Scholarlink

<p>Effective forecasting of the flood depth continues to be the essential element of the proper implementation of disaster management, preservation of infrastructure, and human life. Although accurate simulations are provided by traditional hydrodynamic models, their significant CPU demands usually make them unusable in real-life situations such as in big urban cities that have dense populations and so high spatial resolution is required. In recent years data driven approaches based on deep learning technology have gained traction as attractive alternatives but current approaches are not sufficiently moving towards simulating the complex spatial-temporal interaction of flood dynamics to enable accurate prediction. In order to resolve these identifications, this paper proposes the use of ICENet- a new breed of hybrid deep learning architecture characterized to enhance flood depth prediction. ICENet combines Inception modules and Convolutional Recurrent Units (CRU), thus making it able to recognize multi-scale spatial maps and important temporal correlations associated with floodwater dynamics. This structural composition overcomes important functional challenges that are critical in the existing methods, especially in time based modeling systematic environmental patterns. To maximize learning efficiency and precision further it is optimized by use of DRAW Optimizer which is a specialty-designed, adaptive weighting procedure inspired by rime ice formation and which is used to modify the fusion layers in the network. It guarantees such accurate blending of the different feature streams and boosts convergence. The connection between the hybrid architecture of the model and the task of its functionality consists in the fact that it allows to combine the spatial complexity and the continuity of time and, as a result, the most complete representation of the scene using satellite or aerial images is possible. In comparison, the given ICENet-DRAW framework is compared and shown to perform much better than those available, such as ConvLSTM, U-Net, and traditional CNN-RNN hybrids, with an accuracy of up to 99% in naming prediction on test data. These findings demonstrate the generalized ability of the proposed model in different flood conditions, and thus the proposed model is a scalable model, which provides a good candidate in solving flood depth estimation problem.</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

A Hybrid Deep Learning Approach for Flood Prediction: Integrating ICENet’s Spatial-Temporal Learning with DRAW Optimizer’s Adaptive Weighting

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • S. Sathees Babu,
  • J. Jeba Sonia,
  • S. Aarthee,
  • R. Chithambaramani,
  • P. Ganeshkumar

摘要

Effective forecasting of the flood depth continues to be the essential element of the proper implementation of disaster management, preservation of infrastructure, and human life. Although accurate simulations are provided by traditional hydrodynamic models, their significant CPU demands usually make them unusable in real-life situations such as in big urban cities that have dense populations and so high spatial resolution is required. In recent years data driven approaches based on deep learning technology have gained traction as attractive alternatives but current approaches are not sufficiently moving towards simulating the complex spatial-temporal interaction of flood dynamics to enable accurate prediction. In order to resolve these identifications, this paper proposes the use of ICENet- a new breed of hybrid deep learning architecture characterized to enhance flood depth prediction. ICENet combines Inception modules and Convolutional Recurrent Units (CRU), thus making it able to recognize multi-scale spatial maps and important temporal correlations associated with floodwater dynamics. This structural composition overcomes important functional challenges that are critical in the existing methods, especially in time based modeling systematic environmental patterns. To maximize learning efficiency and precision further it is optimized by use of DRAW Optimizer which is a specialty-designed, adaptive weighting procedure inspired by rime ice formation and which is used to modify the fusion layers in the network. It guarantees such accurate blending of the different feature streams and boosts convergence. The connection between the hybrid architecture of the model and the task of its functionality consists in the fact that it allows to combine the spatial complexity and the continuity of time and, as a result, the most complete representation of the scene using satellite or aerial images is possible. In comparison, the given ICENet-DRAW framework is compared and shown to perform much better than those available, such as ConvLSTM, U-Net, and traditional CNN-RNN hybrids, with an accuracy of up to 99% in naming prediction on test data. These findings demonstrate the generalized ability of the proposed model in different flood conditions, and thus the proposed model is a scalable model, which provides a good candidate in solving flood depth estimation problem.