DEVELOPMENT OF MACHINE LEARNING MODEL FOR GROUND WATER PREDICTION

Abstract

The primary source of pure/fresh water for drinking purposes is groundwater. It is one of the most valuable resources which society consumes through the domestic, industrial, and agricultural water supply. This bulk consumption of groundwater has led this research to develop a proper groundwater prediction model for the prediction of new and healthy resources of groundwater from the surface of the earth. Many surface and subsurface, drilling and radar methods are used for groundwater prediction, but all these methods are costly. From the previous research for predicting groundwater depth, factors like meteorological data, geographic information system (GIS) maps, lithology data etc., were selected with the usability of Machine Learning (ML) techniques. The present study dealt with adding new parameters for predicting groundwater levels. As per the literature, the change of gravity can be computed with the help of the change in terrestrial water storage. Due to temporal variation in gravity, gravity and time were the consecutive parameters used in the model development. The present study used five models: Polynomial Regression, Random Forest, XG boost, K Nearest Neighbourhood, and SVM- Radial Basic Function for predicting groundwater depth. The respective model has been deployed based on the data, and hyperparameter tuning was procured using the cross-validation k-fold method. RMSE metrics and R2 score were used to check the models' accuracy. The research is done in 3 different groups, a single parameter (gravity), a double parameter (gravity and time) and meteorological parameters. Few gravity data samples (~177) were collected using a relative gravimeter conjugated with 284 meteorological samples and the water depth. The models are trained and tested in the ratio of 0.85:0.15, respectively. Comparative evaluations were done between the models, followed by the groups. XG boost model has performed well in every group and performed best with a double parameter, followed by Random Forest. This study offers a practical approach for GWL prediction with a handful of data.