PUBLICATIONS. Journal of Geophysical Research: Atmospheres. Rain detection and measurement from Megha-Tropiques microwave sounder SAPHIR

Transcription

1 PUBLICATIONS Journal of Geophysical Research: Atmospheres RESEARCH ARTICLE Key Points: Rain retrieval from Megha-Tropiques SAPHIR microwave sounder Separate algorithms are developed for identification and retrieval from SAPHIR In the absence of MADRAS, rain from SAPHIR can serve as a supplementary to fulfill the science objectives of the MT mission Correspondence to: A. K. Varma, avarma@sac.isro.gov.in Citation: Varma, A. K., D. N. Piyush, B. S. Gohil, P. K. Pal, and J. Srinivasan (2016), Rain detection and measurement from Megha-Tropiques microwave sounder SAPHIR, J. Geophys. Res. Atmos., 121, , doi:. Received 7 FEB 2016 Accepted 19 JUL 2016 Accepted article online 26 JUL 2016 Published online 13 AUG American Geophysical Union. All Rights Reserved. Rain detection and measurement from Megha-Tropiques microwave sounder SAPHIR Atul Kumar Varma 1, D. N. Piyush 2, B. S. Gohil 1, P. K. Pal 1, and J. Srinivasan 2 1 Atmospheric and Oceanic Sciences Group, Space Applications Centre (ISRO), Ahmedabad, India, 2 Centre for Atmospheric and Oceanic Science, Divecha Centre for Climate change, Indian Institute of Science, Bangalore, India Abstract The Megha-Tropiques, an Indo-French satellite, carries on board a microwave sounder, Sondeur Atmosphérique du Profil d Humidité Intertropical par Radiométrie (SAPHIR), and a microwave radiometer, Microwave Analysis and Detection of Rain and Atmospheric Structures (MADRAS), along with two other instruments. Being a Global Precipitation Measurement constellation satellite MT-MADRAS was an important sensor to study the convective clouds and rainfall. Due to the nonfunctioning of MADRAS, the possibility of detection and estimation of rain from SAPHIR is explored. Using near-concurrent SAPHIR and precipitation radar (PR) onboard Tropical Rainfall Measuring Mission (TRMM) observations, the rain effect on SAPHIR channels is examined. All the six channels of the SAPHIR are used to calculate the average rain probability (P R ) for each SAPHIR pixel. Further, an exponential rain retrieval algorithm is developed. This algorithm explains a correlation of 0.72, RMS error of 0.75 mm/h, and bias of 0.04 mm/h. When rain identification and retrieval algorithms are applied together, it explains a correlation of 0.69 with an RMS error of 0.47 mm/h and bias of 0.01 mm/h. On applying the algorithm to the independent SAPHIR data set and compared with TRMM-3B42 rain on monthly scale, it explains a correlation of 0.85 and RMS error of 0.09 mm/h. Further distribution of rain difference of SAPHIR with other rain products is presented on global scale as well as for the climatic zones. For examining the capability of SAPHIR to measure intense rain, instantaneous rain over Phailin cyclone from SAPHIR is compared with other standard satellite-based rain products such as 3B42, Global Satellite Mapping of Precipitation, and Precipitation Estimation from Remote Sensing Information using Artificial Neural Network. 1. Introduction In order to study the tropical convective systems and their dominance in defining the energy and water cycle in the atmosphere, several atmospheric variables like precipitation, surface evaporation, net radiation at the top of the atmosphere, and vertical profile of temperature and humidity play a crucial role, and hence, it is imperative to measure these variables simultaneously. This is planned to be achieved through a joint Indo- French collaborative Megha-Tropiques (MT) satellite mission. MT satellite was launched on 12 October 2011 and is purposely designed to study the convective systems, water cycles, and energy exchanges in the tropical atmosphere through a host of onboard instruments such as Microwave Analysis and Detection of Rain and Atmospheric Structures (MADRAS), Sondeur Atmosphérique du Profil d Humidité Intertropical par Radiométrie (SAPHIR), Scanner for Radiation Budget, and Radio Occultation Sensor for Atmosphere. To facilitate frequent observations over the tropics, the MT is placed in a low inclination orbit (20 ) at 867 km altitude that provides the coverage over the global tropical region with 14 orbits a day and 7 days of repeat cycle. The MADRAS onboard MT is a five-frequency, nine-channel, self-calibrating, microwave radiometer that sweeps a swath of 1700 km in a conical scanning with local incidence angle of 53.5 on the surface of the earth. The SAPHIR is a six-channel microwave sounder that carries out cross-track scanning of ±50 on both sides of the nadir resulting variable incidence angle and variable pixel size over a swath of about 1700 km [Roca et al., 2015, Desbois et al., 2003]. The specifications of SAPHIR are given in Table 1. SAPHIR operates at frequencies around water vapor resonance line at GHz and is designed to provide clear-sky atmospheric humidity profiles. Frequencies near GHz have very high atmospheric absorption. The emitted radiations from the environment, reaching to the satellite, are dominated by contribution from broad atmospheric layers whose thickness and mean altitude vary with operating frequency as well as with humidity and temperature in the atmosphere. Using observations from SAPHIR channels, algorithms are developed to derive Atmospheric Temperature and Humidity profiles [Gohil et al., 2013]. VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9191

