Scholarlink

<p>Compound extreme heat and drought (CEHD) events have exerted severe impacts on human society and ecosystems. However, accurate prediction of such events remains challenging due to the complexity of the climate drivers and limitations in current prediction methods. Using three machine learning (ML) methods, this study builds effective seasonal prediction models for the CEHD days (CEHDDs) in China during the peak summers of 1983–2022. The six leading empirical orthogonal function (EOF) modes of CEHDDs are highly predicted by the ML methods based on four precursors: sea surface temperature (SST), sea ice, snow cover, and soil moisture. Almost all methods demonstrate significant skills in predicting the principal components (PCs) of the six modes at even 13–24 months in advance. These predicted PCs are then projected onto the observed spatial patterns to reconstruct the predicted spatial fields of CEHDDs in China. The prediction skill varies interannually and spatially, with higher skills in hotspot years and regions, and vice versa. The anomaly correlation coefficient skills reach approximately 0.8 at 13–24 months lead in hotspot years of 2016 and 2022. Further analysis reveals that SST, snow cover, and sea ice provide substantial contributions to the CEHDD prediction across 1–24 months lead, whereas the influence of soil moisture diminishes beyond one year. This study provides new insights for seasonal prediction and predictability of CEHD.</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

Long-lead seasonal prediction of compound extreme heat and drought events in China using machine learning methods

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Jinjin Li,
  • Ruixi Zhu,
  • Yingxiang Guo,
  • Zhenjie Li,
  • Song Yang,
  • Kaiqiang Deng,
  • Zhenchen Liu,
  • Tuantuan Zhang

摘要

Compound extreme heat and drought (CEHD) events have exerted severe impacts on human society and ecosystems. However, accurate prediction of such events remains challenging due to the complexity of the climate drivers and limitations in current prediction methods. Using three machine learning (ML) methods, this study builds effective seasonal prediction models for the CEHD days (CEHDDs) in China during the peak summers of 1983–2022. The six leading empirical orthogonal function (EOF) modes of CEHDDs are highly predicted by the ML methods based on four precursors: sea surface temperature (SST), sea ice, snow cover, and soil moisture. Almost all methods demonstrate significant skills in predicting the principal components (PCs) of the six modes at even 13–24 months in advance. These predicted PCs are then projected onto the observed spatial patterns to reconstruct the predicted spatial fields of CEHDDs in China. The prediction skill varies interannually and spatially, with higher skills in hotspot years and regions, and vice versa. The anomaly correlation coefficient skills reach approximately 0.8 at 13–24 months lead in hotspot years of 2016 and 2022. Further analysis reveals that SST, snow cover, and sea ice provide substantial contributions to the CEHDD prediction across 1–24 months lead, whereas the influence of soil moisture diminishes beyond one year. This study provides new insights for seasonal prediction and predictability of CEHD.