Scholarlink

<p>Satellite precipitation products such as GPM IMERG can show reduced accuracy over complex terrain and high-humidity monsoon regions, where orographic enhancement and convective variability complicate retrievals. In karst-dominated southern China, precipitation is highly heterogeneous and responds rapidly to rainfall events, posing challenges for direct use of raw satellite estimates. This study develops an error-decomposition–based machine learning framework, separating IMERG–gauge discrepancies into missed, false, overestimation-hit, and underestimation-hit components, linked to meteorological, topographic, and temporal features. Three ML models were trained and benchmarked against quantile mapping and geographically and temporally weighted regression. Spatially separated bidirectional validation assessed transferability between karst and non-karst subregions. All models substantially reduced RMSE (≈ 40–45%) and MAE (≈ 45–50%) relative to raw IMERG, maintaining low residual bias across regions. SHAP analysis highlights relative humidity, elevation, and seasonality as key error drivers. By explicitly modeling error mechanisms, the framework provides process-aware, physically interpretable, and spatially transferable correction for satellite precipitation in complex terrain.</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

Error-Decomposition Machine Learning for IMERG Precipitation Correction in Karst Regions

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Fuwan Gan,
  • Xinsong Zhang,
  • Xiang Diao,
  • Xianci Zhong,
  • Yang Gao

摘要

Satellite precipitation products such as GPM IMERG can show reduced accuracy over complex terrain and high-humidity monsoon regions, where orographic enhancement and convective variability complicate retrievals. In karst-dominated southern China, precipitation is highly heterogeneous and responds rapidly to rainfall events, posing challenges for direct use of raw satellite estimates. This study develops an error-decomposition–based machine learning framework, separating IMERG–gauge discrepancies into missed, false, overestimation-hit, and underestimation-hit components, linked to meteorological, topographic, and temporal features. Three ML models were trained and benchmarked against quantile mapping and geographically and temporally weighted regression. Spatially separated bidirectional validation assessed transferability between karst and non-karst subregions. All models substantially reduced RMSE (≈ 40–45%) and MAE (≈ 45–50%) relative to raw IMERG, maintaining low residual bias across regions. SHAP analysis highlights relative humidity, elevation, and seasonality as key error drivers. By explicitly modeling error mechanisms, the framework provides process-aware, physically interpretable, and spatially transferable correction for satellite precipitation in complex terrain.