利用全卷积网络提取Sentinel-2影像高低层建筑区

面向Sentinel-2多光谱数据，依据影像地物空间结构和地表实际功能类型将建筑区分为高层建筑区和低层建筑区，构建了一种基于全卷积网络模型的高、低层建筑区快速提取技术。在此基础上，以雄安新区及其周边为试验区，选取2017年3月获取的4景Sentinel-2多光谱数据进行试验验证和分析。结果表明：本文技术能够实现大范围区域内高层和低层建筑区的快速提取，总体平均提取精度达到95.30%，其中高层建筑区平均提取精度为99.22%，低层建筑区平均提取精度为91.38%，该技术明显优于现有基于纹理结构的高低层建筑区提取方法。通过对提取结果进行统计分析发现：约4.4×10⁴ km²的研究区内高层和低层建筑区分别约为94和7351 km²；雄安新区中心三县内高层和低层建筑区分别约为1.25和312.24 km²。本文技术具有很好的推广性，结合Sentinel-2数据大幅宽高频次观测特点和更多类型建筑区样本，可以实现大范围地表多类型建筑区动态监测。

Extraction of high-rise and low-rise building areas from Sentinel-2 data based on full convolution networks

This paper proposes a fully convolutional networks based method to intelligently exploit Sentinel-2 data for high-rise and low-rise building areas extraction. The building areas are divided into high-rise building areas and low-rise building areas according to their spatial structure in Sentinel-2 data and their real functional types. Four Sentinel-2 data covering the Xiong'an New Area and its surroundings in early 2017 is selected for experimental verification and analysis. The results show that the proposed algorithm can extract the high-rise and low-rise building areas from Sentinel-2 data in an effective and efficient manner. Overall accuracy of the two types of building areas is about 95.30%, of which the high-rise building areas accuracy is about 99.22%, and the accuracy of other building areas is about 91.38%. Compared with the existing texture-based method, the proposed method is more robust and fast. The high-rise and low-rise building areas in the study area covering about 44 000 km² are about 94 and 7351 km², respectively. The high-rise and low-rise building areas in the three core counties of Xiong'an New Area are about 1.25 and 312.24 km², respectively. Our method can be easily extend to extract finer types of building areas given proper training samples of finer types of buildings areas and can realize large scale dynamic monitoring with the large imaging width and the high-frequency observation of Sentinel-2 data.