Scholarlink

<p>Climate change assessments in West Africa are hindered by limited spatial resolution of Global Climate Models (GCMs), which inadequately capture regional precipitation patterns crucial for adaptation planning. This study evaluates deep-learning architectures for downscaling West African summer monsoon precipitation from coarse-resolution climate models to high-resolution regional projections. Using a Perfect Prognosis framework, we compared five neural network architectures (CNN, UNet_3-32, UNetPP_3-32, UNet_3-24, UNetPP_3-24) trained with ERA5 boundary conditions and CHIRPS observations (1981–2005) to downscale 200&#xa0;km CanESM2 outputs to 25&#xa0;km resolution. Scaling Delta Mapping was applied to ensure physically consistent predictor-predictand relationships. Our assessment showed that the nested UNetPP_3-32 model achieved the highest overall performance (Taylor Skill Score = 0.98, Pattern Correlation Coefficient = 0.98), with 91% of grid points showing added value over raw GCM outputs. Regional analysis revealed marked improvements along the Guinea coast, where complex topography and land-sea interactions influence precipitation, while the Sahel region showed consistently high performance across all UNet architectures (P(AV &gt; 0) &gt; 0.90). The downscaled models demonstrated exceptional skill in simulating extreme precipitation, with Taylor skill scores reaching 0.953 for heavy rainfall days (R20&#xa0;mm). Notably, architectural complexity requirements varied by precipitation characteristic, with the simpler CNN model performing unexpectedly well for frequency-based indices. Spatial analysis reveals that performance improvements concentrate in meteorologically complex zones, particularly in coastal and orographic regions. Our framework demonstrates that deep learning approaches effectively bridge the resolution gap between global and regional scales, providing an essential tool for enhancing climate impact studies in West Africa and beyond.</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

DEEP learning downscaling of West Africa summer monsoon precipitation flow: exploring added value

  • 期刊论文

  • 发表日期：

    2025年10月23日

查看全文
  • Precious Ebiendele,
  • Koji Dairaku,
  • Paul Adigun

摘要

Climate change assessments in West Africa are hindered by limited spatial resolution of Global Climate Models (GCMs), which inadequately capture regional precipitation patterns crucial for adaptation planning. This study evaluates deep-learning architectures for downscaling West African summer monsoon precipitation from coarse-resolution climate models to high-resolution regional projections. Using a Perfect Prognosis framework, we compared five neural network architectures (CNN, UNet_3-32, UNetPP_3-32, UNet_3-24, UNetPP_3-24) trained with ERA5 boundary conditions and CHIRPS observations (1981–2005) to downscale 200 km CanESM2 outputs to 25 km resolution. Scaling Delta Mapping was applied to ensure physically consistent predictor-predictand relationships. Our assessment showed that the nested UNetPP_3-32 model achieved the highest overall performance (Taylor Skill Score = 0.98, Pattern Correlation Coefficient = 0.98), with 91% of grid points showing added value over raw GCM outputs. Regional analysis revealed marked improvements along the Guinea coast, where complex topography and land-sea interactions influence precipitation, while the Sahel region showed consistently high performance across all UNet architectures (P(AV > 0) > 0.90). The downscaled models demonstrated exceptional skill in simulating extreme precipitation, with Taylor skill scores reaching 0.953 for heavy rainfall days (R20 mm). Notably, architectural complexity requirements varied by precipitation characteristic, with the simpler CNN model performing unexpectedly well for frequency-based indices. Spatial analysis reveals that performance improvements concentrate in meteorologically complex zones, particularly in coastal and orographic regions. Our framework demonstrates that deep learning approaches effectively bridge the resolution gap between global and regional scales, providing an essential tool for enhancing climate impact studies in West Africa and beyond.