The change detection of Multi-temporal remote sensing images plays a significant role in geography national condition monitoring, such as deforestation detection, vegetation phenology change, disaster detection, military targets detection, aquatic monitoring and so on. Since the different spectral information can reveal and recognize the different features of various ground objects, exploiting spectral information to detect change area has become a key point in change detection domains.Various research methods have been proposed to resolve change detection problem by using spectral information. However, the existing change detection methods have three limitations: first, the classification accuracy post-classification method is not high enough. Second, the state-of-the-art methods failed to make full use of the image spectrum information. Third, change threshold is a key problem for change detection. To overcome these limitations, in this paper, researches have been done from three aspects: first, a semi-supervised Euclidean embedding dimensionality reduction method is exploited to the improve classification accuracy. Second, a semi-supervised distance metric learning method is exploited to make full use of the spectral information. Third, to solve the change threshold problem, a jointly dictionary learning method,which mainly uses reconstruction error to distinguish change or unchange area, is exploited to improve detection accuracy. Besides, the jointly dictionary learning method has also proposed a threshold selection strategy to reduce the dependence of manual intervention.In experiments, both hyperspectral images and multi-spectral images are exploited to test the three proposed methods. By comparing our methods with the state-of-art methods in details, research results show that all of the proposed methods receive good performance. |