LAUSR

Abstract Owing to the rapid development of remote sensing technology, various types of data can be easily acquired at present. However, it has become an important but more challenging task for effectively highlighting changes occurring on the land surface from these available data. In this paper, we propose an iterative feature mapping network learning framework for identifying multiple changes with focus on multi-source images, which are often obtained from sensors with different imaging modalities. Firstly, high-level and robust feature representations are extracted from multi-source images via unsupervised feature learning. Then, on this basis, an iterative feature mapping network is established to transform these features into a common high-dimensional feature space. It aims to learn more discriminative features by shrinking the difference between the paired features of unchanged positions while enlarging that of changed ones. Note that the network parameters are learned by optimizing a well-designed objective function, and the whole learning process is fully unsupervised. Finally, based on a hierarchical tree for clustering analysis, all possible change classes can be detected accurately. In addition, the proposed framework is found to be also suitable for change detection in homogeneous images. The impressive experimental results obtained over different types of remote sensing images demonstrate the effectiveness and robustness of the proposed model.