ESSegNeXt: A Novel Building Extraction Framework Based on Generalized SamGeo and SegNeXt Models Using High-Resolution Remote Sensing Images

Accurate building extraction is crucial for effectively monitoring the spatiotemporal dynamics of urban development. While high-resolution remote sensing imagery provides essential data support, existing deep learning-based methods for building extraction commonly face challenges such as poor edge regularity and loss of details due to the diversity and complexity of buildings. To address these limitations, this study proposed the ESSegNeXt framework, which integrated an extended SamGeo model for coarse-grained building segmentation and utilized edge feature extraction for fine-grained dual supervision. Additionally, an edge-preserving loss function was introduced to enhance the boundary detection capability of the SegNeXt model using edge features extracted by SamGeo during cyclic validation. To evaluate the effectiveness and generalizability of the proposed framework, three sets of experiments were conducted using Jilin-1 satellite imagery covering Wuhan City, UAV imagery covering Nanjing City, and the Inria public dataset. The experimental results demonstrated that ESSegNeXt significantly improved performance across the three datasets. Specifically, for the Jilin-1 satellite imagery, the IoU, mFscore, and mIoU were increased by 1.16%, 0.77%, and 1.98%, respectively, while edge preservation was enhanced by 16.82%. On the UAV dataset, the IoU, mFscore, mIoU, and mean edge IoU were improved by 0.91%, 0.35%, 0.59%, and 12.52%, respectively. For the Inria dataset, the corresponding improvements were 0.35%, 0.22%, 0.37%, and 8.03%, respectively. The proposed ESSegNeXt framework effectively preserves building edge details, making it particularly suitable for building extraction in complex urban environments. It also offers reliable technical support for applications in smart city development, land resource monitoring, and related fields.