2 Table 1. Specifications of SAPHIR Microwave Sounder Channel Center Frequency (GHz) Band Width (MHz) ΔT (K) Sensitivity at 300 K Absolute Calibration (K) Over K Polarization S ± /2 ±1 H S ± /1.5 ±1 H S ± /1.5 ±1 H S ± /1.3 ±1 H S ± /1.3 ±1 H S ± /1.0 ±1 H Rain is an important atmospheric parameter and is most fundamental to achieve science objectives of the MT mission. The instrument onboard MT mission, which is specifically designed to measure rain, is MADRAS. MT being the constellation partner of the GPM mission, the science products from MADRAS were to be used as input for GPM especially for the tropical regions. GPM is an international satellite mission specifically designed to precipitation measurements from microwave sensors. However, after few months of successful operation, some anomaly was observed in the MADRAS observations due to a suspected electrical interference which resulted into random channel mixing. A methodology is worked out by the Centre National d Etudes Spatiales and Indian Space Research Organisation Project teams for realignment of the data. With this additional processing, a significant amount of data is found to be recoverable. However, the MADRAS instrument is declared nonoperational due to another anomaly in the scanning mechanism since 26 January 2013 (please see Thus, we herein make an attempt to measure rain from SAPHIR observations which may help fulfill science objectives of the MT mission, especially in the absence of the MADRAS observations. For rain measurement, passive microwave radiometers operating below 90 GHz frequency (e.g., Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI), Special Sensor Microwave Imager (SSM/I), Advanced Microwave Scanning Radiometer EOS (AMSR-E), and MADRAS) are commonly used [e.g., Grody et al., 2000; Staelin and Chen, 2000; Ferraro et al., 2005; Varma et al., 2015]. Frequencies above 90 GHz are available on a number of sensors like Special Sensor Microwave Temperature 2 on board the DMSP satellite; AMSU (Advance Microwave Sounder Unit) onboard Aqua and NOAA 15, 16, and 17; MHS (Microwave Humidity Sounder) onboard METOP [Aumann et al., 2003] Humidity Sounder for Brazil on Aqua [Lambrigtsen and Calheiros, 2003]; and MADRAS and SAPHIR onboard MT. Because of their different sensitivity to specific layers, the frequencies above 90 GHz are mainly used for the profiling of the atmospheric water vapor and temperature [Lambrigtsen and Kakar, 1985; Wilheit, 1990] or to the retrieval of total precipitable water [Wang et al., 1989]. However, in the presence of precipitating clouds, the microwave radiances interact with them through scattering and absorption processes and make difficult to profile atmospheric water vapor. There are a number of studies that have demonstrated the sensitivity of these channels to the cloud properties [Liu and Curry, 1999; Weng et al., 2003; Burns et al., 1997]. In the past, a great deal of work has been carried out to retrieve precipitation using channels close to the water vapor absorption line at GHz from sensors like AMSU, MHS, and Special Sensor Microwave Imager/Sounder (SSMIS). Laviola and Levizzani [2011], and Laviola et al. [2013] use AMSU-B window channels and the water vapor absorption band to estimate rain using regression analysis. In their algorithm, both window channels are used for rain identification and water vapor absorption channels for rain estimation. One of the main hurdles in using 183 GHz for rain retrieval is warm rain. Because of the sensitivity of these frequencies near absorption band that is nearly blind to lower or warm rain clouds, it causes a lot of errors in the low rain retrieval. Laviola and Levizzani [2011], and Laviola et al. [2013] used 89 GHz and 150 GHz to counter this problem. Surussavadee and Staelin [2008a, 2008b] used numerical weather prediction model and AMSU observations as a training data for their neural network-based global precipitation rate retrieval algorithm. Many other researchers [e.g., Mugnai et al., 2013a, 2013b; Sanò et al., 2013] have also used the water vapor absorption band to retrieve precipitation. An attempt has been made to estimate rain from SAPHIR. Balaji et al. [2014] use weather prediction model data along with Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) and precipitation radar (PR) profile over oceans as an input to the neural network for rain estimation associated with two different cyclones in Indian region from SAPHIR. They concluded that VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9192

