Multi-source data fusion to enhance groundwater levels prediction: merging monitoring networks and orbital remote sensing datasets

Spatial statistics provides a principled framework for analyzing environmental variables that exhibit spatial dependence, enabling inference and prediction in systems governed by heterogeneous processes. In many hydrogeological applications, the most informative perspective emerges from fusing complementary datasets, for example, sparse groundwater observations and spatially exhaustive remote sensing products. This data fusion is rarely straightforward because data sources often differ in sampling design, uncertainty, and, crucially, spatial support (the area or footprint represented by a measurement). When observations collected at one support are used to predict at another, the change-of-support problem can induce biased variances and degraded predictions if scale effects are ignored. Here, we integrate groundwater levels from a monitoring network with multi-resolution remote sensing covariates to improve groundwater depth mapping while explicitly accounting for support differences. The study targets groundwater level prediction in Southeast Brazil, where relief compartments and land-use patterns generate strong spatial heterogeneity in recharge and water consumption. We combine in situ groundwater table depths observed at 56 monitoring locations with (i) geomorphological information derived from the 30 m TanDEM‑X dataset and (ii) land-surface water consumption represented by 10 m evapotranspiration estimates from SAFER (Simple Algorithm for Evapotranspiration Retrieving). These covariates encode terrain-driven controls and land-use effects that are not fully captured by point measurements alone. Spatial dependence within and across variables is modeled using the Linear Model of Coregionalization (LMC), enabling coherent estimation of direct and cross-variograms. To ensure consistency across supports, we address support homogenization by regularizing point-support variances and cross-structures to a common block support defined on the prediction grid. This regularized LMC is then used within a collocated block cokriging (CBCK) framework, which applies collocated block covariates to enhance block-scale groundwater predictions. Model performance demonstrates substantial gains from explicitly treating change of support and incorporating multi-resolution covariates. CBCK yields reliable groundwater depth predictions with root mean squared error (RMSE) of 0.41 m, markedly outperforming ordinary block kriging (OBK) estimations (RMSE = 2.89 m) and improving upon prior CBCK implementations that relied on coarser (500 m) evapotranspiration inputs (RMSE = 0.49 m). Beyond accuracy improvements, the resulting maps better reflect the coupling between land-use water demand, terrain-driven controls, and groundwater levels, supporting groundwater management decisions relevant to agronomic planning and ecosystem sustainability. The proposed methodology is transferable to other aquifer systems and can be adapted to alternative remote sensing products and field measurements to explore climate, land use, and hydrogeology interactions across spatial scales.