Evaluation of satellite rainfall products to estimate extreme flow events over the Kulfo watershed in Ethiopia

The variability and changes in climate brought about by global warming have raised the uncertainty around extreme hydrological occurrences on a global scale (Abbas et al. 2022a, b; 2023; Ullah et al. 2023). The escalating global environmental crisis is marked by climate change and resource depletion through extreme climate events (Jiang et al. 2024). Most of the developing countries are agriculture-based economies and have direct exposure to climate change and extreme events (Waseem et al. 2022). The accelerated process of agricultural commercialization enhanced the amount of fossil fuel burnings both in grain and cash crops, which increased the agricultural carbon share to 14% in net global emissions (Abbas et al. 2022a, b). This accelerated agricultural process leads to changes in climate and also alters the volumes and patterns of rainfall, runoff, and runoff coefficient (Mehta et al. 2022; Mehta et al. 2023).

Rainfall is one of the main climate parameters associated with extreme climate events, and heavy or reduced precipitation yields extreme precipitation or drought, respectively (Anvari et al. 2022; Masood et al. 2023; Najafzadeh & Anvari 2023). Extreme precipitation and drought indexes are commonly used to monitor and quantify the intensities and trends of these extreme climate events (ECEs) (Abbas et al. 2021; Vélez-Nicolás et al. 2022). Precipitation is the main input parameter required to obtain these monitoring indexes; therefore, precipitation data with high spatial and temporal resolutions are prerequisites for ECE analysis. Also, precipitation is an important hydrological parameter used for watershed management, flood forecasting, and climatological assessment (Masood et al. 2023). Thus, monitoring and projection of precipitation changes are of great importance to both disaster prevention and ECE mitigation (Buttafuoco et al. 2014; Kumar et al. 2023a, b).

Runoff is an important hydrological process and can cause negative effects such as soil erosion and excessive flooding over the river basin area. Changes in land use and land cover are dynamic processes and can strongly influence runoff potential in the long run (Mehta et al. 2023; Verma et al. 2023). Floods have had an impact on the natural environment even before humans arrived (Gohil et al. 2024). They are now considered natural hazards and a major global problem that harms human lives (Skilodimou et al. 2021). Floods have historically had significant implications on individuals and communities, including loss of life, damage to homes and infrastructure, crop devastation, and disruption to social affairs, as well as livestock losses (Kumar & Mehta 2021; Mangukiya et al. 2022; Kumar et al. 2023a, b).

Rain gauges, radars, and satellites are common tools for precipitation measurement (Michaelides et al. 2009). In situ gauge observations provide direct measurements of surface precipitation, but their areal coverage is small and usually insufficient for the accurate characterization of the spatial variability of precipitation, which has high spatial heterogeneity (Jongjin et al. 2016). Satellite remote sensing, which provides nearly global coverage, is a satisfactory means of compensating for the above limitations. Many satellite-based rainfall products have been generated to meet various hydro-meteorological needs. The current satellite precipitation products (SPPs) include Climate Prediction Center Morphing (CMORPH), the Tropical Rainfall Measuring Mission (TRMM), Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (GPM IMERG), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) series, and Climate Hazards Group Infrared Precipitation with Station Data (CHRIPS) (Ashouri et al. 2016).

Because such products have global (or quasi-global) orientation, the performances of SPPs are expected to vary from place to place. Satellite rainfall estimates offer several advantages compared to the conventional methods but can also be prone to multiple errors. The rainfall detection capability of SPPs can be affected by local climate and topography (Xu et al. 2013). Therefore, the performance of SPPs should be examined for a particular area before using the products for any application (Hu et al. 2014; Tote et al. 2015; Kimani et al. 2017). In developing countries, the availability of ground measuring stations is extremely limited with a scarce density of the hydro-meteorological network and uneven distribution, making it challenging for water resource development. Rainfall measurement is typically accomplished using rain gauge stations. However, there are small numbers of stations available, especially in the mountainous regions of Ethiopia. In mountainous regions, rainfall is extremely variable and changes in rainfall distribution can occur over short distances and within a short period of time (Gebere et al. 2015). In addition to the sparse network distribution issue, gathering available information from the existent surface observation network and performing ground surveys also are common problems. Among Ethiopian rivers facing this problem, the Kulfo River encompasses a series of challenges. Existing basic data for the watershed is very limited, and also, most of the available data are missing.

