Deep Difference Representation Learning for Multi-spectral Imagery Change Detection

Change detection is an ongoing hot topic in multi-spectral imagery applications, how to exploit the available spectral information effectively for change detection is still an open question. Considering the noise interference and redundancy of multi-spectral imagery, it is important and necessary to learn more abstract and robust feature from raw spectrums for change detection application. In this paper, a deep difference representation learning model is proposed for multi-spectral change detection. In this model, two stacked denoising autoencoders are established, one for learning more abstract features from raw spectrums blocks, and the other for learning difference representations from the stacked change feature. The former is used to weaken noise interference and reduce redundancy, while the latter has the ability to highlight changes and suppress unchanged pixels. The experimental results on real multi-spectral data demonstrate the feasibility, effectiveness and robustness of the proposed deep difference representation learning model on multi-spectral change detection task.