一种基于光谱归一化的丘陵地区植被覆盖度反演方法  

An Inversion Method of Retrieval of Vegetation Coverage in a Hilly Area Based on Spectral Normalization

作　　者：胡铁泷 段利 蒋良群[1] 王杰[1] HU Tieshuang;DUAN Li;JIANG Liangqun;WANG Jie(College of Land and Resources,China West Normal University,Nanchong 637009,Sichuan China)

机构地区：[1]西华师范大学国土资源学院,四川南充637009

出　　处：《山地学报》2019年第5期778-786,共9页Mountain Research

基　　金：四川省教育厅自然科学重点项目(17ZA0387,15ZA0150);中国科学院战略性先导科技专项(A类)(XDA19040504);南充市应用技术研究与开发专项项目(17YFZJ0014);西华师范大学英才基金项目(17YC124)。

摘　　要：由于影像空间分辨率的限制,利用遥感影像反演植被覆盖度时,像元内通常存在植被与其他地物混合的现象.此外,受到物理属性、地形、阴影等因素的影响,植被内部存在较大的光谱差异.混合像元的存在,以及植被内部光谱变化较大都将导致植被覆盖度反演精度降低.本研究基于Sentinel 2A遥感影像,提出了一种基于光谱归一化的光谱混合分析方法,以期解决植被内的光谱差异以及与其他地物的混合问题.首先,对端元矩阵与遥感影像进行归一化预处理,以减弱植被内的光谱变化;然后,采用全约束最小二乘法(FCLSU)、部分约束最小二乘法(CLSU)、扩展线性混合模型(ELMM)三种混合像元分解算法来定量分析植被与其他地物的混合状态.在验证解混算法精度时,采用无人机高分影像分类结果作为植被覆盖度参考影像,并对归一化前后的精度进行对比.光谱归一化前,ELMM和CLSU的R和RMSE都接近0.903和0.353,FCLSU的R和RMSE为0.869和0.434.光谱归一化后,三种算法的R和RMSE都接近0.91和0.2.试验结果表明:端元和影像进行归一化后,降低了光谱变异性,三种算法的解混精度在整体上提高较大,且对四川丘陵地区的植被覆盖度的反演结果接近真实值.Because of the fixed spatial resolution in remote sensing images,there is usually a technical issue of decomposing mixed pixels,which comprise both vegetation element and other endmember fractions,degrading interpretation of vegetation coverage by image inversion analysis.By applications to Sentinel 2 A remote sensing images,in this study it proposed a spectrum mixture analysis method based on spectral normalization in order to compromise the intra-class spectrum variability of vegetation and pixel mixing with other ground objects.Firstly,endmember matrix and remote sensing images were normalized to reduce spectrum variability of vegetation.Then,it applied three unmixed algorithms,full constrained least squares unmixing(FCLSU),partial constrained least squares unmixing(CLSU)and extended linear mixed model(ELMM)to quantitatively analyze the mixed state of vegetation with other ground objects.To verify the accuracy of our proposed unmixing algorithm,the classification of vegetation coverage obtained by UAV high-resolution images was referred as control group,and their accuracy before and after the normalization was carefully compared for accuracy evaluation.It found that before spectrum were normalized,the R and RMSE of ELMM and CLSU were both close to 0.903 and 0.353,and the two values from FCLSU were 0.869 and 0.434,whereas after the normalization,the R and RMSE of the three algorithms were updated to 0.91 and 0.2.The study suggested that after normalization of endmember and images the spectrum variability was considerably reduced by the three algorithms,with a great improvement in unmixing accuracy,and the inversion results of vegetation coverage in Sichuan hilly region was verified by field survey.ELMM and FCLSU had the highest accuracy,and ELMM was the best for its promotion in operational efficiency.This research would provide a substantial reference for pixel unmixing technology in the application of remote sense science to practices.

关 键 词：植被覆盖度 Sentinel2A遥感影像 端元变化 光谱归一化 光谱混合分析 

分 类 号：TP791[自动化与计算机技术—检测技术与自动化装置]