Abstract
Estimation of the flow generated in the Brahmaputra river basin is important for establishing an effective flood prediction and warning services as well as for water resources assessment and management. But this is a data scarce region with few and unevenly distributed hydrometeorological stations. Five high-resolution satellite rainfall products (CPC RFE2.0, RFE2.0-Modified, CMORPH, GSMaP, and TRMM 3B42) were evaluated at different spatial and temporal resolutions (daily, dekadal, monthly, and seasonal) with observed rain gauge data from 2004 to 2006 to determine their ability to fill the data gap and suitability for use in hydrological and water resources management applications. Grid-to-grid (G-G) and catchment-to-catchment (C-C) comparisons were performed using the verification methods developed by the International Precipitation Working Group (IPWG). Comparing different products, RFE2.0-Modified, TRMM 3B42, and CMORPH performed best; they all detected heavy, moderate, and low rainfall but still significantly underestimated magnitude of rainfall, particularly in orographically influenced areas. Overall, RFE2.0-Modified performed best showing a high correlation coefficient with observed data and low mean absolute error, root mean square error, and multiple bias and is reasonably good at detecting the occurrence of rainfall. TRMM 3B42 showed the second best performance. The study demonstrates that there is a potential use of satellite rainfall in a data scarce region.
1. Introduction
Spatial distribution and the amount of rainfall are important for water resources assessment and for establishing an effective flood prediction and warning services and drought monitoring. However, in many regions the number of ground measuring stations is very limited and unevenly distributed, making water resources assessment and flood prediction difficult [1]. In mountainous areas with a limited or no rain gauge network, as in the Himalayan region, satellite-based rainfall estimation can provide information on rainfall occurrence, amount, and distribution [2–4]. Several high-resolution global and regional satellite-based rainfall products are available from different operational agencies and research and academic institutions [5–7]. They include the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Centre Rainfall Estimates Version 2.0 (CPC-RFE2.0) [8], NOAA CPC Morphing Technique (CMORPH) [9], Global Satellite Mapping of Precipitation (GSMaP) [10], and Tropical Rainfall Measuring Mission (TRMM) [6, 11–13], which are available at a high spatial and temporal resolution. These products provide an opportunity to develop near real-time flood predictions and other water resource management applications in data sparse regions using rainfall estimates. However, satellite-based rainfall data have uncertainty and this affects the accuracy of predictions when they are used in rainfall-runoff models for flow simulation [14, 15].
Satellite rainfall estimates (SRE) from different products have been extensively validated with ground data around the world [7, 16, 17], including the Hindu Kush Himalayan (HKH) region [18, 19]. The spatial distribution of NOAA’s CPC-RFE2.0 SRE has been verified separately for the eastern part of the HKH (governed by the summer monsoon) and the western part (governed by the winter monsoon) [18], and country and basinwide verifications have been done for Nepal [1], Bangladesh [20], and India. Verification at three levels (country, physiographic, and basin) at 176 rainfall stations has shown that CPC-RFE2.0 and GSMaP MVK+ underestimate rainfall over Nepal [19]. Islam et al. [21] compared TRMM product with observed rainfall data on a daily basis and found that the trend with TRMM was similar to the trend with observed rainfall, but the actual rainfall was generally underestimated in most days although also overestimated in a few days. Duncan and Biggs [22] assessed the seasonal accuracy of satellite-derived precipitation estimates (TRMM-3B42) over Nepal and showed that the SRE underperformed in estimating extreme rainfall events and did not detect rainy days well. Although most of the satellite-based rainfall products have been verified in this region individually [18, 19, 21, 22], very few studies included an intercomparison of different satellite products. Apart from that, studies evaluating the performance of SRE over complex topography of Brahmaputra river basin are still very limited.
SRE products are still an emerging capability; although they are improving, they are generally not yet precise enough for many hydrological applications because of their certain limitations [16, 23]. The comparison between different satellite rainfall products at the same spatiotemporal resolution can give significantly different results in terms of hydrological modelling application so each satellite product must be evaluated individually in order to be used for hydrological application [24]. In some cases, the products may require additional local improvement (for example, ingestion of rain gauge data or bias correction according to topography) to become useful in hydrological applications. Local adjustments were found to be essential in several studies [1, 25–27].
The Brahmaputra basin was chosen because it has intense seasonal rainfall with rugged terrain, large unpopulated areas, complex transboundary issues with few meteorological stations, and real-time rain gauge data which are scarce, unevenly distributed, and poorly maintained [3]. The analysis was done both for the whole Brahmaputra river basin as a homogenous region and for individual catchments. Investigation of the performance of SRE at catchment level is important because the conceptual and (semi-) distributed hydrological model relies on subbasin or catchment average of hourly or daily rainfall [28]. The final aim was to determine the operational viability of products within the basin and identify a product that could fill the data gap resulting from the scarcity of ground observations and be used in water resources assessment and hydrological applications.
