【摘要】 针对传统分类方法在高分遥感影像城镇绿地提取效果不理想的问题，提出了一种改进的语义分割模型U-Net来更加高效精准地提取城镇绿地区域。使用高分二号影像制作样本数据集，同时对U-Net网络模型改进，采用不同深度的ResNet作为其主干网络提取图像的语义信息，另外加入了注意力机制模块，细化提取的特征图，提高网络的分类性能。实验结果表明：对比经典语义分割网络SegNet、PSPNet、U-Net,加入注意力机制Res-UNet在预测效果和评价指标均有提升，表现最好的是Res152-UNet,其PA值为90.53,MIoU值为80.06,预测效果图接近人工标注。改进U-Net模型能够高效地对遥感影像信息进行识别提取，得到高精度的提取结果，该方法对于高分遥感影像城镇绿地提取具有一定应用意义。更多还原

【Abstract】 In view of the traditional classification method in high resolution remote sensing images, a modified semantic segmentation model U-Net is proposed to extract urban green space more efficiently and accurately. First, GF-2 image is used to make sample data sets, and the U-Net network model is improved. ResNet with different depths is used as the backbone network to extract the semantic information of the images. In addition, the attention mechanism module is added to refine the extracted feature map and improve the classification performance of the network. The experimental results show that: compared with the classical semantic segmentation network SegNet, PSPNet, and U-Net, Res-UNet adding the attention mechanism improves the prediction effect and evaluation index. The best performance is Res152-UNet, with PA value 90.53,the MIoU value 80.06,and the prediction effect map is close to the manual annotation. The improved U-Net model can efficiently identify and extract the remote sensing image information, and obtain the high-precision extraction results. This method has certain application significance for the urban green space extraction of high resolution remote sensing images.更多还原