沈晶灵　杨晓竹

摘 要：针对传统监督分类对高空间分辨率遥感影像中建筑物信息提取精度较低的问题，改进SegNet模型，利用U-Net模型中的跳层连接结构补充SegNet模型中解码器层的目标细节，加入空洞空间金字塔池化模块增强网络对多尺度目标的捕捉能力。利用改进后的SegNet、全卷积神经网络、支持向量机和最大似然法，对遥感影像中建筑物的提取结果进行对比分析。以法国国家信息与自动化研究所航空图像标记数据集为数据源，对分类结果进行定性和定量分析。在有限的迭代次数和实验区域内，改进后的SegNet的Kappa系数在80%以上，总体精度超过90%，在边缘细节的分类效果更精细，改进后的SegNet对遥感图像中建筑物的提取效果更好、精度更高，具有可行性和有效性。

关键词：SegNet;高分影像;建筑物提取;ASPP模块

中图分类号：P237 文献标识码：A 文章编号：1003-5168（2021）25-0006-05

High-Scoring Image Building Extraction Based on the Improved SegNet

SHEN Jingling YANG Xiaozhu

（College of Surveying and Mapping and Geographic Science， Liaoning University of Engineering and Technology，

Fuxin Liaoning 123000）

Abstract： In view of the low accuracy of building information extraction in high spatial resolution remote sensing images by traditional supervision classification， this paper improves the SegNet model， used the jumper connection structure in the U-Net model to supplement the target details of the decoder layer in the SegNet model， and adds ASPP module to enhance the network's ability to capture multi-scale targets.The results of the extraction of buildings in remote sensing imagery are compared using the improved SegNet， Full Convolutional Neural Networks， Support Vector Machine， and Maximum Likelihood Classification.Using the INRIA aerial image marker data set as the data source， through qualitative and quantitative analysis of the classification results， the improved SegNet Kappa coefficient is more than 80% in the limited iteration and experimental area， the overall accuracy is more than 90%， and its classification effect in the edge details is more fine.It is shown that the improved SegNet is more effective， more accurate and feasible for the extraction of buildings in remote sensing images.

Keywords： SegNet;high-scoring imagery;building extraction;ASPP module

高分辨率遙感图像的特征识别和检测是当前科学领域的研究热点[1]，其提取的信息内容丰富，包括地物信息与纹理信息[2]。近年来，大量学者对高分辨率遥感影像中的地物信息提取等问题进行了研究，并提出了K均值[3]、迭代自组织的数据分析法（Iterative Self-Organizing Data Analysis Method，ISODATA）[4]、面向对象的分类方法[5]、最大似然法[6]、最小距离法[7]以及支持向量机[8-9]等方法。然而，在空间分辨率较高的遥感图像中，由于这些方法依赖于光谱特征，导致提取精度较低。随着分辨率的提高，图像中的噪声信息越来越突出。异物同谱和同物异谱等现象给影像质量带来的问题日益突出，给在高分辨率遥感图像中进行目标精确提取造成了很大影响[2]。……