This paper describes the performance of SRE product (RFE2.0-Modified) modified at the International Centre for Integrated Mountain Development (ICIMOD) by merging CDC RFE2.0 with local ground observed data. Altogether, three different global and two regional satellite rainfall products (CMORPH, GSMaP, CPC-RFE2.0, RFE2.0-Modified, and TRMM 3B42) were compared over the Brahmaputra basin using the satellite precipitation verification metrics developed by the IPWG [5]. Each of the products was verified individually by comparison with the gauge-observed-interpolated rainfall data over a three-year period, and then the performance of the different estimates was compared. Three spatial verification methods (visual verification, continuous statistics, and categorical statistics) and two approaches, G-G and C-C [20, 23, 29], were applied in the comparisons. The remaining of this paper is organized as follows. Section 2 introduces the data and methods, study area, preparation of RFE2.0 Modified, and rainfall data including a brief description of observed and satellite rainfall products, followed by preparation of data for validation and verification of satellite rainfall estimates. The results are discussed in Section 3 and Section 4 concludes the overall findings of this paper.
2. Data and Methods
This section of the paper describes the study area as well as the approaches that have been taken for this study. The main steps in the approach were (a) preparation of RFE2.0-Modified; (b) rainfall data; (c) and data for comparison and (d) verification of satellite rainfall.
2.1. Study Area
The Brahmaputra basin is one of the largest river basins in the world and extends across parts of four countries: China, India, Bhutan, and Bangladesh. The river originates as the Yarlung Tsangpo from the great glacier mass of Chemayungdung in the Kailas range in the southern part of Tibet Autonomous Region in China at an elevation of 5,300 masl and travels 1,995 km through China, 983 km through India, and 432 km through Bangladesh, before it empties into the Bay of Bengal through a joint channel with the Ganges and the Meghna [30] (Figure 1).
The Brahmaputra river drains an area of around 573,000 sq.km [31] including the territory of Tibet of China (50.50%), Bhutan (7.80%), India (33.60%), and Bangladesh (8.10%) [30]. The tributaries that originate in Bhutan join the main trunk in India. In China, the river passes through the Yarlung Tsangpo Canyon, which is thought to be the deepest canyon in the world (source: http://en.wikipedia.org/wiki/Yarlung_Tsangpo_Grand_Canyon; http://www.china.org.cn/english/MATERIAL/185555.htm). As the river enters India, it makes a very rapid descent to the plains where it becomes very wide, in places as wide as 10 km. After entering Bangladesh, the Brahmaputra splits into two branches near Bahadurabad. The much larger branch continues south with the name Jamuna and meets with the Ganges river near Aricha; the smaller branch, which was the main channel in the past, flows southeast to join the Meghna river near Dhaka.
The basin comprises such diverse environments as the cold dry plateau of Tibet, the rain-drenched Himalayan slopes, the landlocked alluvial plains of Assam, and the vast deltaic lowlands of Bangladesh [32]. Immerzeel [33] categorized the Brahmaputra basin into three different physiographic zones: Tibetan Plateau (>3,500 mean sea level (masl), 44.4%), Himalayan belt (100–3500 masl, 28.6%), and floodplain (<100 masl, 27%). These physical features play a significant role in the climate and rainfall pattern of the basin. The basin has a mean elevation of 3944 m above masl with the highest/lowest elevations at 8586 masl (peak of Kanchenjunga)/0 masl. The Brahmaputra basin, excluding the Tibetan portion, forms an integral part of the southeast Asian monsoon regime with a mean annual rainfall of 2,300 mm. Distribution of rainfall over the basin varies from 1,200 mm in parts of Nagaland to over 6,000 mm on the southern slopes of the Himalaya [30]. This basin is heavily influenced by monsoon rainfall. Around 70–80% of annual rainfall falls during the monsoon season (June to September) and 15–20% during the premonsoon season (March to May). The basin has a large north-south precipitation gradient with annual rainfall in 2004 ranging from 284 mm in the north at Jiangxi in China to 11,039 mm in the south at Dorokha in Bhutan.
The average annual and flood peak discharge of the Brahmaputra, observed at Bahadurabad, are about 20,200 m3/s and 70,000 m3/s, respectively [31] (Table 1). The peak discharge can reach about 120,000 m3/s in a catastrophic flood season.
2.2. Preparation of RFE2.0-Modified
The merging algorithm of CPC-RFE2.0 defines the analysis of daily precipitation in two steps. First, to reduce the random error inherent in the individual data sources, the three kinds of satellite data (GPI cloud-top IR, SSM/I, and AMSU) are combined linearly through the Maximum Likelihood Estimation Method, in which the weighting coefficients are inversely proportional to the individual error variance. This provides the shape of precipitation. Since the shape of precipitation contains bias passed through from original individual satellite data, a second step is introduced to remove the bias by blending the shape of precipitation with the gauge data using the method of Reynolds [34]. In this blending process, the gauge data are used to define the magnitude of the precipitation field [35].
