Semantic Segmentation for Disaster Response: Evaluating CNNs and Transformers for Flood Mapping

This study aims to enhance the precision of semantic segmentation in remote sensing by evaluating advanced deep learning models on high-resolution datasets, addressing a critical need in geoscience applications. Accurate spatial identification of flood-affected areas is vital for timely disaster response, yet traditional methods often fail to capture the intricate patterns and scales of flood events. Advanced architectures like Convolutional Neural Networks (CNNs) and transformer models have proven transformative in overcoming these limitations.Using high-resolution imagery from the ISPRS dataset, this research compares CNNs and transformers, including the Vision Transformer (ViT), to identify the most effective architecture. While CNNs excel in extracting localized features, they struggle with capturing long-range dependencies. Transformer models, leveraging self-attention mechanisms, address this gap by modeling complex spatial relationships and global contexts, crucial for segmenting large-scale flood scenarios. Additionally, a novel transformer-based framework will be introduced to further enhance segmentation accuracy to detect flooding.To test robustness, the best-performing model is applied to flood detection tasks using lower-resolution datasets, simulating real-world disaster scenarios where data quality varies. Flood detection through advanced deep learning is essential given the growing frequency of climate-driven disasters. These models enable precise and timely mapping of inundated areas, critical for effective resource allocation, evacuation planning, and post-disaster recovery. Transformers’ ability to process fine-grained and large-scale spatial features complements CNNs, delivering more reliable and detailed flood mapping.Focusing on coastal and urban flooding from Hurricane Milton, the study demonstrates the practical utility of these models in diverse scenarios. By optimizing model selection for flood detection, this research advances remote sensing methodologies, bridging the gap between theoretical advancements and real-world applications, and contributing to disaster preparedness and climate resilience efforts.