Reconstructing high-resolution groundwater level data using a hybrid random forest model to quantify distributed groundwater changes in the Indus Basin

High-resolution, continuous groundwater data is important for place-based adaptive aquifer management. This information is unavailable in many areas due to spatial sparsity of and temporal gaps in groundwater monitoring. This study advances the ability to generate high-resolution (1 km(2)), temporally continuous estimates of groundwater level (GWL) changes by incorporating 1 km(2) covariates and existing piezometer observations into predictive modeling. We employed a hybrid machine learning (ML) model, primarily using the geographically weighted random forest (RFgw) model. To assess the performance of the RFgw model, we conducted a comprehensive comparison with the SGS geostatistical method and non-spatial ML models (RF and XGBoost). The framework was implemented across the Indus Basin using biannual (July and Oct) GWL data from piezometers and local covariates from 2003 to 2020. The RFgw model demonstrated superior accuracy in predicting GWLs, improving R-2 by 10 %, 17 %, and 22 % compared to SGS, RF, and XGBoost, respectively. Notably, SGS, RF, and XGBoost substantially underestimated the GWL in deeper wells (7-11 m), whereas RFgw showed a much smaller underestimation (up to similar to 3 m). The 90 % prediction interval revealed that RFgw had less uncertainty (1-3 m) followed by RF (2-5 m), and SGS and XGBoost (up to 8 m) for most testing piezometers. Incorporating high-resolution covariates into RFgw predictive modeling provided reliable estimates of GWL changes for unmonitored sites. Using the reconstructed GWL data, we examined the GWL changes in head (i.e., upstream) and tail (i. e., downstream) farms within canal distributaries, illustrating faster groundwater drawdown in tail farms (e.g., 0.82 m/yr) than head farms (0.02 m/yr in the Hakra canal distributary). Densely populated urban areas (e.g., Lahore, Multan, and Faisalabad) had the highest GWL decline (e.g., up to 0.9 m/yr). The framework can be used in other groundwater-stressed regions to support better aquifer management in the face of limited in-situ observations.