optimum value of 1. For a detailed discussion of the verification scores see Stanski et al. [1989] or the Web site of the WWRP/WGNE (see staff/eee/verif/verif_web_page.html). 3. Theoretical Background and Conceptual Design [12] The following sections present the theoretical background (section 3.1) together with the conceptual design (section 3.2) of the proposed technique for precipitation 2of19

3 Table 1. Overview of the Subareas of Differing Precipitation Processes and Rainfall Intensities According to the Conceptual Model of Rainbands Introduced by Houze [1993] Conceptual Model precipitation from convective cores in connection with narrow cold-frontal rainbands convective stratiform precipitation in connection with convective cores within narrow cold-frontal rainbands enhanced advectivestratiform precipitation produced by embedded generating cells in connection with wide coldfrontal rainbands and warm-frontal rainbands intermediary precipitation from clouds of intermediary character between convectively and advectively dominated precipitation areas advective-stratiform background precipitation from clouds within the frontal rainband Radar very high rainfall intensities with thunderstorms moderate to high rainfall intensities possibly with thunderstorms moderate rainfall intensities light to moderate rainfall intensities light rainfall intensities Radar Reflectivity (dbz) Rainfall Rate (mm/h) >46.0 > to to to to to to to to 0.4 process and rainfall intensity differentiation of the detected rain area Theoretical Background [13] On the basis of the results of the Cyclonic Extratropical Storms project, Houze [1993] summarized the conceptual model of rainbands dominated by different rainfall processes and leading to differing rainfall intensities within extratropical cyclones. According to this conceptual model, the rainfall area is separated into the following subareas of different precipitation processes and rainfall intensities: (1) precipitation from convective cores in connection with narrow cold-frontal rainbands (very high precipitation intensities from Cb); (2) convective-stratiform precipitation from Ns in connection with convective cores (Cb) within narrow cold-frontal rainbands (moderate to high precipitation intensities); (3) enhanced advective-stratiform precipitation from Ns with ice particles from embedded shallow generating cells, that fall from above ( seederfeeder effect) in connection with wide cold-frontal rainbands and warm-frontal rainbands (moderate precipitation intensities); (4) precipitation from clouds of intermediary character between convectively dominated and advectivestratiform background precipitation areas (light to moderate precipitation intensities); and (5) advective-stratiform background precipitation (light precipitation intensities). Following the conceptual model of rainbands in connection with extratropical cyclones it can be stated that narrow coldfrontal, wide cold-frontal and warm-frontal rainbands are dominated by convective precipitation processes. On the other hand the advective-stratiform background and the intermediary precipitation area are of advective-stratiform character. [14] The above mentioned five subareas of differing precipitation processes and rainfall intensities can also be identified in the radar product of the German Weather Service (DWD) [Deutscher Wetterdienst, 2005]. Table 1 gives an overview of the five subareas of differing precipitation processes and intensities according to the conceptual model of rainbands, introduced above, together with the corresponding radar reflectivity and the characteristic rainfall intensity Conceptual Design [15] According to the precipitation processes in connection with extratropical cyclones, convectively dominated precipitation areas are characterized by a large vertical extension and a cold cloud top rising high into the atmosphere. As a result, the established relationship between cloud top temperature and rainfall probability and intensity can be applied for the detection and classification of these precipitation areas. At the same time a major part of the precipitating cloud areas in connection with extratropical cyclones are not necessarily connected to cold cloud top temperatures. As a consequence, a threshold for cloud top temperature does not seem to be effective for the detection and classification of these precipitation areas. [16] For this reason, several authors successfully used optical and microphysical cloud parameters derived from multispectral data of new generation satellite systems for an improved rain area delineation [e.g., Nauss and Kokhanovsky, 2007, 2006; Lensky and Rosenfeld, 2003a, 2003b]. They could show that cloud areas with a high cloud water path (large a ef together with high t) possess a high amount of cloud water and are characterized by a higher rainfall probability than cloud areas with a low cloud water path (CWP). However, beside the optical and microphysical cloud properties the cloud phase in the upper cloud parts has also to be incorporated due to two reasons: (1) effective rain formation processes in connection with extratropical cyclones are mainly coupled to ice particles in the upper part of the clouds and the seeder-feeder effect [Houze, 1993] and (2) CWP-rain relationships are blurred especially in the ice phase by the unknown shape of ice crystals. [17] Considering the dominant precipitation processes for convective and advective-stratiform precipitation areas within extratropical cyclones in connection with an aspired precipitation process and rainfall intensity differentiation within the rain area the following conceptual design is formulated: convective rain clouds with higher rainfall intensities are characterized by a larger vertical extension and a cloud top reaching higher into the atmosphere, and advective-stratiform precipitation areas with a higher cloud water path and a higher amount of ice particles in the upper cloud regions are characterized by higher rainfall intensities. Hence, it is expected that information about the cloud top height, the cloud water path and the cloud phase in the upper cloud parts enable the separation of the rainfall area into subareas of different precipitation processes and rainfall intensities. MSG SEVIRI provides an enhanced spectral resolution, which allows to infer information about the 3of19

