Assessment of satellite precipitation products at different time scales over a cyclone prone coastal river basin in India

IMD

Indian Meteorological Department

TRMM

Tropical Rainfall Measurement Mission

CMORPH

Climate Prediction CenterMORPHing

PERSIANN

Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks

CDR

Climate Data Record

IMERG

Integrated Multi-satellitE Retrievals for Global Precipitation Measurement

GSMaP

Global Satellite Mapping of Precipitation

APHRODITE

Asian Precipitation – Highly-Resolved Observational Data Integration Towards Evaluation

GPCC

Global Precipitation Climatology Centre

SWAT

Soil and Water Assessment Tool

VIC

Variable Infiltration Capacity

GXAJ

Grid-Xinanjiang model

RRI

Rainfall-Runoff-Inundation

CREST

Coupled Routing and Excess STorage

Precipitation data play a vital role in many hydroclimatic applications, such as water resources development projects (Zhu et al. 2020), management practices (Beck et al. 2017; Dinku 2020; Kumar & Singh 2022), planning for hydropower, flood and drought monitoring (Tan & Santo 2018; Chen et al. 2020b; Singh et al. 2020; Suseno & Yamada 2020; Tarnavsky & Bonifacio 2020; Yu et al. 2020), and meteorological investigations (Mishra 2019; Niyogi et al. 2020; Thakur et al. 2020; Ojha et al. 2021). The accurate precipitation estimation at fine resolution is critical owing to the spatiotemporal variability of rainfall (Guntu et al., 2020). Indeed, gauge-based precipitation measurement is more accurate and reliable. However, a requirement for an adequate number of stations in any region elevates the cost of installation, operation, and maintenance (Agarwal et al. 2018). Besides, the assumption of uniform distribution of precipitation is questionable over large areas, especially in mountainous regions (Tan & Santo 2018). Also, ground-based measurements often have missing precipitation records. The missing precipitation data must be avoided in hydrological models, as the same requires continuous precipitation data.

Nowadays, many precipitation products are available with different spatial and temporal resolutions for the globe (Maggioni et al. 2016). These products can be grouped as gauge-only products, only satellite products, and a combination of satellite–gauge products. In recent decades, Satellite Precipitation Products (SPPs) have provided additional benefits such as freely available data with a high spatial and temporal resolution, global coverage, cost-effective coverage of mountains and oceanic regions, and, more importantly, uninterrupted data (Xue et al. 2013; Serrat-Capdevila et al. 2014; Ren et al. 2018). The SPPs are used in many applications such as weather prediction, agriculture monitoring, and hydrological models (Bitew & Gebremichael 2011; Belabid et al. 2019; Dandridge et al. 2019; Jamro et al. 2019; Dars et al. 2020; Krisnayanti et al. 2021) such as Soil and Water Assessment Tool (SWAT) (Rathinasamy et al., 2014; Tuo et al. 2016; Le et al. 2020; Wang et al. 2020), Variable Infiltration Capacity (VIC) (Xie & Xiong 2011; Shah & Mishra 2016; Beria et al. 2017; Shayeghi et al. 2020), Grid-Xinanjiang model (GXAJ) (Yuan et al. 2019), Rainfall-Runoff-Inundation (RRI) (Tam et al. 2019), and Coupled Routing and Excess Storage (CREST) (Tang et al. 2019).

Many studies evaluated the performance of precipitation products at a large basin scale (Saha et al. 2010; Fuka et al. 2014; Zhang et al. 2021). For instance, Tang et al. (2019) evaluated the four SPPs using the SWAT model. They found that the Multi-Source Weighted-Ensemble Precipitation (MSWEP) performed relatively better than the other products in the Mekong River Basin in South-East Asia. Try et al. (2020) concluded that Tropical Rainfall Measuring Mission (TRMM), Asian Precipitation – Highly-Resolved Observational Data Integration Towards Evaluation (APHRODITE), and Global Precipitation Climatology Centre (GPCC) datasets were suitable for rainfall-runoff and flood inundation modeling at the same basin. Dhanesh et al. (2020) evaluated Climate Hazards Group InfraRed Rainfall with Station (CHIRPS) and Climate Forecast System Reanalysis (CFSR) using the SWAT model and found that CHIRPS outperformed CFSR at several basins with different climatic zones. However, only a few studies investigate the performance of precipitation products at a local or regional scale (Bitew & Gebremichael 2011; Yuan et al. 2017). This highlights the need to investigate the applicability of various gridded data to such basins.

