Spatial Modeling of Groundwater Potential in the North of Minas Gerais, Brazil: An Integrated Approach Using Machine Learning and Environmental Data

Abstract

In arid and semi-arid regions, like the North of Minas Gerais (NMG) in Brazil, groundwater serves as a crucial resource. Due to the anticipated surge in demand for these resources, devising effective strategies for managing and analyzing water resources is vital. This study aims to model the spatial distribution of potential groundwater areas in the NMG by evaluating six Machine Learning Algorithms based on water flow data from 4,028 tubular wells (Groundwater Information System - SIAGAS). The modeling was supported by environmental covariates connected with water dynamics (climate, geology, relief, soil, and vegetation). The covariate selection technique (RFE- Recursive Feature Elimination) selected the most important ones. The Random Forest (RF) model was the most efficient in the prediction (R2 0.16 and an RMSE of 17.50 m3/h). The model captured the influence of critical environmental covariates. The central and western regions of the NMG exhibited the highest groundwater potential, with flow values from tubular wells in these areas 620% higher than the eastern regions. This disparity can be attributed to the significant presence of psamitic, and carbonate sedimentary rocks characterized by high porosity and fissures, extensive plateaus (recharge zones), and higher rainfall levels observed in the central and western regions. The mapping results can serve as a valuable tool for public management, especially to define areas suitable for groundwater use in the NMG. We encourage future studies for advances and improvements in groundwater modeling processes in the region.