A two-stage task mechanism network for building change detection method in remote sensing images

Building change detection (CD) based on deep learning is an important means of remote sensing building change detection. However, this method couples feature extraction and feature fusion comparison into a whole, which weakens the feature extraction ability for remotely sensed data under different time phases, resulting in a large number of errors and omissions in building CD. To address these problems, we propose a two-stage task mechanism network (TSTM-Net) for building CD. This method divides the networks into two stages of tasks. In the first stage, the bottom-level features from different temporal phases are directly connected for feature fusion and comparison, guiding the learning of features across different temporal phases and obtaining initial features. In the second stage, features of different scales in different temporal phases are compared and learned through multiscale residual fusion to achieve feature optimization and building change information mining in different temporal phases and complete the building change information detection. In addition, to better complete network training, we introduced Bayesian weighted loss functions in the two-stage task mechanism to automatically adjust loss function weights for different tasks. Through comparative analysis with different methods, such as Changer and MDFFNet, in the WHU-CD and GZ-CD datasets, we see that compared to other methods, our method can successfully detect more complete building change information and reduce omissions of small-sized building changes.