人类的生存和经济发展都依赖于生物多样性，保护生物多样性已经成为当今国际社会最关注的重要环保问题之一。中国作为生物多样性最为丰富的国家之一，越来越重视生物多样性的保护。生物多样性评估是生物多样性保护工作的关键。通过对生物多样性信息的采集、整理、计算而对生物多样性进行评估，可以为保护区的日常管理和运作提供参考。作为一种宏观视角的监测技术，遥感对地观测技术因其快速性、可重复性的特点，在生物多样性监测和评估方面的应用日益广泛。在保护和履行我国生物多样性保护责任的措施日益加强的背景下，生物多样性保护工作也面临着所需技术和方法的挑战。

本文以内蒙古自治区锡林郭勒盟锡林浩特市草原典型区为例，开展植物多样性评估工作；以实地调查数据为依据，计算不同测度的生物多样性指数以反映研究区植物多样性的真实现状；同时利用无人机高光谱数据，探讨基于卷积神经网络的草种识别方法，实现基于遥感数据的生物多样性评估及制图；本研究为高效、细致地进行生物多样性评估工作提供技术依据，同时为研究区生物多样性的保护、管理和恢复提供重要的科学参考。主要结论如下：

（1）实地调查的结果显示，研究区共包含植物21科，40属，48种。通过对各科、属所含种的统计发现，该区域禾本科所含的种数最多，共包含9个物种；其次为菊科、豆科；藜科、百合科、唇形科、蔷薇科，余下的各科各含有1个物种。从地块角度看，1号地块共存在植物28种，以黄蒿、野韭为主；2号地块中总共存在植物24种，以野韭、黄蒿、栉叶蒿为主；3号地块中总共存在植物20种，以野韭、针茅为主；4号地块中总共存在植物31种，以黄蒿、栉叶蒿为主；5号地块中总共存在植物34种，以栉叶蒿、羊草等；6号地块中总共存在植物36种，地块内大部分面积植被相对低矮，其中栉叶蒿的重要值最大，但大部分栉叶蒿处于幼苗阶段，少部分植被茂密的地方以针茅和羊草为主。

（2）6个地块中，6号地块的分布的物种最多，物种丰富度和多样性最高；2号地的物种丰富度最低。从均匀度指标上看，3号地的Pielou均匀度最高，1号地均匀度最低。根据6各地块之间的距离，将地块分为两组，1组中，物种的相似性随着距离的变化出现低-高-低的变化趋势，2组中物种的相似性随着距离的增加而不断下降，物种相异性的变化与之相反。

（3）本文选用植物的生殖株高、营养株高以及生活周期数据作为计算功能多样性的依据，分别计算了6个地块的功能均匀度指数FEve以及功能离散度指数FDiv，从结果上看，4号地块FEve值最高，说明其物种性状在空间上的分布最为规律，1号地块FEve值最低，其物种分布可能存在间隙。6号地块的FDiv值最低，说明地块内物种对资源的竞争较强，对资源的有效利用能力相对低下。

（4）传统的SVM模型和2D CNN模型的草种识别效果校对较差，主要原因在于二者所训练的特征较为单一，SVM模型仅考虑光谱特征，2D CNN仅考虑空间特征。3D CNN模型、HybridSN模型以及HybridSN+模型对各地块的草种的识别结果可以看出，三种模型均能够有效的识别出地块内的重要物种，相比较而言，3D CNN模型对于相对密集且较小的物种的识别容易出现错误，HybridSN模型在一定程度上能够弥补3D CNN的这一缺陷。HybridSN+在HybridSN模型上进行优化，在同等情况下，其对相对较小的物种个体的识别效果更好，总体精度有一定的提升。SVM模型的识别总体精度OA为85.54 %，平均精度AA为84.12 %，Kappa系数为82.92 %；2D CNN模型的识别总体精度OA为88.07%，平均精度AA为86.48%，Kappa系数为85.27%；3D CNN模型是识别总体精度OA为92.44 %，平均精度AA为91.34 %，Kappa系数为90.33%；HybridSN模型的识别结果总体精度OA为94.39 %，平均精度为93.65 %，Kappa系数为93.33 %。HybridSN+模型的识别结果总体精度OA为94.65 %，平均精度为94.27 %，Kappa系数为93.99 %。草种识别效果从优到劣依次是：HybridSN+模型、HybridSN模型、3D CNN模型、2D CNN模型、SVM模型。

（5）利用HybridSN+模型对草种的分类结果计算各地块的Shannon-Wiener多样性指数、Simpson多样性指数、Pielou均匀度指数并制图，并以此为基础从均值、相关性以及RMSE三个方面进行分析，结果表明基于分类计算得到的多样性指数与地面调查结果之间存在较强的相关性和一致性，两种方法计算得到的指数之间的差异较小，在一定程度上表明，基于无人机高光谱数据草种分类计算结果得到的多样性指数能够反映研究区的植物多样性现状。

本研究为高效、细致地进行生物多样性评估工作提供技术依据，同时为研究区生物多样性的保护、管理和恢复提供重要的科学参考。

Biodiversity is a prerequisite for the survival of human beings and the foundation of economic development, and the protection of biodiversity has become one of the most important environmental issues of international concern today. As one of the countries with the most diverse biodiversity, China is increasingly prioritizing the protection of biodiversity. Biodiversity assessment is crucial in biodiversity conservation work. By collecting, organizing, and calculating biodiversity information to assess biodiversity, it can provide reference for the daily management and operation of protected areas. As a macroscopic monitoring technology, remote sensing technology, with its characteristics of rapidity and repeatability, is increasingly widely used in biodiversity monitoring and assessment. Against the backdrop of increasingly strengthened measures to protect and fulfill China's biodiversity protection responsibilities, the work of biodiversity protection also faces challenges in the required technology and methods.

