Unsupervised change detection of satellite images using low rank matrix completion

Traditional unsupervised change detection methods need to generate a difference image (DI) for subsequent processing to produce a binary change map. In addition, few methods explore global structures. This Letter presents a novel unsupervised change detection approach based on low rank matrix completion. Other than generating a DI, the changed pixels are modeled as the estimated missing values for matrix completion, where the changed pixels are represented by a sparse term. A common low rank matrix is recovered by two temporal images. The changed pixels are separated out from the low rank matrix, in which the local information is introduced via graph cuts. The global and local structures are utilized in our model. Experimental results validate the effectiveness of the proposed approach. The proposed method is a new view for change detection.