Evaluation of multiple bias correction methods with different satellite rainfall products in the Main Beles Watershed, Upper Blue Nile (Abbay) Basin, Ethiopia

In Ethiopia, the assessment of water resource issues is difficult in data-sparse areas because estimating water availability necessitates an understanding of rainfall variability both spatially and temporally (Belay et al. 2019). Good-quality precipitation data estimates are critical to the accuracy and dependability of any hydrologic study, whether it is for flood-forecasting, drought-monitoring, water resource management, or climate change effect assessment (Hamlet & Lettenmaier 1999). Rain gauges provide a direct physical measurement of surface precipitation, although they are subject to certain errors due to location, spatial scale (point), wind, and mechanical flaws, among other things (Derin & Yilmaz 2014). This implies that observed rainfall data are better than satellite rainfall estimation, but the lack of sufficient rain gauge observations has an impact on water resource assessment due to the spatial variability of rainfall.

Because precipitation is the most important atmospheric input to land surface hydrology models, accurate precipitation inputs are critical for accurate hydrologic prediction. Ground-based precipitation measurements are sparse or non-existent in many remote parts of the world, particularly in developing countries, due to the high cost of establishing and maintaining infrastructure. Inconsistencies in instrumentation and administrative limitations on data access for rivers that cross international borders further impede the effective use of hydrological models to support reliable flood and drought diagnosis and forecasting (Habib et al. 2014; Hall et al. 2014; Revilla-Romero et al. 2015). Both of these constraints apply to the Main Beles sub-basin of the Upper Blue Nile Basin.

Precipitation estimates based on satellites with high spatial and temporal resolution and wide coverage of the area offer a potential alternative source of impact data for hydrological models in areas where traditional in situ precipitation observations are not always available. The increasing availability of high-resolution (and near real-time) satellite-based rainfall estimates has significant potential in applications like hydrological analysis for engineering design, assimilation of precipitation data into forecast models, flood forecasting, and water resource management in general (Liu et al. 2012; Zambrano-Bigiarini et al. 2017; Moges et al. 2022). These applications could have far-reaching implications for various developing countries, such as Ethiopia, which lacks ground-based rain gauges and does not have radar capabilities to measure representative rainfall magnitude. However, there are errors in satellite-based rainfall estimations, which raise a number of scientific concerns. How accurate are satellite-based rainfall products? Is it possible to employ high-resolution satellite rainfall products (SRPs) for hydrological purposes? The hydrological community faces a challenge due to a lack of understanding about the accuracy of those satellite products, particularly in complex terrain and less gauged areas (Funk et al. 2015). The fact that rain gauge-based rainfall data are point data that represent an area defined by a limited radius surrounding the device's location is a prominent aspect (Collischonn et al. 2008). Furthermore, the density of measuring stations is uneven between areas, and their positions favor accessible lower-lying places (López López et al. 2018). As a result, a high-resolution spatial dataset that can efficiently capture variations in spatial precipitation is required.

For the goal of reanalyzing satellite-derived precipitation data, the Delta, Empirical Quantile Mapping (EQM), and Quantile Mapping (QM) bias correction methods were applied. Furthermore, the Main Beles River basin was selected since it is one of the tributary river basins in the Upper Blue Nile Basin (Yasir et al. 2014). The Main Beles River basin encompasses five metrological stations with a range of climates and physical attributes. Accurate precipitation data for the Main Beles River basin will aid in the prediction of extreme occurrences such as floods and droughts.

