Scholarlink

<p>Efficient disaster management relies on the accurate identification and classification of disasters. This has remained an issue for recent methods as they are subject to a lot of low image quality and complex backgrounds. The work in this study thus proposes an intelligent hybrid deep learning framework that incorporates in its pre-processing techniques such as median filtering, histogram equalization, and SA-SSOA-enhanced Otsu segmentation, with LBP, LTP, and GLCM utilized as feature extraction methods. A tri-level architecture integrating CNN, RNN, and GRU with Multi-Head Attention Networks is developed to improve spatial–temporal representation. As a result of its superior performance, it achieves accuracy of 98.15% and 99.53% for both such data splits as 70/30 and 80/20. Further evaluation metrics confirm the robustness of the system in terms of precision (97.53, 98.21), F1-Score (97.60, 98.24), specificity (98.54, 98.96), sensitivity (97.95, 98.22), and Matthews Correlation Coefficient (97.28, 98.84). The proposed method outperforms all baseline models—CNN, LSTM, and ResNet-50, as it produced very low False Positive Rates (2.86, 1.24) and False Negative Rates (2.18, 0.93). Higher reliability, precision, and generalization have been set up by this hybrid model as the new standard for disaster classification important for real-world emergency response systems.</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 novel multi-model fused classification for classifying the natural disaster

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Gourav Mondal,
  • Rajesh Kumar Dhanaraj

摘要

Efficient disaster management relies on the accurate identification and classification of disasters. This has remained an issue for recent methods as they are subject to a lot of low image quality and complex backgrounds. The work in this study thus proposes an intelligent hybrid deep learning framework that incorporates in its pre-processing techniques such as median filtering, histogram equalization, and SA-SSOA-enhanced Otsu segmentation, with LBP, LTP, and GLCM utilized as feature extraction methods. A tri-level architecture integrating CNN, RNN, and GRU with Multi-Head Attention Networks is developed to improve spatial–temporal representation. As a result of its superior performance, it achieves accuracy of 98.15% and 99.53% for both such data splits as 70/30 and 80/20. Further evaluation metrics confirm the robustness of the system in terms of precision (97.53, 98.21), F1-Score (97.60, 98.24), specificity (98.54, 98.96), sensitivity (97.95, 98.22), and Matthews Correlation Coefficient (97.28, 98.84). The proposed method outperforms all baseline models—CNN, LSTM, and ResNet-50, as it produced very low False Positive Rates (2.86, 1.24) and False Negative Rates (2.18, 0.93). Higher reliability, precision, and generalization have been set up by this hybrid model as the new standard for disaster classification important for real-world emergency response systems.