Improving Machine Learning based Groundwater Level Estimation using Geological Features

Abstract

Estimation of Groundwater level is crucial for managing water resources. Forecasting groundwater level changes can help determine the efficient utilisation of groundwater resources and drive water conservation efforts, especially in arid regions. Existing works have used machine learning techniques to estimate groundwater levels using meteorological data. However, they have restricted the scope of their research to areas with abundant, continuous time-series data. In this paper, we aim to address the issue of sparse data in estimating groundwater levels. This study explores a data-driven approach and thus does not introduce a new machine learning model. We expand the input parameters to incorporate geological and demographic data along with traditional meteorological data. We have collected data of the Kutch region in Gujarat, spanning 11 years with varying data availability at monitoring sites. Using techniques like Random Forest Regression and Neural Networks, we can improve the estimation of groundwater levels compared to using traditional features. We also analyse causal effects of different values of Geological parameters by extending the concept of treatment effect and provide interpretability of the estimation models. The results presented here indicate that factors like soil type and depth are essential in estimating groundwater level and can improve performance on sparse time-series data. The treatment effect analysis also provides results that conform to existing knowledge, thereby bridging the semantic gap between computer science and hydrogeology domains.