4 Figure 1. Overview of the stepwise classification scheme for precipitation process and rainfall intensity differentiation. cloud water path, the cloud phase in the upper cloud parts and the cloud top. On the basis of the conceptual design introduced above, the possibility to differentiate the rain area into subareas of differing rainfall intensities is investigated in the present study. 4. Process Separation and Rainfall Intensity Differentiation With Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager [18] On the basis of the theoretical background presented in section 3 together with the conceptual design, the following section describes the implementation of the precipitation process separation and rainfall intensity differentiation technique for MSG SEVIRI. [19] As mentioned in the introduction, the focus of the present study lies on the differentiation of precipitation processes and rainfall intensities within the rain area. In order to develop and evaluate such a satellite-based differentiation technique, the data of the ground-based radar network of the German Weather Service are used. This means that the rain area detected by the radar network is taken as basis for the investigations. Additionally, the definition of the rainfall intensity classes separated by the developed satellite scheme follows the dbz classification of the DWD radar data (see Table 1). [20] The proposed satellite-based technique aims to separate subareas of (1) convective and advective precipitation processes and (2) to subdivide the convective and advective rain areas in classes of differing rainfall intensities. For the differentiation of the rainfall area in process/intensity classes, a stepwise classification algorithm similar to the cascade method proposed by Capacci et al. [2008] is applied (see Figure 1). After the separation of the convective and advective-stratiform precipitation processes (section 4.1) both subareas are divided into regions of differing rainfall intensities (section 4.2). First, the convective cores together with the convective-stratiform areas are separated from the cloud fields of enhanced advective-stratiform precipitation. Then the convective cores are distinguished from the convectivestratiform precipitating cloud areas (section 4.2.1). Finally, the intermediary precipitation fields are separated from the advective-stratiform background precipitation area (section 4.2.2) Precipitation Process Separation [21] First, the rainfall area is divided into convective and advective-stratiform precipitation areas. The technique follows the suggestion of Houze [1997] and Steiner et al. [1995] to detect the convective rain area and denote the remaining subarea as stratiform plus intermediary. [22] As introduced in section 3.2, the detection of convective precipitation areas is based on information about the cloud top height. Beside the cloud top temperature in the 10.8 mm channel (BT 10.8 ) the brightness temperature difference between the water vapor (WV) and the IR channels are considered to gain information about the cloud top height. This is based on the observation that the brightness temperature difference between a WV and an IR channel allows a more reliable identification of deep convective clouds [see also Tjemkes et al., 1997; Schmetz et al., 1997; Reudenbach et al., 2001]. [23] Concerning the two WV and two IR channels of SEVIRI, the studies of Thies et al. [2008c; also Detection of high rain clouds using water vapor emission Transition from Meteosat First (MVIRI) to Second Generation (SEVIRI), submitted to Advanced Space Research, 2007] revealed that the four possible WV-IR combinations show different sensitivities to the cloud top height at different levels, which provide useful information for the detection and classification of convectively dominated precipitation areas. To incorporate the different sensitivities on cloud top height into the detection scheme, the two channel differences DT WV6.2 IR10.8 and DT WV7.3 IR12.1 have been chosen. To gain information about the cloud top temperature BT 10.8 is also included. This was done to consider the maximum possible information content concerning the rising top of a convective cloud within the troposphere and the interaction of the emitted radiation with the water vapor above cloud top level. [24] To use the information about cloud top height from the two WV-IR channel differences and BT 10.8 for the process separation, the confidences of the convectively dominated rainfall areas are computed as a function of the combined values of the three parameters DT WV6.2 IR10.8, 4of19

