Abstract

Floods are among the most devastating natural disasters, causing widespread damage to infrastructure, homes, and human lives. Rapid assessment of flood severity is critical for effective disaster response and resource allocation. This study explores several deep learning approaches for flood water level classification using UAV imagery. A curated dataset of 2,000 UAV images from diverse regions, including India, the United States, and Brazil, was developed and augmented to improve generalization. Multiple architectures were evaluated, including pre-trained CNNs, ResNet50v2, MobileNetv2, Vision Transformers, and Swin Transformers, with and without the Convolutional Block Attention Module (CBAM) and adaptive learning strategies. Experimental results reveal that integrating Vision Transformers with CBAM achieves the highest classification accuracy of 90.6%, while a hybrid CNN–Vision Transformer model further improves performance to 92.3%. These findings highlight the potential of attention-based hybrid models for precise flood severity mapping. The proposed framework can aid rescue teams and disaster management authorities by prioritizing high-risk areas, enabling faster response and optimized allocation of resources during emergency operations.

Keywords: Flood water level; Pre-trained CNN; CBAM; Vision Transformer; Swin transformer; Hybrid vision transformer