Recognition of Different Disastrous Events Using Convolutional Neural Network

Abstract

Natural disasters or crises include earthquakes, land sliding, floods, and typhoons causing great damage to manmade structures such as buildings. To efficiently manage natural disasters, it is important to develop an automatic disaster recognition system based on deep learning algorithms such as Convolutional Neural Network. This research aims to introduce the application of Deep learning algorithms in the Recognition of different disasters such as building collapse, burning buildings caused by earthquake and fire. In this work, a single deep convolution neural network (DRN) was implemented based on two main characteristics Dilated convolution in which convolution is applied on an input image using defined gaps to capture more contextual information fine details and Residual connection in which the input layer is not only connected to adjacent layer but maybe the summation of previous layers to reduce the problem of vanishing gradient. Besides DRN Capsule Network was also implemented on complex disastrous event classification problem which removes the drawback occurs in traditional Convolutional Network by considering the spatial hierarchies between simple and complex dataset. Dilated Residual Network is trained and tested on publically available datasets of disasters that are NWPU-RESICS45, BOWFire, and Satellite image of Hurricane Damage, and Accident Image Analysis dataset that achieved testing accuracy of 92.06%, 76%, 98.15%, and 93.16% respectively. As there is a lack of a single disastrous event dataset so images of disasters were collected using different software: Zip downloader, Image downloader uses corresponding keywords like collapse building, fire, road accidents, and 4K video downloader. Disastrous Event dataset consists of four classes having 10,000 images each. DRN and Capsule network was applied on Disastrous Event dataset and the result showed models achieved 95.67% and 94.7% testing accuracy. The results proved that the proposed methodologies are efficient enough and can be generalized for other disaster classification problems.