Remote Sensing Image Building Segmentation Integrating Detail and Global Information

A novel architectural segmentation network model for remote sensing imagery, integrating local detail enhancement and global information processing, is proposed to address challenges such as segmentation confusion of color-similar objects, adhesion between buildings and grasslands, and the difficulty in distinguishing buildings from roads in urban environments. Initially, a local detail enhancement module is employed to enhance the model’s capability in recognizing features of buildings and achieving precise detail segmentation in remote sensing images. Subsequently, a global information fusion module is utilized to integrate various features, thereby augmenting the overall segmentation performance of the model. Finally, the incorporation of skip connections and multi-convolutional attention modules enables the model to not only improve the accuracy in capturing details but also effectively utilize the global contextual information of the images. Testing results on the WHU and custom remote sensing image datasets demonstrate that the model achieves an accuracy of 0.9845, a recall of 0.9419, an F1 score of 0.9486, and a mean intersection-over-union of 0.9420. The model effectively resolves segmentation challenges caused by factors such as roads, trees, and shadows, enabling clearer and more accurate delineation of building boundaries.