Building Damage Analysis on Post Hurricane Satellite Image by Fusing Segmentation Features Using Deep Learning

Using satellite images to evaluate building damage after a hurricane has a positive influence on emergency services in disaster response. Convolutional Neural Network has artificial neurons that respond to surrounding units within certain coverage areas thus has excellent performance for large-scale image processing and binary classification. In this paper, it is demonstrated that using a pre-trained model (U-Net) and a relatively shallow convolutional neural network achieves the state-of-the-art classification performance. Moreover, this paper did interpretation (guide Grad-CAM) to figure out the most decisive part for the proposed model to specify an output. As a result, this paper attained 98.45% and 99.18% accuracy in the balanced and unbalanced testing sets, respectively, and F1 scores of 98.43% and 96.27%. Experiment results also indicated that models without first stage feature extraction mainly focused on building envelope while those with U-Net paid attention to more essential features.