3 SAPHIR channels are sensitive to the precipitation. Goyal et al. [2014] have attempted to estimate rain from SAPHIR frequencies using collocated SAPHIR and TRMM-3B42 observations. Hong et al. [2005] have studied the sensitivity of the microwave brightness temperatures to hydrometeors at GHz frequencies and found that frequencies near water vapor absorption band at GHz (e.g., AMSU-B channels) are nearly independent of surface emissivity because of atmospheric opacity at these frequencies. Several other researchers [e.g., Muller et al., 1994; Hong et al., 2005; Holl et al., 2010] have also studied the effect of cloud ice and precipitation on AMSU-B water vapor channels. A depression in the satellite-measured brightness temperature at frequencies near water vapor absorption line ( GHz) resulted from scattering due to hydrometeors on the top of the cloud, and the depression increases for frequencies away from the water vapor absorption line [Burns et al., 1997]. Thus, for SAPHIR, we expect higher depression in ± 11 GHz channel, as it is farther away from the absorption line. Bennartz and Bauer [2003] have studied the sensitivity of the frequencies near GHz water vapor absorption line to ice particle scattering in different environmental conditions and reported little sensitivity to precipitating clouds below 8 km of altitude, which they attributed to the peaking of the weighting functions at higher altitudes. The study by Bennartz and Bauer [2003] was conducted for middle- and higher-latitude conditions and for 183 ± 7, 183 ± 3, and 183 ± 1 GHz channels; we expect a higher sensitivity of SAPHIR 183 ± 11 GHz channel to precipitating clouds even at lower altitudes. In this paper, we demonstrate a procedure to detect and determine precipitation amount using SAPHIR frequencies near water vapor absorption line at GHz (Table 1). The rain measured from SAPHIR provides an alternate and independent set of measurements to fulfill the MT Science objectives, especially after failure of MADRAS radiometer. Rain from SAPHIR using the present algorithm is operationally produced and disseminated from our website, In the coming sections of this paper we will be discussing the use of SAPHIR observations in rain identification and retrieval. Section 2 describes all the data which have been used in the current study. In sections 3 and 4 rain identification and retrieval method will be discussed, and in the last section the conclusion of this study is presented. 2. Measurements Used in the Study SAPHIR on MT collects microwave radiances from the environment at six channels around water vapor absorption band at GHz. These channels are referred to as S1, S2, S3, S4, S5, and S6 with central operating frequency of ± 0.2 GHz, ± 1.1 GHz, ± 2.8 GHz, ± 4.2 GHz, ± 6.8 GHz, and ± 11.0 GHz, respectively. S1 is nearest to the peak of the GHz absorption band making its absorption strength highest; its weighing function peaks very high around 8 to 10 km in the atmosphere. On the other hand Channel 6 lies on the edge of the absorption band; thus, it looks deeper in the atmosphere as only large amount of water vapor can saturate this channel-resulting peak of the weighting function around 2 to 3 km. The SAPHIR sweeps a swath of ~1700 km while carrying out cross-track scanning of ±50 on both sides of the nadir. The SAPHIR swath data set composes of 182 pixels with 91 pixels along both sides of the nadir. It provides variable local incidence angle along the swath at the surface of the earth. The spatial resolution also varies along the swath from 10 km at nadir to about 22 km at pixel locations 1 and 182. In the present study, we have used SAPHIR swath observations for the year We also used Tropical Rainfall Measuring Mission (TRMM) precipitation radar (PR) observations of surface rain rates. The TRMM satellite carried on board a TRMM Microwave Imager (TMI) working at 10.65, 19.35, 21.3, 37.0, and 85.5 GHz. All the frequencies, except 21.3 GHz, received radiation in both vertical and horizontal polarizations. It also carried Ku band precipitation radar (PR) which works on 13.8 GHz frequency. The PR scans ± 17 from the nadir with 49 positions result in a 247 km swath, with a spatial resolution of 5 km. In the present study, we use precipitation radar measurements taken from TRMM standard swath product 2A25, which contains the rainfall rates from surface to 20 km altitude with 250 m vertical resolution. The TRMM sensor package has been described by Kummerow et al. [1998], and PR rain measurement algorithm is provided by Iguchi et al. [2000] and Iguchi et al. [2009]. Since 7 October 2014 TRMM is declared nonoperational. Other data sets that we have used include IR and microwave 3-hourly merged rain product referred as TRMM 3B42 [Huffman et al., 2007], TRMM 3B42-RT, and monthly rain product TRMM 3B43. The TRMM level 3 products are generated by merging MW and IR observations; 3B42-RT is the merged product of all the microwave observations from different sensors like AMSU-B, MHS, SSMI, and AMSR-E to TMI, whereas 3B42 is VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9193

4 infrared precipitation estimates using the calibrated microwave precipitation. We also use SSMIS oceanic rain rates [Hilburn and Wentz, 2008], Global Satellite Mapping of Precipitation (GSMaP) rain product [Aonashi et al., 2009], and PERSIANN (Precipitation Estimation from Remote Sensing Information using Artificial Neural Network) [Hsu et al., 1997; Hong et al., 2004]. The rain from SSMIS used in this study is purely a microwave estimate, whereas PERSIANN algorithm uses IR data from geostationary platforms (GOES 8, GOES 10, GMS-5, Metsat-6, and Metsat-7) and rainfall estimates from low-orbital satellites like TRMM; NOAA 15, 16, and 17; and DMSP F13, F14, and F15. GSMaP is an hourly rain product which is generated by combining MW-IR algorithm with GPM-GMI, TRMM-TMI GCOM-W-AMSR2, and DMSP series SSMIS, AMSU, and Geostationary IR data. 3. Rain Detection Rain identification algorithm is an integral part of any satellite passive microwave-based rain retrieval algorithm. At microwave frequencies (say, from 6 GHz to 190 GHz) variations in the surface emissivity, surface and atmospheric temperature, water vapor, cloud liquid water, or wind speed may produce the brightness temperature (T b ) variations of the same order as that due to rain rates. Apart from that the uncertainties observed in the brightness temperature versus rain rate relationship due to horizontal and vertical rain variability within satellite instantaneous field of view makes it extremely difficult to differentiate low rain rates from the background field [Varma et al., 2004]. The rain identification over the land is even more complex due to strong and highly varying surface emissivity, especially at low microwave frequencies. Varma and Pal [2012] have provided a detailed account of the problem of precipitation detection by passive microwave measurements. Here we intend to develop a rain identification algorithm using SAPHIR brightness temperature measurements. Because of high atmospheric opacity at frequencies near the water vapor absorption line ( GHz), the measurements are nearly independent of the surface emissivity [Hong et al., 2005]. There are also some studies [Kawamoto and Suzuki, 2013; Varma and Liu, 2006] which suggest that cloud structure over land is different from the oceans. Clouds over land usually develop much deeper than that over the oceans, and therefore, they may exhibit different brightness temperature depression for a given rain rate [You and Liu, 2012]. However, unlike passive microwave radiometers with a conical scan mechanism that are traditionally used for rain measurement, e.g., SSM/I, TMI, and MADRAS, the SAPHIR incidence angle at each scan position differs and that has to be incorporated in any retrieval algorithm. Ideally, the pixels that are equidistant from the nadir on a scan line should have the same incidence angle. But such pixels may still have asymmetry resulting from radio frequency interference problem of the instrument [Buehler et al., 2005]. Because of the shape and size of the footprint that is a function of incidence angle, with higher incidence angle and associated larger footprint away from nadir for SAPHIR, the rain variability within the footprint and partial beam-filling problem makes the rain retrieval more erroneous [Varma et al., 2004]. Also, the humidity sounder channels have large scan-dependent biases as we move off nadir [John et al., 2013]. Thus, we have attempted the development of the present algorithm by considering all 182 pixels independently, so that the bias of one scan position does not affect the retrieved values across the scan. In order to develop a rain identification algorithm, we use concurrent observations from SAPHIR and PR. The data set of concurrent observations is prepared by collocating SAPHIR and PR observations over space and time. The use of collocated observations of an active and passive instrument does not have a long history. Holl et al. [2010, 2014] use collocated observations of different sets of active and passive instruments to retrieve cloud properties. Gong and Wu [2013] use cloudsat and MHS collocations for ice water path and cloud top height retrieval. Many researchers [Seto et al., 2005; Masunga and Kummerow, 2005; Kida et al., 2009] use collocated TRMM-TMI and PR observations for their studies. It is understood that the radar retrievals are themselves subject to errors and limitations, especially in shallow, light, and/or frozen precipitation [Petty and Li, 2013]. The TRMM-PR rain retrieval algorithm may not be considered to be ideal as the atmospheric and cloud physical models that are used for the quantitative evaluation are not very accurate and also because of the estimation error in the surface scattering model used in the retrieval [Iguchi et al., 2009]. Since we use PR rain observations in the present study, the errors in the PR measurements will likely percolate to our algorithm. For collocation of SAPHIR and TRMM-PR, we have considered circular footprint of SAPHIR that linearly varies from 10 km at nadir to 22 km at the edge of the scan. All the PR pixels having their center falling within the VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9194