In the Indian scenario, the different SPPs were evaluated for their accuracy and uncertainty by comparing them with Indian Meteorological Department (IMD) gridded data (Sunilkumar et al. 2015; Shah & Mishra 2016; Beria et al. 2017; Himanshu et al. 2018; Prakash et al. 2018; Shukla et al. 2019; Thakur et al. 2020; Yeggina et al. 2020). For instance, Sunilkumar et al. (2015) reported on the evaluation and intercomparison of the different SPPs (TRMM-3B42, PERSIANN, CMORPH) that Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) and Climate Prediction Center MORPHing method (CMORPH) severely underestimated the northeast monsoon. The performance of TRMM3B42 was found to be relatively better than the others. However, Yeggina et al. (2020) found that TRMM-3B42V7 notably overestimates the light precipitation and underestimates extreme precipitation over a small river basin in southern India. Similarly, during the hydrological evaluation of TRMM-3B42V7 using the SWAT model over a small catchment in southwestern India, Himanshu et al. (2018) found that the product underestimated the heavy rainfall amount and showed limited skill on a daily scale and fairly good prediction on a monthly scale. Shukla et al. (2019) revealed that TRMM-3B42V7 is limited in capturing rainfall during the monsoon season over the Himalayan region. However, applying bias correction resulted in satisfactory improvement. Due to significant errors in the product over the Himalayan region, Bharti & Singh (2015), Himanshu et al. (2018), and Shukla et al. (2019) recommended using bias correction methods and improving the retrieval algorithms. Prakash et al. (2018) compared TRMM, Integrated Multi-satellite Retrievals for Global precipitation measurement (IMERG), and Global Satellite Mapping of Precipitation (GSMaP) during the monsoon of 2014 and reported that both IMERG and GSMaP showed improvement over TRMM Multi-satellite Precipitation Analysis (TMPA) in the estimation of low to moderate rainfall distribution across the country. However, the results were preliminary observations solely based on the monsoon season of 2014. Hence, further studies on additional monsoon season were recommended. Most of these studies have focused on quantitative evaluation, and less work has been reported on these products' ability to drive hydrological models.

Therefore, it will be of great interest to the hydrologic community to understand the accuracy of data and its ability to drive hydrological models. Further, it will be helpful to investigate the performance of these precipitations in capturing extreme events in a cyclone-prone region.

This study evaluates three satellite precipitation products (GPM-IMERG, TRMM3B42, and PERSIANN-CDR) over the Vamsadhara River Basin (VRB). The basin is in a tropical climatic regime with semi-arid and semi-humid characteristics. The mainstream originates in the Kalahandi district of Odisha state and flows through approximately 254 km before joining the Bay of Bengal at Kalingapatnam, Andhra Pradesh. Due to the monsoonal winds from the Bay of Bengal branch, the basin often experiences various cyclones during the monsoon season. The specific objectives of the study are:

• (i)

  To evaluate three satellite precipitation products for viz. TRMM-3B42V7, IMERG-GPM V6 (0.1° × 0.1°) and PERSIANN-CDR (0.25° × 0.25°) at different temporal and spatial scales during 2000–2018 at cyclone-prone river basin and

• (ii)

  to assess their ability to drive hydrological models for streamflow simulation at daily and monthly time scales.

The rest of the manuscript has the following structure. Section 2 describes the data and methods applied in the study, and the results from the study are presented in Section 3. The discussion based on the analysis of the results and concluding remarks are provided in Sections 4 and 5, respectively.

VRB received an annual rainfall of 1,281 mm during 2000–2018 (Yeditha et al. 2020), and most of the rainfall is covered during the southwest monsoon months of June to September. The Indian summer monsoon has two branches from the high-pressure regions of the Arabian Sea and the Bay of Bengal. Due to the monsoonal winds from the Bay of Bengal branch, the basin often experiences various cyclones during the monsoon season. This often invokes various socio-economic losses. For instance, the highly severe tropical cyclonic storm ‘Fani’ recently hit the state of Odisha in 2019. This was the strongest cyclone after the 1999 Odisha cyclone.

Similarly, the region was hit by the ‘HUD HUD’ cyclone in 2015. As for the soil characteristics of the basin, around 94% of the basin is covered by sandy clay loam soil.

In this study, three satellite precipitation products (SPPs), namely, 1) GPM-IMERGV6 data, 2) TRMM 3B42 data, and 3) PERSIANN-CDR data, are used for evaluation. The gauge-based IMD gridded dataset is used as a benchmark for evaluating these SPPs. Detailed information about the precipitation data products is shown in Table 1.

The Global Precipitation Measurement (GPM) project was developed by the combined effort of the Japan Aerospace Exploration Agency (JAXA) and the National Aeronautics and Space Administration (NASA) of the United States. GPM is a widely used precipitation product throughout the world, and it is also an updated precipitation product of the TRMM. GPM has two primary sensors: one is GPM Microwave Imager (GMI), which helps measure precipitation droplet size, type, and intensity; the second one is a Dual-frequency Precipitation Radar (DPR), which measures the internal structure of clouds (Hou et al. 2014). The GPM microwave imager has more frequency channels than TRMM, and the precipitation radar is more advanced (Hou et al. 2014). More importantly, spatial coverage (60° S–60° N) and the spatiotemporal resolution (30 min and (0.1° × 0.1°)) have been improved compared to the previous SPPs.

