DisasterRes-Net: A framework for analyzing social media images in disaster response

Social media platforms generate vast real-time data that assist government and relief organizations in formulating rapid and efficient disaster response strategies. Compared to the texts, images in social media tend to provide more factual and precise information. However, in the literature, limited studies focus on utilizing social media images to develop efficient and effective disaster response action plans. In this work, we address three disaster-related problems: (i) identification of the informative images, (ii) image-based damage classification, and (iii) image-level damage detection and evaluation. We present a framework named DisasterRes-Net (Disaster Response Network) that analyzes the images shared on the social media platforms during emergency situations. This framework comprises two main modules. The first module combines the deep and handcrafted features to segregate relevant images and classify them based on the severity of the depicted damage. Whereas, the second module introduces a novel and less computationally intensive approach that leverages sparse visual features for detecting and evaluating the extent of the damage. Furthermore, we create a large collection of labeled dataset for the purposes of damage severity identification and informativeness classification. The proposed framework achieves a significant accuracy of 85% for identifying the informative images and of 78% for the severity classification, respectively. Additionally, we also present a damage distribution map for damage detection and a damage evaluation metric for evaluating the extent of damage. The results confirm that the DisasterRes-Net provides valuable insights to decision-makers involved in disaster response and rescue efforts.