Currently, the operational CPC-RFE2.0 algorithm uses less than 10 rain gauge stations from global telecommunication system (GTS) networks for the Brahmaputra basin. An improved version of CPC-RFE2.0 was developed at ICIMOD by blending 24 hours of accumulated rainfall data of 33 gauge stations of the Brahmaputra basin in the CPC-RFE2.0, including 3 stations in Bangladesh, 7 in India, 11 in China, and 12 in Bhutan. We selected these 33 stations from available 90 stations according to the criteria of minimum density of precipitation given by WMO [36] to ensure a consistent distribution of stations in each country. For blending, detailed information about each local gauge station (name, latitude, longitude, elevation, and others) was added to a master NOAA GTS rain gauge file to calculate the gauge-to-gauge distances and update the number of new local stations in the master file. Finally, the program was run to blend the daily precipitation data from local stations for final precipitation estimates. The final precipitation estimates retain the station’s rain gauge value, while, as distance from a station increases, the estimates rely more heavily on satellite derived precipitation. This indicates that the spatial pattern of the CPC-RFE2.0 does not change, only the magnitude of the CPC-RFE2.0 changes. If short-term convective rainfall events which were not well observed by the CPC-RFE2.0 were measured at a particular station, this would be included in the RFE2.0-Modified. This improved CPC-RFE2.0 (final precipitation estimates) which is called RFE2.0-Modified. The reader is referred to the CPC-RFE2.0 training manual [37] for more details about these methods. The effect of additional local gauges on the products was clear (Figure 2).
(a) RFE2.0-Modified from 2006-05-12
(b) CPC-RFE2.0 from 2006-05-12
2.3. Rainfall Data
Gauge-observed rainfall data of the study area for the period 2004 to 2006 was provided by the regional partners in the “Application of satellite rainfall estimations in the HKH region" project [38]. Based on continuous observed data availability, 2004–2006 period was selected for the present study. Altogether, 90 gauge stations (8 in Bangladesh, 41 in Bhutan, 19 in China, and 22 in India) were used (Figure 1). For the general evaluation with densities of contributing gauges per basin ranging from one station per 1328 km2 to 6814 km2, daily observed rainfalls were averaged over basin by using inverse distance weighted (IDW) interpolation technique.
Five satellite rainfall products were used in the study: CMORPH, GSMaP, CPC-RFE2.0, RFE2.0-Modified, and TRMM 3B42. Three of these products are from NOAA. CPC-RFE2.0 produces 24 hours of precipitation estimates on a 0.1° latitude/longitude grid over South Asia (70°E–110°E; 5°N–35°N) in a real-time basis. It is based on the combination of daily GTS rain gauge data, advanced microwave sounding unit (AMSU) satellite precipitation estimates, special sensor microwave/imager (SSM/I) satellite rainfall estimates, and geostationary operational environmental satellite (GOES) Precipitation Index (GPI) cloud top infrared (IR) temperature precipitation estimates. The three satellite estimates are first combined linearly using predetermined weighting coefficients and then merged with station data to determine the final rainfall. The CPC technique is capable of estimating rainfall from convective (cold) as well as stratified (warm) clouds [35, 39]. RFE2.0-Modified is the modified version of CPC-RFE2.0, obtained by merging CPC-RFE2.0 with locally observed rain gauge data. CMORPH uses high quality passive microwave satellite sensors, which are then propagated by motion vectors derived from more frequent geostationary satellite IR data. In effect, IR data are used as a means to transport the microwave-derived precipitation features during periods when microwave data are not available at a location [9]. The spatial and temporal resolutions of CMORPH are 0.1° and 30 minutes (half hourly). In this study, half hourly data was summed to daily to match the frequency and magnitude of observed rainfall product and other satellite rainfall products.
Another satellite rainfall product that we considered in our study is TRMM 3B42. The National Aeronautics and Space Administration (NASA) produces this product and is available at 3 hourly intervals to the research community. This product contains the output of TRMM Algorithm 3B42, which is to produce tropical rainfall measuring mission (TRMM) merged high quality (HQ)/infrared (IR) precipitation and root-mean-square (RMS) precipitation-error estimates. The combined instrument rain calibration algorithm (3B-42) uses an optimal combination of 2B-31, 2A-12, SSMI, AMSR, and AMSU precipitation estimates (referred to as HQ) to adjust IR estimates from geostationary IR observations. The 3B-42 estimates are scaled to match the monthly rain gauge analyses used in 3B-43 [6]. The output is rainfall for 0.25 × 0.25 degree for every 3 hours. In this study, TRMM Version 6 has been used (ftp://trmmopen.gsfc.nasa.gov/pub/merged/; http://gcmd.nasa.gov/records/GCMD_GES_DISC_TRMM_3B42_daily_V6.html) and 3 hourly TRMM 3B42-V6 data was summed to a daily interval to match the observed rainfall product and other satellite rainfall products.
