基于高分辨率遥感影像的建筑物提取方法研究

Building extraction method based on high resolution remote sensing images

为了提高基于高分辨率遥感影像的建筑物提取精度及边缘精细化程度,本文结合U-Net模型与SegNet模型的优势,提出了一种U-SegNet组合模型方法。该组合模型首先以SegNet模型为基础,通过有效融合高层次与低层次特征图实现建筑物边界的准确定位;其次,通过构建的训练样本库微调权重,提高网络高层次视觉特征的输出效果。将本文提出的方法应用于杭州市某区域影像数据的建筑物提取中,检验本文方法的有效性与优越性。试验结果表明,本文提出的方法可有效提取试验区建筑物,相较于目前主流的基于多特征提取方法与基于易康软件提取方法,该方法在召回率、准确率与精确率上分别提高了2.34%、7.22%与5.33%。

In order to improve the accuracy and edge refinement of building extraction based on highresolution remote sensing images,this paper proposes a combined USegNet model by fully utilizing the advantages of UNet model and SegNet model.Firstly,accurate positioning of building boundaries is achieved by this combined model by effectively integrating highlevel and lowlevel feature maps.Secondly,the output effect of highlevel visual features of the network is improved by constructing a training sample library to fine tune the weights.The method proposed in this paper is applied to the building extraction of image data in an area of Hangzhou City,and the effectiveness and superiority of the method are tested.The test results show that the method proposed in this paper can effectively extract buildings in the testing area.Compared with the current mainstream multifeature extraction methods and Yikang software extraction methods,the proposed method has an average improvement of 2.34%,7.22%,and 5.33%in recall,accuracy,and precision,respectively.