Evaluating the Contribution of Remote Sensing Data Products for Regional Simulations of Hydrological Processes in West Africa using a Multi-Model Ensemble

Abstract

Water is a crucial resource for human health, agricultural production and economic development. This holds especially true in West Africa, where large parts of the population work as self-sustaining farmers. Accurate knowledge of available water resources is therefore essential to properly manage this valuable commodity. Hydrologic modeling is seen as a key aspect in generating predictions of available resources. However, the overall availability of <em>in situ</em> data for model parametrization in West Africa has been steadily declining since the 1990s. When observations are available, they often contain errors and gaps. This lack of data severely hinders the application of hydrologic models in the region. Nowadays, many global and regional remote sensing and reanalysis data products exist which may be used to overcome these problems. A thorough analysis of the contribution of these products to regional simulations of hydrologic processes in West Africa has so far not been conducted. The purpose of this study is to close this gap. The study area spans from 3 to 24° latitude and -18 to 16° longitude and encompasses, among others, the Niger, Volta, and Senegal river basins. This study focuses on three key aspects, namely how the performance of remotely sensed and reanalyzed products can be validated without the availability of <em>in situ</em> data for the region; to what extent semi-distributed hydrologic models of the region can be parameterized and validated using these data; and how a fully distributed, grid-based model can be set up, calibrated and validated for sparsely-gauged river basins using multivariate data inputs.<br /> Comparisons of remote sensing and reanalysis precipitation products for the region show strong<br /> variability. A hydrologic evaluation was conducted, during which the skill of each precipitation dataset to accurately reproduce observed streamflow in HBV-light simulations was tested. Best results are achieved by products which include satellite infrared and microwave measurements as well as bias-correction based on <em>in situ</em> observations. Averaged Nash-Sutcliffe Efficiencies (NSE) of 0.66 were reached during the calibration of the CMORPH CRT and PERSIANN CDR products over six subbasins.<br /> In a next step, three SWAT models were set up for the region using multiple remote sensing and reanalysis data products and then calibrated and validated against observed river discharge with global and local approaches. While streamflow results differ within models and model regions, they are mostly satisfactory with coefficient of determination (R²) values of 0.52 and 0.51 for calibrations and 0.63 and 0.61 for validations. In a multivariate validation framework, the skill of the model in simulating variables not included in the calibration is further evaluated against remote sensing observations of actual evapotranspiration, soil moisture dynamics, and total water storage anomaly. Here, it has been shown that the models perform robustly and reach a good agreement in relation to observations.<br /> Furthermore, the grid-based mHM model was applied to several river basins in the south of the study area. After the quality of precipitation and evapotranspiration inputs was tested, a multivariate calibration was conducted. Models were calibrated using discharge observations<br /> (Q) and, to further constrain model boundary conditions, discharge in combination with remote sensing actual evapotranspiration observations (Q/ET). Finally, the quality of the simulations was tested against streamflow data as well as against remote sensing actual evapotranspiration, soil moisture, and total water storage anomaly data. Streamflow simulations performed well with averaged Kling-Gupta Efficiencies (KGE) of 0.53 for the first (Q) and 0.49 for the second (Q/ET) calibration. Further variables tested during the multiobjective validation were within good predictive ranges, especially during the Q/ET calibration. When SWAT and mHM model results are compared against each other and against external data products, results show that while both models perform robustly, mHM predictions outperform SWAT results.<br /> This study furthers the understanding of the contribution of remote sensing, reanalysis and global data products in regional simulations of hydrologic processes in West Africa. Specific modeling strategies and routines were developed to further increase predictive capabilities of hydrologic models of the region using these freely-available datasets.