In this study, we used GSMaP (the GSMaP_MVK+ product) which is a JAXA product. The GSMaP_MVK(+) algorithm is a combination of the CMORPH technique and Kalman filter. The IR data are used as a means to move the precipitation estimates from microwave observation during periods when microwave data are not available at a location. The microwave sensors which used are TRMM/TMI, Aqua/AMSR-E, and DMSP/SSMI (F13, 14, 15) for the GSMaP_MVK product; in addition to these, AMSU-Bs are included in the GSMaP_MVK+ product [10, 40]. The spatial and temporal resolutions of GSMaP are 0.1° and 60 minutes (hourly). In this study, hourly GSMaP data was summed to a daily total to match the observed rainfall product and other satellite rainfall products. These products are described briefly in Table 2.
2.4. Data Preparation
Satellite rainfall products are very important for regional and global hydrological studies, particularly for remote regions and developing countries [1, 14, 24, 29] because they provide large area coverage, high temporal and spatial resolution, and free access to near real-time data through the internet [25]. To better understand the impact of precipitation inputs on hydrological applications, the accuracy of satellite precipitation should be assessed against the reference data considering basin average precipitation [23] (in our term catchment-to-catchment or C-C) and G-G comparison [29]. Considering this aspect and to determine the performance of G-G and catchment wise rainfall, the point observed rainfall was converted to a continuous rain gauge-based gridded rainfall product using IDW interpolation techniques through ArcGIS. Tong et al. [41] elucidates that IDW is the most widely used interpolation technique in the rainfall surface preparation and it returns closer magnitude of the observed rainfall data at the gauge location without extensive efforts; hence, we considered it to be suitable for our present analysis. Thiemig et al. [42] highlighted that the IDW precipitation field showed a rather homogeneous distribution ranging over the whole basin. During IDW interpolation, we took variable radius of 200 km as maximum distance search radius with at least 6 rain gauge stations needed and power of 2 for the exponent in the relationship of weights. Hence, IDW generated gridded rainfall data was used for the present verification and comparison and two types of data were prepared from the observed and satellite rainfall gridded data for G-G comparison and data for C-C comparison.
(i) Data for Grid-to-Grid Comparison. The spatial resolution of products other than TRMM is 0.1° × 0.1° while the resolution of TRMM is 0.25° × 0.25° in world geodetic system (WGS) 1984 coordinate system. The binary format of satellite rainfall estimates were converted to raster rainfall data and projected in Lambert Azimuthal Equal Area projection system. The projection gave different spatial resolution in meter distance for different products. Using the spatial analyst tool in ArcGIS, all SRE were resized for consistency, to the same resolution as CPC-RFE2.0 (10,728 m) so that all products maintained the same number of grids in the study area (4,602 grids, each 115 km2) in the Lambert projection system. The resizing of the resolution was done in a way such that any grid or raster cell of a particular satellite product completely coincides with the corresponding grid or raster cell of another product, that is, maintaining the same analysis extension in ArcGIS. It should be noted here that the same resolution and extension were also maintained for the interpolated gridded surface rainfall of gauge data. After completing these background tasks, including gauge data interpolation, converting binary rainfall data to raster data followed by resizing all gridded data into a same resolution, rainfall value at the centre of each grid was extracted from all data sets (observed-interpolated and satellite rainfall data). So there were a total of 4,602 grid rainfall data for each day for each data set available for the verification purpose. Each satellite data set was then verified with observed-interpolated data on a daily basis followed by a summarization over month and season by means of averaging daily performance and the performance of each product was compared with one another. This is how the G-G analysis was done and the process is illustrated in Figure 3.
(ii) Data for Catchment-to-Catchment Comparison. To determine the performance of the satellite-based rainfall estimates over the subbasin or catchment scale in the Brahmaputra Basin, catchment average daily, dekadal (10 daily accumulation), and monthly satellite-based rainfall were compared with gauge interpolated rainfall.
Mei et al. [23] stated that catchment average rainfall approach allows a more direct inference on the hydrological impact of the satellite rainfall estimation error and, similarly, size of catchments also influences the satellite rainfall errors. A total of 217 subbasins or catchments (Figure 1) were delineated in the study area using the shuttle radar topography mission (SRTM) derived DEM data and an ArcGIS hydrological analysis tool. The minimum, maximum, and average sizes of those catchments were 98, 8,982, and 2,401 km2, respectively. The ArcGIS spatial analyst tool was used to generate the average daily rainfall in each catchment from the gauge interpolated rainfall and satellite rainfall data sets as well. The analysis was based on basin average rainfall rather than the usual pixel-based comparison as elucidated by Mei et al. [23]. The conceptual and (semi-) distributed hydrological model relies on catchment average rainfall data; comparison of catchment average rainfall thus gives an idea of how useful the selected satellite rainfall products are in a hydrological modelling study. The catchment average comparison between observed and satellite rainfall data has been referred to as C-C comparison in this paper.
