A machine learning downscaling framework based on a physically constrained sliding window technique for improving resolution of global water storage anomaly

초록

Terrestrial water storage anomaly (TWSA) and groundwater storage anomaly (GWSA) data are of great importance for hydrological research and water resource management. However, products derived from the Gravity Recovery and Climate Experiment (GRACE) mission and its Follow-On mission (GRACE-FO) are constrained by the satellite design and variation in processing strategies among different institutions, resulting in multiple suboptimal products. These products not only have a coarse spatial resolution but also suffer from a continuous 11-month gap from July 2017 to May 2018, which represent significant limitations for related research. To address these challenges, the Bayesian-based three-cornered hat (BTCH) method was initially employed to fuse low-uncertainty GRACE-derived TWSA and GWSA products with a coarse resolution (0.5 degrees). degrees ). Subsequently, based on the fused TWSA/GWSA products and multi-source datasets, a machine learning downscaling framework based on a physically constrained sliding window (MLDF-PCSW) was developed. This framework not only enhanced the spatial resolution of the products but also filled the data gap, ultimately producing a high-resolution water storage anomaly product (HWSA v1.0), a long-term, continuous dataset on a global scale (excluding Greenland and Antarctica; variables: TWSA and GWSA; spatial resolution: 0.05 degrees; degrees ; temporal resolution: monthly; period: April 2002 to December 2022). The results of uncertainty analysis using the three-cornered hat (TCH) method indicated that the TWSA derived from the Center for Space Research (CSR) exhibited the lowest uncertainty (23.91 mm) of the three original products, and the uncertainty decreased by 35.42% (15.44 mm) after data fusion using the BTCH method. The HWSA v1.0 dataset based on MLDF-PCSW exhibited high correlation coefficients (CCs) (0.999 and 0.999 for TWSA and GWSA, respectively) and a low root mean square error (RMSE) (0.68 mm and 1.24 mm for TWSA and GWSA, respectively) with the original TWSA and GWSA products, while also providing more detailed TWSA and GWSA spatial information. Independent validation for TWSA_DYY based on the water balance method has been done across six basins and the results are great, with the CCs between 0.69 and 0.92. The GWSA from the MLDF-PCSW demonstrated acceptable correspondence with global in-situ groundwater observations, with a mean CC of 0.27. The other four previously released gap-filled TWSA products also exhibited outstanding performance, with the average CC reaching 0.98 for the 11-month gap. Collectively, these results suggest that the HWSA v1.0 dataset based on the MLDF-PCSW has significant potential as a robust data foundation for related research.

키워드