Groundwater Quality Prediction Using Proximal Hyperspectral Sensing, GIS, and Machine Learning Algorithms

The primary objective of the present study is to assess the suitability of groundwater for drinking purposes using the Water Quality Index (WQI), Geographic Information System (GIS), and proximal sensing techniques. For this purpose, 272 groundwater samples collected before and after the monsoon period were analyzed for various hydrochemical parameter concentrations. To explore the relationship between water spectral reflectance and WQI, a spectroradiometer was used to measure the reflectance of each sample under laboratory conditions. Four machine learning procedures i.e., Support Vector Regression (SVR), M5P, Random Forest (RF), and Gaussian Process Regression (GPR) were used to predict WQI using resampled spectral reflectance. The WQI analysis revealed that approximately 30% of the groundwater samples fell within the poor to extremely poor-quality category in both periods, while 55.88% (Pre-monsoon) and 55.14% (Post-monsoon) of samples were classified under moderate quality, indicating marginal suitability for drinking purposes. The spectral curves of WQI indicated a lower reflectance for water samples with lower WQI values, whereas higher reflectance values were associated with higher WQI values. Performance evaluation of machine learning models demonstrated that SVR outperformed M5P, RF, and GPR, achieving the highest Correlation Coefficient (CC = 0.9964) and the lowest Root Mean Square Error (RMSE = 7.0313 mg/L) and Mean Absolute Error (MAE = 3.5056 mg/L) in the training phase, while maintaining similar performance in the testing phase (CC = 0.9964, RMSE = 7.6748 mg/L, MAE = 4.0297 mg/L). These findings highlight the potential of integrating spectral reflectance with machine learning models for accurate groundwater quality assessment and spatial mapping.