Enhancing Building Change Detection with UVT-BCD: A UNet-Vision Transformer Fusion Approach

Abstract Building change detection (BCD) is particularly important for comprehending ground changes and activities carried out by humans. Since its introduction, deep learning has emerged as the dominant method for BCD. Despite this, the detection accuracy continues to be inadequate because of the constraints imposed by feature extraction requirements. Consequently, the purpose of this study is to present a feature enhancement network that combines a UNet encoder and a vision transformer (UVT) structure in order to identify BCD (UVT-BCD). A deep convolutional network and a section of the vision transformer structure are combined in this model. The result is a strong feature extraction capability that can be used for a wide variety of building types. To improve the ability of small-scale structures to be detected, you should design an attention mechanism that takes into consideration both the spatial and channel dimensions. A cross-channel context semantic aggregation module is used to carry out information aggregation in the channel dimension. Experiments have been conducted in numerous cases using two different BCD datasets to evaluate the performance of the previously suggested model. The findings reveal that UVT-BCD outperforms existing approaches, achieving improvements of 5.95% in overall accuracy, 5.33% in per-class accuracy, and 8.28% in the Cohen's Kappa statistic for the LEVIR-CD dataset. Furthermore, it demonstrates enhancements of 6.05% and 6.4% in overall accuracy, 6.56% and 5.89% in per-class accuracy, and 6.71% and 6.23% in the Cohen's Kappa statistic for the WHU-CD dataset.