Exploring the role of deep neural networks for post-disaster decision support

Disaster management operations are information intensive activities due to high uncertainty and complex information needs. Emergency response planners need to effectively plan response activities with limited resources and assign rescue teams to specific disaster sites with high probability of survivors swiftly. Decision making becomes tougher since the limited information available is heterogenous, untimely and often fragmented. We address the problem of lack of insightful information of the disaster sites by utilizing image data obtained from smart infrastructures. We collect geo-tagged images from earthquake-hit regions and apply deep learning method for classification of these images to identify survivors in debris. We find that deep learning method is able to classify the images with significantly higher accuracy than the conventionally used machine learning methods for image classification and utilizes significantly lesser time and computational resources. The novel application of image analytics and the resultant findings from our models have valuable implications for effective disaster response operations, especially in smart urban settlements. •We analyze images of earthquake-hit regions from social media.•We use deep learning to identify survivors in images of debris.•Deep learning helps us classify images more accurately than machine learning.•Deep learning models are less susceptible to noise than machine learning models.