5 Figure 2. Overview of the introduced scheme for precipitation process separation as part of the stepwise classification. DT WV7.3 IR12.1, and BT The use of confidences has been successfully applied for rain area delineation [e.g., Thies et al., 2008a, 2008b] and is employed here to separate the different precipitating cloud areas. Figure 2 gives an overview of the introduced scheme for the separation of the two precipitation processes, which is described in the following. [25] The computation of the pixel based confidences of the convective rain areas is done by a comparison of DT WV6.2 IR10.8, DT WV7.3 IR12.1, and BT 10.8 with the ground-based radar for the training data set (altogether 1559 scenes, see section 2). For the calculation of the confidences of the convective precipitation areas the radar classes between 7.0 and 27.9 dbz are considered as advective-stratiform and the radar classes with a reflectivity higher than 28.0 dbz are considered as convectively dominated (see Table 1). [26] Figures 3a and 3b show the confidences of the convective precipitation areas as a function of the two WV-IR differences as well as of BT 10.8 calculated with equation (1): Confidence A ðx 1 ; x 2 Þ ¼ N A ðx 1 ; x 2 N A ðx 1 ; x 2 ÞþN B x 1 ; x 2 Þ ð Þ ; ð1þ where Confidence A denotes the confidences of the convectively dominated precipitation area. N A and N B are the frequencies of the convectively dominated pixels and the advective-stratiform pixels, respectively; x 1 and x 2 denote the respective channel differences (DT WV6.2 IR10.8, DT WV7.3 IR12.1 ) and BT 10.8 combined for the calculation of the confidences. [27] As can be seen in Figure 3a high DT WV6.2 IR10.8 values together with high DT WV7.3 IR12.1 values reveal high convective rainfall confidences. The same holds true for low BT 10.8 together with high DT WV6.2 IR10.8 values (Figure 3b). This corroborates the findings of Thies et al. [2008c; B. Thies et al., submitted manuscript, 2007] that DT WV6.2 IR10.8 and DT WV7.3 IR12.1 increase with increasing cloud top height. At the same time, the higher the cloud reaches into the atmosphere the lower becomes BT Thus, to make use of the combined information content about cloud top height available from the two WV-IR channel differences and BT 10.8 for the separation of convective precipitation fields, the confidences of convective rain areas are computed as a function of the combined values of the three parameters as shown in equation (2): Confidence A ðx 1 ; x 2 ; x 3 Þ ¼ N A ðx 1 ; x 2 ; x 3 Þ ð Þ ; ð2þ N A ðx 1 ; x 2 ; x 3 ÞþN B x 1 ; x 2 ; x 3 where Confidence A denotes the confidences of the convective precipitation areas. N A and N B are the frequencies of the convectively dominated pixels and the advectivestratiform pixels, respectively, and x 1, x 2, and x 3 denote the three variables DT WV6.2 IR10.8, DT WV7.3 IR12.1, and BT 10.8 combined for the calculation of the confidences of the convective precipitation areas. The confidence threshold appropriate for the separation of the convective precipitation fields is determined by optimizing the ETS (see section 2). Different confidence threshold values between 0.1 and 0.7 were used to separate the satellite-based convective rain area. The ETS for the separated convective fields based on the different confidence levels were calculated in comparison with ground-based radar data. The detected convective rain area using a confidence threshold of 0.15 yields the maximum ETS of and is taken as the minimum threshold. Pixels with value combinations possessing a confidence higher than the minimum threshold are classified as convectively dominated. The remaining area is referred to as advective-stratiform precipitating cloud area. [28] To illustrate the different classification steps, Figure 4 shows an example for the separation between convective and advective-stratiform precipitation processes for the 18 July 2004, 2145 UTC. Figures 4a and 4b display the WV-IR differences within the rain area detected by the radar, the BT 10.8 are depicted in Figure 4c. For each pixel within the rain area a value combination of the three variables (DT WV6.2 IR10.8, DT WV7.3 IR12.1, BT 10.8 ) is obtained. The confidences for the convective area as a function of the three variables have been calculated using the training data set (see equation (2)). On the basis of these calculations, a confidence value is assigned to the respective value combination of each pixel. The resulting confidences are depicted in Figure 4d. In the next step the confidences of each pixel are compared to the minimum confidence threshold for the separation of the convective area of Pixels 5of19

6 Figure 3. Calculated confidences of the convective precipitation areas in contrast to the advectivestratiform precipitation areas as a function of two variables (a) DT WV6.2 IR10.8 versus DT WV7.3 IR12.1 and (b) DT WV6.2 IR10.8 versus BT 10.8, where WVis water vapor. Calculated confidences of the convective cores together with convective-stratiform precipitation areas in contrast to the enhanced advective-stratiform precipitation cloud fields as a function of two variables (c) DT WV6.2 IR10.8 versus DT WV7.3 IR12.1 and (d) DT WV6.2 IR10.8 versus BT Calculated confidences of convective cores in contrast to the convective-stratiform precipitation areas as a function of two variables (e) DT WV6.2 IR10.8 versus DT WV7.3 IR12.1 and (f) DT WV6.2 IR10.8 versus BT with a confidence higher than the minimum threshold are classified as convectively dominated. The classification result can be seen in Figure 4e. Figure 4f shows the corresponding convective area classified by the radar data. The differentiation of the convective rain area into subareas of differing rainfall intensities is performed analogously to the presented example Intensity Differentiation [29] After the separation of the rain area into convective and advective-stratiform precipitation areas, the division into subareas of different rainfall intensities is described in the following Subdivision of the Convective Precipitation Areas [30] As for the detection of the convective precipitation area the classification into the three subareas of different rainfall intensities (precipitation from convective cores, convective-stratiform precipitation, enhanced advectivestratiform precipitation; see section 3.1) is accordingly done by means of confidences of the respective subarea and an 6of19

