Zoom out CNNs features for optical remote sensing change detection

In this paper, we propose a novel unsupervised optical remote sensing change detection (CD) based on pre-trained convolutional neural network (CNN) on ImageNet dataset and superpixel (SLIC) segmentation technique. The proposed approach can be divided into three steps. First, bi-temporal images are stacked, and Principal Component Analysis (PCA) is applied to extract three higher uncorrelated channels, which will be later segmented into superpixels. Second, we zoom out each region into three levels and fit them separately into a pre-trained CNN. Third, we extract features of different zooming levels that represent the same region (superpixel) and concatenate them. We compare the concatenated features to get the final change map. The experimental results demonstrate the efficacy of the proposed approach.