Scholarlink

<p>Flooding remains a significant and recurring challenge in disaster-prone regions such as Bangladesh, where early detection and real-time monitoring are hindered by limited technological and infrastructural resources. This study presents a lightweight, real-time flood severity classification system based on MobiFloodViT, a vision transformer architecture optimized for mobile and embedded platforms. A diverse dataset comprising 2,327 images was compiled using drones, CCTV footage, and smartphone cameras during the 2024 flood event in Bangladesh. The images were categorized into four classes: No Flood, Heavy Rain, Waterlogged, and High Flood. The MobileViT model achieved a test accuracy of 90%, with high precision, recall, and F1-scores across all categories. Comparative analysis against other lightweight models, including MobileNetV2, MobileNetV3, EfficientNet-Lite, and SqueezeNet, demonstrated MobileViT’s superior performance. A real-time prototype system, incorporating Jetson Nano and existing CCTV infrastructure, was implemented and validated in field conditions, confirming the model’s practical effectiveness. The proposed system offers a scalable, non-intrusive, and cost-efficient solution for real-time flood detection and severity classification, with significant potential for deployment in resource-constrained environments to support early warning and disaster response initiatives.</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

Real-time flood severity classification using MobileViT and visual data in resource-constrained environment of Bangladesh

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Ashif Mahmud Joy,
  • Sadman Rafi,
  • Ayesha Siddiqua,
  • Arpita Roy,
  • Sanjida Afroz Shimu,
  • Md. Ahsan Arif

摘要

Flooding remains a significant and recurring challenge in disaster-prone regions such as Bangladesh, where early detection and real-time monitoring are hindered by limited technological and infrastructural resources. This study presents a lightweight, real-time flood severity classification system based on MobiFloodViT, a vision transformer architecture optimized for mobile and embedded platforms. A diverse dataset comprising 2,327 images was compiled using drones, CCTV footage, and smartphone cameras during the 2024 flood event in Bangladesh. The images were categorized into four classes: No Flood, Heavy Rain, Waterlogged, and High Flood. The MobileViT model achieved a test accuracy of 90%, with high precision, recall, and F1-scores across all categories. Comparative analysis against other lightweight models, including MobileNetV2, MobileNetV3, EfficientNet-Lite, and SqueezeNet, demonstrated MobileViT’s superior performance. A real-time prototype system, incorporating Jetson Nano and existing CCTV infrastructure, was implemented and validated in field conditions, confirming the model’s practical effectiveness. The proposed system offers a scalable, non-intrusive, and cost-efficient solution for real-time flood detection and severity classification, with significant potential for deployment in resource-constrained environments to support early warning and disaster response initiatives.