Identifying Soil Profiles and Groundwater Quality using Machine Learning Models with and without Clustering Methods

Abstract

This study outlines the machine learning-based prediction methodology for subsurface lithology and groundwater quality using the K-Nearest Neighbors algorithm on well data collected in the study region. The well data set includes spatial coordinates, well depth, lithology, and important water quality parameter values, such as Total Dissolved Solids, calcium, magnesium, turbidity, chloride, and pH. The machine learning-based prediction methodology using the K-Nearest Neighbors algorithm on well data collected in the study region was found to have prediction accuracy of 92.4% for lithology classification and 89.1% for groundwater quality parameter classification. Absolute prediction values were also obtained for the water quality parameter values, with Total Dissolved Solids varying from 420 to 980 mg/L and pH varying from 7.1 to 8.2, matching well with the observed values. On comparing with the WHO and BIS standards for drinking water, some of the well values were found to be above the maximum permissible limits for Total Dissolved Solids, calcium, and magnesium. This is due to the spatial variations in groundwater quality. The study proves that the K-Nearest Neighbors algorithm is effective in capturing spatial and feature-based similarities, thus being useful for hydrogeology-based prediction problems in relatively homogeneous regions.