Machine learning models to predict groundwater level in a Semi-arid river catchment, Central India

Abstract

We applied different machine learning algorithms to predict the groundwater water fluctuation in a semi-arid river basin. Precipitation, temperature, evaporation, relative humidity, soil type, and groundwater lag were modeled as input features. Feature importance analysis of the input features indicate that the groundwater lag is the most relevant whereas the soil type is the least relevant input features. We applied the backward elimination approach to eliminate the less relevant features in mapping groundwater. Using the relevant input features we trained different machine-learning models (random forest, decision tree, neural network, linear regression, ridge regression, support vector regression, k-nearest neighbors, recurrent neural network). All these algorithms predicted the groundwater level with a high correlation of coefficient (R) ranging from 0.83 to 0.91 and a Root Mean Square Error (RMSE) ranging from 1.61 to 2.17 m.  We found that the random forest algorithm outperforms the other algorithms with a RMSE of 1.61 and R = 0.91.