Multi-scale Features Fusion Network for Unsupervised Change Detection in Heterogeneous Optical and SAR Images

Change detection (CD) in heterogeneous remote sensing image applications has become an issue of increasing concern in, as they cannot be compared directly with traditional homogenous CD methods. To solve feature loss problem and generating better representations to accommodate regions of various sizes in heterogeneous images CD, a multi-scale features fusion network (MFFN) is proposed. Firstly, multi-scale representative deep features can be extracted to distinguish difference in high-dimension feature space. Then, hierarchical features from the original image pairs can be fuse to generate a difference image with more explicit semantic information owing to the strategy of multi-scale features fusion, which can better adapt different scale of changes in heterogeneous remote sensing images. It is noteworthy that the experimental results on both heterogeneous and homogeneous data set confirm the effectiveness of the proposed method.