5 Figure 1. Probability distribution of SAPHIR brightness temperatures for a scan position # 45 for rainy (line joining open circles) and nonrainy (line joining solid circles) conditions for channels (a) S1, (b) S2, (c) S3, (d) S4, (e) S5, and (f) S6 over ocean and 2-D PDF as a function of combination of SAPHIR S5 and S6 channels for (g) rainy and (h) nonrainy conditions. SAPHIR footprint with a time difference of 10 min are considered collocated. On an average the PR pixels which were being collocated vary from 5 to 20 for the SAPHIR pixels (10 km to 22 km). The average rain from collocated PR observation over a SAPHIR footprint is calculated by weighted averaging. The weight varies linearly from 1 at the center to 0.5 at the edge of the SAPHIR footprint [Varma and Liu, 2006]. This way we generate a large data set of collocated SAPHIR-PR observations of several months encompassing major seasons VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9195

6 Figure 2. Probability distribution of P R for rainy (line joining open squares) and nonrainy (line joining open circles) over (a) ocean and (b) land from training data sets for all pixel locations. of the year. We use 8 months (January 2013 to August 2013) of collocated observations of SAPHIR and TRMM-PR. We then randomly choose 60% of the collocated points for training and the rest 40% for testing of the algorithm. Given the temporal variability of the precipitation [Piyush et al., 2012] and noncircular footprint of the SAPHIR away from nadir, collocation of two data sets in spatial and temporal domain may not be the best possible yet keeps the procedure of collocation simple while providing a large collocated data set of about 12 million observations over the ocean and 4.5 million over land. The rain identification and retrieval scheme have been performed separately for the land and oceans. We further divided data set into rainy and nonrainy sets for 182 swath positions of the SAPHIR. We examine the probability distribution function (PDF) of brightness temperatures measured at all six channels of SAPHIR for rainy and nonrainy conditions for all scan locations. Figures 1a to 1f show PDFs of SAPHIR brightness temperatures for the channels, S1 to S6 for scan position # 45 for rainy and nonrainy conditions over the oceans. The PDF of the brightness temperatures for rainy cases is shown as a line joining open circles and that for nonrainy cases is shown as a line joining solid circles. It may be observed that there is a large overlap of PDFs of brightness temperatures from all the channels for rainy and nonrainy cases and thus making it difficult to identify a threshold value of brightness temperature for clearly demarcating rainy and nonrainy conditions. Despite of having large overlap area between PDFs of rainy and nonrainy cases, Figures 1a 1f show clear and distinct peaks of PDFs for two cases in all the six channels of SAPHIR, which suggests some tangible sensitivity of the brightness temperatures to rain. We then computed 2-D PDF to check the combination effect of these channels for rain and no-rain conditions. Here we show synergy of S5 and S6 for rainy (Figure (1)g) and nonrainy (Figure (1)h) cases. From this figure we can see that for nonrainy cases the PDF density varies from 275 K to 290 K, whereas for rainy cases it is from 260 K to 285 K. This study encouraged us to use all the six channels for rain identification. Figure 1 shows PDFs of brightness temperatures for a scan position; however, there exists similar qualitative behavior of PDFs of brightness temperatures for rainy and nonrainy cases at other scan positions as well. We thus take advantage of the sensitivity of all the six SAPHIR channels toward rainfall as provided by the probability distribution of brightness temperatures for rainy and nonrainy cases, separately for all 182 pixel locations (please note that Figures 1a 1f are just an example for pixel location # 45 over the oceans), and calculate the average probability of a given pixel as rainy (P R ) and nonrainy (P NR ) from all the six channels. First, we calculate probability of all the six channels for rain and no-rain conditions, separately. Then we take average of all the six VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9196