2.5. Verification of Satellite Rainfall Estimates
There are many methods of spatial verification available that can be used to compare rain gauge measurements with SRE. In this study, the statistical measures used to compare the satellite estimations with the ground truth (rain gauge) data were taken from the results of the 3rd Algorithm Intercomparison Project of the Global Precipitation Climatology Project (GPCP) ([5, 43–45]; http://cawcr.gov.au/projects/verification/). The spatial verification methods included visual verification, continuous statistics (MAE, RMSE, and Mbias), and categorical statistics (POD and FAR) and were based on daily, dekadal (10 daily accumulation), monthly, and seasonal accumulation rain gauge and satellite estimated data. The continuous statistics were used to evaluate the performance of the satellite products in estimating the amount of rainfall whereas categorical statistics were used to access rain detection capabilities. These categorical statistics are very much important if SRE products will be used in modelling of floods because of precipitation detection. Both POD (hits) and FAR (misses) help to understand the hydrological consequences of the sources of errors in SRE products [16].
Both G-G and C-C verification were carried out as shown in Figure 3. G-G analysis was carried out using daily, monthly, and seasonal rainfall over the entire study area whereas C-C analysis was carried out using daily, dekadal, and monthly average rainfall in each catchment.
3. Results
3.1. Visual Comparison of Daily Rainfall Estimates
Despite being subjective in nature, simple visual comparison of mapped estimates and observations (eyeball verification) is one of the most effective verification methods [5]. The basinwide daily rainfall distribution of SRE and observed-interpolated rainfall map was compared visually for June 14, July 8, August 21, and September 2, 2004. The dates were chosen to test the performance of the SRE in times of heavy rainfall and correspond to the days with maximum basinwide rainfall in monsoon months in 2004. Figure 4 shows the rainfall distribution maps for July 8, the heaviest rainfall day in the heaviest monsoon, as an example.
(a) Observed-interpolated
(b) CPC-RFE2.0
(c) RFE2.0-Modified
(d) CMORPH
(e) GSMaP
(f) TRMM 3B42
In general, there was a good detection of rainfall distribution for most of the verified days despite some discrepancies; this can be attributed to the fact that most of high altitude areas suffer most from the rain gauge insufficiency problem [17]. Stisen and Sandholt [25] elucidate that this might be the issue of interpolation uncertainty due to low gauge density that could not properly capture rainfall pattern influenced by orography. Another possible source of error is that all daily precipitation stations in this domain are measured at 03Z to 03Z which is not consistent with the SRE daily accumulation. This generates a 3:15-hour bias of the rainfall accumulation. Close analysis showed that all SRE rainfall patterns were consistent with the observed rainfall in as much as heavy rainfall was detected in the southwestern, central, central-south, and southeastern parts of the basin and moderate to low rainfall in the north-central and northwestern part of the basin. RFE2.0-Modified and TRMM 3B42 corresponded significantly well with the observed-interpolated data in terms of distribution, followed by CMORPH, CPC-RFE2.0, and GSMaP showing large discrepancies with the observed daily rainfall map. All the satellite rainfall maps showed a clear underestimation of daily rainfall (Figure 4). One of the possible explanations of this underestimation of the SRE product is due to warm orographic rainfall that cannot be detected by microwave (MW) and IR sensor. Furthermore, IR cannot solve the multiple layers of raining clouds during monsoon [16]. One of the possible reasons for this behaviour of the SRE product could be the surface snow and ice screening procedure embedded in the algorithm [9]. MW sensors largely fail to discriminate between frozen hydrometers and surface snow and ice [46, 47]. Nevertheless, the SRE products show the distribution of the daily rainfall reasonably well.
3.2. Grid-to-Grid Comparison
In order to get an impression of the spatial distribution of the differences between the SRE and the interpolated rain gauge in the entire basin and not only at the gauge pixels, the different SRE were compared from the validation images on a pixel to pixel basis [17]. Under this comparison, the whole Brahmaputra was considered as a single homogenous region.
(i) Daily Rainfall. Table 3 shows the results of comparison of the different SRE products over Brahmaputra basin on July 8, 2004; RFE2.0-Modified performed best among all the satellite rainfall products. In terms of rainfall detection, it can be seen from Table 3 that CMORPH has the lowest POD (0.74) performance among the SRE products indicating slight rainfall detection problem compared to other SRE products. But FAR is zero for all SRE products.
