基于无人机高光谱影像的三江源草种精细识别研究

草种精细识别对三江源区草地生态系统退化监测具有重要意义。基于无人机高光谱遥感系统，获取三江源草地退化典型区的高光谱影像。在对原始光谱特征利用XGBoost进行优化选择的基础上，结合扩展形态学属性剖面特征，利用稀疏多项式逻辑回归与自适应稀疏表示两种分类方法分别对影像上的不同可食与毒杂草种进行精细识别，在此基础上提出形状自适应的后处理方法对识别结果进行平滑处理。结果表明：①利用XGBoost方法选择出重要性高的光谱特征能提升高光谱数据的识别效果并节省运行时间；②利用空间—光谱特征的识别方法相较于仅利用光谱特征的方法可以有效改善草种识别效果，使总体精度提升4%~5%；③利用两种稀疏表示方法在小样本的情况下对草种精细识别的精度分别达到94.07%、93.15%，利用形状自适应后处理方法能有效提高多种毒杂草种的识别精度，使得总体精度分别提升约1.64%和1.12%。基于特征挖掘的稀疏表示分类方法能实现高精度的无人机高光谱影像草种精细识别，为更大范围的草原物种精细识别提供了技术支撑。

Researches on Grass Species Fine Identification based on UAV Hyperspectral Images in Three-River Source Region

Fine identification of grass species is of great significance for grassland ecosystem degradation monitoring in the Three Rivers Source Region. Based on the UAV hyperspectral remote sensing system， the hyperspectral image of the typical grassland degradation area of Three-River Source Region was obtained. Firstly， using the obtained UAV hyperspectral image， the optimal bands combination were selected using XGBoost， the extended morphological attribute profile features were extracted and were combined with the selected spectral features. Secondly， sparse multinomial logistic regression and adaptive sparse representation methods were adopted to identify different grass species. Finally the shape adaptive based post-processing method was proposed to smooth the identification results. The results showed that： （1） Using the XGBoost method to select important spectral features can improve the identification result and save running time； （2） the spatial-spectral feature based method can effectively improve the identification result of grass species and the overall accuracy were improved by 4%~5% compared with the method of using only spectral features； （3） using two sparse representation methods，the overall accuracy of fine identification of grass species in the case of limited samples was 94.07% and 93.15% respectively， and the identification accuracy of various poison weed species was improved effectively by using shape adaptive post-processing method， which improved the overall accuracy by about 1.64% and 1.12%， respectively. The feature mining based sparse representation classification methods can achieve high-precision grass species fine identification of UAV hyperspectral images， and provide technical support for a wider range of grassland species fine identification.