7 probabilities and normalized them to find a pixel to be rainy (P R ) and nonrainy (P NR ). We then normalize the P R and P NR to make their sum unity. Now we examine the P R by plotting the distribution of P R from all pixel locations calculated separately from rainy and nonrainy SAPHIR-PR collocated training data sets (Figure 2a) for the ocean and for the land (Figure 2b). The distribution of P R from the nonrainy data set is shown as a line joining open circles and that for rainy data set is shown as a line joining open squares. We still find a large overlap between two plots in Figures 2a and 2b. This is probably because detection of low rain rates in passive microwave observations has been always probabilistic [Conner and Petty, 1998]. The changes in the brightness temperatures are significant (30 70 K) in response to geophysical parameters other than precipitation and liquid water [Petty, 1994]. The frequencies that are being used in the present study measure the scattered radiation by the ice particles present on the top of the clouds. In the scattering-based microwave algorithms for rain retrievals, the basic premise is always that higher Scattering causes more depression in the brightness temperatures which provide more rain below the clouds. Thus, it can be assumed that in the presence of rain, the effect of water vapor present in the higher altitudes above the clouds is not much. The SAPHIR frequency channels are sensitive to the cloud ice, and hence, they will be more appropriate for estimation of rain from convective clouds. Because the SAPHIR channels are not directly sensitive to the core of the clouds or the falling rain, there is a possibility of underestimation at higher rain rates, especially when brightness temperatures tend to saturate. This problem is typical to all scattering-based algorithms using high-frequency channels (e.g., 85 GHz and above). Assuming a typical convective cloud with height reaching to tropopause, the weighting functions of S6 and S1 are such that under raining conditions the S6 is least affected by the water vapor and S1 the most; and both S1 and S6 are not affected by the surface. However, over the oceans we notice a long tail in the distribution of P R associated with the nonrainy data set and a double peak in the distribution associated with rainy data set (Figure 2a). Thus, it is possible that in the SAPHIR-PR collocated data set a large number of pixels which are considered as nonrainy are actually rainy and vice versa. The other peak might be resulting from the other geophysical parameters including water vapor and cloud liquid in the absence of rain. However, the overlapping area in Figures 2a and 2b is partly represented by nonprecipitating clouds and light rain, which has less contribution to total rain amount. The other reason for overlap area is due to sensitivity of TRMM-PR in low rain regime (~0.7 mm/h) [TRMM Precipitation Radar Team, 2011] and collocation errors arising out due to different observation geometries and time bin of 10 min between PR and SAPHIR. Thus, we do not take P R = 0.5 as the threshold value to identify rain/no rain, and a more appropriate value of the P R threshold has been worked out. We examined different threshold values of P R and tried to identify the most appropriate value of P R which is optimized for minimal error in the identification of rain/no rain. This procedure led us to find a P R value of 0.60 as the most appropriate value for ocean and 0.63 for land. According to Figure 2a, the P R threshold of 0.6 results in about 9% false rain pixels from nonrainy pixels and 34% missing rain from rainy pixels over oceans, whereas over land with P R = 0.63 there is about 21% false rain pixels from nonrainy and 29% missing rain from rainy pixels. We also calculate probability of detection (POD) and false alarm for the whole data, i.e., training as well as test data set, and found that for ocean, detection ratio is 0.61 having false alarm 0.20, whereas it is 0.77 and 0.29, respectively over land. We further examine how these wrongly identified pixels with P R thresholds would have likely ramification on the rain identification and measurement. Figures 3a and 3b show the average rain with vertical bars of standard deviation (SD) as a function of P R from the rainy data set of ocean and land, respectively. As expected the average rain is small in lower values of P R and increases as we go to higher values of P R. Over oceans, the average rain at P R = 0.6 is 0.11 mm/h with SD of 0.48 mm/h. We follow the same procedure over land and find that average rain at P R = 0.63 is 0.09 mm/h with SD of 0.65 mm/h. Thus, the rainy points that are classified as nonrainy would contribute only to the very low rain regime (say, < ~0.70 mm/h). We also examined the quantiles of average rain rates with the P R values from rainy data set. Quantiles are points which divide the range of a probability distribution into contiguous intervals with equal probabilities. Here we compute 4-quantiles which is also called as quartile. Figures 3c and 3d show the first, second, and third quartiles for ocean and land, respectively. The second quartile represents the median (open circles) of the data set. The median divides the whole data set into two equal halves in terms of numbers. The third quartile is shown as open squares, below which 75% of data exist. These figures (3c and 3d) are in agreement with Figures 3a and 3b and suggest that the wrongly classified rainy pixels would have contribution in very low rain rates. We further examine the nonraining point that turns raining with the P R thresholds. Figure 3d shows the cumulative VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9197

8 Figure 3. Average rain rate (mm/h) with vertical bars of standard deviation (SD) as a function of P R from the rainy data set over (a) oceans and (b) land and 4-quantiles (quartiles) of average rain rate as a function of P R with the same data set for (c) ocean and (d) land. (e) Cumulative probability distribution of rain rates from the nonrainy point that turns rainy with the P R > 0.63 (open circles) for land and P R > 0.6 (open triangles) for oceans. VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9198

