Attention-Enhanced Dual-Direction Domain Adaptation for Emergency-Oriented Cross-Domain Multihazard Disaster Classification

Abstract

Rapid multihazard disaster recognition is a key technique for automatic interpretation of immediate remote sensing imagery and plays an important role in emergency response, damage assessment, and risk mitigation. Deep convolutional neural networks (CNNs) have shown strong representation ability for disaster classification. However, their heavy dependence on labeled training samples and sensitivity to cross-domain distribution shifts severely restrict their practical application to newly affected regions, heterogeneous sensors, and different spatial resolutions. This study proposes a semi-supervised domain adaptation (DA) framework, named attention-enhanced dual-direction DA (AEDDA), for rapid cross-domain disaster recognition with scarce target-domain labels. AEDDA first embeds a convolutional attention module to enhance disaster-relevant representations and suppress complex background interference. Then, a CORrelation ALignment (CORAL)-based global distribution alignment (GDA) module is constructed to reduce second-order statistical discrepancies between labeled samples and unlabeled target samples. To further preserve class-level semantic consistency, AEDDA maintains dynamic source and target memory banks and performs bidirectional prototype alignment between cross-domain category centers. In addition, a class-adaptive dual-reliability strategy jointly evaluates prediction entropy and source-prototype consistency, and reliability-weighted target prototypes limit the influence of uncertain or semantically inconsistent pseudo labels. Extensive experiments on a multisource disaster benchmark constructed from xBD, WorldView-3 (W3), GeoEye-1 (G1), and Sentinel-2 (S2) imagery under 12 cross-domain transfer tasks demonstrate that AEDDA achieves superior performance over representative DA methods, with an average accuracy of over 95%. Our code is available at https://github.com/luxury-ray/AEDDA