Change Detection Using Unsupervised Sensitivity Disparity Networks

At present, algebraic operation methods in the field of change detection still holds the dominant position. However, in the face of disturbance features, due to the characteristics of poor expansibility, the performance of algebraic operation methods varies greatly in different scenes, and cannot meet the requirements of practical application. In this paper we propose a change detection model based on Sensitivity Disparity Networks (SDNs) for performing change detection in Bi-temporal Hyper-spectral images captured by AVIRIS sensor and HYPERION sensor over time. The SNDs consist of two deep learning models, Unchanged Sensitivity Networks (USNet) and Changed Sensitivity Networks (CSNet), they have sensitivity disparity in changed and unchanged pixels, and thus to generate effective argument region. Next, we re-evaluate the change probability of argument region, and merge the change result of the argument region with that by one of the SDNs. The detected Binary Change Map (BCM) of the scheme is thus obtained. To train and evaluate the proposed schema we employ two Bi-temporal Hyper-spectral image datasets which contain challenging pseudo-changed features (PCFs) and pseudo-invariant features (PIFs) cause by various external interference factors. The proposed schema outperforms the existing state-of-the-art algorithms on tested datasets. Experimental results show that the proposed schema has good universality and adaptability.