9 Figure 4. False positive (red) and false negative (blue) rain difference from SAPHIR and 3B42 for 10 July 2013 for rain rate > 0.25 mm/h. probability distribution of rain rates for such pixels. The rain rate is plotted along abscissa calculated from an algorithm discussed in the next section. From Figure 3d, we find that for oceans (open triangles) about 85% of the nonraining pixels that give false rain produce an average rain of less than 0.15 mm/h. About 95% of such pixels produce an average rain of less than 0.29 mm/h, whereas for land about 58% of the nonraining pixels give an average rain of 0.15 mm/h, and 95% of such pixels produce an average rain less than 0.59 mm/h and none of them yield average rain of more than 0.95 mm/h. Thus, with the P R -based rain/no-rain threshold, the rainy pixels that are classified as producing missing rain and the nonrainy pixels that are classified as producing false rain have effect only in the low rain regime (i.e., < 1 mm/h). Given the errors in the rain estimation by the microwave radiometers [Varma and Liu, 2010; Varma and Pal, 2012; Wolff and Fisher, 2008] and the error involved in the spatial-temporal collocation of the SAPHIR-PR data set, this error in the rain identification can be considered as insignificant. With a P R threshold of 0.60 and 0.63 for ocean and land, respectively, we generate rain/no-rain map from SAPHIR and also from TRMM 3B42 standard data set. Figure 4 shows the difference of rain map from SAPHIR and 3B42 for 10 July Due to rain identification problem in very low rain regime, we have deliberately avoided low rain rates of < 0.25 mm/h. This figure shows the false positives (red) and false negatives (blue) SAPHIR rain compared with 3B42. Based on the SAPHIR rain determined using retrieval algorithm discussed in the next section. Figure 4 shows that rain area by SAPHIR matches reasonably well with that from 3B42 keeping in mind the fact that 3B42 is available every 3 h interval, and SAPHIR is a swath product averaged daily to generate daily average rain rate. We also compute some statistics using 1 month (July 2013) of data from SAPHIR and 3B42 for land and ocean separately. FAR (false alarm rate), Hit Rate, POD (probability of detection), and missing rain are calculated using 1 month of data of SAPHIR and 3B42. Over the ocean H = 0.77, POD = 0.80, FAR = 0.40, and missing rain = 0.33, and over land H = 0.92, POD = 0.71, FAR = 0.45, and missing rain = Rain Retrieval In the previous section, rain identification in SAPHIR observations is discussed. In this section, we make an attempt to retrieve the rainfall rate. For development of the retrieval algorithm, we first tried to examine the behavior of SAPHIR frequency channels with respect to rain rate. For that we use radiative transfer model developed by Liu [1998]; we simulate brightness temperatures using a scattering-based microwave radiative transfer model for all the six channels of SAPHIR. For the simulation, we use numerical weather prediction model data and TRMM-TMI observations. Atmospheric profiles (RH, T, and P) are taken from model data, and hydrometeor profiles are from TMI 2A12 profiles. In the model simulations we take ice particle size from 50 to 2000 μm and ice particle shape as six bullet rosettes. All the simulations are carried out for 50 incidence angle. From the past studies [e.g., Eymard et al., 2001; Bennartz and Bauer, 2003] we know that channels that are farther away from the absorption line are most affected by the rain. We take the difference of simulated brightness temperature of S1 and S6 and plotted in Figure 5. Figure 5 shows the variation of brightness temperature difference (ΔT b ) with rain rate (mm/h), which appears close to exponential in nature. We follow the same procedure on land also and find that the exponential relation holds well over land too. For rain retrieval algorithm, we use the SAPHIR and PR collocated data set as referred in the previous section. First, we separate the data over land and ocean and then divided them further into two parts rainy and VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9199

10 Figure 5. Radiative transfer simulation-based variation of the difference of brightness temperature from SAPHIR channels S1 and S6. nonrainy. We, however, use only rainy pixels and referred this data set as a training data set. As discussed above, from radiative transfer simulations, we find an exponential relationship between rain rate and the difference of brightness temperatures measured at S1 and S6. We thus use the following form of the equation for ocean and land and for each pixel location R ¼ a þ b*e c*δt b (1) where R is rain rate in (mm/h) and a, b, and c are regression coefficients. We calculate the values of a, b and c separately over the land and oceans and at each of the pixel locations. The regression coefficients are calculated using PR-SAPHIR collocated data set at each of the SAPHIR pixel locations. The exponential relationship is fitted using least squares approach. In order to determine the accuracy of the rain estimation algorithm, we apply equation (1) to the training data set of collocated SAPHIR and PR observations of rainy pixels and calculate rain rate from SAPHIR observations at each pixel location and compared with the PR rain rates. This provides a correlation of 0.72, RMS error of 0.75 mm/h, and bias of 0.04 mm/h for a total number of observations of about 3.2 million over land and ocean. When we consider all the collocated SAPHIR and PR observations from the training data set composing of rain as well as no-rain observations and subject them to rain identification algorithm followed by rain retrieval algorithm, we find a correlation of 0.69 with an RMS error of 0.47 mm/h and bias of 0.01 mm/h for a total number of 6.3 million observations. If this algorithm is used operationally, the SAPHIR observations subjected to the rain retrieval algorithm would be preclassified as rainy or nonrainy and hence the statistics provided in the latter case provides the algorithm accuracy. In order to test the algorithm, we apply this algorithm to the independent test data set and compare the rain from SAPHIR with that from PR. With the test data set we find a correlation of 0.68, RMS error of 0.84 mm/h, and bias of 0.05 mm/h. These results are very encouraging given the high temporal variability of precipitation [Piyush et al., 2012] and possible imperfection in preparing collocated data set of SAPHIR and PR as discussed earlier. These results appear to be more encouraging in view of similar comparison that we carried out between TMI and PR for a period of 10 days from 21 to 30 May 2009 that shows a correlation of 0.61, RMS error of 2.23 mm/h, and bias of 0.1 mm/h for a total number of observations of 1.88 million. Here it may be VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9200

11 Figure 6. Difference of PDFs of daily averaged rain rate (mm/h) from SAPHIR and TRMM-3B42 for May to note that PR and TMI are on the same orbiting platform with less than a minute of time difference between their observations, whereas we allowed 10 min of time difference between SAPHIR and PR observations. We also examined the distribution of daily averaged rain rates; for that we compared the PDF of daily averaged rain over tropics from SAPHIR with TRMM-3B42 for May 2013 (Figure 6). We have plotted the difference of PDFs of SAPHIR and 3B42 rain rates in Figure 6. Figure 6 shows a good agreement of the PDF of the daily averaged rain from SAPHIR and 3B42 except in the low rain regimes which is possibly due to rain identification error as discussed in section 3. The daily averaged rain from SAPHIR, 3B42, and SSMIS is also compared with each other. The comparison shows correlations of 0.68, 0.51, and 0.54 with RMS differences of 0.29, 0.58, and 0.56 between SAPHIR-3B42, SAPHIR-SSMIS, and SSMIS-3B42, respectively. It is also to note here that for 1 month of data, averaged SAPHIR pixels varies from 800 to 1200 pixels per grid box, whereas TRMM-PR averaged only 200 to 300. Further, we have carried out a qualitative comparison of the monthly averaged rain rates from SAPHIR, TRMM-3B42, and SSMIS for August Figure 7 shows the monthly average gridded ( ) maps of precipitation from SAPHIR, TRMM-3B42, and SSMIS for the month of August Figure 7. Monthly averaged gridded ( ) rain rates (mm/h) from (a) SAPHIR, (b) TRMM-3B42, and (c) SSMIS for the month of August The rain from SSMIS is only over oceans. VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9201

