Refined deep features for unsupervised change detection in high resolution remote sensing images

With the rapid development of remote sensing technologies, a growing number of high-resolution remote sensing images are providing more opportunities for detailed earth surface observation such as land-use classification, objects detection and land-cover change detection. Deep learning-based methods have been extensively studied for change detection in high-resolution images. However, most of the existed deep learning-based methods are in a supervised manner which needs a large collection of ground truth images, which are not available in many application scenarios of which lack ground truth images. In this paper, we conduct an experiment using a refined deep feature-based method for unsupervised change detection in high-resolution remote sensing images and evaluate its effectiveness. Firstly, deep features of two bi-temporal images are extracted by a pre-trained VGG16 network. Then the deep change features are refined by variance ranking-based method to retain relevant features while discard irrelevant features. Finally, the Otsu segmentation method is applied on the refined deep change features to produce the final change maps.