The GPM team has developed GPM-IMERG precipitation products. The IMERG algorithm first calibrated the microwave precipitation estimates (derived from the constellation of the microwave satellite sensors) using the Global Precipitation Climatology Project (GPCP) monthly estimates to reduce bias. The algorithm uses the Climate Prediction Centre Morphing-Kalman Filter (CMORPH-KF) quasi-Lagrangian time interpolation to estimate finer time and space scales. The CMORPH-KF quasi-Lagrangian interpolation uses both the inter-calibrated microwave estimates and the PERSIANN-CCS estimates (Infra-Red estimates) to generate precipitation estimates at half-hourly temporal resolution. Currently, three IMERG products are available: early run, late run, and final run, with a latency period of 4 h, 14 h, and 3.5 months, respectively. The first two products are helpful in real-time application, whereas the latter can be used for water balance studies. All these products are available at 0.1° × 0.1° spatial resolution over the fully global domain.

Presently, IMERG is at its Version 06 stage (https://gpm.nasa.gov/missions/two-decades-imerg-resources). A new aspect of the data processing in Version 06 IMERG is that the algorithm combines the data collected during the TRMM era (2000–2014) with the precipitation estimates collected during the GPM era. Due to this, IMERG has now been available since June 2000 (https://gpm.nasa.gov/data/directory). The ‘Final run’ of IMERG combines the GPCC Monitoring product, the V8 Full Data Analysis, for most of the time (currently 2000–2019).

The Tropical Rainfall Measuring Mission (TRMM) is a joint mission developed by the Japan Aerospace Exploration Agency (JAXA) and the National Aeronautics and Space Administration (NASA) to measure tropical precipitation. The TRMM Multi-Satellite Precipitation Analysis (TMPA) product is obtained from the combination of multi-satellite (microwave and infrared satellites) and GPCC monthly gauge precipitation, and more details are available about TMPA in Huffman et al. (2007). TRMM 3B42 version 7 (TRMM-3B42V7) is one type of TMPA product, and it covers a latitude from 50°N to 50°S with a resolution of 0.25° from 1997 to 2018. The TRMM 3B42 daily precipitation data were obtained from this website https://disc.gsfc.nasa.gov/datasets/TRMM_3B42_Daily_7/summary.

PERSIANN-CDR is retrospective satellite precipitation data developed by the National Oceanic and Atmospheric Administration (NOAA) which is widely used to study climatology and hydrological studies (Ashouri et al. 2015). The PERSIANN-CDR algorithm is used for generating rain rates from satellite imagery using an artificial neural network (ANN) model, which entirely depends on the combination of Infrared (IR) and Passive Microwave (PMW) observations developed by geostationary and low earth orbit satellites, respectively (Nguyen et al. 2018). This product has been available daily from 1983 to the present, covering between 60° S and 60° N with 0.25° resolution. PERSIANN-CDR products have been downloaded from this website https://chrsdata.eng.uci.edu/#tabPERSIANNCDR, and for more details on the PERSIANN-CDR product, the readers can refer to Ashouri et al. (2015) and Nguyen et al. (2018).

In the present study, the IMD gridded data (IMD 4) having a spatial resolution of 0.25° × 0.25° has been used as ground validation. The temporal resolution of the data set is daily. The data sets were developed based on the rainfall records from 6,955 rain gauges across the country (Pai et al. 2014). Around 30 rain gauges fall within the study region considered. These rain gauges were partly maintained by IMD, Agromet Department, and mainly by the State Government. Using the Shepard Interpolation method, the collected rainfall from the rain gauges has been converted into the gridded data set. For detailed information, the readers can refer to Pai et al. (2014). The data set is available from 1901 to the present; however, for this study, the data set for 2000–2018 is considered. The datasets can be obtained from the Indian Meteorological Department, Pune, https://imdpune.gov.in/Clim_Pred_LRF_New/Grided_Data_Download.html. Several studies confirm the reliability of the IMD gridded dataset (Venkata Rao et al. 2020; Yeggina et al. 2020). Hence, the IMD gridded data is used as a benchmark for comparing the three satellite precipitation products.

The spatial resolution of the satellite precipitation products, TRMM, and PERSIANN, was 0.25 degrees, the same as the IMD dataset. However, the IMERG dataset has a higher spatial resolution of 0.1 deg. The IMERG data was sampled at the grid locations of the IMD data by using bilinear interpolation to have the same spatial resolution for comparison.