12 Figure 8. (a) Quantitative difference of average monthly rain rates (mm/h) for the month of August 2013 from SAPHIR and 3B42 and (b) corresponding probability distribution of the difference of the rain rates for land and oceans. Here we use SSMIS rain product [Hilburn and Wentz, 2008] which provides rain only over the oceans. As it can be seen from Figure 7 there is a good qualitative comparison among three different rain products. The monthly rain of August shows the global rain patterns as expected from the climatology. The northern shift of Intertropical Convergence Zone and the Indian summer monsoon is well captured by SAPHIR (Figure 7a). We also attempted a quantitative comparison of the monthly rain rates from SAPHIR, 3B42, and SSM/I in latitude-longitude grid for the months of August 2013 and January The comparison between SAPHIR and TRMM-3B42 is found to be the most encouraging with correlation coefficients of ~0.85 and ~ 0.81 and RMS differences of 0.08 and 0.09 mm/h for about 0.33 million valid points for the months of August 2013 and January 2014, respectively. On the other hand, the comparison of SSMIS monthly rain with monthly rain from SAPHIR and TRMM-3B42 shows corresponding correlations of 0.66 and 0.61 for the months of August 2013 and January 2014, respectively. Next, we tried to find the quantitative difference between SAPHIR and 3B42 monthly average rain rates in latitude-longitude grid for August Figure 8a shows the rain rate difference of SAPHIR from 3B42. The differences are in the range from 0.5 to VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9202

13 Figure 9. Zonal monthly averaged rain rates from SAPHIR (dash line), 3B43 (solid line), and PERSIANN (dotted line) for the month of July ( ) over (a) land and (b) ocean and also for the month of January ( ) over (c) land and (d) ocean. The grey colored error bars represent the uncertainty range in the 3B43 (solid line) observations. 0.5 mm/h. In order to examine these differences more precisely, in Figure 8b we show the probability distribution of the difference of rain rate from Figure 8a for both land and oceans separately. It is evident from Figure 8b that the probability distributions of the difference of rain rates from SAPHIR and TRMM-3B42 are approximately Gaussian in nature and peak near at 0 mm/h for both land and oceans. In order to show another meaningful comparison of the quality of rain estimation using SAPHIR with that from TMI and PR, we make monthly average comparison of rain over land from SAPHIR with TMI and PR in latitude-longitude grid and compared with the similar statistics between PR and TMI provided by Wang et al. [2009]. Wang et al. [2009] using 10 years of TRMM data carried out the comparison study and reported a correlation of 0.78 between 2A12 and 2A25 monthly mean rain rates. Based on 10 years of TRMM observations over land, they have reported a correlation of 0.78 between 2A12 and 2A25 monthly mean rain rates. We have also compared SAPHIR monthly rain rates with TMI and PR for the month of November 2012 and found a correlation of 0.85 between SAPHIR and 2A12, and 0.70 between SAPHIR and 2A25 observations. For the same month, we find a correlation of 0.71 between 2A12 and 2A25. While comparing the statistics, it has to be kept in mind that TMI and PR have near-concurrent observations VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9203

14 Figure 10. Distribution of rain rates from (a) SAPHIR, (b) 3B42, (c) GSMaP, and (d) 3B40-RT over a tropical cyclone Phailin on 11 October 2013 between 2050 UTC and 2100 UTC. (within 1 min), whereas SAPHIR may have large time differences with TMI and PR observations. Nevertheless, the SAPHIR performs well when compared to TMI and PR. Further, to examine the performance of SAPHIR rain algorithm for different climatic zones over land and oceans in two contrasting months of July and January, zonal averaged rain rates from SAPHIR, TRMM-3B43, and PERSIANN (Precipitation Estimation from Remote Sensing Information using Artificial Neural Network) using 3 years ( ) of data are plotted in Figure 9. The shaded portion in Figures 9a and 9b shows plots for the month of July for land and oceans, respectively. Figures 9c and 9d are similar to Figures 9a and 9b but for the month of January. Figures 9a 9d show that SAPHIR rain is well within the uncertainty of 3B43, except in theregionsoflowrainratespossiblyduetotheerrorsintherainidentificationinthelowrainregimeasdiscussed in section 3. Here it may be noted that both 3B42 and PERSIANN are available at every 3 h time interval over 50 S to 50 N globally whereas SAPHIR having swath coverage with 4 5 observations at any given location in a day. We have further presented rain from SAPHIR associated with a tropical cyclone Phailin that persisted over the Bay of Bengal from 9 to 12 October This was recorded as a category 5 cyclone with a maximum wind speed of about 215 km/h. Phailin made its landfall on 12 October A cyclonic system is an excellent atmospheric phenomenon to critically evaluate the performance of the retrieval algorithm as several types of clouds and rain structures are present in the system. For comparison purposes, we have analyzed all the available rain products over that region which is coincident with the SAPHIR. We use 3B42 and 3B40-RT, both 3-hourly averaged rain rate products from NASA, and GSMaP, an hourly averaged rain rate product from JAXA, for the comparison. 3B42 [Huffman et al., 2007] is a latitude-longitude gridded infraredmicrowave merged product which is fine tuned with surface observations, 3B40-RT is an intermediate product of 3B42 which is derived from passive microwave observations only, and GSMaP [Aonashi et al., 2009] is also a merged product in latitude-longitude grid and is produced hourly. The 3B42 and 3B40- RT are having instantaneous rain values which are averaged over every 3 h. All the comparison performed here are in rectangular grid. We calculate the rain from SAPHIR and presented in Figure 10a. VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9204

