Disentangling and integrating spatiotemporal features: Deep learning-based downscaling of groundwater storage anomalies from GRACE and GRACE-FO satellites

Xinjiang, China, is one of the most representative arid regions of the world. The research focuses on enhancing the spatial resolution of GRACE-derived Groundwater Storage Anomaly (GWSA) data from 0.5° to 0.1° using a deep learning downscaling framework that decouples and fuses spatio-temporal features. It evaluates three downscaling models—semi-supervised variational autoencoder regression (SSVAER), geographically neural network weighted regression, and geographically and temporally neural network weighted regression (GTNNWR)—using auxiliary variables like evapotranspiration (ET), land surface temperature (LST), and normalized difference vegetation index (NDVI). The performance of these models was further validated against groundwater well observations. The GTNNWR model performed best, increasing the correlation between downscaled GWSA and well observations from 0.47 to 0.57. The analysis reveals that from 2002 to 2023, Xinjiang's groundwater declined at a significant rate of 5.03 ± 9.42 mm yr⁻¹. The driving factors of groundwater changes were decoupled: long-term trends were primarily driven by precipitation, while seasonal variability was strongly influenced by ET. This study provides an effective approach for high-resolution groundwater monitoring, which is crucial for sustainable water resource management in arid regions like Xinjiang. • Decouple and integrated spatiotemporal features in GRACE data. • Spatiotemporal feature fusion improves GWSA downscaling. • GTNNWR was superior in improving correlation with in-situ groundwater data. • Novel framework for high-resolution groundwater can be used in arid regions.