Scholarlink

<p>The integration of Explainable Artificial Intelligence (XAI) into hydrology and hydrogeology addresses the black-box nature of most machine learning (ML) and deep learning (DL) models, which restricts their interpretability and acceptance. This review provides the first systematic and critical synthesis of XAI applications in hydrology and hydrogeology, covering over 180 peer-reviewed studies published up to 2024, organized across nine domains: (i) evapotranspiration, (ii) glacio-hydrology, (iii) precipitation, (iv) rainfall–runoff, (v) river and streamflow, (vi) sediment transport, (vii) floods and droughts, (viii) water quality, and (ix) groundwater. The review analyzes post-hoc methods like SHAP, LIME, PDP, and ALE, discussing their advantages, disadvantages, and suitability for various hydrological domains. Due to its methodological abundance and ease of use, SHAP has become the most popular. Results indicated that XAI not only advances transparency and trust in AI-based models, but XAI also advances the understanding of physical processes. Remaining difficulties include computational scalability, domain unevenness, and lack of integration with physics-informed models.</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

Explainable Artificial Intelligence in Hydrology: A Review

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Mohammad Zounemat-Kermani,
  • Marwan Kheimi

摘要

The integration of Explainable Artificial Intelligence (XAI) into hydrology and hydrogeology addresses the black-box nature of most machine learning (ML) and deep learning (DL) models, which restricts their interpretability and acceptance. This review provides the first systematic and critical synthesis of XAI applications in hydrology and hydrogeology, covering over 180 peer-reviewed studies published up to 2024, organized across nine domains: (i) evapotranspiration, (ii) glacio-hydrology, (iii) precipitation, (iv) rainfall–runoff, (v) river and streamflow, (vi) sediment transport, (vii) floods and droughts, (viii) water quality, and (ix) groundwater. The review analyzes post-hoc methods like SHAP, LIME, PDP, and ALE, discussing their advantages, disadvantages, and suitability for various hydrological domains. Due to its methodological abundance and ease of use, SHAP has become the most popular. Results indicated that XAI not only advances transparency and trust in AI-based models, but XAI also advances the understanding of physical processes. Remaining difficulties include computational scalability, domain unevenness, and lack of integration with physics-informed models.