Faculty of Electronic Information and Electrical Engineering, Dalian University of Technology
National Key Research and Development Program of China(2016YFC0400903), Fundamental Research Funds for the Central Universities(DUT20LAB114，DUT2018TB06)
Remote sensing change detection plays an important role in the supervision and management of land resource utilization. A change detection approach based on collaborative clustering and weighted-attention sparse autoencoder is proposed, which aims at the disadvantage of human intervention in training samples in supervision change detection, the problem of unbalanced data and the phenomenon of "salt and pepper" in change detection based on pixel-level. Fuzzy c-means and K-means are adopted to cluster difference image for training data and data to be classified. Meanwhile, the gray level co-occurrence matrix feature is considered in the samples, and the synthetic minority oversampling technique is utilized to expand the changed samples for solving issues of sample imbalance. Through the layer-wise weight-attention module that enhances the learning of positive weights and weakens the impact of negative weights, the classification performance of autoencoder is improved, and classification results of which in the mapping space of the superpixel segmentation boundary of the difference image eliminate "salt and pepper" noises for generation of change detection map according to the specific constraints. Compared with the state-of-the-art autoencoders in the standard handwritten data set, the proposed autoencoder shows better classification performance and stability. The change detection approach achieves the balance of missing detection and false detection, which increases the accuracy of change detection and reduces the human intervention at the same time.