洱海东岸海滨三种典型湿地植被光谱特征分析与识别建模  

Spectral Characteristic and Identification Modelling of Three Typical Wetland Vegetation Along the Seashore of the East Coast of the Erhai Lake

作　　者：李璇 甘淑[1,2] 袁希平 杨敏 龚伟圳 LI Xuan;GAN Shu;YUAN Xi-ping;YANG Min;GONG Wei-zhen(Faculty of Land Resources and Engineering,Kunming University of Science and Technology,Kunming 650093,China;Application Engineering Research Center of Spatial Information Surveying and Mapping Technology in Plateau and Mountainous Areas Set by Universities in Yunnan Province,Kunming 650093,China;West Yunnan University of Applied Sciences,Dali 67100,China;West Yunnan University of Applied Sciences,Key Laboratory of Mountain Real Scene Point Cloud Data Processing and Application for Universities,Dali 671006,China)

机构地区：[1]昆明理工大学国土资源工程学院,云南昆明650093 [2]云南省高校高原山区信息测绘技术应用工程研究中心,云南昆明650093 [3]滇西应用技术大学,云南大理671009 [4]滇西应用技术大学/云南省高校山地实景点云数据处理及应用重点实验室,云南大理671006

出　　处：《光谱学与光谱分析》2024年第9期2439-2444,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62266026);云南省科技厅基础研究专项(202201AU070108)资助。

摘　　要：利用高光谱数据对湿地植被进行识别历来是植被遥感研究的重点之一。高光谱遥感数据包含更加细致的植被光谱特征,为高光谱植被识别提供了强有力的手段。以洱海东岸海滨为研究区,测取了3种典型湿地植被(菰、芦、槐叶蘋)的高光谱数据作为目标样本。对原始光谱进行一阶微分、包络线去除变换并分析其光谱特征,采用连续投影(SPA)、竞争性自适应重加权采样(CARS)两种特征变量选择算法选取原始光谱及其变换光谱中的特征波长,最后基于全波段数据以及特征波长选取后的数据分别建立支持向量机(SVM)、随机森林(RF)、径向基(RBF)神经网络的识别模型。结果表明:SPA与CARS算法对高光谱数据都有良好的降维效果,选取出的特征波长数量在5~18之间。对比组合不同的光谱变换处理与特征波长提取方法进行模建实验,包络线去除-SPA-SVM模型识别三类目标样本表现最好,其识别精度为0.9375,此时选取用于输入建模的特征波长数量仅为10个,占全波段的4.7%,极大的降低了模型的运算时间,而且选取的特征波长中,70%都位于特征吸收带内,其分布可以较好的反应植被化学成分差异导致的光谱吸收特征规律。实验结果表明利用光谱变换、特征选择后建模的高光谱植被识别是可行的,可以为其他湿地植被识别方法提供参考。The identification of wetland vegetation using hyperspectral data has traditionally been one of the focuses of vegetation remote sensing research.Hyperspectral remote sensing data contains more detailed spectral features of vegetation,providing a powerful means for identifying hyperspectral vegetation.In this paper,the hyperspectral data of three typical wetland vegetation species,Zizania latifolia;Phragmites australis Salvinia natans;They were measured as target samples in the study area of the east coast of Erhai Lake.The original spectra were transformed by first-order differentiation and envelope removal and analysed for their spectral features.The feature wavelengths in the original spectra and their transformed spectra were selected using two feature variable selection algorithms,namely,successive projection(SPA)and competitive adaptive reweighted sampling(CARS),and the support vector machine(SVM)and random forest(RF)were finally established based on the full-wavelength data as well as the feature wavelengths after the selection-Recognition models.The results show that both SPA and CARS algorithms have a good dimensionality reduction effect on hyperspectral data,and the number of selected feature wavelengths is between 5and 18.Comparing the combination of different spectral transform processing and feature wavelength extraction methods for modelling experiments,the envelope removal-SPA-SVM model performs the best in identifying the three types of target samples,with a recognition accuracy of 0.9375.At this time,the number of feature wavelengths selected for input modelling is only 10,which accounts for 4.7%of the full wavelength range,which greatly reduces the model's computation time and of the selected feature wavelengths,70%of the selected characteristic wavelengths are located in the characteristic absorption bands.Their distribution can better reflect the spectral absorption characteristic law caused by the differences in the chemical composition of vegetation.The experimental results show that the hyp

关 键 词：高光谱 湿地植被 光谱变换 特征波段选择 支持向量机 随机森林 

分 类 号：O433.4[机械工程—光学工程]