Identification of undamaged buildings after the event of disaster using Deep Learning

Abstract

As direct response, or recovery and security operations, it is of paramount importance to establish precisely the location and assess the extent of damage to a building as quickly as possible after a tragic event. The automation of damage analysis may enhance the capability of administration to provide the help. For the same, convolutional neural-networks are being used by recent proposals to perform image classification of building damage depending on the amount and type of damage to be detected. Furthermore, the use of up/down-sampling images during CNN preparation helps in better damage recognition. However, a number of challenges has been observed in convolutional neural-networks-based methods such as multi-resolution images of damaged areas. Furthermore, recent convolutional neural-networks-based models are having very complex architecture which increases the requirement of computational power. Therefore, in this paper, a simple convolutional neural-network model has been presented which effectively identifies the damage and undamaged buildings after the natural disaster. The presented method has been compared with recent convolutional neural-network models. The experimental results shows that the simple convolutional neural-network outperforms the existing models with a 99.2% validation accuracy.