The hydrological model requires both spatial and temporal data as input. The temporal data are meteorological (such as temperature, solar radiation, relative humidity, and wind speed) and hydrological observation data (discharge). The meteorological data were collected from the Climate Forecast System Reanalysis (CFSR) reanalysis product. These data were downloaded from https://globalweather.tamu.edu/. The daily discharge data at the Kasinagar gauging station was downloaded from https://indiawris.gov.in/wris/. The spatial information includes the digital elevation model, land use and cover, and soil characteristics. Table 2 describes the spatial, weather, and discharge input data used in developing the SWAT model.

Statistical analysis is used to evaluate satellite precipitation products with reference precipitation data. For this study, we have used (i) continuous statistical metrics (Correlation Coefficient (CC), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Percentage Bias (PB), Standard Deviation (SD)), (ii) categorical statistical metrics (Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI)), and (iii) empirical distribution function and (iv) hydrological model-based investigation for evaluating the different precipitation products such as TRMM-3B42V7, GPM-IMERGV6, PERSIANN-CDR with reference of IMD precipitation product throughout 2000–2018. All the evaluations were done at different temporal scales (daily, monthly, seasonal, and yearly) and spatial (at grid and basin-scale) scales.

The categorical statistical metrics such as Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI) have been used for the detection of precipitation capability of satellite precipitation products with reference precipitation products depending on rainy and non-rainy days. POD represents the hit rate, which means the ratio of correctly detected by satellite as a rainfall event to the actual total rainfall events measured by reference data. It varies from 0 to 1 (1 is the ideal value). FAR is known as a false rate. It represents the ratio of satellite-detected falsely as a rainfall event to the total rainfall events detected by satellite. It varies from 0 to 1 (0 is the perfect value). CSI is the ratio of correctly detected rainfall events by satellite to the total number of hits, misses, and false alarms. CSI values range from 0 to 1 (1 is a suitable value), as shown in Table 3.

Table 3

Details of the continuous and categorical statistical metrics used for evaluating the precipitation products

Statistic Metric .	Equation .	Perfect Value .	Description .
Nash-Sutcliffe Coefficient Efficiency (NSE)		1	It measures the model performance
Correlation Coefficient (CC)		1	It represents the degree of agreement in between observed and measured data
Root Mean Square Error (RMSE)		0	It represents the magnitude of average error
Percent Bias (Pbias)		0	It represents the deviation between observed and measured data
Mean Absolute Error (MAE)		0	It measures the average difference in between observed and estimated data
Probability of Detection (POD)	POD = H/(H + M)	1	It represents the detected rainfall events correctly by satellite out of rainfall events by reference data.
False Alarm Ratio (FAR)	FAR = F/(F + H)	0	It represents the false, detected rainfall event by satellite when reference data showed no rainfall
Critical Success Index (CSI)	CSI = H/(H + M + F)	1	It shows more accuracy when correct negative values are not considered

Statistic Metric .	Equation .	Perfect Value .	Description .
Nash-Sutcliffe Coefficient Efficiency (NSE)		1	It measures the model performance
Correlation Coefficient (CC)		1	It represents the degree of agreement in between observed and measured data
Root Mean Square Error (RMSE)		0	It represents the magnitude of average error
Percent Bias (Pbias)		0	It represents the deviation between observed and measured data
Mean Absolute Error (MAE)		0	It measures the average difference in between observed and estimated data
Probability of Detection (POD)	POD = H/(H + M)	1	It represents the detected rainfall events correctly by satellite out of rainfall events by reference data.
False Alarm Ratio (FAR)	FAR = F/(F + H)	0	It represents the false, detected rainfall event by satellite when reference data showed no rainfall
Critical Success Index (CSI)	CSI = H/(H + M + F)	1	It shows more accuracy when correct negative values are not considered

View Large

Apart from the metrics discussed above, the empirical distribution is a handy feature for evaluating the distribution of the total rainfall into different precipitation intensity categories. As per World Meteorological Organization (WMO) standard Geneva, 2012, daily rainfall is divided into seven classes based on rainfall intensity, namely 0–1 mm/day, 1–2 mm/day, 2–5 mm/day, 5–10 mm/day (high moderate rain), 10–20 mm/day (low heavy rain), 20–50 mm/day (high heavy rain), >50 mm/day (extreme rain) Setti et al. (2020a, 2020b).

To evaluate SPPs quantitatively, we have used statistical accuracy indicators such as CC, SD, MAE, Pbias, and RMSE. CC described the degree of linear correlation between the satellite and observed precipitation products, ranging from −1 to 1, representing perfect correlation, either negative or positive correlation, respectively (CC value 0 represents no correlation). The RMSE, MAE, and Pbias measure the error and bias of satellite precipitation products with reference precipitation products, respectively (see Table 3). When the RMSE value is smaller, the SPPs data are close to the observed precipitation data. Pbias has positive and negative values representing the underestimation and overestimation by the SPPs, respectively.

