Abstract: To break through the bottlenecks of high dependence on samples and multi-scale feature fusion in previous remote sensing semantic segmentation,this paper proposes a remote sensing semantic segmentation method VF-UNet++,which integrates the visual foundation model SAM with the U-Net++ architecture.This method innovatively constructs a dual-stream feature collaborative modeling framework and designs a dual-stream feature interactive fusion mechanism to achieve cross-scale feature fusion.Meanwhile,aiming at the characteristics of remote sensing scenarios,it proposes an adaptation strategy of parameter freezing and domain knowledge injection,which effectively improves the generalization ability of the model under the condition of limited samples.Experiments based on the Inria Aerial Image Labeling Dataset demonstrate that VF-UNet++ outperforms the comparison models in metrics such as recall,F1-score,and mIoU.This method effectively tackles the challenges in transferring and adapting visual foundation models to the remote sensing domain,offering a reference paradigm for intelligent remote sensing interpretation under low-sample conditions.Additionally,it overcomes the limitation of inadequate fusion between global semantic features and local detailed features,achieving a dual enhancement in segmentation accuracy and model robustness within complex remote sensing scenarios.

Key words: remote sensing semantic segmentation, dual-stream feature collaborative modeling, domain adaptation optimization, channel attention mechanism