Temporal Gap Filling and Model-Based Spatial Downscaling of GRACE-Based Groundwater-Storage Anomalies Using Gaussian Process and Random Forest Models

Groundwater-storage anomalies (GWSA) derived from the Gravity Recovery and Climate Experiment (GRACE) mission provide valuable information for regional groundwater monitoring. Improving the spatial representation and temporal continuity of GRACE-derived GWSA is important for supporting groundwater assessment at subregional scales. A sequential framework combining Gaussian Process (GP) temporal gap filling and Random Forest (RF) spatial downscaling was developed for GWSA reconstruction in Henan Province, China, during 2002–2022. The GP model was used to reconstruct missing observations and characterize temporal variations, while the RF model statistically redistributed the GRACE-based GWSA signal using multi-source hydroclimatic predictors. The resulting dataset comprises model-derived GWSA estimates on a 1 km output grid constrained by the coarse spatial support of GRACE observations and the relationships learned from the auxiliary variables. Therefore, the 1 km grid spacing should not be interpreted as an independent 1 km resolving capability for groundwater-storage variations. Agreement with the parent GRACE-based GWSA product was used to assess coarse-scale reconstruction consistency rather than independent fine-scale accuracy. Comparison with groundwater-level anomalies from 63 monitoring wells yielded a correlation coefficient of 0.88, indicating temporal agreement at the sampled locations. Because the groundwater-level observations were not converted into storage anomalies using specific yield, this comparison does not establish absolute GWSA accuracy or independently validate the model-derived fine-scale spatial patterns. The reconstructed estimates revealed pronounced spatial heterogeneity in groundwater-storage changes, with persistent depletion concentrated in northern Henan, where groundwater decline rates exceeded 20 mm yr−1. Overall, the framework improved the temporal continuity and spatial representation of GRACE-based groundwater-storage estimates while retaining the fundamental spatial constraints of satellite gravimetry. The results demonstrate the potential of integrating GRACE observations, machine learning, and multi-source Earth observation data to support regional groundwater assessment.