15 The corresponding rain from 3B42, GSMaP, and 3B42-RT is shown in Figures 10b 10d, respectively. Over the study area, the MT has an instantaneous observation time of ~2050 UTC, whereas all the other products are at 2100 UTC and averaged over the time as discussed above. From these figures it can be seen that SAPHIR picks up the intensity and pattern of rain over the cyclonic area very well, as exhibited by other rain products. When rain rates are compared quantitatively between SAPHIR and 3B42, and between SAPHIR and GSMaP, we find correlations of 0.67 and 0.75, RMS differences of 1.10 and 2.29 mm/h, and biases of 0.10 and 0.65 mm/h, respectively. A similar comparison of rain from GSMaP and 3B42 shows a correlation of 0.72, RMS difference of 2.1 mm/h, and bias of 0.26 mm/h. Here it may be to note that all the rain products with which SAPHIR rain is compared are the merged products of MW, IR, and ground observations (3B42) and supposed to be more accurate. The comparison of SAPHIR rain with these rain products for Phailin cyclone is carried out only to show the validity of SAPHIR rain in case of extreme rain events. The discrepancies between different products are possibly due to their different observation time. Nevertheless, SAPHIR rain shows a good agreement with TRMM products. Acknowledgments Radiative transfer model is developed and provided by Guosheng Liu, FSU (USA). SAPHIR Level-1 (brightness temperatures) and Level-2 (geophysical products) data are available through and are provided in accordance with MOSDAC data distribution policy. TRMM data (3B42, TMI, and PR) from NASA, SSMIS data from Remote Sensing Systems and Global Hydrology Resource Center (USA), and GSMaP data from JAXA are acknowledged. The Precipitation-PERSIANN CDR used in this study was acquired from NOAA s National Centers for Environmental Information. Authors are also thankful to The Director, Space Applications Centre (SAC), Ahmadabad for his encouragement. Authors would like to thanks anonymous reviewers for their helpful comments. 5. Conclusion and Discussion The study presented herein provides a detailed account of rain detection and measurement algorithm from SAPHIR. The algorithm is in two parts: the first part deals with rain identification, and the second provides the rain amount at identified pixels. We have carried out extensive study to examine the worthiness of the algorithm. The statistics presented herein shows that the algorithm works well for rain identification as well as for rain estimation at instantaneous, daily, and monthly bases. The comparison at instantaneous basis with the independent data set (test data set) gives us confidence that the algorithm works satisfactorily. The measurement of rain from SAPHIR over the Phailin cyclone around the time when it attained the highest intensity and its qualitative and quantitative comparison with other standard rain products further strengthen the confidence that SAPHIR is able to provide intense rain satisfactorily. Due to the smaller wavelength, the scattering signal from the hydrometeors is strong at SAPHIR channels. The major contribution of scattering comes from the ice droplets on the top of the clouds. It is possible that the algorithm may not work efficiently for rain coming from warm clouds. This is generally true for any scattering-based algorithm from passive microwave radiometers. This may be one of the reasons for uncertainty in the rain identification in the low rain regime as discussed in section 3 where the other possible reasons are discussed as well. Though SAPHIR is not specifically designed for rain estimation, yet we have shown its use for rain identification as well as for rain estimation. The best instrument on MT for rain estimation is MADRAS, but after few months of operation MADRAS observations first suffered from channel mixing and later by a snag in the scanning mechanism, as a result of which MADRAS operation is completely stopped from January In the absence of MADRAS, the rain from SAPHIR can serve as a supplementary to the MADRAS rain to fulfill the science objectives of the MT mission. SAPHIR-based rain identification may even help flagging off the humidity profile measurements from it. The MT is a GPM constellation satellite; failure of MADRAS sensor has been a setback for the GPM user community which may now receive an impetus with the availability of SAPHIR rain. The SAPHIR rain product from the present algorithm is operationally available in near real time from our site with domain name References Aonashi, K., J. Awaka, M. Hirose, T. Kozu, G. Liu, S. Shinge, S. Kida, S. Seto, N. Takahashi, and Y. N. Takayabu (2009), GSMaP passive microwave precipitation retrieval algorithm: Algorithm description and validation, J. Meteorol. Soc. Jpn., 87A, Aumann, H. H., et al. (2003), AIRS/AMSU/HSB on the Aqua mission: Design, science objectives, data products, and processing systems, IEEE Trans. Geosci. Remote Sens., 41(2), Balaji, C., C. Krishnamoorthy, and R. Chandrasekar (2014), On the possibility of retrieving near-surface rain rate from the microwave sounder SAPHIR of the Megha-Tropiques mission, Curr. Sci., 106(4), 587. Bennartz, R., and P. Bauer (2003), Sensitivity of microwave radiances at GHz to precipitating ice particles, Radio Sci., 38(4), 8075, doi: /2002rs Buehler, S. A., M. Kuvatov, and V. O. John (2005), Scan asymmetries in AMSU-B Data, Geophys. Res. Lett., 32, L24810, doi: / 2005GL Burns, B. A., X. Wu, and G. R. Diak (1997), Effects of precipitation and cloud ice on brightness temperatures in AMSU moisture channels, IEEE Trans. Geosci. Rem. Sens., 35, VARMA ET AL. RAIN RETRIEVAL FROM SAPHIR 9205