面向高分辨率遥感图像分割的级联精炼网络OA

To address the issues of detail degradation,edge blurring,and structural target distortion aris-ing from repeated downsampling operations in Transformer-based models for high-resolution remote sens-ing image semantic segmentation,this paper proposed a cascaded refinement network for high-resolution remote sensing imagery.The network employed Swin-Transformer as the backbone encoder and sequen-tially constructed three core modules on the decoder side,namely the dual feature aggregation module,the dynamic wavelet fusion upsampling module,and the geometric structure perception module.The dual fea-ture aggregation module enhanced the semantic complementarity across multiple encoder layers through joint channel-spatial collaborative modeling.The dynamic wavelet fusion upsampling module performed frequency-domain reconstruction and fine-grained detail recovery via adaptive multi-wavelet basis fusion.The geometric structure perception module refined the terminal features through geometric calibration by integrating graph-structural priors with multi-scale spatial modeling,thereby improving boundary sharp-ness and target integrity.Extensive experiments on three public remote sensing benchmarks,including the ISPRS Potsdam dataset,the ISPRS Vaihingen dataset,and the LoveDA dataset,demonstrate that the proposed network achieves mean Intersection over Union scores of 87.13%,83.86%,and 54.26%,re-spectively.The method exhibits superior performance in delineating complex building contours,preserv-ing small-scale targets,and handling cross-scene segmentation tasks,which validates the effectiveness of the cascaded decoding refinement strategy in mitigating progressive feature degradation and enhancing the segmentation quality of remote sensing imagery.