Scholarlink

<p>Floods are among the most devastating natural disasters, causing extensive property damage and loss of life. An effective flood early warning system can significantly reduce such damage and save lives. However, data related to flood characteristics are often non-linear and uncertain, which leads to poor predictive accuracy in conventional hydrological models. Traditional models struggle to handle these complexities due to their rigid assumptions, limited adaptability to regional variability, susceptibility to input uncertainties, and inability to capture abrupt changes caused by extreme weather events. This study aims to enhance the prediction of flood occurrence and danger levels in non-tidal rivers by proposing a hybrid deep learning (DL) model that integrates a Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) network, and an Attention mechanism—collectively referred to as the CNN-LSTM-Attention model. Furthermore, we incorporate Explainable AI (XAI) tool SHAPASH, to interpret the model’s predictions. The proposed model is benchmarked against several state-of-the-art DL algorithms using a dataset that includes water level, rainfall, discharge, and other relevant hydrological variables. Experimental results demonstrate that the CNN-LSTM-Attention model outperforms all baseline models in predicting flood occurrence and water danger levels. It achieves 98.63% accuracy with 99% precision, recall, and F1 score for water danger level prediction, and 97.26% accuracy with 95% precision, 97% recall, and 96% F1 score for flood prediction. From a global explainability perspective, SHAPASH indicates that water level is the most influential feature in flood prediction, followed by discharge and rainfall. Local explainability results show that predictions are primarily driven by key features, with lesser influential variables contributing marginally. This study provides a robust and interpretable flood prediction framework, offering valuable insights for hydrologists and climate scientists.</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

Enhancing flood forecasting performance using effective and transparent explainable hybrid deep learning model

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Mahmudul Hasan,
  • Md. Fazle Rabbi,
  • Md Amir Hamja,
  • Kanij Fatema,
  • Md Mahedi Hassan

摘要

Floods are among the most devastating natural disasters, causing extensive property damage and loss of life. An effective flood early warning system can significantly reduce such damage and save lives. However, data related to flood characteristics are often non-linear and uncertain, which leads to poor predictive accuracy in conventional hydrological models. Traditional models struggle to handle these complexities due to their rigid assumptions, limited adaptability to regional variability, susceptibility to input uncertainties, and inability to capture abrupt changes caused by extreme weather events. This study aims to enhance the prediction of flood occurrence and danger levels in non-tidal rivers by proposing a hybrid deep learning (DL) model that integrates a Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) network, and an Attention mechanism—collectively referred to as the CNN-LSTM-Attention model. Furthermore, we incorporate Explainable AI (XAI) tool SHAPASH, to interpret the model’s predictions. The proposed model is benchmarked against several state-of-the-art DL algorithms using a dataset that includes water level, rainfall, discharge, and other relevant hydrological variables. Experimental results demonstrate that the CNN-LSTM-Attention model outperforms all baseline models in predicting flood occurrence and water danger levels. It achieves 98.63% accuracy with 99% precision, recall, and F1 score for water danger level prediction, and 97.26% accuracy with 95% precision, 97% recall, and 96% F1 score for flood prediction. From a global explainability perspective, SHAPASH indicates that water level is the most influential feature in flood prediction, followed by discharge and rainfall. Local explainability results show that predictions are primarily driven by key features, with lesser influential variables contributing marginally. This study provides a robust and interpretable flood prediction framework, offering valuable insights for hydrologists and climate scientists.