DD-UNet: A Deep Learning-Based Network for Building Extraction by Enhancing Multiscale Information Utilization

The rapid, precise, and automated extraction of buildings from high-resolution remote sensing imagery is crucial for intelligent remote sensing information retrieval, particularly in applications such as urban planning. However, existing hybrid convolutional neural network (CNN)-Transformer methods usually emphasize feature fusion in the encoder or decoder, while the skip connections that transfer multilevel spatial details are still insufficiently exploited. To address this limitation, this study proposes an innovative deep learning model, namely, the dual attention transformer (DAT) and decoder-guided recalibration attention (DRA) U-shaped network (DD-UNet). Building upon the conventional UNet architecture, DD-UNet integrates transformer-based global attention mechanisms into the skip connection stages, aiming to enhance the model's capability to effectively utilize multiscale information. The DD-UNet comprises three core components: the conv-encoder, the DAT skip connection, and the DRA decoder. First, the conv-encoder extracts hierarchical representations through convolutional operations. Second, the DAT module within the DAT skip connection processes spatial and channel information from the encoder-derived feature maps to generate new feature representations. Finally, the DRA module in the DRA decoder fuses the novel feature maps from the encoder and skip connection to eliminate semantic discrepancies. Extensive experiments are conducted to compare DD-UNet with existing high-performing networks across three public datasets. The results demonstrate that DD-UNet consistently outperforms existing approaches across all evaluation metrics. The superior performance of DD-UNet highlights the effectiveness of the proposed method in enhancing multiscale information processing, especially for accurate building segmentation under complex environmental conditions.