Multitemporal Hyperspectral Image Change Detection by Joint Affinity and Convolutional Neural Networks

To improve performance of change detection (CD) using multitemporal hyperspectral images (HSI), this paper designs a novel spectral-spatial feature descriptor based on joint affinity (JA) and proposes a deep learning framework embedded with it. JA preserves context information that can increase accuracy and robustness of the CD model. It completes the task of CD in three steps, dimensionality reduction which is optional, JA tensor construction and binary classification by convolutional neural networks (CNN). Thus, the proposed method is denoted by JA-CNN. Experiments on one set of real-world data and two sets of synthetic data show that JA-CNN outperforms several state-of-the-art methods for CD. In addition, it is more robust than its counterpart which does not consider spatial information when applied to severely corrupted HSIs.