Gridded dataset for groundwater storage estimation in the Duero using a Deep Learning model

Gridded dataset supporting the estimation of monthly groundwater storage change (GWSC) in the Duero River Basin (Spain) with a spatio-temporal deep-learning model, developed within the STARS4Water project (EC grant 101059372). It contains: (i) Deep Learning model inputs — static and dynamic spatiotemporal data layers directly or indirectly related to groundwater dynamics, from both local and global sources (soil properties, geology, permeability, groundwater bodies, hydrogeological regions, water bodies, springs, river/canal network, abstraction, precipitation, potential evapotranspiration, maximum temperature, CORINE land cover and population); (ii) Deep Learning model outputs — corrected groundwater storage changes, both observations and model predictions; and (iii) simulated future climate scenarios (RCP 4.5 and RCP 8.5) of groundwater storage change up to 2040. The model (SpatialTransformer_Concat_AR) is a hybrid spatio-temporal transformer combining static descriptors, dynamic exogenous drivers and the autoregressive history of the target through a gated fusion, producing 12-month forecasts. Test performance: R² = 0.556, MAE = 0.031 m. All layers are 11 km monthly GeoTIFFs in EPSG:4326 with per-band date labels; nodata = −9999; GWSC in metres.