This article takes the grassland typical area of Xilinhot City, Xilingol League, Inner Mongolia Autonomous Region as an example to carry out plant biodiversity assessment. Based on field investigation data, different measures of biodiversity indices were calculated to reflect the true status of plant diversity in the study area. At the same time, using drone hyperspectral data, a grass species recognition method based on convolutional neural network was explored to achieve biodiversity assessment and mapping based on remote sensing data. This study provides technical support for efficient and detailed biodiversity assessment work and is an important scientific reference for the protection, management, and restoration of biodiversity in the study area. The main conclusions are as follows:

(1) The results of field investigation show that the research area contains a total of 21 families, 40 genera and 48 species of plants. By analyzing the number of species contained in each family and genus, it was found that the family Poaceae has the most species in this area, with a total of nine species; followed by Asteraceae and Fabaceae; Amaranthaceae, Liliaceae, Lamiaceae, and Rosaceae. The remaining families each contain one species. From the perspective of land plots, there are a total of 28 plant species in plot 1, with dominant species being yellow gentian and wild chives. Plot 2 has a total of 24 plant species, with dominant species being wild chives, yellow gentian, and serrated dock. Plot 3 has a total of 20 plant species, with dominant species being wild chives and needle grass, and the vegetation growth and coverage are relatively good. Plot 4 has a total of 31 plant species, with dominant species being yellow gentian and serrated dock. Plot 5 has a total of 34 plant species, with dominant species being serrated dock, bluegrass, etc. Plot 6 has a total of 36 plant species, with most of the vegetation in the plot being relatively low and short, with serrated dock having the highest importance value. However, most of the serrated dock is in the seedling stage, and in areas where vegetation is dense, needle grass and bluegrass are dominant species.

(2) Among the six plots, the species distribution is highest in plot 6, with the highest species richness and diversity; plot 2 has the lowest species richness and plot 1 has the lowest diversity. In terms of evenness, plot 3 has the highest Pielou evenness and plot 1 has the lowest evenness. Based on the distances between the six plots, they are divided into two groups. In group 1, the similarity of species shows a low-high-low trend with the change of altitude, while in group 2, the similarity of species decreases continuously with the increasing distance, and species dissimilarity is the opposite.

(3) This paper uses the reproductive plant height, nutritional plant height, and life span data of plants as the basis for calculating functional diversity. The evenness index FEve and the diversity index FDiv were calculated for six plots. From the results, it can be seen that plot 4 has the highest FEve value, indicating that the distribution of its species traits is the most regular in space, while plot 1 has the lowest FEve value, suggesting that there may be gaps in its species distribution. Plot 6 has the lowest FDiv value, indicating that the competition for resources among species in the site is relatively strong, and their ability to effectively utilize resources is relatively low.

(4) The traditional SVM model and 2D CNN model have poor accuracy in grass species recognition, mainly because the features they are trained on are relatively single. The SVM model considers only spectral features, while the 2D CNN model considers only spatial features. As a result, both models have poor recognition of individuals with mixed distributions and short plant height. As for the 3D CNN model, HybridSN model and HybridSN+ model, the recognition results of grass species in each plot show that all of the three models can effectively identify important species in each plot. Comparatively, the 3D CNN model is prone to errors when identifying relatively dense and small species, while the HybridSN model can partially compensate for this deficiency. The HybridSN+ model is an optimization of the HybridSN model which demonstrates better recognition of relatively smaller species, with an overall increase in accuracy under equivalent conditions.The overall accuracy (OA), average accuracy (AA), and Kappa coefficient of the SVM model are 85.54%, 84.12%, and 82.92% respectively; As for the 2D CNN model, they are 88.07%, 86.48%, 85.27%; As for the 3D CNN model, they are 92.44%, 91.34%, 90.33%; As for the HybridSN model, they are  94.39%, 93.65%, 93.33%, as for the HybridSN+ model, they are 94.65 %, 94.27%, 93.99%. In terms of grass species recognition accuracy, the models rank from highest to lowest as follows: HybridSN+ model, HybridSN model, 3D CNN model, 2D CNN model, and SVM model.

(5) The HybridSN+ model was used to calculate the Shannon-Wiener diversity index, Simpson diversity index, and Pielou evenness index for each plot based on the classification results of grass species. Maps were then created based on these indices. The analysis was based on mean, correlation, and RMSE, and the results showed strong correlation and consistency between the diversity indices calculated based on classification and ground survey results. The difference between the indices calculated by the two methods was small, which suggests that the diversity indices obtained from grass species classification using unmanned aerial vehicle hyperspectral data can reflect the plant diversity status in the study area to a certain extent.

This study provides technical support for the efficient and detailed assessment of biodiversity, and provides important scientific references for the conservation, management, and restoration of biodiversity in the study area.