Therefore, Climate Prediction Center Morphing (CMORPH) (Joyce et al. 2004; Haile et al. 2013; Yang et al. 2021), Climate Forecast System Reanalysis (CFSR) (Bao & Zhang 2013; Dile & Srinivasan 2014), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) (Nguyen et al. 2019), and Tropical Rainfall Measuring Mission (TRMM) Multi-satellite Precipitation Analysis (TMPA_3B42V7) (Prakash et al. 2016; Vu et al. 2018) SRPs are employed in this study. These SRPs (PERSIANN, CMORPH, TMPA_3B42V7 and CFSR) were selected based on the following criteria: public domain dataset, long-term data availability, reasonable spatiotemporal resolution, near-real-time availability and their wide applicability in Africa (Dinku et al. 2007; Thiemig et al. 2012; Dembélé & Zwart 2016; Haile et al. 2013; Prakash et al. 2015).

Some of the previous studies which investigated the Main Beles River Basin involved hydrological responses of a catchment to climate change (Yimer et al. 2009; Ebrahim et al. 2013; WaleWorqlul et al. 2018), hydrological responses to land use/cover changes (Woldesenbet et al. 2017), hydropower (Annys et al. 2019), Soil erosion (Kebede et al. 2021), soil management (Molla et al. 2022), assessment of water resource and forecasting (Hamza & Getahun 2022) and spatial and temporal variability in hydrological responses (Lemann et al. 2018). It is still a common practice that rainfall is interpolated from point rain gauge measurements in water resources and other related studies. However, such a method is expensive, particularly for developing countries like Ethiopia, because many gauges are needed to capture typically large rainfall variability, so often, gauge numbers are compromised. As such, interpolation can create high bias. Satellite rainfall estimates can fill this gap by providing spatio-temporal data coverage, although processed satellite rainfall products have often substantial biases when compared with true, gauge measurements, mainly due to different spatial resolutions. Besides, the performance of satellite products is different in different areas due to geographical and climatic variations. To fill this gap in a given study area, it is therefore essential to evaluate the error of different satellite rainfall estimates against the rain gauge network of that area, based on that choose an optimal satellite rainfall product for that particular area, and finally remove the bias of the satellite rainfall product.

This research looks at the capability and limitations of SRPs (CMORPH, CFSR, PERSIANN, and TMPA_3B42V7), as well as the performance of three bias correction methods (Delta, EQM, and QM) for the Main Beles basin with five metrological stations like Pawe, Dangla, Shahura, Durbete, and Yismala. By comparing statistical properties and Taylor diagrams (Ayoub et al. 2020; Centella-Artola et al. 2020), the performance of SRPs and bias correction approaches have been evaluated.

The Main Beles catchment is located in the western part of the Upper Blue Nile Basin, Ethiopia. It is one of the tributary river basins of the Upper Blue Nile Basin, Ethiopia. This is situated on the plateau of the north-western highlands of Ethiopia near Lake Tana in a geographic location from a latitude of 10°50′N to 11°50′N and a longitude of 36°10′E to 37°00′E. The topography of the area is mostly flat with altitudes between 999 m and 2,724 m above sea level.

The CMORPHing approach (Climate Prediction Center (CPC) MORPHing technique) generates worldwide precipitation studies with extremely high spatial and temporal resolution. This technique exclusively uses precipitation estimations acquired from low-orbiter satellite microwave observations, whose features are transmitted via spatial propagation information derived purely from geostationary satellite IR data (Joyce et al. 2004). Algorithms for the Satellite Program series of special sensor microwave/imagers (SSM/I) (Ferraro 1997), Advanced Microwave Sounding Unit (AMSU-B) (Ferraro et al. 2000), and TRMM Microwave Image (TMI) are used to generate these estimates (Kummerow et al. 2001). From December 3, 2002 to the present, CMORPH provides daily precipitation products with a grid resolution of 0.07277° latitude/longitude (8 km at the equator), a temporal resolution of 30 min, and a geographical coverage of 60°S–60°N latitude globally (Joyce et al. 2004). https://climatedataguide.ucar.edu/ has CMORPH data.