In this study, the Soil Water Assessment Tool (SWAT) model was used for the hydrologic model-based evaluation of the precipitation products. The satellite precipitation products were used as input precipitation for a calibrated model, and the streamflow simulation was compared at different time scales. Details about the SWAT model are provided in section 2.4.1. To evaluate the effect of four precipitation products on streamflow, two scenarios were generated: Scenario A and Scenario B. In scenario A, first, the model was calibrated with IMD precipitation data; then, the same model was run with other precipitation products such as GPM-IMERGV6, TRMM-3B42V7, and PERSIANN-CDR. For this purpose, first, the SWAT model was developed with IMD data; after calibration and validation, the model was driven using other precipitation products for the validation period. In scenario B, individual models were calibrated and validated for each precipitation product. The results of scenario B are discussed in the supplementary material.

Based on data availability, such as observed streamflow data and precipitation products, the total period is divided into three parts: warm-up period (2000–2001), calibration period (2002–2010), and validation period (2011–2016). Additionally, the models were calibrated and validated at daily and monthly time scales using the daily observed streamflow and mean monthly streamflow for both scenarios.

The Agriculture Research Service of the United States Department of Agriculture developed the SWAT model. The SWAT model can be used for predicting the different hydrological components like erosion, water balance, vegetation growth, evapotranspiration, and water quality of watersheds by solving processes based on the water balance equation in daily time steps. The surface runoff is estimated by using the Soil Conservation Service (SCS) -Curve Number (CN) method. The SWAT model is simulated percolation using the storage routing technique, which assumes that percolation occurs when the water content exceeds the field capacity and the layer below is unsaturated (Abbaspour et al. 2015). Also, the SWAT model has three methods for estimating Potential Evapotranspiration (PET). These are Priestley Taylor (Priestley & Taylor, 1972), Hargreaves (Hargreaves & Samani 1985), and Penman-Monteith (Monteith, 1965). In this study, the Penman-Monteith method is used for estimating potential evapotranspiration. SWAT divides the whole basin into sub-basins and Hydrological Response Units (HRUs) based on the unique land use, soil, and slope combination (Setti et al. 2018).

The QSWAT (version 1.3) tool used to develop the Vamsadhara River Basin model interfaced with QGIS 2.6.1. All spatial input data was re-projected and resampled to 30 m spatial resolution for the SWAT model. The generation of stream network and delineation of watersheds was done using the TauDEM tool in QSWAT. The entire study area was divided into 26 sub-basins using a threshold of 150 km² then subdivided into 543 Hydrological Response Units (HRUs). The HRUs are the smallest component in the SWAT model and are unique land use, soil, and slope combinations. The sensitivity analysis, auto-calibration, and validation were done by using SWAT CUP 2012 software with the SUFI-2 algorithm (please refer to Abbaspour et al. (2007) and Setti et al. (2020a, 2020b) for more details). Numerous iterations, with 2000 simulations, were run for model calibration of the IMD precipitation product using SWAT CUP. This study adopts a global sensitivity method for sensitivity analysis (Abbaspour et al. 2007).

In this study, we evaluated the three satellite precipitation products (TRMM-3B42V7, GPM-IMERGV6, PERSIANN-CDR) compared to IMD gridded precipitation products using statistical metrics such as CC RMSE, MAE, RB, and SD, as shown in Table 4. As per statistical metrics of all precipitation products at a daily time scale, GPM-IMERGV6 has a better correlation value (0.48) compared to TRMM-3B42V7 (0.44) and PERSIANN-CDR (0.36) precipitation products. The GPM-IMERGV6 product has the lowest error value in terms of RMSE (9.5 mm/d) and MAE (3.8 mm/d) compared to TRMM-3B42V7 and PERSIANN-CDR data. Thus, based on statistical metrics, GPM-IMERGV6 is close to the IMD precipitation product. PERSIANN-CDR data show inferior performance compared to the TRMM-3B42V7 and GPM-IMERGV6. The three satellite products (TRMM-3B42V7, IMERGV6, PERSIANN-CDR) seem to overestimate precipitation amount compared to IMD precipitation amount for the period of 2000–2018 as the negative percentage bias (%) range from 3.8% (GPM-IMERGV6) to 22.4% (PERSIANN-CDR). The standard deviation was low (9 mm) for TRMM-3B42V7 and was 9.5 and 12 mm for GPM-IMERGV6 and PERSIANN satellite precipitation products. Overall, the IMERGV6 and PERSIANN-CDR products showed the best and worst performance, respectively, based on five statistical metrics, as shown in Table 4.

