Abstract: To solve the problem that the process of traditional remote sensing image scene change detection is complex and severely depends on the classification performance, a scene change detection method considering both global and local scene similarity is proposed in this paper. Firstly, the two temporal remote sensing images are cropped into image patches with fixed size to construct scene pair database, and divided into training and testing dataset. Secondly, the two-branch convolutional neural network integrating scene global and local similarity is constructed to carry out similarity learning of scene pairs. Then, the trained network is used to extract the similarity of scene pairs of the training dataset, and the best similarity threshold is determined via traversal. Finally, the testing dataset is categorized into changed and unchanged scenes based on the similarity threshold to obtain the final change detection results. The experimental results show that the overall accuracy and Kappa of the proposed method is 0.94 and 0.88, respectively, which outperforms traditional detection after classification approach and is a simple yet effective change detection method.

Key words: remote sensing image scene, change detection, feature similarity, convolutional neural network