7 Figure 4. Example for the separation between convective and advective-stratiform precipitation process areas for 18 July 2004, 2145 UTC. (a, b) WV-IR differences (DT WV6.2 IR10.8, DT WV7.3 IR12.1 ). (c) BT (d) Assigned confidences on a pixel basis. (e) Classification result based on the confidences. (f) Corresponding convective area classified by the radar data. appropriate minimum threshold. Figure 5 gives an overview of the introduced scheme for the division of the convective rain areas into subareas of differing rainfall intensities. [31] The differentiation of the convective precipitation area into subareas of differing rainfall intensities also relies on the WV and IR channels available during daytime and nighttime. For that reason, the computation of the pixel based confidences by comparing the two WV-IR channel differences and BT 10.8 with ground-based radar data is applied for the whole training data set consisting of 1559 precipitation scenes from January to August 2004 (see section 2). [32] The classification is based on the conceptual model that convective rain clouds with higher rainfall intensities are characterized by a larger vertical extension and a cloud top reaching higher into the atmosphere (see section 3.2). 7of19

8 Figure 5. Overview of the introduced scheme for the division of the convective rain areas into subareas of differing rainfall intensity. First, the convective cores together with the convectivestratiform precipitation areas are separated from the areas of enhanced advective-stratiform precipitation. The calculation of the confidences of convective cores together with the convective-stratiform precipitation areas in contrast to the areas of enhanced advective-stratiform precipitation is done in a comparison with the ground-based radar data, where radar classes with reflectivities greater than 37.0 dbz are contrasted to the radar classes with reflectivities between 28.0 to 36.9 dbz (refer to Table 1). [33] Figures 3c and 3d show the calculated confidences of convective cores together with convective-stratiform precipitation areas in contrast to the enhanced advectivestratiform precipitation cloud fields as a function of the two WV-IR differences and BT 10.8, respectively, calculated in analogy with equation (1). As can be seen the combinations of high DT WV6.2 IR10.8 and high DT WV7.3 IR12.1 values possess high confidence levels (Figure 3c). The same holds true for low BT 10.8 together with high DT WV6.2 IR10.8 (Figure 3d). This is in accordance with the assumption introduced above that higher values of DT WV6.2 IR10.8 and DT WV7.3 IR12.1 as well as low values of BT 10.8 signify higher rainfall intensities. [34] To make use of the combined information content available from DT WV6.2 IR10.8, DT WV7.3 IR12.1, and BT 10.8 the confidences of convective cores together with convective-stratiform precipitation areas are computed in contrast to the enhanced advective-stratiform precipitation areas as a function of the combined values of the three parameters analogously to equation (2). [35] The determination of the appropriate confidence threshold to separate the areas of convective cores and convective-stratiform precipitation from the enhanced advective-stratiform precipitation areas is done by maximizing the ETS (refer to section 4.1). The confidence threshold of 0.19 yields to the maximum ETS of and is used as the minimum threshold. Pixels with value combinations possessing a confidence higher than the minimum threshold are classified as convective cores and convective-stratiform precipitating cloud fields. The remaining convective rain areas are classified as enhanced advective-stratiform precipitation. [36] In the next step the convective cores are separated from the convective-stratiform precipitation areas. For this purpose the confidences of convective cores (radar reflectivity greater than 46.0 dbz) are calculated in contrast to the convective-stratiform precipitation areas (radar reflectivities between 37.0 and 45.9 dbz) following equation (1) analogously to the procedure presented above. Figures 3e and 3f show the calculated confidences of the convective cores in contrast to the convective-stratiform precipitation areas as a function of the two WV-IR differences and BT 10.8 respectively. In accordance to the assumption introduced above, the combination of high DT WV6.2 IR10.8 and high DT WV7.3 IR12.1 values possess high confidence levels (Figure 3e). The same holds true for low BT 10.8 values together with high DT WV6.2 IR10.8 values (Figure 3f). 8of19

9 Figure 6. Overview of the introduced scheme for the division of the advective-stratiform precipitation areas into the subareas of intermediary and advective-stratiform background precipitation during daytime and nighttime. [37] To make use of the combined information content available from the two WV-IR channel differences and BT 10.8, the confidences of convective cores are computed as a function of the combined values of the three parameters as shown in equation (2). [38] The determination of the appropriate confidence threshold to separate the convective cores from the convective-stratiform precipitation areas is done by maximizing the ETS again (refer to section 4.1). The confidence threshold of 0.23 yields to the maximum ETS of and is used as a minimum threshold. Pixels with value combinations possessing a confidence higher than the minimum threshold are classified as convective cores. The remaining rain areas are classified as convective-stratiform precipitating cloud fields Subdivision of the Advective-Stratiform Precipitation Areas [39] After the differentiation of the convective precipitation areas the further division of the advective-stratiform precipitation areas is described in the following. Because of different band combinations available day and night, the subdivision of the advective-stratiform precipitation areas is done by two separate daytime and nighttime algorithms. Figure 6 gives an overview of the proposed scheme for the division of the advective-stratiform precipitation areas into the subareas of intermediary and advective-stratiform background precipitation during daytime and nighttime Daytime Algorithm [40] Because no fast enough operational technique for the explicit retrieval of optical and microphysical properties (a ef, t and CWP), of water and ice clouds is currently available for MSG SEVIRI, the authors decided to use the original reflectance of the mm (VIS 0.6 ) and mm (NIR 1.6 ) SEVIRI channels, inherently encompassing this information. [41] Information about the cloud phase in the upper parts of the cloud are gained by considering the brightness temperature difference between the 8.7 mm channel 9of19

