Scholarlink

<p>Accurate precipitation forecasts at subseasonal lead times (1–5 weeks) remain challenging because numerical weather prediction (NWP) models exhibit systematic biases and limited predictability. Here we introduce ReST, a deep-learning-based post-processing framework that improves global subseasonal precipitation forecasts by explicitly incorporating geographic and seasonal conditioning. The model integrates Spatially Adaptive Denormalization (SPADE) and Feature-wise Linear Modulation (FiLM) within a unified spatio-temporal adaptive modulation architecture embedded in a U-Net backbone. ReST is trained using 20 years (2000–2019) of GEFSv12 reforecasts and evaluated against a global station-based precipitation dataset derived from 63,588 stations. Compared with raw forecasts and conventional post-processing approaches, including quantile mapping, random forest, and a Res34-Unet model, ReST consistently improves forecast skill and produces more spatially coherent precipitation fields. Skill gains are largest during Weeks 1–2, while predictability declines rapidly beyond Week 2. These results highlight the importance of spatially explicit conditioning for correcting global precipitation forecast biases at subseasonal timescales.</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 with spatio-temporal conditioning improves global subseasonal precipitation forecasts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Gyu-Ho Noh,
  • Kuk-Hyun Ahn

摘要

Accurate precipitation forecasts at subseasonal lead times (1–5 weeks) remain challenging because numerical weather prediction (NWP) models exhibit systematic biases and limited predictability. Here we introduce ReST, a deep-learning-based post-processing framework that improves global subseasonal precipitation forecasts by explicitly incorporating geographic and seasonal conditioning. The model integrates Spatially Adaptive Denormalization (SPADE) and Feature-wise Linear Modulation (FiLM) within a unified spatio-temporal adaptive modulation architecture embedded in a U-Net backbone. ReST is trained using 20 years (2000–2019) of GEFSv12 reforecasts and evaluated against a global station-based precipitation dataset derived from 63,588 stations. Compared with raw forecasts and conventional post-processing approaches, including quantile mapping, random forest, and a Res34-Unet model, ReST consistently improves forecast skill and produces more spatially coherent precipitation fields. Skill gains are largest during Weeks 1–2, while predictability declines rapidly beyond Week 2. These results highlight the importance of spatially explicit conditioning for correcting global precipitation forecast biases at subseasonal timescales.