Learning disaster-scenes using deep learning methods

Structures and infrastructure systems in modern society are designed and constructed with different levels of reliability and resilience to extreme hazards. It has been witnessed in numerous disaster events that built objects manifest different degree of damage, ranging from intact to collapse, even at locations that endure approximately the same level of hazard intensities. This further imply that given a complex disaster scene, for the decision making upon rapid recovery and monitoring of damaged structures, one immediate task is to rapidly identify and mark the damage levels of individual structures. With the advances in data collection tools today, e.g. digital cameras, LiDAR, smart phones, and their use through professional reconnaissance or crowdsourcing, collection of such perishable data become unprecedently easy today; however, the data understanding has become a bottle neck. Given such deluge of visual data collected after an extreme event, an autonomous and rapid understanding approach is demanded to understand the mechanics in disaster scenes. In this paper, we propose an artificial intelligence assisted approach through the use of deep learning, particularly a convolutional neural network-based semantic object method. This paper summarizes the data preparation, the training and validation of the deep learning framework, and evaluates quantitatively the performance of the predictive model for identifying both the causal types of disasters and the damage levels that they cause.