This study focused on the evaluation of satellite rainfall products (SRPs) to estimate extreme flow events over the Kulfo watershed in Ethiopia. The specific aims of the present study are to assess the suitability of SRPs for the Kulfo watershed, to evaluate the performance of SRPs at different time-scales against the observed data, select the appropriate product types at the Kulfo watershed, and to examine SWAT models’ performance using observed and SRPs in capturing the variability of extreme flow events over the Kulfo watershed. Most of the studies carried out in different parts of the world, as well as in Ethiopia, to evaluate the performance of SRPs showed that SRPs have potential use in prediction and modeling, but their performance depends on the type of SRP, watershed area, hydro-climatic regions, length of study periods, spatial and temporal resolutions, and topography of the area. However, these studies did not cover the Kulfo watershed and most of the studies are limited to short periods (usually 3 to 9 years). Therefore, this study was intended to evaluate the performance of widely used, easily available SRPs (namely, TRMM, CMORPH, and CHRIPS) in the Kulfo watershed to capture the gauged data scarcity problem.

According to the Ethiopian Mapping Authority, the Kulfo River is formed from the junction of the Gulando and Titika Rivers. The tributaries Baba, Gulando, and Yeremo drain the upper part of the basin, whereas the tributaries Wombale and Majale drain the middle part of the watershed. The tributaries Korzha, Ambule, Titika, and Kulfo make up the lower part of the catchment area.

Historical records of daily datasets were collected from the available meteorological stations (Table 1). In this study, the daily precipitation data during 1988–2015 was used as a benchmark for the evaluation of SRPs.

In order to have reliability in this research work, having relevant information or data was mandatory. These data are meteorological data (rainfall, temperature, relative humidity, wind speed, and sunshine hour) from the Ethiopian Meteorological Institute (EMI) of Ethiopia, hydrological data (stream flow) from the Ministry of Water Irrigation and Energy (MoWIE) of Ethiopia, SPP data from TRMM, CMORPH, and CHRIPS and other spatial data, which means a digital elevation model (DEM) of 30 × 30 m SRTM was obtained from the USGS (earth explorer.usgs.gov) website in raster form. Geographical coordinates, catchment area, and other related spatial data were processed and delineated from the 30 × 30 m DEM using arc GIS 10.3 version. The land use/land cover data were collected from the Ministry of Agriculture (MoA) of Ethiopia and GIS department.

A time series of hydrological data may exhibit jumps and trends owing to what (Yevjevich & Jeng 1969) called as inconsistency and non-homogeneity. Inconsistency is a change in the amount of systematic error associated with the recording of data. It can arise from the use of different instruments and methods of observation. Homogeneity is a change in the statistical properties of the time series. Its causes can be either natural or manmade. These include alterations to land use, relocation of the observation station, and implementation of flow diversions.

In order to extract the satellite rainfall data of the study area and to export the data in to Excel for each pixel, MATLAB R2013a and Panoply for NetCDF, HDF, and GRIB data viewer version 4.5.1 were used. MATLAB R2013a was used to extract and export the rainfall data of the study area and the period for satellite rainfall estimates in to Excel. Finally, with the use of these software and computer programs, daily satellite-based rainfall estimates were extracted in a suitable format for further analysis.

MATLAB is an abbreviation for ‘matrix laboratory’ and R2013a was the model number. MATLAB is a proprietary multi-paradigm programming language and numeric computing environment developed by MathWorks. MATLAB can be used as a tool for simulating various electrical networks, but the recent developments in MATLAB make it a very competitive tool for artificial intelligence, robotics, image processing, wireless communication, machine learning and data analytics. The MATLAB configuration for a processor desktop with a 64 bit processor CPU will be minimum 2-core with AVX2 instruction set support, the RAM of the computer will be minimum 8 GB, and the free disk space will be minimum 6 GB. The MATLAB setting is to open the preferences window to view and change; reproducible computation is the one that gives the same results every time it runs. The limitations of MATLAB is that compiled applications can run only on operating systems that run MATLAB, can be used only under control sources, and accessibly for Enter commands is only at the command line.

Panoply is a NASA-developed data viewer for netCDF, HDF, and GRIB files. It allows users to quickly and easily open a satellite data file, examine the contents, and make a very basic plot of the data. It is helpful for troubleshooting if you encounter errors when using Python to work. Panoply is a cloud-based, end-to-end managed data platform that allows you to start getting insights from your data in a matter of minutes.

The spatial patterns of the SPPs were evaluated and compared with rain gauge data at daily, monthly, and seasonal scales. Both the satellite and gauge rainfall data were collected at different temporal scales, and first, the daily data were aggregated to monthly and seasonal scales. As more than 85% of the total annual rainfall occurs during the wet season (June–September), seasonal comparison was considered only for this period (Gebremichael & Bitew 2011). The ability to replicate the observed rainfall by the products was done during this common period between all satellite and station rainfall. Considering the given climatic variability, complex topographical characteristics, and hydrological working units of the watershed, the performance of these products was evaluated using two approaches, namely point-to-pixel and aerial-averaged rainfall comparison.

