Scholarlink

<p>Satellite-derived precipitation products are indispensable for climatological analysis in gauge-sparse regions, yet they often exhibit substantial systematic biases that vary across precipitation intensity regimes, seasons, and topographic settings. This study proposes a regime-aware bias-correction framework specifically designed for multi-source satellite precipitation in monsoon-dominated tropical climates. The framework integrates five sequential modules: (1) error structure decomposition across precipitation intensity regimes, (2) capacity-controlled machine learning correction conditioned on regime membership, (3) physics-informed consistency enforcement, (4) ensemble-based variation tracking with reliability stratification, and (5) multi-dimensional robustness validation encompassing spatial transferability, temporal generalization, and regime-specific stability. Applied to GPM IMERG (1998–2014) and ERA5 (1989–2014) products over Peninsular Malaysia, with bias correction trained against daily observations from 64 rain gauge stations, the framework achieved stable correction performance across heterogeneous precipitation conditions. Ensemble-based reliability stratification effectively differentiated high-confidence predictions (RMSE: 1.27&#xa0;mm/day) from low-reliability estimates (RMSE: 6.80&#xa0;mm/day). Leave-one-station-out cross-validation confirmed spatial transferability with moderate degradation (mean RMSE increase: 7%), while cross-season validation demonstrated bounded temporal generalization (RMSE increase: 11%). The robustness envelope delineates conditions of reliable correction—moderate intensities, lowland terrain, established monsoon periods—and identifies contexts requiring cautious interpretation. By prioritizing reliability over peak accuracy, this framework offers a practical, transferable approach for constructing analysis-ready precipitation climate datasets in challenging observational environments.</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

A regime-aware bias-correction framework for multi-source satellite precipitation in monsoon-dominated regions: capacity-controlled modeling and physics-informed validation

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Ziqi Liang,
  • Wan Zurina Wan Jaafar,
  • Yik Ng Cia,
  • Ahmed Elshafie,
  • Juneng Liew,
  • Wan Zurina

摘要

Satellite-derived precipitation products are indispensable for climatological analysis in gauge-sparse regions, yet they often exhibit substantial systematic biases that vary across precipitation intensity regimes, seasons, and topographic settings. This study proposes a regime-aware bias-correction framework specifically designed for multi-source satellite precipitation in monsoon-dominated tropical climates. The framework integrates five sequential modules: (1) error structure decomposition across precipitation intensity regimes, (2) capacity-controlled machine learning correction conditioned on regime membership, (3) physics-informed consistency enforcement, (4) ensemble-based variation tracking with reliability stratification, and (5) multi-dimensional robustness validation encompassing spatial transferability, temporal generalization, and regime-specific stability. Applied to GPM IMERG (1998–2014) and ERA5 (1989–2014) products over Peninsular Malaysia, with bias correction trained against daily observations from 64 rain gauge stations, the framework achieved stable correction performance across heterogeneous precipitation conditions. Ensemble-based reliability stratification effectively differentiated high-confidence predictions (RMSE: 1.27 mm/day) from low-reliability estimates (RMSE: 6.80 mm/day). Leave-one-station-out cross-validation confirmed spatial transferability with moderate degradation (mean RMSE increase: 7%), while cross-season validation demonstrated bounded temporal generalization (RMSE increase: 11%). The robustness envelope delineates conditions of reliable correction—moderate intensities, lowland terrain, established monsoon periods—and identifies contexts requiring cautious interpretation. By prioritizing reliability over peak accuracy, this framework offers a practical, transferable approach for constructing analysis-ready precipitation climate datasets in challenging observational environments.