基于多源遥感数据亚热带森林乔木物种多样性估测研究

发布时间：2018-05-20 04:33

本文选题：森林物种多样性 + GF-2　；参考：《中国林业科学研究院》2017年博士论文

【摘要】：生物多样性是维持生态系统健康和提供服务功能的基础。生物多样性监测是生物多样性保护的重要组成部分,使得各国政府和民间社会发现问题、制定解决方案,评估在实现“生物多样性公约”目标方面的行动和进展。机载高光谱数据具有高光谱和高空间分辨率特点,对估测森林物种多样性的分布具有较大潜力。但是,高空间分辨率的特点有可能成为限制高光谱数据对森林物种多样性估测精度的因素,特别是当像素尺寸小于树种识别尺度(例如树冠)时。机载激光雷达是一种主动遥感方式,由于其具有探测三维植被结构的能力,在生态和生物多样性保护研究方面愈发重要。由于不同物种间的竞争关系,不同的物种趋向于占据森林冠层中不同空间,因此使用LiDAR提取的森林垂直结构参数估测森林物种多样性成为可能。本文的目标是评估高分二号(GF-2)多光谱卫星数据、机载高光谱数据、机载激光雷达点云数据和模拟星载大光斑激光雷达波形数据对研究区森林乔木物种多样性的解释、预测能力。本文探索了森林乔木物种α多样性指数如物种丰富度、Shannon-Wiener和Simpson多样性指数、Pielou均匀性指数与遥感数据提取的森林垂直结构参数、光谱指数和纹理特征的关系。由于多源遥感数据提取的参数众多,本文使用随机森林(Random Forest,RF)方法筛选出与森林乔木物种多样性最相关的变量集,然后比较了四种不同机器学习模型(随机森林、SVR、KNN和Cubist)对森林乔木物种多样性的建模能力。根据各模型表现,选择了SVR算法对研究区森林乔木物种多样性进行估测研究,并使用十折交叉验证的方式考察SVR模型对各物种多样性的估测精度。论文的主要结果与结论如下:(1)机器学习方法有助于从海量多源遥感数据源中选择适合当前试验区的主要特征,并在线性不可分的情况下充分探索样本中各特征值的分布情况,提高多源遥感数据源对森林乔木物种多样性分布的解释能力。(2)高分辨率遥感数据提取的森林冠层光谱异质性特征对两个试验区Shannon-Wiener、Simpson物种多样性指数和Pielou物种均匀性指数的分布均具有较好的解释能力(R~20.54),对森林α物种多样性的估测具有重要意义。(3)激光雷达提取的特征从垂直结构层面对云南普洱试验区森林乔木物种多样性分布具有较好的解释能力(R~20.54)。激光雷达数据的加入,可以提高高分辨率光学数据对森林乔木物种丰富度、Simpson多样性指数及Pielou均匀性指数的估测精度。其中,结合机载激光雷达数据和高光谱数据提取特征具有最好的估测精度,尤其是对物种丰富度的估测精度(R~2=0.74,RMSE=3.36)高于其他数据源。(4)GF-2号数据对Shannon-Wiener和Simpson物种多样性指数、Pielou物种均匀度指数的估测精度与机载高光谱的估测精度类似。在使用遥感数据对森林乔木物种多样性间接估测时,遥感数据对这三种多样性指数的解释能力并没有随着空间及光谱分辨率的大幅提高而提高。(5)结合模拟大光斑波形数据提取的森林垂直结构参数与GF-2多光谱数据提取的植被指数和纹理信息对普洱试验区森林乔木物种多样性的分布比单一数据源具有较好的估测精度(物种丰富度:R~2=0.58,RMSE=4.2;Shannon-Wiener多样性指数:R~2=0.7,RMSE=0.47;Simpson多样性指数:R~2=0.73,RMSE=0.164;Pielou均匀性指数:R~2=0.68,RMSE=0.15)。对我国陆地生态系统碳监测卫星的生物多样性应用具有参考价值。
[Abstract]:Biodiversity is the basis for the maintenance of ecosystem health and service functions. Biodiversity monitoring is an important component of biodiversity conservation. It makes governments and civil society identify problems, formulate solutions, and assess the action and progress in achieving the objectives of the Convention on biological diversity. Airborne Hyperspectral Data The characteristics of hyperspectral and high spatial resolution have great potential for estimating the distribution of forest species diversity. However, the characteristics of high spatial resolution may be a factor limiting the accuracy of hyperspectral data to estimate forest species diversity, especially when the pixel size is smaller than the tree recognition scale (such as the crown). It is an active remote sensing method. Because of its ability to detect the three-dimensional vegetation structure, it is becoming more and more important in the study of ecological and biological diversity protection. Because of the competitive relationships among different species, different species tend to occupy different spaces in the forest canopy. Therefore, the forest objects are estimated using the forest vertical structure parameters extracted by LiDAR. The objective of this paper is to evaluate the high score two (GF-2) multi spectral satellite data, Airborne Hyperspectral Data, airborne laser radar point cloud data and simulated satellite large spot laser radar waveform data to explain the species diversity of forest tree species in the study area. This paper explores the alpha diversity of forest tree species. The relationship between the species richness, the Shannon-Wiener