10 Figure 7. Calculated confidences of the intermediary precipitation areas in contrast to the advectivestratiform background precipitation areas as a function of two variables (a) VIS 0.6 versus NIR 1.6 and (b) DT versus DT for daytime scenes. Calculated confidences of the intermediary precipitation areas in contrast to the advective-stratiform background precipitation areas as a function of two channel differences (c) DT versus DT , (d) DT versus DT , and (e) DT versus DT for nighttime scenes. ( mm) and the 10.8 mm channel ( mm) (DT ) together with the brightness temperature difference between the 10.8 mm channel and the 12.1 mm channel (11 13 mm) (DT ) [see Strabala et al., 1994]. [42] To use the implicit information about the CWP and the cloud phase, inherent in two channels VIS 0.6,NIR 1.6 and the two channel differences DT , DT for a division of the advective-stratiform precipitation area, the confidences of the intermediary precipitating cloud fields (radar reflectivity between 19.0 and 27.9 dbz) are calculated in contrast to the advective-stratiform background precipitation areas (radar reflectivity between 7.0 and 18.9 dbz) as a function of the value combinations of the four variables VIS 0.6,NIR 1.6, DT , and DT The computation of the pixel based confidences is done by a comparison of the four value combinations with groundbased radar data for daytime precipitation events of the training data set (altogether 850 scenes, see section 2). [43] Figures 7a and 7b show the calculated confidences of the intermediary rain areas as a function of VIS 0.6 and NIR 1.6 (Figure 7a), as well as a function of DT and DT (Figure 7b) computed analogously to equation (1). As can be seen in Figure 7a high confidences coincide with high values of VIS 0.6 and low values of NIR 1.6, indicating a higher CWP. High values of VIS 0.6 are char- 10 of 19

