A feature enhancement network combining UNet and vision transformer for building change detection in high-resolution remote sensing images

Building change detection (CD) is significant for understanding ground changes and human activities. Deep learning has become the mainstream approach for building CD. However, the detection accuracy remains insufficient due to limitations in feature extraction. Therefore, this paper proposes a feature enhancement network, FENET-UEVTS, to improve the accuracy of building detection, which combines a UNet encoder and a vision transformer structure to detect building changes. It can enhance the ability to detect irregular buildings and distinguish changes between adjacent buildings in different locations. The model combines a deep convolutional network with a part of vision transformer structure, which has a robust feature extraction ability for various types of buildings. We design a spatial-channel attention mechanism module (SCAM) that takes into account both the spatial and channel dimensions to enhance the detection ability of small-scale buildings. We also develop a u-shaped residual module (USRM) and a strengthened feature extraction module (SFEM) to improve the feature extraction capability for buildings with different shapes and edge details. A self-attention feature fusion module (SAFFM) is proposed to facilitate the full convergence and integration of different feature information. The SAFFM can better distinguish buildings of various shapes and sizes to prevent false detection and missed detection. To minimize information loss, a cross-channel context semantic aggregation module (CCSAM) is designed to perform information aggregation in the channel dimension. To evaluate the performance of our model, we conducted numerous experiments on three CD datasets. The results demonstrate that our proposed model outperforms eight other state-of-the-art (SOTA) algorithms in F1-score, overall accuracy, and KAPPA coefficient, achieving up to 91.83 %, 87.65 %, and 93.29 % F1-score on three widely used public datasets, i.e., LEVIR-CD, WHU-CD, and CDD dataset.