A spatially promoted SVM model for GRACE downscaling: Using ground and satellite-based datasets

Abstract

Satellite-based terrestrial water storage changes have been recorded using the Gravity Recovery and Climate Experiment (GRACE) satellite which causing it an important dataset in hydrology and other related fields. GRACE dataset is widely utilized in many studies, but its coarse spatial resolution is a limiting drawback. Machine-learning approaches (e.g., ANN and SVM) are commonly applied in spatially downscaling. However, their input formation, which is in vector form, is a limitation of considering neighbor relations between the gridded-based inputs, specifically in spatial downscaling. Thus, developing an appropriate, simple, fast, and novel model to spatially downscale GRACE resolution is initially necessary for its utilizations. In this study, a Spatially Promoted Support Vector Machine (SP-SVM) model is innovatively proposed for GRACE downscaling from 0.5° to 0.25°. This promotion is investigated utilizing the distances between the unknown target points (with 0.25°) and their surrounding GRACE-valued points (0.5°), called their Distance Effect Coefficient (DEC), as the SP-SVM model input. In addition, the efficiencies of different in-situ and satellite-based datasets (fifteen variables from May 2005 to August 2020) are evaluated as the inputs of the GRACE downscaling models. After finding the most influential datasets, showing the best correlation with the GRACE, their best combinations in GRACE downscaling are identified. Based on the results, the set of PERSIANN-CDR without delay, the in-situ evaporation with a 1-month delay, and the soil moisture in 0–10 cm depth with a 1-month delay show the best performance in GRACE downscaling. The results of GRACE downscaling by the SP-SVM approach are also compared with the ones based on a usual statistical SVM (S-SVM) model, consisting of an intermediate bias interpolation to improve the estimations through a bias correction step. The results show that the SP-SVM model outperforms the common statistical SVM-based. Thus, compared with the usual S-SVM approach, the proposed SP-SVM (linear) model could be used as a simpler and more accurate model for downscaling any variable in a hierarchical process.