Table 5 presents the distribution of the rainfall events in terms of the rainfall intensity classifications. Among the seven classes, the very light rain (0 to 1 mm/d) class has the highest probability, and the extreme rain (>50 mm/d) class has the lowest probability in both the SPPs and gridded precipitation product.

It is observed that the GPM-IMERGV6 and TRMM-3B42V7 have frequent very light rain indicating a high false alarm rate. However, the PERSIANN-CDR (943 days) have a more or less similar number of rainy days with the IMD (930 days) (see Table 6). The IMD receives 1.6% of the total rainfall in terms of precipitation amount, while the GPM, TRMM-3B42V7, and PERSIANN-CDR receive 2.8, 2.1, and 1%, respectively.

In this rainfall category, the GPM and TRMM-3B42V7-based estimates are close to IMD regarding the number of rainfall days and the amount of precipitation. However, PERSIANN-CDR data has fewer rainy days and underestimates the magnitude of light rain and low, moderate events.

16.8% (IMD) of the total precipitation falls in this rainfall category. For the GPM, TRMM-3B42V7 and PERSIANN-CDR, 16.9%, 16.7%, and 9.9% fall in this category, respectively. This indicates that the GPM and TRMM-3B42V7 could capture the magnitude of rainfall in this category. However, the PERSIANN-CDR receives noticeably less rainfall.

Based on the values shown in Table 5, TRMM-3B42V7 (514 days) and GPM-IMERGV6 (446 days) have a slight tendency for false alarms and misses, respectively, concerning the IMD data (472 days). But the estimates for PERSIANN-CDR seem to have a higher tendency for false alarms (584 days). This means that the PERSIANN-CDR has more frequent low heavy rains, contributing to the overall high overestimation of the magnitude of rainfall.

In this category also, the GPM (250 days, 29.4%) and TRMM-3B42V7 (268 days, 29.9%) were close to the IMD (230 days, 27.9%) in terms of the number of rainy days and the percentage (out of the total rainfall depth) of the rainfall received. However, the PERSIANN-CDR (516 days and 41.3%) data-based estimates overestimate the precipitation. A similar pattern was observed for extreme rain, where PERSIANN-CDR data showed significant false alarms and overestimations. This frequent heavy and extreme rainfall resulted in overestimated rainfall at the basin.

In the case of average Critical Success Index (CSI) values, as shown in Figure 7, the PERSIANN-CDR data shows a low success rate compared to the TRMM-3B42V7 and GPM-IMERGV6 for all three thresholds.

Overall, from the analysis of the daily scale, it can be summarized that the TRMM-3B42V7 and GPM-IMERG V6 estimates are close to IMD based on the statistical analysis and detection metrics. In contrast, the PERSIANN-CDR estimates seem to overestimate and provide false alarms, particularly during high rainfall days.

Comparing the average monthly precipitation for 19 years between IMD gridded and satellite precipitation products, there is a tendency for slight overestimation during monsoon months (June to September) using the GPM-IMERGV6 and TRMM-3B42V7 data. But the PERSIANN product tends to highly overestimate precipitation during monsoon months and underestimate from November to March. Patterns in mean absolute error (MAE) for all satellite precipitation products followed the average monthly precipitation pattern of precipitation products. The higher MAE values occurred in monsoon months when precipitation is high, with lower MAE in summer months when precipitation is low.

The Indian Meteorological Department classifies four seasons in India, namely spring (March to May), summer (June to September), fall (October to December), and winter (January and February). The SPPs were evaluated separately at each season.

On watershed-scale analysis, it is observed that all the products performed well in the fall and summer seasons, as indicated by higher correlation coefficients. All the products overestimated the mean rainfall during the summer season and underestimated it during the fall and winter. The errors were highest in the PERSIANN product and lowest in IMERG during the spring and summer. The systematic error in IMERG was considerably reduced during the spring season. The TRMM performed relatively better than IMERG during the winter season, as indicated by higher CC and lower percent bias. The detailed results are shown in Table 6.

The annual total precipitation data is obtained from the accumulated total monthly precipitation. The correlation between IMD gridded data and satellite precipitation products was 0.62, 0.62, and 0.23 by TRMM-3B42V7, IMERGV6, and PERSIANN-CDR products, respectively, as shown in Table 4. The PERSIANN-CDR data had more significant error values of RMSE, and MAE is 756 mm/yr and 679 mm/yr, respectively, compared to other data sets. GPM-IMERG V6 data product has better error values of 182 mm/yr for RMSE and 160 mm/yr for MAE than the TRMM-3B42V7 product, as shown in Table 4. TRMM-3B42V7 and IMERGV6 data products have slight differences with RMSE and MAE error values, but PERSIANN-CDR data shows the worst performance with the highest error values. The standard deviation for all products is 187, 169, and 304 for TRMM-3B42V7, IMERGV6, and PERSIAN data, respectively. Based on this analysis at the annual time scale, the GPM-IMERGV6 product performs better than the other two products, such as TRMM-3B42V7 and PERSIANN-CDR data products; TRMM-3B42V7 data shows moderate performance.