Rainfall over a complex topography like the Kulfo watershed is largely subjected to small-scale variability, which implies that the evaluation of such satellite products should be at the smallest possible spatial and temporal scales (Thiemig et al. 2013). Accordingly, in the first approach, all SRPs from the corresponding grid cell were compared to the ground observed data within the satellite box. The variance of SPPs was smoother in space and time as these products were represented by spatial averages over the pixels. For this analysis, the SPPs were extracted for the location of each rainfall station and their performance was evaluated using statistical indices. It was assumed that the amount of point rainfall was uniform in the area of the pixel, which may not be necessarily true.

The second approach was based on the aerial rainfall comparison at different spatial scales. Representative sub-watersheds from lowland and highland areas with an average elevation of 1,400 and 3,000 m.a.s.l. were considered in order to account for the effect of topography. Satellite products will be validated at sub-watershed and watershed levels by comparing the spatially aggregated pixel values against the corresponding interpolated observed rainfall from gauge stations using the inverse distance weighting (IDW) method (Ruelland et al. 2008).

The performance of different products was evaluated using quantitative, categorical, and graphical measures. Several statistical indicators were used to quantify the consistency between the SPPs and observation data on different temporal scales: the Pearson linear correlation coefficient (R), root-mean-square error (RMSE), bias (Bias), relative RMSE (RRMSE), centralized RMSE (CRMSE), relative bias (RB), false alarm ratio (FAR), probability of detection (POD), frequency bias index (FBI), normalized missed rainfall volume (NRMV), normalized false alarm satellite rainfall volume (NFASRV), and equitable threat score (ETS) (Wilks 2011). The rain occurrence of each SPP was evaluated by comparing the probability density function (pdf) occurrence of daily precipitation of SPP to the rain gauges for the Kulfo watershed.

R measures the linear agreement between rain gauge observations and satellite precipitation. RMSE measures the absolute average error magnitude and assigns a greater weight to larger errors. RRMSE normalizes RMSE when daily precipitation estimates are the mean daily precipitation observed on the ground. When RRMSE is more than 50%, such precipitation estimates are considered unreliable. This particular threshold was used in previous studies. RB denotes the degree of overall overestimation or underestimation. For more detailed information about these statistical indices, the reader can refer to Moriasi et al. 2007; Anjum et al. 2016; and Chen & Li 2016.

The categorical technique is an assessment technique of satellite estimation/model forecast using a contingency table that reflects the frequency of ‘Yes’ and ‘No’ of the satellite estimation/forecast model (see Table 2).

A dichotomous estimate says, ‘Yes, an event will happen’, or ‘No, the event will not happen’. By using this table for daily precipitation, a set of statistical indices are shown as follows: POD responds to the question of what fraction of the observed ‘Yes’ events was correctly estimated/forecasted. The perfect score is 1.

Calibration is an effort to better parameterize a model to a given set of local conditions, thereby reducing the prediction uncertainty. Model calibration is performed by carefully selecting values for model input parameters (within their respective uncertainty ranges) by comparing model predictions (output) for a given set of assumed conditions with observed data for the same conditions (Arnold et al. 2012). The complex processes occurring in watersheds coupled with the uncertainty inherent in hydrologic modeling parameters, inputs, and measured data require that hydrologic models be calibrated and validated to minimize the predictive errors (Abbaspour et al. 2015). To calibrate and validate the hydrologic setup of the SWAT model, the parameters are automatically calibrated by using SUFI2 for the first 10 years until the model simulation result is within the acceptable range of the model performance measures.

A split sample procedure using the monthly stream flow data from the Kulfo watershed gauging station for the periods 1991–2008 and 2009–2015 is used for calibration and validation, respectively. The first 3 years were used as the warm up period to mitigate unknown initial conditions. As a result, the first 3 years were excluded from this analysis. For the same period of time, independent calibrations and validations were undertaken for satellite rainfall estimates.

The performance of a model must be judged by the extent to which it satisfies its objective of simulating the real world phenomena (accuracy), or the extent to which the achieved level of accuracy persists through different samples of data (consistency), and to the extent to which it can sustain the achieved level of accuracy when subjected to diverse application tests other than those used for calibrating the model (versatility). The performance of the SWAT model was evaluated using statistical measures to determine the quality and reliability of predictions when compared to the observed values. Regression coefficient (R²) and Nash and Sutcliff simulation efficiency (E_NS) were used to evaluate the model performance for both simulation cases (satellite rainfall-based and in situ rainfall-based) (Nash & Sutcliffe 1970).

