Scholarlink

Forecasting harmful algal blooms (HABs) with certainty and accuracy is essential to protect marine ecosystems and public health. Building on our previously developed HGS-At-LSTM method that demonstrated state-of-the-art predictive accuracy, this study introduces an additional layer of robustness through uncertainty quantification (UQ). We investigated three complementary approaches, Monte Carlo (MC) Dropout, Fuzzy Logic, and Dempster-Shafer (DS), within the baseline model, enabling the new framework to output not just predictions but also their respective confidence. Each UQ technique is integrated as a post-processing module of the basic model in order to evaluate its ability to recognize various aspects of predictive uncertainty. Experimental results on three HAB datasets reveal that although all uncertainty-aware methods yield better interpretability, the DS-based extension provides the best trade-off between accuracy and uncertainty representation. With \(\text {R}^{2}\) scores of about 0.98 and significantly lower error measures compared to the other UQ methods, the proposed DS-HGS-At-LSTM framework provides a trustworthy and accurate solution for robust HAB forecasting.

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

From Prediction to Confidence: Post-processing Uncertainty Quantification for HAB Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Abir Loussaief,
  • Raïda Ktari,
  • Yessine Hadj Kacem

摘要

Forecasting harmful algal blooms (HABs) with certainty and accuracy is essential to protect marine ecosystems and public health. Building on our previously developed HGS-At-LSTM method that demonstrated state-of-the-art predictive accuracy, this study introduces an additional layer of robustness through uncertainty quantification (UQ). We investigated three complementary approaches, Monte Carlo (MC) Dropout, Fuzzy Logic, and Dempster-Shafer (DS), within the baseline model, enabling the new framework to output not just predictions but also their respective confidence. Each UQ technique is integrated as a post-processing module of the basic model in order to evaluate its ability to recognize various aspects of predictive uncertainty. Experimental results on three HAB datasets reveal that although all uncertainty-aware methods yield better interpretability, the DS-based extension provides the best trade-off between accuracy and uncertainty representation. With \(\text {R}^{2}\) scores of about 0.98 and significantly lower error measures compared to the other UQ methods, the proposed DS-HGS-At-LSTM framework provides a trustworthy and accurate solution for robust HAB forecasting.