DEEP NO LEARNING APPROACH FOR UNSUPERVISED CHANGE DETECTION IN HYPERSPECTRAL IMAGES

Abstract. Unsupervised deep transfer-learning based change detection (CD) methods require pre-trained feature extractor that can be used to extract semantic features from the target bi-temporal scene. However, it is difficult to obtain such feature extractors for hyperspectral images. Moreover, it is not trivial to reuse the models trained with the multispectral images for the hyperspectral images due to the significant difference in number of spectral bands. While hyperspectral images show large number of spectral bands, they generally show much less spatial complexity, thus reducing the requirement of large receptive fields of convolution filters. Recent works in the computer vision have shown that even untrained networks can yield remarkable result in different tasks like super-resolution and surface reconstruction. Motivated by this, we make a bold proposition that untrained deep model, initialized with some weight initialization strategy can be used to extract useful semantic features from bi-temporal hyperspectral images. Thus, we couple an untrained network with Deep Change Vector Analysis (DCVA), a popular method for unsupervised CD, to propose an unsupervised CD method for hyperspectral images. We conduct experiments on two hyperspectral CD data sets, and the results demonstrate advantages of the proposed unsupervised method over other competitors.