Strategic imputation of groundwater data using machine learning: Insights from diverse aquifers in the Chao-Phraya River Basin

Abstract

Effective groundwater monitoring is essential for sustainable water management, particularly in data-sparse regions. To address inconsistencies in groundwater level data, we developed a machine learning framework for robust data imputation, tested in the Chao-Phraya River (CPR) Basin, a region facing significant groundwater challenges due to high population density and ecological importance. Our study evaluated five models—K-Nearest Neighbors (KNN), Multiple Imputation by Chained Equations (MICE), Multilayer Perceptron (MLP), Random Forest (RF), and Soft Imputation (SI) —to fill gaps in monthly groundwater level data across various locations, aquifer depths, and data loss scenarios. Results show that MICE perform well in high-density well environments, while SI excels with lower well density, maintaining Pearson correlation coefficients (R) above 0.80 and RMSE values below 6 even at 10% data loss. The Coefficient of Variation (COV) analysis also confirmed that imputed data remains stable and reliable. However, the study also reveals a significant decrease in model performance in regions with fewer wells, as indicated by increased RMSE and reduced R. Our findings indicate that machine learning models are capable of handling groundwater level observations with missing data. The well density in a region has a significant impact on these model's performance. Imputation techniques should be tailored to each aquifer's specific characteristics and surroundings in order to get accurate groundwater data.