The TRMM Multi-Satellite Precipitation Analysis (TMPA) product's latest post-real-time data, TMPA 3B42V7, are superior to all prior versions (Yong et al. 2014). Since January 1998, the 3B42V7 dataset has covered the global latitude belt from 50°S to 50°N, with a spatial resolution of 0.25° by 0.25° and a temporal resolution of 3 h (Huffman et al. 2007). The Goddard Earth Sciences Data and Information Services Center (https://mirador.gsfc.nasa.gov) provides daily precipitation data from 2000 to 2015. The daily 3B42V7 data are consolidated into monthly mean values and include a timeframe of 0:00 UTC–24:00 UTC with the gauge observation.

The Center for Hydrometeorology and Remote Sensing (CHRS) at the University of California, Irvine (UCI) developed the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) system, which uses neural network function classification/approximation procedures to quantify an estimate of rainfall rate at each 0.25° × 0.25° pixel of the infrared brightness temperature image provided by geostationary satellites (Hsu et al. 1997). When independent rainfall estimates are available, an adaptive training feature makes it easier to update the network parameters. The PERSIANN system was originally built on geostationary infrared imagery, but it was later expanded to include both infrared and visible imagery. The PERSIANN algorithm is used to generate global rainfall using the geostationary long-wave infrared imagery (Sorooshian et al. 2014). From 2000 to the present, PERSIANN provides daily precipitation products with a spatial resolution of 0.25° 0.25° and a spatial coverage of 60°S–60°N latitude, near real-time with a 2-day delay (Nguyen et al. 2018). https://chrsdata.eng.uci.edu/contains PERSIANN data.

CFSR is a third-generation reanalysis product. It is a global high-resolution system that includes the atmosphere, ocean, land, surface, and sea ice to provide the best assessment of the state of these coupled domains throughout this time period. The CFSR incorporates (1) atmosphere–ocean connection during the creation of the 6-h riddle field, (2) an interactive model of sea ice, and (3) satellite radiation assimilation. The global CFSR atmosphere has a resolution of 38 km (T382) with 64 levels. With 40 levels and data available from 1979-01-01 to 2014-07-31, the global ocean is at 0.25° near the equator, extending to 0.5° beyond the tropics. There are four soil levels in the global land surface model and three levels in the global sea ice model. The CFSR atmospheric model includes observed CO₂ fluctuations, as well as aerosol and other trace gas and solar variations. The scenario evaluated will include estimations of changes in the climate of the Earth's system as a result of these variables. The current CFSR will be extended in the future as a real-time operating product from the research data archive (RDA) summary on CFSR (Sharp et al. 2015; Liu et al. 2018; Wei et al. 2021). CFSR data are available at https://globalweather.tamu.edu/.

In this study, four SRPs (CMORPH, CFSR, TMPA3B42V7, and PERSIANN) were investigated to identify the best product to use for the spatial and temporal assessment of any hydrologic study, either related to flood-forecasting, drought-monitoring, water resource management, or climate change impact assessment at daily and monthly time scales. The evaluation was carried out using data for the period from 2003 to 2016 for those SRPs but did not include CFSR SRPs for intercomparison because the variation of availability of rainfall data records from 1979-01-01 to 2014-07-3. Three bias correction approaches were used for the intercomparison of each satellite's product like Delta, EQM, and QM.

There are numerous statistical bias correction methods for precipitation and temperature; the Statistical Downscaling of General Circulation Models (SDGCM) tool is available on the website https://agrimetsoft.com/sd-gcm for downscaling. In this tool, there are three statistical downscaling models: the Delta, the QM and the EQM. The tool is very useful that can carry out the downscaling project for a list of stations in one run and perform both monthly and daily data compared to other tools (Martinez-Villalobos & Neelin 2019). This tool is better to improve the fitting of climate model simulations to observations in the control period, in order to enhance reliability (Boé et al. 2007; Oruc 2022).

The original QM method is a non-parametric bias correction BC method generally applicable to all possible distributions of rainfall (Ringard et al. 2017; Luo et al. 2018).

