Scholarlink

<p>Terrestrial Water Storage Change (TWSC) data from Gravity Recovery and Climate Experiment (GRACE) / its Follow-On (GFO) missions are essential for understanding hydrological processes, yet its coarse resolution limits basin-scale applications. This study proposed a Graph Convolutional Network (GCN) downscaling framework that leverages spatial topological relationships between grid cells, and compared its results with those of Convolutional Neural Networks (CNN) and Artificial Neural Networks (ANN) to verify its superiority. Finally, the high-resolution (0.1°) TWSC was generated to quantify hydrological dynamics and vegetation interactions in the Pearl River basin (2003–2024). The closed-loop validation results demonstrated that GCN significantly outperforms both benchmarks, achieving Root Mean Square Error reductions of 1.83&#xa0;mm versus CNN and 10.23&#xa0;mm versus ANN. At grid and sub-basin scales, GCN delivered superior accuracy in 56% of grid points, confirming the advantage of topological modeling. Ecological analysis revealed that TWSC-dependent ecosystems are mainly concentrated in the western part and the areas with low population density. The linkage between vegetation growth and TWSC on the annual scale is more pronounced than that during the growing season, with croplands/grasslands exhibiting higher TWSC sensitivity than forests/shrubs. This study verified the applicability and superiority of graph-based deep learning for TWSC downscaling, and offered a feasible technical reference for hydrological and ecological research in tropical and subtropical basin regions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing GRACE/GFO spatial downscaling accuracy via graph convolutional networks: a comparative study and ecological response analysis in the Pearl River basin

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Yuheng Lu,
  • Lilu Cui,
  • Bo Zhong,
  • Chuanjiang Luo,
  • Haoyang Guo,
  • Xiaochuan Qu

摘要

Terrestrial Water Storage Change (TWSC) data from Gravity Recovery and Climate Experiment (GRACE) / its Follow-On (GFO) missions are essential for understanding hydrological processes, yet its coarse resolution limits basin-scale applications. This study proposed a Graph Convolutional Network (GCN) downscaling framework that leverages spatial topological relationships between grid cells, and compared its results with those of Convolutional Neural Networks (CNN) and Artificial Neural Networks (ANN) to verify its superiority. Finally, the high-resolution (0.1°) TWSC was generated to quantify hydrological dynamics and vegetation interactions in the Pearl River basin (2003–2024). The closed-loop validation results demonstrated that GCN significantly outperforms both benchmarks, achieving Root Mean Square Error reductions of 1.83 mm versus CNN and 10.23 mm versus ANN. At grid and sub-basin scales, GCN delivered superior accuracy in 56% of grid points, confirming the advantage of topological modeling. Ecological analysis revealed that TWSC-dependent ecosystems are mainly concentrated in the western part and the areas with low population density. The linkage between vegetation growth and TWSC on the annual scale is more pronounced than that during the growing season, with croplands/grasslands exhibiting higher TWSC sensitivity than forests/shrubs. This study verified the applicability and superiority of graph-based deep learning for TWSC downscaling, and offered a feasible technical reference for hydrological and ecological research in tropical and subtropical basin regions.