Nusrat Sultana, Anika Islam, Yousra Tabassum, Jahanggir Hossain Setu, Asif Mahmud, Ashraful Islam, M. Ashraful Amin
2024 7th Asia Conference on Cognitive Engineering and Intelligent Interaction (CEII 2024)
IEEE, pp. 1-5
Floods are among the most devastating natural dis-asters. It causes extensive damage to infrastructure, agriculture, and most importantly, human life. Timely and accurate flood detection is very important for minimizing these impacts. This study explores the use of Deep Learning (DL) models to enhance flood detection through image segmentation of Synthetic Aperture Radar (SAR) imagery. Three segmentation models are compared namely, U-Net, X-Net, and Segment Anything Model 2 (SAM 2), using the NASA IMPACT ETCI 2021 flood detection dataset. The results show that SAM 2 performs better than U-Net and X-Net across key metrics. SAM 2 improves accuracy by almost 1 % over U-Net and 1.64% over X-Net. It also demonstrates a higher precision rate, reducing false positives with a 1.25 % increase compared to U-Net and a 2.32% improvement over X-Net. Moreover, SAM 2 significantly enhances recall, detecting 8.71 % more flood-affected areas than both U-Net and X-Net. In terms of segmentation quality, SAM 2 achieves a higher Dice coefficient and Jaccard index, indicating more accurate flood zone identification and boundary detection, especially in smaller or complex areas, SAM 2 proves to be the most reliable model for flood detection.
Assistant Professor
Department of Computer Science and Engineering
Independent University, Bangladesh (IUB)Human-Computer Interaction, AI for Social Good, AI for Public Health, AI for Impact