基于LightGBM和多光谱—SAR多特征综合的中亚干旱区城市不透水面提取  

LightGBM based Impervious Surface Area Extraction of Cities from Arid Areas in Central Asia Using Synthesized Multi-features of Multispectral-SAR Images

作　　者：刘希鸣 阿里木·赛买提[1,2,3] 王伟 吉力力·阿不都外力 LIU Ximing;SAMAT Alim;WANG Wei;ABUDUWAILI Jilili(State Key Laboratory of Desert and Oasis Ecology,Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,Urumqi 830011,China;University of Chinese Academy of Sciences,Beijing 100049,China;Research Center for Ecology and Environment of Central Asia,Chinese Academy of Sciences,Urumqi 830011,China)

机构地区：[1]中国科学院新疆生态与地理研究所荒漠与绿洲生态国家重点实验室,新疆乌鲁木齐830011 [2]中国科学院大学,北京100049 [3]中国科学院中亚生态与环境研究中心,新疆乌鲁木齐830011

出　　处：《遥感技术与应用》2023年第2期319-331,共13页Remote Sensing Technology and Application

基　　金：国家自然科学基金面上项目“样本与特征迁移的中亚典型城市覆被精细分类方法研究”(42071424);中国科学院青年创新促进会项目(2018476)共同资助。

摘　　要：不透水面是评价城市化水平和城市生态环境的重要指标,是近年来城市遥感研究中的热点方向之一。与湿润、半湿润区相比,干旱区城市植被覆盖度较低,不透水面与裸土、荒漠之间相似的光谱特征导致传统基于光学影像的亚像元分解法与光谱指数法在干旱区不透水面提取的适用性降低。针对该问题,提出一种多光谱与合成孔径雷达(SAR)影像多特征综合的方法以增大不透水面与其他地物覆盖类型之间的特征差异,从而提取干旱区城市不透水面。以阿斯塔纳、塔什干和杜尚别3个中亚城市为研究区,哨兵2号和哨兵1号影像为数据源,通过LightGBM算法对多光谱和SAR图像的空间特征、SAR的极化特征进行分类并提取不透水面。研究对比了不同特征组合以及不同分类方法的不透水面提取结果,实验结果表明:多光谱与SAR影像多特征综合的方法能有效提高干旱区不透水面提取精度,明显改善干旱区其他土地覆盖类型错分为不透水面的问题;LightGBM算法与XGBoost、HistGBT等基于梯度提升决策树的算法和随机森林等方法相比能获取更高的精度,更适用于干旱区不透水面提取。这表明基于LightGBM以及多光谱和SAR多特征联合的方法能够有效提取中亚干旱区城市不透水面。Impervious surface is an important factor indicates the level of urbanization and the urban ecological environment,and it is one of the current research hotspots in urban remote sensing.Compared with humid and semi-humid areas,urban vegetation coverage in arid areas is relatively low,the similar spectum between impervious surface and barren area makes the traditional optical image-based spectral mixing analysis method and spectral index method not suitable for the impervious surface extraction of cities in arid areas.In response to this problem,a method for impervious surfaces extraction of cities from arid areas in Central Asia using synthesized multi-features of multispectral-SAR images is proposed to improve the mixclassifiation between impervious surfaces and bare soil,so as to extract impervious surface in arid area.In detials,Sentinel-2 and the dual-polarization SAR image of Sentinel-1 are selected for three Central Asia cities,Astana,Tashkent and Dushanbe.The spatial characteristics of multi-spectral and SAR images,and the polarization characteristics of SAR are feeded to LightGBM algorithm to classify and extract impervious surface.This paper compares the impervious surface extraction results of different feature combinations and different classification methods.Experimental results indicated that the multi-feature synthesis method of multispectral and SAR images proposed can effectively improve the accuracy of impervious surface extraction in arid areas,indicating the improvement in the misclassification of impervious surface and other land cover types in arid areas;the LightGBM algorithm has higher accuracy than XGBoost,HistGBT and other algorithms based on gradient boosting decision trees and random forest algorithm,and it is more suitable for extraction of impervious surface in arid area.This shows that the method based on LightGBM and the combination of multispectral and SAR multi-features can effectively extract the urban im-pervious surface in the arid area of Central Asia.

关 键 词：不透水面 部分重建的形态学属性剖面 LightGBM 干旱区 中亚 

分 类 号：P237[天文地球—摄影测量与遥感] TU984[天文地球—测绘科学与技术]