Improving damage classification via hybrid deep learning feature representations derived from post-earthquake aerial images

ABSTRACT One of the crucial problems after earthquakes is how to quickly and accurately detect and identify damaged areas. Several automated methods have been developed to analyse remote sensing (RS) images for earthquake damage classification. The performance of damage classification is mainly depending on powerful learning feature representations. Though the hand-crafted features can achieve satisfactory performance to some extent, the performance gain is small and does not generalise well. Recently, the convolutional neural network (CNN) has demonstrated its capability of deriving more powerful feature representations than hand-crafted features in many domains. Our main contribution in this paper is the investigation of hybrid feature representations derived from several pre-trained CNN models for earthquake damage classification. Also, in this study, in contrast to previous works, we explore the combination of feature representations extracted from the last two fully connected layers of a particular CNN model. We validated our proposals on two large datasets, including images highly varying in scene characteristics, lighting conditions, and image characteristics, captured from different earthquake events and several geographic locations. Extensive experiments showed that our proposals can improve significantly the performance.