Weakly supervised building change detection based on DeepCut and Temporal Invariant

Building change detection for remote sensing images is crucial for urban development and illegal building management. Semantic segmentation networks (SSNs) are widely used for building change detection due to their multi-scale feature extraction capabilities but require large amounts of pixel-level labeled data, which is time-consuming to annotate. To reduce annotation time, pixel-level pseudo-labels using Graph Cuts for Interactive Foreground Extraction (GrabCut) can be generated by only selecting bounding boxes for changed building regions. However, GrabCut faces challenges in remote sensing images due to complex environments. Additionally, changes in seasons and lighting can alter building appearances, leading to "pseudo-changes" in detection results. Surrounding objects close to buildings can also interfere with feature fusion, making it difficult to accurately extract changed building features. To address these issues, this paper proposes DeepCut and the Temporal Invariant Change Detection Network (TICDN). DeepCut enhances Grab-Cut by incorporating deep features that provide higher-level semantic information, resulting in more accurate segmentation of building parts. TICDN introduces adversarial learning to reduce the impact of pseudo-changes caused by seasonal and lighting variations, ensuring robust building features. It also features a Correlation Fuse Module (CFM) for feature fusion, which adaptively calculates the correlation between pre- and post-change features. This module effectively eliminates the influence of irrelevant objects, such as roads and trees, and accurately extracts changed building features. Extensive comparative experiments and ablation studies on the WHU and LEVIR-CD datasets demonstrate the superior performance and broad applicability of these methods in building change detection. The code is available at https://github.com/YuchunHuang/Weakly-Supervised-Change-Detection.