Change detection with incorporating multi-constraints and loss weights

The past decades have seen significant progress in change detection for remote sensing images, particularly in urban change analysis and land management. The current mainstream change detection models mainly adopt the network structures of the Combination of Siamese networks and UNet. However, as the previous research mentioned, there are still several challenges to be tackled. On the one hand, A higher number of false positive samples than is the case in false negative samples, in other words, false pixel always occurs in the background region. On the other hand, speckle noise is also a serious problem to be resolved. To address these limitations, our work proposes a novel method, named the Change Detection model with incorporating Multi-Constraints and loss weights (CDMC). We introduce the similarity constraint and the boundary complementation information constraint into SUNet for the first time and design the adaptive dual focus loss module and adaptive weighted loss module. Similarity constraint is composed of two multilayer perceptrons that guide background consistency in deep features and suppress changes that are not of interest, such as environmental factors i.e., location and lighting. The boundary complementary information constraint is realized by adding boundary classes, which enhances the model's attention to the position with strong uncertainty around the target of interest. The adaptive dual focus loss module uses trainable parameters to set weights for different classes, which effectively prevents the model from paying too much attention to the background class and alleviates the class imbalance. Adaptive weighted loss module weights the importance of similarity constraint and the boundary complementation information constraint. The devised approach was assessed with five mainstream advanced change detection approaches on four open-source datasets, including BCDD, LEVIR-CD, DSIFN, and S2Looking. The experimental results depicted that the proposed CDMC can achieve a state-of-the-art accuracy, convergence speed and stability. Through analyzing statistical results and visualization results, the proposed CDMC reduces the number of false positive samples and the uncertainty around the target. Additionally, the nonparametric test results of Friedman ranking also verify that CDMC is more optimal than the other five baseline models.