11 acteristic for higher t and lower values of NIR 1.6 are the result of larger a ef as the absorption increases with increasing particle size. Figure 7b shows that ice clouds, where DT values are greater than coincident DT values, are characterized by high confidence levels. On the other hand, for water clouds DT values are greater than DT values [compare Strabala et al., 1994]. These areas are characterized by lower confidence levels. [44] To summarize, value combinations representative for large CWP and ice particles in the upper cloud parts are characterized by higher confidence levels for the intermediary precipitation area with higher rainfall intensities. This corroborates the assumption, formulated above, that raining cloud areas characterized by a higher CWP and the existence of more ice particles in the upper regions produce higher rainfall intensities. [45] To make use of the combined information about the CWP and the cloud phase provided by the parameters VIS 0.6,NIR 1.6, DT , DT , the confidences of the intermediary precipitation areas are computed as a function of the combined values of the four variables as shown in equation (3) using the above mentioned 850 scenes: Confidence A ðx 1 ; x 2 ; x 3 ; x 4 N A ðx 1 ; x 2 ; x 3 ; x 4 Þ Þ ¼ N A ðx 1 ; x 2 ; x 3 ; x 4 Þþ N B ðx 1 ; x 2 ; x 3 ; x 4 Þ ; ð3þ where Confidence A denotes the confidences of the intermediary rain area. N A represents the frequencies of intermediary precipitating pixels, while N B stands for the frequencies of the pixels with advective-stratiform background precipitation; x 1, x 2, x 3 and x 4 denote the four variables VIS 0.6,NIR 1.6, DT , DT combined for the calculation of the confidences of the intermediary rain areas. [46] The threshold of the calculated confidences appropriate for a delineation of the intermediary precipitation fields is determined by optimizing the ETS (refer to section 4.1). The confidence threshold of 0.34 yields the maximum ETS of and is therefore used as minimum threshold. Pixels with value combinations possessing a confidence higher than the minimum threshold are classified as intermediary precipitation. The remaining areas are classified as advective-stratiform background precipitation Nighttime Algorithm [47] As for daylight, there is no operational retrieval for MSG SEVIRI at hand, that can explicitly compute the CWP necessary for rainfall intensity differentiation during nighttime. Anyhow, several case studies have shown that implicit information about a ef and t and therefore the CWP is available in the emissive channels during nighttime as well [e.g., Huang et al., 2004; Lensky and Rosenfeld, 2003a, 2003b; González et al., 2002; Ackerman et al., 1998; Baum et al., 1994; Ou et al., 1993]. On the basis of the findings of these case studies, Thies et al. [2008a] used radiative transfer calculations to demonstrate that implicit information about the CWP is available from the four SEVIRI channel differences DT , DT , DT , DT [48] This is due to differing sensitivities of the respective channels on microphysical and optical cloud properties, which is explained here exemplarily for DT A large CWP is the product of a large effective particle radius (a ef ) and a high optical thickness (t). Large particles have a higher emission in the 3.9 mm channel compared to smaller particles. This is due to the increased scattering for smaller particles which reduces the cloud emissivity. Therefore, the brightness temperature in the 3.9 mm channel is higher for larger particles. This dependence on particle size is much less distinct in the 10.8 mm channel, which leads to higher DT for larger particles. For thinner clouds the emission in the 10.8 mm channel is higher than in the 3.9 mm channel. As a result, the 3.9 mm transmittance is larger than the 10.8 mm transmittance, which implies a larger transmissivity of radiance from lower levels of the former wavelength [see Lensky and Rosenfeld, 2003b]. Thus, for clouds with a small CWP (small a ef with low t), the brightness temperature of the 3.9 mm channel is larger than that of the 10.8 mm channel and DT reaches the highest values. Clouds with high CWP (large a ef with high t) result in medium to high difference values. For further explanations refer to Thies et al. [2008a] and the literature cited therein. [49] Thies et al. [2008a] also demonstrated that there is accordance between channel difference values indicating higher cloud water paths and channel difference values indicating higher rainfall confidences of cloud areas. They successfully used this relation between CWP and rainfall confidence to separate raining from nonraining cloud areas. [50] To make use of the implicit information about the CWP and the cloud phase inherent in the four channel differences DT , DT , DT ,andDT for the division of the advective-stratiform precipitation areas during nighttime, the confidences of the intermediary precipitation areas are calculated in contrast to the advectivestratiform background precipitation areas as a function of the four channel differences (DT , DT , DT , DT ) by using equation (1). The computation of the pixel based confidence of the intermediary precipitation areas is analogous to the daytime scheme and is done by a comparison of the SEVIRI channel differences with ground-based radar data for nighttime precipitation events of the training data set (altogether 709 scenes, see section 2). [51] The confidences of the intermediary precipitation areas as a function of two different channel differences calculated with equation (1) are depicted in Figures 7c 7e. For the combination of DT versus DT (Figure 7c) higher confidences can be found for medium DT and medium DT values. These intervals coincide with those for a larger CWP. Lower confidences are characterized by high DT and high DT , which correspond to a lower CWP [see Thies et al., 2008a]. Regarding the combination of DT versus DT (Figure 7d), higher confidences are indicated for medium DT and large DT values. These intervals correspond to a larger CWP. Lower confidences can be found for high DT and small DT values, which coincide with a lower CWP [see Thies et al., 2008a]. Concerning the combination of DT and DT (Figure 7e) higher confidences are indicated for medium DT and small DT values, which coincide with a larger CWP. Lower confidences can be 11 of 19

12 Table 2. Overview of the Calculation of the Confidences and the Thresholds Used for the Separation of the Respective Subareas, Together With the Optimized Equitable Threat Score and the Corresponding Radar Reflectivities a Confidence Confidence A N A N B N A (dbz) N B (dbz) threshold ETS > to Convectively dominated precipitation area convectively dominated precipitation area advectivestratiform precipitation area Convective cores and convectivestratiform precipitation area convective cores and convectivestratiform precipitation area enhanced advectivestratiform precipitation area > to Convective cores convective cores convectivestratiform precipitation area > to Intermediary rain area intermediary rain area advectivestratiform background precipitation area a Abbreviation is as follows: ETS, equitable threat score to to (daytime) 0.23 (nighttime) (daytime) (nighttime) found for high DT and high DT values, which correspond to a lower CWP [see Thies et al., 2008a]. [52] To summarize it can be stated that value combinations indicating higher CWP values are characterized by higher confidence levels for the intermediary precipitation area with higher rainfall intensities. This corroborates the conceptual model, introduced in section 3.2, that rain areas with a higher CWP are characterized by higher rainfall intensities and possess higher confidence levels. [53] To make use of the combined information about the CWP and the cloud phase provided by the four channel differences DT , DT , DT , and DT , the confidences of the intermediary rain area are computed as a function of the combined values of the four variables as shown in equation (3) using the above mentioned 709 nighttime scenes. The confidence threshold appropriate for rain area separation is determined analogously to the daytime scheme leading to a confidence threshold of 0.23 (ETS of ), which is chosen as the minimum threshold. [54] Table 2 shows the calculation of the confidences of the respective subareas together with the threshold values, the optimized ETS value as well as the corresponding radar reflectivity thresholds. [55] To illustrate the separation between the area of advective-stratiform background precipitation and the area of intermediary precipitation, Figure 8 shows the scene from 18 July 2004, 2145 UTC. Figures 8a 8d display the channel differences used for the intensity differentiation (DT , DT , DT , DT ) within the rain area detected by the radar. For each pixel within the rain area a value combination of the four channel differences is obtained. The confidences for the intermediary area as a function of the four variables have been calculated using the training data set (see equation (3)). On the basis of these calculations, a confidence value is assigned to the respective value combination of each pixel. The resulting confidences are depicted in Figure 8e. Please note that the already classified convective area (refer to Figure 4e) is masked out. In the next step the confidences of each pixel are compared to the minimum confidence threshold of 0.23 determined for the separation of the intermediary precipitation. Pixels with a confidence higher than the minimum threshold are classified as subareas of intermediary precipitation. The classification result together with the classified convective are can be seen in Figure 8f. Figure 8g shows the corresponding classification result obtained from the radar data. The separation between the area of advective-stratiform background precipitation and the area of intermediary precipitation during daytime is performed analogously, but considers the reflectance in the VIS 0.6 and NIR 1.6 together with the channel differences DT , DT Appraisal of the Introduced Rain Area Differentiation Technique [56] For an appraisal of the new proposed technique for precipitation process separation and rainfall intensity differentiation, the new scheme is applied to the validation data set consisting of precipitation events from January to August 2004 (altogether 1396 scenes, see section 2). The precipitation events chosen for the evaluation study are independent from the above mentioned precipitation events of the training data set used for algorithm development. Because of the differing information content concerning the cloud water path (CWP) between day and night (see section 4.2.2), the comparison study for is done separately for the 720 daytime and the 676 nighttime scenes. [57] The comparison is realized by calculating standard validation scores for dichotomous data sets (see section 2) on a pixel basis for each scene in comparison with corresponding ground-based radar data from the German 12 of 19

