Detecting Building Changes Using Multimodal Siamese Multitask Networks From Very-High-Resolution Satellite Images.

Two main issues are faced when using very-high-spatial-resolution (VHR) satellite images for building change detection: 1) the boundaries of detected changes are hard to be consistent with the ground truth and 2) detected changes are easily affected by different viewing angles of bitemporal images, leading to noticeable false changes. To deal with these issues, this study develops a new Siamese change detection network [i.e., Siamese multitask change detection network (SMCD-Net)] based on a multitask learning framework to improve building change detection, particularly in the geometric aspect. Boundary information is formulated as an auxiliary task to constrain the learning of high-level semantic features. To enhance the identification of real changes from false changes, we model the directional relationships between buildings and their shadows by fuzzy sets, and incorporate the relationship information into SMCD-Net, leading to a network variant, labeled as SMCD-Net-m. Experiments were conducted on three datasets: a publicly available dataset, a Chinese GaoFen-2 dataset, and a French Pleiades dataset. We compared our methods with seven other methods, i.e., object-based Siamese network, ChangeStar, ChangeFormer, BIT, STANet, FC-Siam-diff, and Siam-NestedUNet. Results showed that the proposed SMCD-Net obtained the best detection results, achieving the lowest global total errors on all datasets. By incorporating directional information, SMCD-Net-m evidently improved detection accuracy, particularly when using bitemporal images with a large viewing angle difference. The improvement was positively correlated with the accuracy of building shadows extracted from VHR images.