Modeling spatial groundwater level patterns of Bangladesh using physio-climatic variables and machine learning algorithms

Abstract

Groundwater monitoring is essential for sustainable groundwater resource management in a country like Bangladesh, where this precious resource is gradually declining due to over-extraction. However, traditional methods of acquiring groundwater level (GWL) data over a large area are time-consuming and expensive. To address this, this study proposes an alternative approach using freely available daily groundwater storage (GWS) gridded data from the Global Land Data Assimilation System (GLDAS) and other data sources such as population, rainfall, temperature, irrigation, and elevation. By employing different regression and machine learning models like multiple linear regression (MLR), regression trees, support vector machines (SVM), Gaussian process regression (GPR), and artificial neural networks (ANN), the study aimed to model GWL data for Bangladesh at a spatial resolution of 0.25° × 0.25°. In-situ weekly GWL data collected from 844 locations across the country were used for model development. The results indicated that GWS data alone was insufficient to estimate the spatial variability and trend of groundwater in Bangladesh. However, the comparison of different models showed that the ANN model performed better, with an overall correlation coefficient (R) of 0.95 and mean squared error (MSE) of 0.64 m2 when estimating GWL using GWS and other data. The study also identified population and rainfall as the most influential factors in determining GWL. The developed ANN model can be utilized to estimate GWL at locations where observation data are unavailable, enabling the monitoring of GWL for sustainable groundwater management in Bangladesh.