Development and applications of Taylor diagrams have roots in climate change studies (Taylor 2001; Smiatek et al. 2016) but also have frequent applications in environmental model evaluation studies (Srivastava et al. 2015). Bhatti et al. (2016) propose the use of Taylor's diagrams for assessing the effectiveness of SRP bias correction schemes. The most effective bias correction schemes will have data that lie near a point marked ‘reference’ on the x-axis, a relatively high correlation coefficient, and a low root mean square difference. Bias correction schemes matching gauge-based standard deviation have patterns that have the right amplitude.

The performance evaluation criteria and the Delta, QM, and EQM methods for refining CMORPH, PERSIANN, TMPA_ 3B42V7, and CFSR SRPs used in study stations in the Main Beles basin for both daily and monthly rainfall analysis are shown in Tables 1,²³–4. The effectiveness of the bias reduction approaches utilized in this study is determined by the rainfall measuring station and the satellite observation. Based on the overall performance of six performance evaluations, the Delta approach improved the quality of CFRS rainfall in all stations for both daily and monthly time scales. In the case of CMORPH SRP, the Delta method was best among EQM and QM on both daily and monthly time scales for all weather stations (Pawe, Shahura, Dangla, Durbete, and Yismala) except Durbete and Dangla at daily base QM best. For example, for the Dangla station (CMORPH), Delta bias correction evaluation criteria of RMSE (average 9.24 mm day⁻¹) were relatively close to zero, whereas EQM and QM showed higher values (average 10.44 and 9.87 mm day⁻¹), respectively, deviating from zero at daily time step and monthly comparisons of this station. Delta bias correction evaluation criteria of RMSE (average 2.47 mm month⁻¹) were relatively close to zero, whereas EQM and QM showed higher values (25 and 3.12 mm month⁻¹; Table 1).

In the case of PERSIANN SRPs, Delta for Pawe, QM for Dangla & Shahura and EQM for Yismala stations are the best BC approaches on daily and monthly time scales (Table 2). For example, for the Pawe station (PERSIANN), Delta bias correction evaluation criteria of RMSE (average 10.13 mm day⁻¹) were relatively close to zero, whereas EQM and QM showed higher values (average 10.85 and 10.99 mm day⁻¹, respectively; Table 2). The TMPA 3B42V7 SRP is best for Delta at the Pawe station both daily and monthly, EQM for Dangla, Yismala, and Shahura stations on a daily time scale but Delta on a monthly time scale, and EQM for the Durbete station on both daily and monthly time scales. For example, EQM bias correction evaluation criteria of the correlation coefficient (r) value of 0.51 were relatively close to one, whereas Delta and QM showed higher values (average 0.49 and 0.51, respectively) deviating from one at daily and monthly time step comparisons of this station. The correlation coefficient (r) is demonstrated to have a comparable value in Durbete station, which is 0.95 substantially close to one at the monthly time scale based on Delta, EQM, and QM bias correction. For all weather stations, the CFSR SRP Delta method performed best among EQM and QM on both daily and monthly time scales (Pawe, Shahura, Dangla, Durbete, and Yismala). For example, the Delta bias correction evaluation criterion of correlation coefficient (r) value for the Shahura station (CFSR) was 0.51, relatively close to one, whereas EQM and QM showed higher values (0.47 and 0.46, respectively), deviating from one at daily and monthly time step comparisons of this station (Table 4).