(ii) Monthly Average of Daily Rainfall Error Statistics. The daily error found in continuous statistical analysis for the gridded rainfall was averaged over a month, and the results for the same month in the three consecutive years again were averaged to give average monthly statistics for the whole period (monthly average). The monthly average of the daily error statistics for the different satellite products from 2004 to 2006 are shown in Figure 5.
RFE2.0-Modified showed the lowest daily root mean square error (RMSE) with values between 7.3 and 11.7 mm/day, an average of 9.7 mm/day during monsoon (June to September) with 11.5, 12, 11.4, and 13.3 mm/day for CPC-RFE2.0, CMORPH, GSMaP, and TRMM 3B42, respectively (upper right panel of Figure 5). The lower left panel of Figure 5 shows the monthly average correlation coefficient of all considered SRE products. Again, RFE2.0-Modified is clearly the best product suggesting that the daily rainfall of this product corresponds well with the observed rainfall. The correlation coefficient is equal to or more than 0.4 in the premonsoon, monsoon, and postmonsoon (April to October) season, though it drops to 0.3 or even less during the dry season for RFE2.0-Modified. The correlation coefficient for other products does not exceed 0.4 in any month of the year. The monthly average multiplicative bias (Mbias) for RFE2.0-Modified for March to September is on average 0.86 (lower right panel of Figure 5) suggesting that RFE2.0-Modified underestimated daily rainfall in comparison with observed data by 14% during March to September. This was the best result, that is, least underestimation of actual rainfall among all satellite products. High Mbias during dry season for all satellite products is the result of very low rainfall amount detection in comparison to observed data. In this analysis, the gauge corrected TRMM 3B42 daily product did not perform well; possibly because of the very limited access to observed rain-gauge data in this region. However, the monthly and seasonal estimates performed well compared to CPC RFE2.0, CMORPH, and GSMaP. In summary, TRMM 3B42 is a better product when a long term average is considered, a finding that is consistent with the findings of previous studies [16, 29].
The daily error categorical statistics for the gridded rainfall were also averaged over a month, and the results for the same month in the three consecutive years were averaged to calculate average monthly statistics for the whole period (monthly average). The monthly average of the daily error categorical statistics values for the different products is shown in Figure 6. CPC-RFE2.0 and RFE2.0-Modified gave relatively similar results, with overall average daily POD values of 0.66 and 0.64, respectively, during monsoon; the other three products were somewhat different. TRMM 3B42 performed slightly better than CMORPH and GSMaP. One possible reason for the slightly better performance could be that TRMM 3B42 used gauge data compared to SRE products that used only remote sensing data. All satellite products had more or less similar results for FAR. The FAR was as expectedly less in June to September than in the other months.
(iii) Error Statistics of Monthly Rainfall. A direct analysis of monthly rainfall was also carried out by summing daily rainfall data to provide the monthly value, analysing the monthly error statistics, and averaging the results for the same month over the three years (2004–2006). The results are shown in Figure 7.
The pattern of variation of the correlation coefficient () and Mbias over the year was similar for all five satellite rainfall products, although there were clear differences in overall performance. Again, RFE2.0-Modified provided the best estimates by a considerable margin. The correlation coefficient for RFE2.0-Modified, TRMM 3B42, and CMORPH was fairly consistent from April to October with average values of 0.83, 0.70, and 0.68, respectively (lower left panel of Figure 7). The performance of CPC-RFE2.0 and GSMaP was less consistent, dropping during July and August and increasing again in September and October being 0.58 and 0.63, respectively. With the exception of CPC-RFE2.0, remaining four products provided correlation coefficient of about 0.5-0.6 in March where RFE2.0-Modified and TRMM 3B42 provided highest (0.6). In February correlation coefficient range between 0.3–0.6 where TRMM 3B42 showed better performance than other satellite rainfall products. This analysis shows that TRMM 3B42 provides a consistent correlation coefficient of monthly rainfall during February to October of 0.6 or higher, which was quite exceptional in comparison to other products. The correlation coefficient was markedly lower during November to January for all products except RFE2.0-Modified. The monthly rainfall of RFE2.0-Modified also performed best in terms of MAE, RMSE, and Mbias and TRMM 3B42 performed second best. However, the performance of these two products during premonsoon and monsoon (April to September) was close, providing the same Mbias of 0.8 and RMSE of 107 and 139 mm/month, respectively. CPC-RFE2.0 also provided Mbias of 0.8, but the RMSE was considerably high in a month, 156 mm. The least performing products were CMORPH and GSMaP which provided the same Mbias of 0.6 but the RMSE was 158 and 169 mm/month.
