DisasterFormer: An Enhanced Transformer with Edge-Guided Attention for Disaster Image Recognition

n the realm of disaster image classification, effective feature extraction is pivotal for accurate and timely response. This paper presents "DisasterFormer," a novel deep learning model designed to enhance disaster image classification. The architecture combines the CAIT transformer with a convolutional neural network (CNN) branch that extracts edge-based features. These features fuel an edge-based soft attention mechanism integrated into the transformer's encoders, guiding patch-level feature extraction with precision. By leveraging edge information, DisasterFormer prioritizes salient details crucial for identifying various disaster scenarios, thus improving both the accuracy and interpretability of classification results. Initial evaluations of the model demonstrate its potential in advancing disaster response technologies by emphasizing relevant features and enhancing model sensitivity. The model was evaluated on a disaster dataset consisting of twelve classes, where it achieved an exceptional performance with an accuracy of 0.8666 and an AUC of 0.9472. Experimental results show the model’s ability to generalize across different and complex disaster images, effectively balancing precision and recall, and with a high MCC of 0.8296. Qualitative analysis also proves the discriminative capability of the proposed model.