13 Figure 8. Example for the separation between the areas of advective-stratiform background precipitation and the areas of intermediary precipitation for 18 July 2004, 2145 UTC. Channel differences used for the intensity differentiation: (a) DT ,(b)DT ,(c)DT , and (d) DT (e) Assigned confidences on a pixel basis. (f) Classification result based on the confidences. (g) Corresponding precipitation areas classified by the radar data. Weather Service without any spatiotemporal aggregation. The statistics were calculated separately for each area of different precipitation processes and rainfall intensities classified by the new scheme (see Table 1). [58] The results for the five subareas are summarized in Table 3. Altogether, the bias indicates a very good consistence between the area identified by the radar and by the satellite technique for all subareas. For the convectivestratiform precipitation area a slightly more pronounced overestimation is indicated. Apart from the area of enhanced advective-stratiform precipitation, which is very slightly overestimated during daytime but more distinctly underestimated during nighttime, the bias for daytime and nighttime scenes is very similar for all subareas. 13 of 19

14 Table 3. Results of the Standard Verification Scores Applied to the Subareas of Differing Rainfall Intensities Classified by the Proposed Satellite a Area (%), Radar Area (%), Satellite Bias POD POFD FAR CSI ETS HSS HKD Convective cores day mean std min max night mean std min max Convective-stratiform precipitation day mean std min max night mean std min max Enhanced advective-stratiform precipitation day mean std min max night mean std min max Intermediary precipitation day mean std min max night mean std min max Advective-stratiform background precipitation day mean std min max night mean std min max a The scores are based on 720 daytime and 676 nighttime precipitation scenes from January to August Mean signifies the average value of the respective index calculated from the whole day and night data set, respectively. Std signifies the standard deviation. Min and Max denote the minimum and maximum value of the respective index determined from the whole day and night data set, respectively. Abbreviations are as follows: CSI, Critical Success Index; ETS, equitable threat score; FAR, false alarm ratio; HKD, Hansen-Kuipers discriminant; HSS, Heidke skill score; POD, probability of detection; POFD, probability of false detection. [59] The POD shows good values for all subareas, signifying that a satisfying percentage of pixels classified as the respective subarea by the radar are consistently identified by the satellite scheme. However, inspecting Table 3 one can state a dependence of the POD on the area percentage of the respective subarea on the whole rain area. From the lowest values for convective cores the POD increases continuously with increasing area percentage until the highest values for the area of advective-stratiform background precipitation. Apart from the last mentioned subarea, the POD always indicates a superior performance during daytime. The subarea of advective-stratiform background precipitation covers a larger part of the whole rain area for the nighttime scenes compared to the daytime scenes, which again points to a dependence on the size of the covered area. The POFD shows a clear relationship with the POD and decreases as the latter decreases. As the POD the False Alarm Ratio (FAR) also shows dependence on the area percentage of the respective subarea on the whole rain area. With increasing area percentage the FAR tends to decrease and reaches lowest values for the area of advective stratiform background precipitation. Altogether both, the POFD and the FAR show a satisfying range of values. [60] The CSI indicates an overall good degree of correctly classified pixels for all subareas. Again, this index shows a dependence on the area percentage of the respective subarea. This is due to the fact that with an increasing size of the respective subarea the probability of correctly classified pixels also rises. This fact has to be accounted for, when interpreting the high CSI values. Overall the CSI indicates a better performance of the satellite technique during daytime. As for the POD and the FAR the CSI shows better values during nighttime for the area of advective-stratiform background precipitation. Again this is due to the higher area percentage of this subarea for nighttime scenes. On the other hand, not everything can be attributed to a dependence on area percentage. Despite a differing area percentage between day and night, the CSI for the convective cores and the 14 of 19