(iv) Evaluation of Seasonal G-G Rainfall. The monsoon season was the primary focus in this study as more than 80% of annual rainfall falls during this period, and it is the most important season for flood prediction and warning [3]. The monsoon or rainy season is important from the agriculture point of view in terms of paddy cultivation. Figure 8 shows the spatial distribution of observed-interpolated and satellite estimated average monsoon rainfall over the period of 2004 to 2006 (overall average of the monsoon season value in each of the three years). The distribution pattern of heavy, moderate, and low rainfall areas shown by RFE2.0-Modified, CMORPH, and TRMM 3B42 corresponded fairly well with that of the observed-interpolated data, but there was an underestimation of the amount. CPC-RFE2.0 and GSMaP were also able to capture the heavy, moderate, and low rainfall areas, but overall correspondence with observed rainfall was very poor. RFE2.0-Modified and CPC-RFE2.0 tended to overestimate the rainfall in the rain shadow areas in the northern part of the basin, but the other SRE products underestimated the rainfall in these areas.
(a) Observed-interpolated
(b) CPC-RFE2.0
(c) RFE2.0-Modified
(d) CMORPH
(e) GSMaP
(f) TRMM 3B42
Figure 9 shows the results of analysis with continuous statistics of the seasonal rainfall given by the different products. RFE2.0-Modified was the best product for estimating monsoon rainfall with an average RMSE of 434 mm/season, correlation coefficient of 0.85, and Mbias of 0.84. TRMM 3B42 was the next best product followed by CMORPH, CPC-RFE2.0, and GSMaP. Although CMORPH had a relatively good correlation coefficient value (0.84), the values of other parameters showed that it did not provide good estimates of monsoon rainfall.
3.3. Catchment-to-Catchment Comparison (C-C)
Evaluating the error propagation of satellite rainfall through the prism of surface hydrology is a very challenging task because it relates too many factors, which include (i) specifications of the satellite rainfall products and its resolution, (ii) scale of the basin, (iii) spatiotemporal scale of the hydrologic variable of interest, (iv) the level of complexity and physical processes represented by the hydrologic model used, and (v) regional characteristics [28]. The C-C analysis aimed to evaluate the performance of satellite products in estimating the amount of rainfall in individual catchments and thus capturing the spatial variation resulting from the complex topography, significant elevation change, and scale rather than the usual pixel-based comparison [23, 25]. These results are particularly useful for understanding the applicability of satellite rainfall for developing hydrological applications.
(i) Evaluation of Daily Rainfall. As the main focus was on the monsoon season and heavy rainfall that might lead to flooding, the analysis compared the performance of catchment values from SRE products compared to the observed rain gauge product on July 8, 2004. The results are shown in Table 4. RFE2.0-Modified performed better than all other satellite rainfall products for all parameters except FAR. Some values (RMSE, Mbias) were better than in the G-G comparison. There was little difference in performance among the other products.
(ii) Monthly Average of Daily Rainfall Error Statistics. The daily error continuous statistics for average daily rainfall in each of the 217 catchments were averaged over a month, and the results for the same month in the three consecutive years again were averaged to give average monthly statistics for the whole period (monthly average). The monthly averages of the daily error statistics for the different satellite products from 2004 to 2006 are shown in Figure 10. RFE2.0-Modified performed the best, followed by TRMM 3B42, but all the SRE showed the same trend of low errors in the dry months and high errors in the wet months. The categorical statistical method is not appropriate for C-C analysis, hence not done for this C-C analysis.
The results of C-C analysis in monthly average of daily error statistics gave almost the same result as G-G analysis. However, C-C analysis provided slightly lower MAE and RMSE for all products than G-G analysis, indicating that averaging rainfall over a larger area (i.e., from grid to catchment) tends to minimize the errors in magnitude, though the changes may happen in big margin and depending on the size of the catchments. On the other hand, the changes in correlation coefficient and Mbias from G-G to C-C analysis in daily rainfall analysis was not significant at all.
(iii) Evaluation of Average Dekadal C-C Rainfall. Herman et al. [39] evaluated 10-day (dekadal) African rainfall estimates created for famine early warning systems and highlighted that 10-day precipitation estimates are generated for drought monitoring purpose, a standard period defined by famine early warning systems and considered appropriate for hydrological applications. Also, to demonstrate the utility in flood forecasting, because depending on upstream basin size, flood routing lag time may vary from daily to dekadal or so. Visual and statistical comparisons were made of gauge-observed and satellite-based catchment average dekadal rainfall estimates (catchment wide 10-daily accumulation) over the period 2004 to 2006. The results are shown in Figure 11. There was general agreement in the overall pattern of rainfall distribution between observed and satellite estimated data, with SRE following the same trend of high and low rainfall intensity as the observed-interpolated rainfall. However, the amount of rainfall was generally underestimated. The RFE2.0-Modified satellite rainfall showed the best correspondence with observed rainfall. The statistical analysis also showed that RFE2.0-Modified provides much better estimates of catchment-wise dekadal rainfall than the other products, with a coefficient of determination () of 0.96, compared to 0.83, 0.84, 0.86, and 0.88 for CPC-RFE2.0, GSMaP, CMORPH, and TRMM 3B42, respectively. Symeonakis et al. [17] highlighted that dekadal sums yielded better results than the respective daily data, which is in agreement with the findings of this study.
