Scholarlink

<p>Urban flood risk forecasting helps to reduce the consequences of disasters, infrastructure safety, and responding to the emergency in mushrooming big cities. As climate became more volatile with unpredictable rain patterns, efficient data-driven intelligent flood prediction systems are more desperately needed than ever before. Nevertheless, the current deep learning and hybrid models tend to lack generalization over different urban landscape because of the redundancy to the input features, low flexibility according to the spatial-temporal dynamics, and poor attention learning. Those difficulties highly impact the accuracy of prediction and the resilience of the model when it comes to structured data in the real world. These weaknesses will be addressed by proposing a hybrid system called STORM-Net (Spatiotemporal Tracking and Observation of Resilient Mapping) to high-precision prediction of urban flood risks in this paper. The most valuable properties of the suggested piece of work reside in the combination of two new modules, i.e. SAFER (Sparse Autoencoder-guided Feature Elimination with Reinforcement) and BRAVE (BuilderFairy Refined Attention Value Estimator). The algorithm of SAFER is smart at filtering out the redundant or noisy features based on the sparse autoencoder and reinforcement-guided evaluation resulting in a very small but highly effective feature selection. Meanwhile, BRAVE trains the attention scaling factors to pay more attention to important spatiotemporal patterns and yield a better generalization capacity across different flood situations in parallel. This is a synergistic design that allows STORM-Net to learn robust representations and at the same time remain efficient and interpretable. To evaluate by experiment, two benchmark data sets were used such as Kaggle Urban Flood Prediction Dataset and UCI Rainfall in Australia Dataset. The developed STORM-Net framework showed better results than ongoing baseline models by reaching a classification accuracy of 98.9%, precision of 98.8%, recall of 98.7, and the F1-score of 98.9. It has also registered the lowest validation loss of 0.28 and the shortest time 3.1&#xa0;s in training which shows that it runs extremely fast and performs with precision in any prediction.</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

STORM-Net for urban flood risk prediction: an AI-based spatiotemporal tracking and mapping approach

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • M. Sathiyamoorthy,
  • P. Subramanian,
  • N. V. S. Sree Rathna Lakshmi,
  • S. Hemalatha,
  • L. Nagarajan,
  • S. Sathish,
  • M. Gnana Prakash

摘要

Urban flood risk forecasting helps to reduce the consequences of disasters, infrastructure safety, and responding to the emergency in mushrooming big cities. As climate became more volatile with unpredictable rain patterns, efficient data-driven intelligent flood prediction systems are more desperately needed than ever before. Nevertheless, the current deep learning and hybrid models tend to lack generalization over different urban landscape because of the redundancy to the input features, low flexibility according to the spatial-temporal dynamics, and poor attention learning. Those difficulties highly impact the accuracy of prediction and the resilience of the model when it comes to structured data in the real world. These weaknesses will be addressed by proposing a hybrid system called STORM-Net (Spatiotemporal Tracking and Observation of Resilient Mapping) to high-precision prediction of urban flood risks in this paper. The most valuable properties of the suggested piece of work reside in the combination of two new modules, i.e. SAFER (Sparse Autoencoder-guided Feature Elimination with Reinforcement) and BRAVE (BuilderFairy Refined Attention Value Estimator). The algorithm of SAFER is smart at filtering out the redundant or noisy features based on the sparse autoencoder and reinforcement-guided evaluation resulting in a very small but highly effective feature selection. Meanwhile, BRAVE trains the attention scaling factors to pay more attention to important spatiotemporal patterns and yield a better generalization capacity across different flood situations in parallel. This is a synergistic design that allows STORM-Net to learn robust representations and at the same time remain efficient and interpretable. To evaluate by experiment, two benchmark data sets were used such as Kaggle Urban Flood Prediction Dataset and UCI Rainfall in Australia Dataset. The developed STORM-Net framework showed better results than ongoing baseline models by reaching a classification accuracy of 98.9%, precision of 98.8%, recall of 98.7, and the F1-score of 98.9. It has also registered the lowest validation loss of 0.28 and the shortest time 3.1 s in training which shows that it runs extremely fast and performs with precision in any prediction.