Table 5 shows that the best bias correction evaluation of both daily and monthly time step comparisons using Delta, EQM, and QM statistical methods corresponds with SRPs (CMORPH, TMPA 3B42V7, PERSIANN, and CFSR) and five rainfall gauge stations. Delta bias correction methods outperformed other bias correction methods (EQM and QM) at the Pawe station for all daily and monthly SRPs (CMORPH, TMPA 3B42V7, PERSIANN, and CFSR). For the Dangla station, EQM was chosen for TMPA 3B42V7 and PERSIANN bias correction on a daily and monthly basis, but Delta was chosen for CFSR bias correction on a daily and monthly basis, and QM was chosen for CMORPH SRPs on a daily and monthly basis. Delta for CMORPH and CFSR bias correction methods and QM for PERSIANN bias correction methods were chosen from the Shahura station for both daily and monthly time scales, but EQM at daily and Delta at monthly time scales were chosen for TMPA 3B42V7 satellite rainfall bias correction. For the Durbete station, EQM for TMPA 3B42V7 and PERSIANN bias correction methods and Delta for CFSR bias correction methods were chosen for both daily and monthly time scales but QM at daily and Delta at monthly for CMORPH satellite rainfall bias correction. For the Yismala station, Delta for CMORPH and CFSR bias correction methods and EQM for PERSIANN bias correction methods were chosen on both daily and monthly time scales, but EQM on a daily time scale and Delta on a monthly time scale were chosen for TMPA 3B42V7 satellite rainfall bias correction.

In this study, the SRPs at two time scales (i.e., daily and monthly) were analyzed to identify satellite-derived rainfall data that could potentially be used for meteorological, hydrological, environmental, and drought-monitoring assessments in the Main Beles basin. Overall, the performance of the four rainfall-based satellite products (i.e., CFSR, CMORPH, TMPA_3B42V7, and PERSIANN) over the Main Beles basin is promising for further application in drought years. The result of comparative evaluation of these four satellite products and bias correction methods showed that CMORPH_ Delta, CMORPH_QM, CMORPH_Delta, CMORPH_EQM, and CMORPH_EQM are the best satellite-derived rainfall data at daily scale, and CMORPH_Delta, TMPA_3B42V7_EQM, TMPA_3B42V7_Delta, TMPA_3B42V7_EQM, and TMPA_3B42V7_Delta are the best satellite-derived rainfall data at monthly time scales for Pawe, Dangla, Shahura, Durbete, and Yismala weather stations, respectively. Thus, this SRPs and bias correction methods were selected in this study for further application in the spatial and temporal assessment of meteorological and hydrological study in the Main Beles basin.

The availability of satellite-derived rainfall products at local and global scales has proved to be beneficial in filling the data gap, particularly in developing countries that have data scarcity. However, evaluating these rainfall products is essential for any application that includes studying water resource problems. In this regard, SRPs are a valuable source of information, particularly in sparsely gauged regions like the Main Beles watershed. In this study, we selected the three bias-corrected (Delta, EQM, and QM) methods and four SRPs (CMORPH, TMPA_3B42V7, PERSIANN, and CFSR) and evaluated the performances of three SRPs (CMORPH, TMPA_3B42V7, and PERSIANN) by comparing them with gauged rainfall data from five independent weather stations across the Main Beles basin for real-time hydrological applications. The statistical approach was used for the performance evaluation at daily and monthly time scales. The evaluation process was undertaken to identify the best SRP and bias correction methods for the spatial and temporal assessment of meteorological data in the basin. After analyzing the results, the following conclusions were drawn.

In general, the bias correction achieved significantly improved the ability of these products to hydrology forecast on a daily and monthly scale. This result is in agreement with previous findings that have suggested that rainfall correction methods have a more significant influence than temperature correction methods. Delta performed better than QM and EQM for the Pawe station for five SRPs at both daily and monthly time scales only when the estimates in the datasets were highly cross-correlated. For the Dangla station, EQM was selected for TMPA_3B42V7 and PERSIANN bias correction on both daily and monthly time scales, but Delta was selected for CFSR bias correction on both daily and monthly time scales and QM at daily and Delta at monthly bias correction method selected for CMORPH SRPs. For the Shahura station, Delta for CMORPH and CFSR and also QM for PERSIANN bias correction methods were selected for both daily and monthly time scales, but EQM at daily and Delta at monthly time scales selected for TMPA_3B42V7 satellite rainfall bias correction. For the Durbete station, EQM for TMPA_3B42V7 and PERSIANN and also Delta for CFSR bias correction methods were selected for both daily and monthly time scales but QM at daily and Delta at monthly for CMORPH satellite rainfall bias correction were selected. For the Yismala station, Delta for CMORPH and CFSR and also EQM for PERSIANN bias correction methods were selected for both daily and monthly time scales, but EQM at daily and Delta at monthly time scales were selected for TMPA_3B42V7 satellite rainfall bias correction. Both the bias correction methods (EQM, QM, and Delta) improved the quality of SRPs, and consequently, they have a direct influence over the accuracy of hydrological simulations.

