Importance of spatial and depth-dependent drivers in groundwater level modeling through machine learning
Abstract. The water and food security of South Asia is embedded in the groundwater resources of the transboundary aquifer system of Indus-Ganges-Brahmaputra-Meghna (IGBM) rivers, which has been subjected to diverse natural and anthropogenic triggers. Thus, understanding the relative importance of such triggers in groundwater level change and developing a prediction framework is essential to sustain future stress. Although a number of studies on groundwater level prediction and simulation exist in the literature, characterization of predictive performances of groundwater level modeling using a large network of ground-based observations (n = 2303) is not yet reported. To identify the spatial and depth-wise predictors influence, here, we used linear regression based dominance analysis and machine learning methods (Support Vector Machine and Artificial Neural network) on long term (1985–2015) GWLs and/or climatic variables in the parts of IGBM basin aquifers. The results from the dominance analysis show that groundwater level change is primarily influenced by abstraction and population in most of the IGBM, whereas in the Brahmaputra basin, precipitation exhibits greater influence. Our results show a large proportion of the observation wells (n > 50 % for ANN and n > 65 % for SVM) demonstrate good correlation (r > 0.6, p 0.65), and normalized root mean square error (RMSEn