Groundwater level estimation in northern region of Bangladesh using hybrid locally weighted linear regression and Gaussian process regression modeling

Abstract

Urban groundwater resources (GWRs) have declined substantially in recent decades, due to rapid urbanization, population growth, groundwater exploitation, land use/land cover changes, and climate change. However, the knowledge about the role of the underlying variables on groundwater level (GWL) fluctuation on a local scale in the drought-prone urban areas of Bangladesh is still not explored. To gain better insight into the relative contributions of underlying factors on GWL fluctuation, this study proposed a novel hybrid ensemble modeling framework based on locally weighted linear regression (LWLR) and four Gaussian process regressions (GPRs), e.g., poly kernel, Pearson universal kernel (PUK), radian basis function (RBF), and normalized poly kernel. The proposed framework has been employed to predict GWL at six wells in the drought-prone local areas of the north-western urban region of Bangladesh, where GWL is declining rapidly. The rainfall, temperature (Tave), soil moisture (SM), normalized difference vegetation index (NDVI), Indian Ocean Dipole (IOD), Southern Oscillation Index (SOI), Nina3.4, and population growth rate for the period 1993–2017 were utilized as inputs to develop GWL models. The best input combination was explored using the best subset regression model and sensitivity analysis, and the optimal input combination was applied in LWLR and GPRs to estimate the monthly GWL fluctuation. On average, the hybrid LWLR-GPR-PUK model, improves the prediction accuracy by 10 to 50% during the training stage and 20 to 70% during the testing stage compared to other models. The proposed modeling tool could be a good alternative to physical law-based models when there is insufficient groundwater data to make them. This is true for drought-prone areas in urban areas where groundwater data is scarce.