ADA-DR: An Adversarial Domain Adaptation Framework for Disaster Response to Address Cold-Start Issue for Multiclass Classification of Disaster Images

Rapidly classifying disaster-related social media (SM) images into humanitarian information categories can significantly aid disaster response efforts. However, the lack of labeled data in the initial hours of a disaster presents a cold-start issue for supervised learning models, making real-time classification difficult. To overcome this challenge, an Adversarial Domain Adaptation framework for Disaster Response (ADA-DR) based on Vision Transformer (ViT) is proposed for the real-time classification of SM images into humanitarian information categories. ADA-DR leverages labeled data from a past disaster (source) along with unlabeled data from an ongoing disaster (target) to learn domain-invariant feature representations through the adversarial learning of ViT. This approach aligns the distributions of the source and target domains, enabling the model to generalize to the target domain and effectively classify target images without requiring labeled data. ADA-DR achieves an average F1 score of 84%. Compared with three baseline approaches-(1) Simple transfer-learning (STL) (2) Discrepancy-based domain adaptation (DDA) and (3) Supervised-learning (SL), ADA-DR significantly reduces domain-shift, outperforming STL by 7% and DDA by 4% in average F1 score while maintaining competitiveness with SL models. Furthermore, ADA-DR shows a 6% improvement in average F1 score over existing methods, making it a promising solution for cold-start multiclass SM disaster image classification.