and Simpson diversity index, the Pielou evenness index and the forest vertical structure parameters, the spectral index and the texture feature extracted by remote sensing data. Because of the many parameters extracted from the multi source remote sensing data, this paper uses the random forest (Random Forest, RF) method to select the species diversity of the forest tree species. The most relevant variables set, and then compare the modeling ability of four different machine learning models (random forest, SVR, KNN and Cubist) for forest tree species diversity. According to the performance of each model, the SVR algorithm is selected to estimate the species diversity of the forest tree species in the study area, and the SVR model is examined by the way of ten fold cross validation. The main results and conclusions of the species diversity are as follows: (1) the machine learning method is helpful to select the main features suitable for the current test area from the mass multi-source remote sensing data source, and to fully explore the distribution of the characteristic values in the sample, and improve the multi source remote sensing data source to forest Joe. The interpretation ability of tree species diversity distribution. (2) the spectral heterogeneity of forest canopy from high resolution remote sensing data has good explanatory ability (R~20.54) for the distribution of Shannon-Wiener, Simpson species diversity index and Pielou species uniformity index (R~20.54), which is important for estimating the diversity of forest species. (3) the characteristics of the laser radar extracted from the vertical structural layer have a good interpretation ability (R~20.54) in the forest species diversity distribution in Yunnan Pu'er test area. The addition of laser radar data can improve the estimation of the species richness, the Simpson diversity index and the Pielou evenness index of high resolution optical data to forest trees. Precision. Among them, combined with airborne lidar data and hyperspectral data extraction features have the best estimation accuracy, especially for species richness estimation accuracy (R~2=0.74, RMSE=3.36) is higher than other data sources. (4) GF-2 data on Shannon-Wiener and Simpson species diversity index, Pielou species evenness index estimation accuracy and machine The estimation accuracy of the high spectrum is similar. When using remote sensing data to estimate the species diversity of forest tree species, the interpretation ability of remote sensing data to these three diversity indices does not increase with the spatial and spectral resolution. (5) the forest vertical structure parameters extracted with simulated large spot waveform data and GF-2 The vegetation index and texture information extracted by spectral data have better estimation accuracy (species richness: R~2=0.58, RMSE=4.2; Shannon-Wiener diversity index: R~2=0.7, RMSE=0.47; Simpson diversity index: R~2=0.73, RMSE=0.164; Pielou evenness index: R~2=: R~2=: R~2=: R~2=: R~2=; Pielou evenness index: R~2=: R~2=: R~2= 0.68, RMSE=0.15). It has reference value for biodiversity monitoring of terrestrial ecosystem carbon monitoring satellite in China.
【学位授予单位】：中国林业科学研究院
【学位级别】：博士
【学位授予年份】：2017
【分类号】：S718.5

【相似文献】