融合通道—空间注意力与组合损失函数的改进U-Net遥感地物分割方法OA

Ground object identification is a key prerequisite for designing seismic acquisition point layouts.The Sichuan Basin in China is characterized by complex terrain and diverse ground object types,and conventional general-purpose deep-learning models often suffer from insufficient adaptability,low boundary-extraction accu-racy,and poor robustness in the segmentation of ground objects from remote-sensing images.To address these limitations,an improved U-Net model integrating a joint channel-spatial attention mechanism and a combined loss function is proposed.First,a standardized system for annotating and managing ground object samples is es-tablished to ensure that the training data are consistent and representative.Second,a dual-path attention module is introduced to enhance the selection and representation of core ground-object features,thus directing the model's attention toward important regions.Meanwhile,the combined loss function is reformulated to effec-tively mitigate the bias caused by class imbalance,shifting the segmentation objective from simple geometric contour fitting to the optimization of semantic consistency.Experimental results on a remote-sensing image test set from the Sichuan Basin show that the proposed model achieves an intersection over union(IoU)of 95.0%.Compared with the conventional U-Net model,the proposed model significantly improves segmentation accu-racy,boundary clarity,and processing efficiency,effectively addressing problems such as the merging of densely clustered buildings,blurred boundaries,and missed detection of minority-class ground objects.This study provides reliable technical support for the fine-grained extraction of ground objects in areas with complex surface conditions and can be widely applied to practical engineering scenarios such as seismic survey point-layout optimization and construction route planning.