基于SDG-Segformer的高分辨率遥感影像城市地物提取与制图:以成都市为例OA

To address the demand for automatic and fine mapping of urban ground objects from high-resolution remote sensing images,and solve the problems of jagged edges,fractures of ground object boundaries and loss of fine target features when the Segformer model processes such images,this study takes Chengdu City as the research object,constructs a high-resolution semantic segmentation dataset covering four typical ground object types(green space,bare land,building and water body),and carries out large-scale extraction experiments based on Google Earth images acquired in the summer of 2020.To address the problem that the Segformer model weakens high-frequency boundary information due to downsampling,which in turn leads to step and jagged artifacts during upsampling recovery,this paper proposes the SDG-Segformer model.In the multi-scale fusion and decoding stages,the model introduces a lightweight edge enhancement framework of"multi-scale feature fusion-spatial detail self-guidance-content-aware upsampling",which improves the continuity and geometric consistency of ground object boundaries.To verify the effectiveness of the proposed method,comparative evaluations were conducted with Mask2former,DeepLabV3+,Swin-Transformer,and the original Segformer under the unified training and pre-training initialization conditions.The results show that the SDG-Segformer achieves excellent performance in the extraction of buildings(F1=94.11%)and water bodies(F1=95.04%),confirming the robustness and engineering applicability of the framework in complex urban landscapes and providing high-precision data support for urban planning,map updating,and resource and environmental monitoring.