The performance of the three SRPs (TMPA_3B42V7, PERSIANN, and CMORPH) was reasonably good in detecting the occurrence of rainfall and in estimating the amount of daily and monthly rainfall in the basin. Comparison of the three SRPs has shown that good agreement (r > 0.4) of CMORPH_ Delta, CMORPH_QM, CMORPH_Delta, CMORPH_EQM, and CMORPH_EQM at daily time scale and at the monthly time scale of agreement (r > 0.88) of CMORPH_Delta, TMPA_3B42V7_EQM, TMPA_3B42V7_Delta, TMPA_3B42V7_EQM, and TMPA_3B42V7_Delta rainfall products and bias correction methods with ground observations of Pawe, Dangla, Shahura, Durbete and Yismala weather stations, respectively, and the overall statistical weighting performances (i.e., Pearson's correlation coefficient (r), MAE, MBE, RMSE, index of agreement (d), and Nash–Sutcliffe Efficiency (NSE) model) are also good. The results depicted that SRPs and bias correction methods of CMORPH_QM (r = 0.538) and TMPA_3B42V7_EQM (r = 0.95) data showed good performance, while PERSIANN_EQM (r = 0.348) and PERSIANN_Delta (r = 0.83) perform worst at daily and monthly time scales, respectively, among the others. This shows that both SRPs and bias correction methods can be used to develop operational drought or flood monitoring, any hydrological study, and early warning system since daily and monthly time scales better identify periods of low or heavy rainfall events in the study area with each metrological station.

Improvements in hydrological forecasts obtained by bias-correcting raw rainfall estimates can help to enhance the operation of reservoirs, planning for irrigation, and construction of hydraulic works, among other things. This process is undoubtedly relevant for forecasting future scenarios in which the pressure exerted by users of water resources increases. Additionally, it is worth mentioning that since these bias correction methods are assumed to be stationary, the correction algorithm and its parameterization can be valid for current climate conditions. However, further research is necessary on this topic to clearly determine the ability of these bias correction methods to correct future raw estimates in the Main Beles basin under different climate scenarios.

The authors would like to extend their gratitude to the Amhara National Metrological Agency (NMA) of Ethiopia for the provision of valuable data for this research. The authors also thankful the reviewers for their thoughtful comments. A.B.N. would like to thank the facilities provided by the Wollo University/Kombolcha Institute of Technology (KIoT) and friends for their appreciative support for doing this research. A.B.N. is also thankful to IWA Publishing for Subscribe to Open (S2O) platform and for supporting developing countries.

A.B.N. and H.W.T. conceptualized the whole article. A.B.N., F.A., and H.W.T. developed the methodology. A.B.N. was involved in data curation and wrote the original draft. H.W.T., A.E., F.A., and G.W. wrote the original draft and the review, and edited the article. A.E. and G.W. supervised the work. All authors have read and agreed to the published version of the manuscript.

The authors did not receive support from any organization for the submitted work.

The data used in this study are available from the authors on reasonable request and some data obtained from: CMORPH (https://climatedataguide.ucar.edu/), TMPA_3B42V7 (https://mirador.gsfc.nasa.gov/), PERSIANN (https://chrsdata.eng.uci.edu/), and CFSR (https://globalweather.tamu.edu/).

The authors declare there is no conflict.