Semisupervised classification of hurricane damage from postevent aerial imagery using deep learning

Aerial images can greatly facilitate rescue efforts and recovery in the aftermath of hurricane disasters. Although supervised classification methods have been successfully applied to aerial imaging for building damage evaluation, their use remains challenging since supervised classifiers have to be trained using a large number of labeled samples, which are not available soon after disasters. However, rapid response is crucial for rescue tasks, which places greater demands on classification methods. To accelerate their deployment, a semisupervised classification method is proposed in this paper using a large number of unlabeled samples and only a few labeled samples that could be rapidly obtained. The proposed approach consists of three steps: segmentation, unsupervised pretraining using convolutional autoencoders (CAE), and supervised fine-tuning using convolutional neural networks (CNN). Leveraging the representation capability of CAE, the learned knowledge from CAE could be transferred to the counterparts of CNN. After pretraining, the CNN classifier is further refined with a few labeled samples to improve feature discrimination. To demonstrate this methodology, a recognition strategy of damaged buildings based on context information using only vertical postevent aerial two-dimensional images is presented in this paper. As a case study, a coastal area affected by the 2012 Sandy hurricane is investigated. Experimental results show that the proposed semisupervised method produces an overall accuracy of 88.3% and obtains an improvement of up to 9% against a CNN classifier trained from scratch. (C) 2018 Society of Photo Optical Instrumentation Engineers (SPIE)