Regionalisation of Aquifer Properties through Machine Learning in the Lake Chad Basin

Data on aquifer properties and characteristics in the Sahel zone are often scarce, hampering the required elaboration of transboundary, as well as integrated water management strategies. However, satellite data may be used to derive various parameters to be linked to, for instance, aquifer productivities and hence exploration potentials. Since knowledge of the continuous aquifer properties is essential for sustainable groundwater management and supply, the best possible regionalization of individual aquifer parameters is an important component that may be enhanced by the use of this data. To meet this demand, machine learning can play an important role with regard to suitable regionalisation methods, as (hydro-)geological processes are non-linear and their causes and effects are heterogeneous.For a cross-border study area of approximately 80,000 km2 in the Lake Chad basin, various machine learning algorithms, such as random forest and support vector machine, were used to carry out a regionalisation of the specific capacity (SPC) and subsequently to derive a map of the potential groundwater productivity. In total, about 1500 data points from several field campaigns served as ground truth data for the SPC. For training, multiple explanatory features, such as elevation, slope, NDVI, gravimetry, lithology etc. originating from a digital elevation model, multispectral satellite images or digitalized geoscientific maps, were used.After tuning all common hyperparameters and testing the model performances with a 5-fold spatial cross-validation, the model predictions were evaluated and compared as well as recommendations for application and transferability of machine learning algorithms under the given conditions were provided.