4. Discussion and Conclusion
The evolution of regional and global SRE products with high temporal and spatial resolution has opened up new opportunities for hydrological applications in data sparse regions. The main purpose of the present study was to evaluate the estimates from three global and two regional SRE products in comparison with observed rain gauge data in the Brahmaputra river basin, in order to determine their operational viability for use in hydrological applications in a region with sensitivity to orographic effects. The evaluation was carried out at daily, dekadal, monthly, and season temporal scales for the period 2004 to 2006 using G-G and C-C approaches, with visual analysis, continuous verification statistics, and categorical verification statistics.
The estimates from the five SRE products generally showed a qualitative agreement with observed rain-gauge data and rainfall events but differences in quantitative values. One possible reason for underestimation of rainfall amount is mainly attributable to warm orographic rain which cannot be detected by the IR as well as MW sensors. IR algorithms use cloud-top temperature thresholds that are too cold for the orographic clouds; leading to underestimation of orographic rain. Passive microwave (PM) algorithms underestimate rainfall from orographic rain, which may not produce much ice aloft [47]. CPC-RFE2.0 and RFE2.0-Modified performed better in the categorical verification statistics and showed good rain/no-rain detection; the other three products performed less accurately in POD and FAR. The average daily Mbias from 2004 to 2006 for RFE2.0-Modified using G-G comparison was 0.86; that is, RFE2.0-Modified underestimates rainfall by 14% on average. RFE2.0-Modified had the lowest values for daily RMSE. The seasonal average error statistics for RFE2.0-Modified showed that rainfall occurrence was underestimated by about 16% in the monsoon, 20% in premonsoon, not at all postmonsoon, and only 0.03% in winter; CMORPH, TRMM 3B42, and CPC-RFE2.0 overestimated rainfall slightly in postmonsoon. CPC-RFE2.0 and RFE2.0-Modified had a positive bias in the rain shadow areas of the trans-Himalaya, one of the possible reasons for overestimation of rainfall amount is that the MW sensor used by CPC-RFE2.0 presumed very cold surfaces and ice cover mountain tops as a rain cloud [47]. Further, IR based techniques may overestimate of rainfall due to misidentification of some cold clouds, such as cirrus, that may not generate any rainfall [16], whereas GSMaP, TRMM 3B42, and CMORPH underestimated the rainfall amount in these areas. One of the possible reasons for this behaviour of SRE products could be the surface snow and ice screening procedure embedded in the algorithm due to the fact that MW sensors largely fail to discriminate between frozen hydrometeors and surface snow and ice, another possible reason for limitations in spatial and temporal sampling by the MW sensors [16].
The SRE estimates were slightly better when the river basin was divided into catchments rather than considering whole Brahmaputra as a single unit (Grid-to-Grid). Time series comparison of C-C basin average dekadal rainfall from 2004 to 2006 showed strong agreement between RFE2.0-Modified and the observed data with a correlation coefficient of 0.96. The potential of RFE2.0-Modified has been shown in a small catchment (the Narayani and the Bagmati river basin) where it was found to be suitable for use in hydrological applications [1, 48]. The other SRE products also performed better but still underestimated the rainfall amount.
In summary, the results indicate that SRE provides reasonable rainfall estimates over the Brahmaputra river basin. RFE2.0-Modified showed the best correspondence with observed rainfall and was the best product in the current evaluation followed by TRMM 3B42, CMORPH, CPC-RFE2.0, and GSMaP. Overall, in the rugged topography of the Brahmaputra river basin, SRE products which incorporated gauge data performed better than the products that only used remotely sensed data. The effect of additional local gauges on the quality of the products was clear in the present study. It also revealed that evaluation of SRE products at monthly and seasonal temporal resolution provided better results which could be considered as useful for overall water resource assessment of the basin.
Conflict of Interests
The authors declare no conflict of interests.
Acknowledgments
This study was carried out under the regional Projects “Application of Satellite Rainfall Estimates in the Hindu Kush Himalaya Region” and “HIMALA” supported by USAID/OFDA. The authors would like to express their sincere gratitude to their regional partners and the NOAA Climate Prediction Center for providing data for the study. They also thank Dr. David Molden, Dr. Eklabya Sharma and Dr. Aditi Mukherji for their encouragement to prepare this paper. They acknowledge the assistance of Dr. Joe Turk for providing an assessment of the analysis and a careful proofreading of the paper.