DEEP LEARNING MODELS FOR NATURAL DISASTER DETECTION & ASSESSMENT

An automated damage assessment provides critical information for decision making and resource allocation for rapid emergency response. Satellite imagery provides real-time, high-coverage information and offer opportunities for large-scale post-disaster damage assessment. Studies have proposed deeplearning based methods for automatic disaster assessment from satellite imagery in order to avoid huge amount of labor and manual work by satellite imagery analysts. The aim of this paper is to investigate novel deep learning architectures that can assess and classify damage with enhanced accuracy using pre- and post- disaster satellite images. Further, the intensity of the damage is accessed using GRAD-CAM visualization on post disaster images. Furthermore, the size of the model is compressed using combination of pruning and quantization based neural network compression techniques to facilitate edge deployment in resource constraint environments such as drones and satellites. Experimental analysis on xBD dataset suggest efficacy of the proposed models in enhancing the performance over state-of-the-art methods for natural disaster assessment.