Handcrafted Features Based Analysis of Social Media Images for Disaster Response

During a disaster images shared on social media enable concerned authorities to obtain rapid situational awareness and to estimate incurred damages. As the volume and velocity of such content are typically high, real-time filtering of irrelevant images and damage assessment has become an urgent need for faster disaster response. However, research in this field has yet to receive sufficient attention and extracting useful information is still challenging. This paper proposes a framework for filtering out irrelevant images and estimating the damages incurred due to a disaster. The framework relies on the fusion of handcrafted feature descriptors to analyze the social media imagery data. The performance of the proposed framework is compared with the existing handcrafted and learned feature descriptors. The results reveal that the proposed framework is able to filter out the irrelevant images with an accuracy of 87.8% (precision 89.6%, recall 90.4% and F1 score 90.0%) and estimate the disaster severity with an accuracy of 76.9% (precision 86.8%, recall 83.8% and F1 score 84.8%). In order to assess the robustness and generalization capability of the proposed framework, we compared it with the state-of-the-art techniques. Moreover, the proposed framework results in less computational cost than the deep learning models while achieving the comparable accuracy. We believe that the proposed framework can be integrated with the social media platform to filter out irrelevant images and assess the fatal impact of a disaster.