This study observed the IMD precipitation product; the average annual precipitation is 1,295 mm/y, and minimum and maximum annual precipitation is 917 mm/y (in 2011) and 1,604 mm/y (in 2018), respectively, for 19 years. The average annual precipitations for the TRMM-3B42V7, GPM-IMERGV6 and PERSIANN-CDR satellite products are 1,408 mm/y, 1,345 mm/y, and 1,974 mm/y, respectively. Interestingly, the PERSIANN-CDR has a huge overestimation of 679 mm/year compared to IMD's average annual precipitation (1,295 mm/y).

Overall, the GPM-IMERGV6 product performs better than the other products based on the statistical metrics at all times.

Before calibrating the model, the sensitivity analysis was performed to identify the sensitive parameters with a global sensitivity analysis method using the SUFI-2 algorithm (part of the SWAT-CUP tool). Out of several parameters, 18 were selected (shown in Table 7) for streamflow simulation based on other literature for the study region (Setti et al. 2020a, 2020b). The model was calibrated at daily and monthly time scales; at each scale, 18 parameters were used for model calibration.

This paper comprehensively analyses the reliability and applicability of the latest satellite precipitation datasets (PERSIANN-CDR, TRMM-3B42V7–3B42, GPM-IMERGV6) against the gauge based gridded datasets reported for a cyclone-prone coastal river basin of peninsular India. For this purpose, the evaluation and comparison of the precipitation estimates were performed at different temporal scales. Following this, the spatial comparison was studied along with the suitability of each product to drive a hydrologic model with the SWAT model at daily and monthly time steps

At the daily scale, the performance of the SPPs varied for different rainfall intensity ranges, even though the PERSIANN-CDR is closer to gauge precipitation data when the precipitation range is very light rainfall, for moderate and heavy rainfall events, 3B43V7 and GPM-IMERG V6 were closer to the gauge estimates than PERSIANN-CDR. All the SPPs investigated in this study showed an overestimation of rainfall and further comparison among the precipitation data sets. The PERSIANN-CDR data tend to have higher FAR values for all rainfall thresholds, indicating a high tendency for false alarms. Similarly, in CSI values, the PERSIANN-CDR data shows a low success rate compared to the TRMM-3B42V7 and GPM-IMERGV6 for different intensity ranges. It was observed that the POD for all SPPs varies with the space and intensity of the rainfall. The POD is high in the northwestern part of the catchment and low in the southern part of the catchment. It was also observed that all the SPPs have lower POD in the eastern part of the catchment, which is prone to frequent cyclones (Setti et al. 2020a, 2020b; Yeditha et al. 2020), particularly for heavy rainfall (>50 mm/day). The FAR analysis revealed that all the SPPs have a significant level of false detection. It would be interesting to explore a detailed evaluation of these products focusing on extreme events associated with a cyclone event at the basin.

Comparing the SPPs at the monthly scale, the correlation between reference data and TRMM-3B42V7, GPM-IMERGV6, and PERSIANN-CDR products was found to be 0.94, 0.93, and 0.78, respectively, indicating the TRMM-3B42V7 and GPM-IMERG V6 data products were better correlated with IMD gridded data compared to the PERSIANN-CDR data product. This shows that the TRMM-3B42V7 and GPM-IMERGV6-based products can capture the overall seasonal pattern in the study region and, therefore, can be used for modeling studies. Comparing satellite and gauged-based precipitation datasets at an annual scale, both TRMM-3B42V7 and GPM-IMERGV6 demonstrate better performance than PERSIANN-CDR in terms of all criteria over the study region. In addition, TRMM-3B42V7 and GPM-IMERGV6 slightly overestimate precipitation, while PERSIANN-CDR shows a substantial overestimation. An interesting finding from this analysis was that the SPP's performance (in terms of statistical analysis and hydrologic model results) was better at a monthly scale than the daily and yearly scales. This could be because the seasonal mean cycle highly dominates the overall variance, thereby giving better results at a monthly scale. An interesting exercise would be to evaluate the performance of these SPPs after removing the seasonal means. Further, Hosseini-Moghari & Tang (2020) suggest that these SPPs are generally adjusted at a monthly scale, showing better results on a monthly scale.

