Detecting three-dimensional changes in buildings plays a vital role in urban monitoring, sustainable development, and disaster management. This research presents an innovative method for detecting multi-class building changes. The proposed approach combines two advanced architectures—a Siamese transformer network and spatial-channel attention mechanisms—to enable automated, precise detection of structural changes. The primary innovation lies in the design of a dual-attention encoder block that simultaneously analyzes local spatial relationships and channel dependencies to identify changes.
A key challenge in change detection is the class imbalance (e.g., unchanged buildings, new constructions, and demolished structures). To address this, methods such as enhanced data augmentation and overlapping patch extraction during preprocessing have been employed. The proposed method was implemented on a stereo dataset from the GeoEye-1 satellite (0.5-meter resolution) and an aerial stereo dataset (0.08-meter resolution).
In experiments, the proposed method achieved Kappa coefficients of 94% and 93% for the satellite and aerial datasets, respectively. This marks a significant improvement ( 3% increase in Kappa coefficient) compared to state-of-the-art methods like ChangeFormer, which achieved 91% for both datasets. By enhancing feature extraction and performing robustly on diverse data, the model emerges as a powerful tool for urban environment monitoring, offering a scalable and reliable solution for urban planning and management.