A Hybrid Model based on Fused Features for Detection of Natural Disasters from Satellite Images

Abstract

Earthquakes, floods, tsunami, and other natural disasters are appearing as worldwide threats because of their widespread destruction that results in thousands of human and economic losses. It is vital for first responders to know the root cause of damages in a region so that the emergency response activities can be planned accordingly and more effectively. We proposed a framework for the detection and recognition of natural disasters from satellite images. In this work, satellite images of six different types of disasters are considered, namely earthquake, volcanic eruption, flood, tsunami, and hurricane. The framework relies on the fusion of wavelet image scattering features and local binary pattern features for constructing the final feature vector. We also compared the accuracy of our framework with the existing state-of-the-art hand-crafted and machine learning models. Simulation results confirm that the proposed framework is able to recognize the type of natural disaster from the satellite images with an accuracy of 99.59% (Kappa coefficient 98.54% and F-Score 99.40%). The proposed approach results in less computational cost while achieving better accuracy compared to the deep convolutional neural network. We believe that the proposed model can be integrated with satellite imagery for locating the geographical regions affected by the multiple natural disaster events at the same time or at short intervals.