Uncertainty analysis was performed to quantify the uncertainty associated with model simulations. During the initialization of model parameters, SUFI2 assumes a large parameter uncertainty and then decreases this uncertainty through the p-factor and the r-factor performance statistics. The range of the p-factor varies from 0 to 1, with values close to 1 indicating a very high model performance and efficiency (Yong et al. 2014; Abbaspour et al. 2015).

The SRPs were evaluated using statistical measures such as Bias, RMSE, CV, and the annual cycle of rainfall. The areal mean annual, observed rainfall amount of the catchment is 1,207.23 mm. When comparing the observed mean annual rainfall to the SRPs, there subsist some differences confirming underestimations for all models having 840 mm for the climate model (see table below). All models underestimate the observed rainfall. Accuracy of the models is not the same in representing the rainfall of the study area, with TRMM performing the best (Bias = −5.78) and CMORPH performing the worst (Bias = −9.87). Overall, SRPs tend to overestimate the inter-annual rainfall variability. CMORPH performs the best (RMSE = 19.54 mm per month), while CHRIPS performs the worst (RMSE = 13.43 mm per month). Table 3 shows the statistical test parameters, indicating the correspondence of the uncorrected climate model.

The study result is reliable with other results conducted in the catchment. Thiemann & Förch (2005) showed the bimodal rainfall pattern for the observed rainfall (Figure 3), and station wise, it showed some deviation with this result, but that is because of the infilling method and the number of years they used was different from this author and they showed it station wise.

Twenty-seven SWAT parameters were considered for model parameterization sensitivity analysis and the most sensitive parameters were identified using the global sensitivity analysis method in SWAT-CUP SUFI2, by following (Abbaspour et al. 2015) the SWAT-CUP user manual. Only eight (top eight) of them were effective for monthly flow simulation analysis.

Eight most sensitive parameters were selected based on t-stat and p-values as shown in Table 4. A t-stat provides a measure of sensitivity (larger in the absolute value is more sensitive), whereas p-values give the significance of sensitivity, a value close to zero is more significant.

Based on t-stat and p-value, from the model output, the first three most sensitive parameters are the SCS runoff curve number (CN2), baseflow alpha factor (ALPHA_BF), and effective hydraulic conductivity of the main channel (CH_K2). Ground water delay (GW_DELAY), soil evaporation compensation factor (ESCO), soil available water capacity (SOL_AWC), the threshold depth of water in shallow aquifer required for return flow (GWQMN), and the threshold depth of water in the shallow aquifer for ‘revap’ to occur (REVAPMN) are the 4th, 5th, 6th, 7th, and 8th sensitive parameters, respectively.

The historical hydrological data for SWAT model calibration and validation include the river stream flow data from hydrological stations. The observed stream flow data of the Kulfo watershed from 1991 to 2015 were used for the calibration and validation of the SWAT model. The SWAT model was calibrated for the period 1991–2008 and validated for 2009–2015 based on the principle that 70% of stream flow for calibration and 30% for validation use the monthly stream flow observation data from gauging stations within the study area.

After sensitive parameters’ identification, calibration of the model was executed to evaluate the performance of model simulation using automatic calibration tool (SWAT-CUP SUFI2) algorithms. The calibration processes considered the eight most sensitive parameters, as shown in Table 3, and their values were varied iteratively within the allowable ranges until a satisfactory agreement between the measured and simulated stream flows was obtained.

The validation result showed that the coefficient of determination (R²) was 0.723 as shown in Figure 7. The validation result showed that there is a very good agreement between the simulated and observed monthly flows.

The impact of climate change extreme flow was analyzed using high- and low-flow analyses. The Q₁₀ value is a robust indicator for high flows and designates a value of river discharge, which only exceeds 10% of the time. A negative trend in Q₁₀ means a reduction in flood risk, and a positive trend represents an increase. A Q₉₀ value is used for identifying low flows, indicating that 90% of the time, the value exceeds. If the value shows a negative trend, it implicates that the low flow is further decreasing and river droughts are likely to occur more often.

In the catchment, the simulated stream flow based on the SPPs of CHRIPS and CMORPH show an increase by 17.923 and 17.829 m³/s, respectively, from observed precipitation (17.4 m³/s), whereas the TRMM SPPs show a decrease from observed precipitation for the high flow (Q₁₀). The same is true for the low flow, CHRIPS and CMORPH show an increase by 9.045 and 8.203 m³/s, respectively, from observed precipitation (1.75 m³/s), whereas the TRMM SPPs show a decrease from observed precipitation for the high flow (Q₉₀).

