Scholarlink

<p>Accurate flood susceptibility modeling is critical for effective disaster risk management, particularly in rapidly urbanizing regions exposed to extreme precipitation. This study applies machine learning models to map flood susceptibility across the São Paulo Metropolitan Region (SPMR), Brazil, and integrates Explainable Artificial Intelligence (XAI) techniques to improve model transparency and interpretability. Among the tested models, Random Forest achieved the highest predictive performance (F1-score: 71.14%), outperforming XGBoost (64.86%), LightGBM (64.77%), and artificial neural networks (38.36%). SHapley Additive exPlanations (SHAP) identified valley depth, land cover (CN_Grid), and geology as the most influential predictors, whereas vegetation indices such as NDVI and NDMI exhibited low explanatory power. Complementary use of Local Interpretable Model-Agnostic Explanations (LIME) and counterfactual analysis enabled local-level interpretability, illustrating how specific environmental conditions influence flood risk. The susceptibility map revealed that 92% of the SPMR is classified as very low susceptibility, but localized high-risk zones exist, particularly in pasturelands, wetlands, and built-up areas. These findings emphasize that topographic and hydrological factors are the primary drivers of flood susceptibility, and that integrating XAI into flood modeling enhances both scientific insight and practical utility. The study demonstrates that interpretable ensemble models offer a reliable and transparent framework for urban flood risk assessment and spatially targeted mitigation planning in the context of climate adaptation.</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

Interpretable machine learning for flood susceptibility mapping in the metropolitan region of São Paulo, Southeast Brazil

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月16日

查看全文
  • Enner Alcântara,
  • Yasmim Carvalho Guimarães,
  • Cheila Flávia Praga Baião,
  • José Roberto Mantovani

摘要

Accurate flood susceptibility modeling is critical for effective disaster risk management, particularly in rapidly urbanizing regions exposed to extreme precipitation. This study applies machine learning models to map flood susceptibility across the São Paulo Metropolitan Region (SPMR), Brazil, and integrates Explainable Artificial Intelligence (XAI) techniques to improve model transparency and interpretability. Among the tested models, Random Forest achieved the highest predictive performance (F1-score: 71.14%), outperforming XGBoost (64.86%), LightGBM (64.77%), and artificial neural networks (38.36%). SHapley Additive exPlanations (SHAP) identified valley depth, land cover (CN_Grid), and geology as the most influential predictors, whereas vegetation indices such as NDVI and NDMI exhibited low explanatory power. Complementary use of Local Interpretable Model-Agnostic Explanations (LIME) and counterfactual analysis enabled local-level interpretability, illustrating how specific environmental conditions influence flood risk. The susceptibility map revealed that 92% of the SPMR is classified as very low susceptibility, but localized high-risk zones exist, particularly in pasturelands, wetlands, and built-up areas. These findings emphasize that topographic and hydrological factors are the primary drivers of flood susceptibility, and that integrating XAI into flood modeling enhances both scientific insight and practical utility. The study demonstrates that interpretable ensemble models offer a reliable and transparent framework for urban flood risk assessment and spatially targeted mitigation planning in the context of climate adaptation.