Scholarlink

<p>Accurate and reliable long-term sea surface temperature (SST) predictions are crucial for mitigating the effects of climate change, managing marine resources, and preparing for extreme weather events, particularly in dynamically complex regions such as the Korean Peninsula. However, traditional models and existing deep learning approaches often suffer from accuracy degradation and the inability to quantify uncertainty over long prediction periods. Therefore, this study proposes a new deep learning framework called the spatiotemporal uncertainty-aware mixture probabilistic network (STUMP-Net), which is designed to provide highly accurate SST predictions over long periods (up to 28 days). STUMP-Net features an innovative Space-to-Depth architecture that integrates a Multi-Scale Unshuffle Encoder to losslessly downsample spatial features and an efficient latent-based transformer to capture complex spatiotemporal dependencies with minimal computational cost. It also incorporates a PixelShuffle decoder with skip connections for high-resolution spatial reconstruction and a mixture density network (MDN) head that provides comprehensive probabilistic uncertainty quantification. Using historical single-mode SST data from the National Oceanic and Atmospheric Administration (NOAA), STUMP-Net demonstrates state-of-the-art performance, outperforming significantly various benchmark models. The proposed model achieved a mean absolute error (MAE) of 0.836 and a root mean square error (RMSE) of 1.112 in 28-day forecasts. Notably, STUMP-Net achieves this superior accuracy with only 0.39&#xa0;million parameters, demonstrating exceptional efficiency compared to existing heavy-weight models. By providing accurate deterministic forecasts and robust uncertainty estimates, STUMP-Net serves as a valuable tool for supporting evidence-based, risk-informed decision-making concerning ocean climate variability and resource management.</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

STUMP-Net: a spatiotemporal uncertainty-aware mixture probabilistic network for long-term SST forecasting on the Korean Peninsula

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Seungwon Oh,
  • Myeong-Taek Kwak,
  • Hyi-Thaek Ceong

摘要

Accurate and reliable long-term sea surface temperature (SST) predictions are crucial for mitigating the effects of climate change, managing marine resources, and preparing for extreme weather events, particularly in dynamically complex regions such as the Korean Peninsula. However, traditional models and existing deep learning approaches often suffer from accuracy degradation and the inability to quantify uncertainty over long prediction periods. Therefore, this study proposes a new deep learning framework called the spatiotemporal uncertainty-aware mixture probabilistic network (STUMP-Net), which is designed to provide highly accurate SST predictions over long periods (up to 28 days). STUMP-Net features an innovative Space-to-Depth architecture that integrates a Multi-Scale Unshuffle Encoder to losslessly downsample spatial features and an efficient latent-based transformer to capture complex spatiotemporal dependencies with minimal computational cost. It also incorporates a PixelShuffle decoder with skip connections for high-resolution spatial reconstruction and a mixture density network (MDN) head that provides comprehensive probabilistic uncertainty quantification. Using historical single-mode SST data from the National Oceanic and Atmospheric Administration (NOAA), STUMP-Net demonstrates state-of-the-art performance, outperforming significantly various benchmark models. The proposed model achieved a mean absolute error (MAE) of 0.836 and a root mean square error (RMSE) of 1.112 in 28-day forecasts. Notably, STUMP-Net achieves this superior accuracy with only 0.39 million parameters, demonstrating exceptional efficiency compared to existing heavy-weight models. By providing accurate deterministic forecasts and robust uncertainty estimates, STUMP-Net serves as a valuable tool for supporting evidence-based, risk-informed decision-making concerning ocean climate variability and resource management.