All the products performed relatively better on a seasonal scale during the wet season (summer and fall) than the dry seasons at pixel and basin scales. Similar results were also reported at the Tibetan Plateau (Alazzy et al. 2017). However, all the SPPs overestimated the amount of rainfall during the summer season. This is probably because raindrops evaporate in the semi-arid climate before reaching the surface (Tesfagiorgis et al. 2011). During the winter season, TRMM is found to be relatively better than the other products. The PERSIANN-CDR performed consistently worse in all the seasons. Our results were consistent with the previous study (Moazami et al. 2016; Chen et al. 2020a; Gadouali & Messouli 2020). Chen et al. (2020a) also reported that the PERSIANN-CDR failed to capture seasonal precipitation trends and performed worse among the other products (CHIRPS and TRMM) over an arid mountainous area in China.

Comparing the performance of the TRMM and GPM products, it was observed that the latter showed a better ability to capture different rainfall intensities more accurately than the former. Further, the error statistics were comparatively lower for GPM than the TRMM- 3B42V7. The improved performance of the IMERG V6 product compared to its predecessor TRMM-3B42V7 can be attributed to the changes, such as in the retrieval algorithm used in Version 6. Huffman et al. (2015) report that the retrieval algorithm of IMERG V06 uses a new morphing approach and involves some refinements in the CMORPH-Kalman Filter. Further, Anjum et al. 2019; Hosseini-Moghari & Tang, (2020) also showed that GPM- IMERG V6 performs better than its predecessor in the Iran region.

The results of streamflow simulation using SWAT under the two scenarios in this study suggest that: In scenario A, TRMM- 3B43V7 and GPM-IMERG V6-derived simulations with IMD benchmarked model parameter values yield comparable performance to benchmark model in the calibration and validation period. By contrast, the PERSIANN-CDR product yields the worst model performance and has little potential for use in hydrological simulation over the study region.

In Scenario B, all SPP-based model simulations show significantly improved performance after recalibration both in the calibration and validation period; however, PERSIANN-CDR still performed worse than the other two products (refer to Table S1 and Figure S1).

On comparing both scenarios, we observed that the model performance improved in Scenario B compared to Scenario A at a daily scale, implying that individual model calibration improved the streamflow simulation. However, stream flow simulations have no significant improvement at the monthly time scale. This could be attributed to the fact that there is good agreement between IMD precipitation data and TRMM and GPM data sets monthly than at the daily scale. These results imply that the application of SPPs for hydrological models needs daily calibration with individual precipitation datasets. In contrast, at the monthly time scale, the IMD driven model is sufficient for streamflow simulation using SPPs.

In this study, the IMD gridded dataset is the ground truth data. Since the IMD dataset is interpolated data sampled at 0.25°, this would undoubtedly add uncertainty to the error estimates obtained from the study and, thereby, the quality of the satellite data products.

Satellite precipitation products (SPPs) are an effective alternative for rainfall data and streamflow simulations at global and regional scales, thereby helping water management practices and decisions. Despite the superiority of the SPPs over the gauge data (in terms of temporal resolution, spatial coverage, and availability of data), their accuracy has to be evaluated at regional scales. This study evaluated three recent precipitation products (IMERGV6, TRMM-3B42V7, PERSIANN-CDR) using a gauge-based precipitation product for 2000–2018 using statistical metrics and a hydrological model over a cyclone-prone basin in southeast India. The salient findings from the study can be summarized as follows.

• 1.

  Of the different precipitation products in the present study, IMERGV6 shows reliability compared to the other products in capturing extreme events and is found to be suitable for driving hydrological models. Further, the accuracy of the SPPs is moderate at daily time scales and good at monthly time scales, indicating that these products can be used reliably for water balance studies at coarser time scales; however, it needs appropriate correction before being applied for short-term flood modeling.

• 2.

  The grid level evaluation of the SPPs showed that the IMERGV6 precipitation product has higher CC, lower RMSE, and higher POD and CSI than TRMM-3B42V7 and PERSIANN-CDR.

• 3.

  The SWAT model calibration and validation using individual precipitation products showed that the satellite data IMERGV6 and TRMM-3B42V7 could drive the hydrological model. The results were comparable to those of the ground-based model. However, the results using the PERSIANN-CDR were not convincing as the satellite data tend to overestimate precipitation in the study region.

Overall, among the three SPPs, GPM-IMERGV6 shows comparatively better performance than the other products. Hence, it could potential alternative for a data-scarce cyclone-dominated river basin in southeast India. Future studies could investigate the spatio-temporal characteristics of specific cyclones over this region and the ability of the SPPs to capture the same.

The Early Career Research Award has supported this research under SERB, India, under grant No. ECRA/2016/01721. The authors are also grateful for the partial support from the Inspire Faculty grant held by Dr Maheswaran from DST. The authors would like to thank the anonymous reviewers.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.