The goodness-of-fit (GOF) tests measure the compatibility of a random sample with a theoretical probability distribution function. In other words, these tests show how well the selected distribution fits to the data. Based on the GOF tests, general extreme value distribution was confirmed. The parameters of distribution were estimated and are shown in Table 5.

According to the result, the recurrent flood magnitude increase as recurrent year in studied period. AM included only one flood from each water year and excluded significantly large floods if several occur in a single water year; this would have the advantages of taking into account other major floods in flood-rich years and preventing the analysis of small or non-flood events in other years.

It is important to take into account about how much magnitude of flow occurs in the given recurrence period while designing any hydraulic structure across the river, particularly near Arba Minch town.

The domain of the study is the Kulfo watershed; for this study, the scope has been limited with respect to the stated objectives. Thus, we evaluate and assess the SRPs, considering gauged rainfall as the reference. Simulating each of the SRPs irrespective of the ground data, comparing each, and finally, selecting the reliable rainfall estimation over the basin were the main targets of the study. The SWAT-2012 model is applied to simulate the stream flow in the basin.

Satellite-based rainfall products with high spatial and temporal resolution and wide coverage provide a potential alternative source of forcing data for hydrological models in regions where conventional in situ precipitation measurements are not readily available. The Kulfo watershed is a good example of a case where the use of satellite-based rainfall can be useful. To overcome the limitations of robust station data, this study uses some of the globally available online high-resolution precipitation data to simulate runoff using the SWAT model. Analyses were limited to the following details: semi-distributed SWAT hydrological model, Kulfo basin (367 km²), and three types of SRPs (TRMM, CHRIPS, and CMORPH). Hydrologic simulation models are a very important way to assess the hydrologic characteristics of a watershed. They are powerful tools for assessing the effects and impacts on a hydrological environment. They can be used to know, predict, and interpret what happened and is happening in the entire basin in time and space. They can be used to learn about the effects of various environmental and other pool-related factors, as well as their effects and adverse effects when threatened. The results show that the usefulness of SPPs as input to the SWAT for monthly stream flow simulation is highly dependent on the product type. The simulation of each precipitation captured the observed hydrographic trend. The simulation based on the new versions of CHRIPS and CMORPH showed consistent and modest skills in their simulations but slightly overestimated the large flood peaks in some years (1997, 1999, 2001, 2002, and 2007). On the other hand, the TRMM simulation showed a reasonable ability in reproducing the hydrograph at the watershed outlet. The Kulfo watershed faced the problem of a good metrological station; this leads to extreme floods that can cause soil erosion, and the destruction of agricultural crops and hydraulic structures. Due to that reason, satellite-based rainfall products’ study was incorporated to estimate an extreme flow event.

The calibration results show that the coefficient of determination (R²) and the NSE are 0.792 and 0.692, respectively. The coefficient of determination (R²) during calibration was found to be 0.792. The calibration result showed that there is a very good agreement between the simulated and observed monthly flows. These performance ratings are very good and good, respectively. The validation period has also shown a good agreement between the monthly measured and simulated flows. The validation result showed that the coefficient of determination (R²) and the NSE are 0.723 and 0.643, respectively. The statistical comparisons of simulations of the Arc SWAT model from various rainfall inputs and the simulations of stream flow from TRMM, CHRIPS, and CMORPH inputs showed comparatively good performance. Simulations based on CHRIPS and CMORPH have nearly similar NSE and R² values (0.69 and 0.853 for CHRIPS and 0.652 and 0.8072 for CMORPH, respectively). There were fair values of NSE (0.61) and R² (0.7752) in the TRMM-based simulation of stream flow. The result showed that CHRIPS overestimates the large flood peaks in some years (1999, 2001, and 2002), while CMORPH overestimates the large flood peaks in two years (1997 and 007). The current results can be useful for agriculture and water management and to know how it contributes to World Climate Research Program's (WCRP's) grand challenges on climate extremes, in order to estimate extreme climate change, extreme flood event to save agriculture land, and the loss of life and structure.

In general, the results show that although there are some uncertainties in these gridded data series (CHRIPS and CMORPH), the use of these satellite precipitations is useful for hydrological studies in the absence of station data. This strongly demonstrates the increasing potential of satellite precipitation forecasting accuracy in reproducing hydrological features. The SWAT model also proves to be a good tool for this type of modeling.

The domain of the study is the Kulfo watershed; for this study, the scope has been limited with respect to the stated objectives. In addition, weather stations must be improved both qualitatively and quantitatively to promote hydrological model performance. Therefore, it is strongly recommended to establish good weather stations and obtain high quality flow data.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.