Semi-supervised change detection method for multi-temporal hyperspectral images

Change detection is one of the most important open topics for multi-temporal remote sensing technology to observe the earth. Recently, many methods are proposed to detect the land-cover change information by multi-temporal hyperspectral images. However, many existing traditional change detection methods failed to utilize the spectral information effectively. Hence the models are not robust enough for more widely applications with “noise” bands. In this case, a semi-supervised distance metric learning method is proposed to detect the change areas by abundant spectral information of hyperspectral image under the “noisy” condition. This paper focuses on semi-supervised change detection method, and proposes a new distance metric learning framework for change detection in “noisy” condition with three mainly contributions: (1) Distance metric learning is demonstrated to be an effective method for revealing the change information by high spectral features. (2) An evolution regular framework is utilized to handle change detection under a “noisy” condition without removing any noise bands, which is impacted by atmosphere (or water) and always removed manually in other literatures. (3) A semi-supervised Laplacian Regularized Metric Learning method is exploited to tackle the ill-posed sample problem, and large unlabeled data is exploited in our method. The proposed method is performed on two multi-temporal hyperspectral datasets. Experimental results show that the proposed method outperforms the state-of-the-art change detection methods under both “ideal” and “noisy” conditions.