15 Figure 9. Relative operation characteristic plot for the subareas of different rainfall intensity classified by the proposed satellite technique for (left) daytime and (right) nighttime scenes: (a) convective cores, (b) convective-stratiform precipitation subareas, (c) subarea of enhanced advective-stratiform precipitation, (d) subarea of intermediary precipitation, and (e) subarea of advective-stratiform background precipitation. intermediary precipitation area does not show significant differences between day and night. [61] Because the ETS, HSS, and HKD consider the probability of correctly classified pixels by chance, these indices do not show such a clear dependence on the area 15 of 19

16 Figure 9. percentage of the respective subarea. Only for the area of advective-stratiform background precipitation with a significantly larger area percentage the indices show higher values. However, within this subarea a relation to the size of the covered area is not noticeable regarding the differences between daytime and nighttime scenes. This holds also true for the other subareas. The overall performance indicated by the three indices can be stated as satisfactory. Apart from the areas of intermediary and advective-stratiform background precipitation the performance during day and night is equivalent. [62] The relative operation characteristic [Jolliffe and Stephenson, 2003; Mason, 1982] plots displayed in Figure 9 underline the good results concerning the detected precipitation area. For the main part of the classified scenes for all subareas the calculated POD is greater than the corresponding POFD signifying that the scheme has skill. Moreover, the results for the classified scenes show a convenient combination of medium POD values together with low POFD values. Altogether, the daytime scheme is somewhat superior to the nighttime scheme. [63] To summarize the results of the comparison study an overall good performance of the proposed scheme can be stated (e.g., HSS between 0.09 and 0.2 during daytime and (continued) between 0.07 and 0.15 during nighttime), especially concerning the high temporal resolution of 15 min and the high spatial resolution of 3 by 3 km. Thus, a processoriented separation of areas with different rainfall intensities according to the conceptual model of rainbands introduced in section 3.1 is possible. [64] The slightly better performance of the daytime scheme for the advective-stratiform precipitation area is probably due to the higher information content about the CWP inherent in the VIS 0.6 and NIR 1.6 channel compared to the four channel differences considered in the nighttime scheme (refer to section 4.2). [65] The better results for the convectively dominated precipitation areas for daytime scenes might be a consequence of the higher area percentage of the respective subarea for the daytime data set compared to the nighttime data set. In general, the lower area fraction of the convectively dominated precipitation areas makes potential spatial misalignments between radar and satellite data more probable. Furthermore, the lower fraction exerts a higher influence on the quality of the calculated scores. This interpretation is supported by the fact that the fraction of the classified convectively dominated precipitation areas and the bias point to a good accordance between the radar 16 of 19

17 Figure 10. Rain area for the scene from (a) 16 August 2004, 1045 UTC and from (b) 30 August 2004, 2345 UTC classified by (left) the proposed satellite-based retrieval scheme and by (right) the radar data. Numbers on the bottom color scale are defined as follows: 1, advective-stratiform background precipitation; 2, intermediary precipitation; 3, enhanced advective-stratiform precipitation; 4, convectivestratiform precipitation; and 5, convective cores (refer to Table 1). data and the satellite-based technique. Spatial misalignments between both techniques are most probably the result of the differing viewing geometry between the satellite scanning the cloud top and the ground-based radar network scanning the cloud bottom and detecting the rainfall intensities at the surface together with the differing spatial resolution and the projection of the radar data to the SEVIRI viewing geometry. The winds within the cloud and below the cloud bottom can displace the raindrops and also cause spatial misalignments. [66] The discrepancy between the POD and FAR for the classified areas can be explained by the differing denominator for both indices. The POFD gives the fraction of the wrongly classified pixels proportional to the pixels classified as not belonging to the respective subarea by the radar data. The advective-stratiform precipitation area covers a great part of the whole rain area leading to a low fraction of the area not belonging to this subarea. This explains the relatively high POFD. In contrast, the FAR gives the fraction of the wrongly classified pixels proportional to all pixels classified as advective-stratiform precipitation areas by the satellite technique. Therefore, the denominator for calculating the FAR is much higher compared to the POFD and leads to low FAR values. Regarding the convective precipitation areas the situation is reversed. The pixels classified as not belonging to these subareas by the radar cover a large fraction of the whole rain area. As a consequence, the calculated POFD is much lower. In contrast, the 17 of 19