基于植被指数的三江源高寒草地植物分类与识别方法研究

随着生态健康检测与保护工作的实践以及研究问题的深入，传统的植物分类手段不能完全满足当前研究的需要。因此为研究快速分类识别草地植物的方法，本研究利用ASD (Analytical spectral devices)地物光谱仪，采集了三江源地区高寒草地常见的阿尔泰葶苈(Draba altaica)、高山风毛菊(Saussurea japonica)和车前状垂头菊(Cremanthodium ellisii)等36种植物的原始光谱数据，并选择了比值植被指数等16种高光谱植被指数，基于支持向量机(Support vector machines，SVM)等3种机器学习算法，构建高寒草地植物光谱分类识别模型。研究结果表明：高寒草地植物的原始光谱均符合绿色植物特征，但由于植物形态特征不同光谱差异主要集中在可见光波段；基于植被指数结合3种算法构建的分类模型，精度依次为随机森林(Random forest,RF)(99.4%)>SVM (93.2%)>K邻近算法KNN (88.0%)，且模型的预测结果都出现了误判情况；相比SVM与KNN，RF为基于植被指数构建模型的最佳算法，同时能对所构建模型参数进行重要性分析，其中RGI和SAVI为提高RF分类模型精度的两个重要参数。

Studying on Plant Classification and Recognition Method for Three-River Source Alpine Grassland Plant Based on Vegetation Index

With the practice of ecological health detection and protection and the deepening of research problems,traditional plant classification methods cannot fully meet the needs of current research. In this study,ASD (Analytical spectral devices,ASD) ground object spectrometer was used. The original spectral data of 36 species of plants such as Draba altaica,Saussurea japonica, and Cremanthodium ellisii,which are common in the alpine grasslands of the Three-River region,were collected. And 16 kinds of hyperspectral vegetation indices such as RVI (Ratio vegetation index) were selected,and a spectral classification and identification model of alpine grassland plants was constructed based on 3 machine learning algorithms including Support vector machines (SVM). Research indicates:the original spectra of alpine grassland plants are consistent with the characteristics of green plants,but the spectral differences are mainly concentrated in the visible light band due to different plant morphological characteristics;the classification model constructed based on the vegetation index combined with three algorithms,the accuracy is RF (99.4%) > SVM (93.2%) > KNN (88.0%),and the prediction results of the model are all misjudged;compared with SVM and KNN,RF is the best algorithm to build a model based on vegetation index,and it can also analyze the importance of the parameters of the built model. Among them,RGI and SAVI are two important parameters to improve the accuracy of the RF classification model.