SECA-Net: A Lightweight Spatial and Efficient Channel Attention for Enhanced Natural Disaster Recognition

Disaster management and detection from satellite imagery and different sources are crucial yet challenging tasks. In this work, we introduce SECA-Net, a compact deep learning architecture designed to enhance disaster image classification. Employing a cascade attention mechanism, SECA-Net refines feature maps to improve accuracy significantly, achieving a validation accuracy of 93.45% on a diverse dataset covering four disaster categories, surpassing traditional models. Despite its lightweight size of 24 MB, SECA-Net is robust and generalizable across various environments, making it ideal for use in resource-constrained settings typical of disaster-affected areas. This model offers a promising solution for effective disaster response, combining efficiency, accuracy, and practical applicability in challenging real-world scenarios.