【作者】 郭颖；

【导师】 李增元；

【作者基本信息】 中国林业科学研究院 ， 森林经理学， 2011， 博士

【摘要】 近年来,综合使用多源遥感数据的森林地上生物量(Forest Above Ground Biomass,FAGB)估测技术已逐渐成为遥感领域的研究热点。由于传统的参数化估计方法不能有效地描述FAGB与多源遥感数据间的非线性关系,所以出现了一些将非参数化方法,如K最近邻(K Nearest Neighbours,KNN)、神经网络(NN)算法等,用于森林AGB多源遥感估测的研究。支持向量机(Support Vector Machine,SVM)也是一种非常有前景的非参数化方法,它基于统计学习理论,可以较好地解决小样本、非线性、过学习、高维度、局部极小点等实际问题。自提出以来,因其独特的优势,使其在众多领域中得到了广泛的研究与应用,然而利用SVM进行森林AGB遥感估测的相关报道还很少,并且也只有少数研究开展了旨在提高估测精度和稳键性的非参数化优化算法。本文全面分析评述了森林AGB遥感估测的主要方法,重点总结了KNN、SVM等非参数化估测方法用于遥感生物物理参数反演方面的研究现状。在对已有的一些非参数化估测方法进行系统深入研究及优化的基础上,确定了NN、KNN用于森林AGB估测的优化方法,并建立了优化的基于SVM的森林AGB遥感估测技术流程,提出了一种将随机特征选择技术与SVM相结合的组合算法。主要研究内容及成果如下:(1)基于SPOT5影像,LiDAR归一化点云统计变量以及相关的一些遥感因子,如纹理、植被指数等,设计了单隐层的错误反向传播NN模型(Back Propagation Neural Network,BPNN)及三层径向基NN模型(Radial Basis Function Neural Network,RBFNN),并利用随机森林(Random Forest,RF)算法进行特征选择。结果表明,特征选择可以提高NN算法的估测效果,RBFNN在估测精度、稳定性上均优于BPNN。(2)基于KNN采用多源遥感数据,包括光学SPOT影像光谱值、相关遥感因子,如纹理、植被指数等及LiDAR归一化点云统计变量进行森林AGB估测,为了与其他研究的非参数化模型进行比较,同样采用RF特征选择对基于KNN的估测方法进行优化,并对KNN算法的众多参数进行优选。结果表明,优化后的KNN算法可以用于森林AGB估测,并且在处理错误样本时具有良好的容错能力。(3)改进了SVM用于森林AGB遥感估测应用中的一些缺陷,建立了SVM在森林AGB遥感估测应用中的技术流程,包括特征选择、最优参数确定、核函数的自动寻优等。在特征选择部分采用了RF算法。结果表明,基于RF算法进行特征选择的估测效果优于F-score的估测效果。在核函数寻优部分,利用留一法,根据均方根误差最小原则选择最优核函数,并依此建立了从5种常用核函数中自动寻优的算法与程序,解决了常规SVM估测方法依赖经验选择核函数的局限性。(4)提出了将SVM和随机特征选取技术相结合的森林AGB组合估测模型。该模型在随机森林特征选择的基础上,利用随机特征选取技术生成一组特征子集,然后基于这组特征子集构建一组SVM,并对每个SVM子模型都进行参数的最优化选择及核函数的自动寻优等处理步骤,最后采用三种不同的方法将每个SVM子模型的估测值进行结合,分别为平均法、加权平均法以及选择法。与现有主流的非参数化方法的估测效果进行了比较,结果表明:由于组合模型在计算过程中会根据不同的样本选择其对应的最优拟合超平面,在某种程度上具有自适应性,整体的估测效果均要优于单一SVM及其他非参数化方法。本文研究工作的创新点如下:(1)建立并实现了基于SVM的森林AGB遥感估测应用技术流程。发展了基于SVM与随机特征选择相结合的森林AGB遥感估测模型。该模型在处理森林AGB遥感估测问题时,无论从预测精度还是稳定性均优于参比模型。(2)实现了基于KNN的森林AGB遥感估测方法的优化。在对KNN算法的参数如K值、距离计算方法及权重计算方法进行优选的基础上,通过结合RF特征选择方法,使估测精度得到提高。(3)将基于RF算法的特征选择方法应用于非线性估测模块中,与传统的F-score算法进行了比较,证实了该方法的有效性,为非线性估测模型的优化提供了一种通用方法。更多还原

【Abstract】 In the last years, the remote sensing community has devoted particular attention to the estimation of forest above ground biomass (AGB) via the analysis of multisource remote sensing data. A major observation in previous research on forest AGB estimation is that conventional parametric statistical pattern recognition methods are not appropriate in forest AGB estimation using multisource remote sensing data, since they cannot be modeled by a convenient multivariate statistical model. In these situations, the use of nonlinear regression techniques based on machine learning methodologies, such as K nearest neighbours(KNN), neural networks (NN), can represent an effective approach to solve such estimation problems. Another promising alternative nonparametric method to neural networks and KNN is the Support Vector Machine(SVM), originating from statistical learning theory, have provided capacities to deal with the few ground measurements, nonlinear, overfitting, highly dimension problems and have already proven their usefullness in many literature. However, few studies have investigated the potential of applying SVM to estimate forest AGB and the peculiarities of the nonparametric methods to improve the robustness and precision of the estimation process.This paper completely appraised various methods on forest AGB estimation and summerized the research status of the nonparametric methods, such as SVM, for biophysical variable estimation from remotely sensed images. On the basis of systematicly and intensely investigation and optimization of some nonlinear tools, the optimized neural networks and KNN algorithms on forest AGB estimation is presented. Furthermore, the technique flow of SVM for estimating forest AGB is introduced. To increase the performance of the algorithms in terms of estimation accuracy and robustness, the algorithm of SVM is modified and combined with random feature selection for optimizing the estimation results.The main works and results are as follows: (1) The method of back propagation neural network (BPNN) with single hidden layer and radial basis function neural network(RBFNN) are designed using the SPOT5 spectral reflectivity, LiDAR points cloud statistical variables and some remote sensing factors such as texture and vegetation index. The results clearly demonstrate that estimation accuracies increased by feature selection based on the random forest(RF) algorithm. Furthermore, compared to BPNN, the RBFNN model provided more accurate, improved robust result on the considered case.(2) The optimal KNN algorithm is established to estimate the forest AGB by using SPOT5 spectral reflectivity, LiDAR points cloud statistical variables and some remote sensing factors such as texture and vegetation index. The results show that the optimal KNN model, based on the selection of features and the determination of optimal parameters, provides a solution to deal with such problems and represents a promising method to address the complex and important problem of containing kinds of noise in training samples. In order to be compared with other nonparametric methods in this paper, the feature selection algorithm is RF too.(3) This paper introduces the basic flow of SVM for estimating forest AGB, which includes: feature selection, the determination of optimal parameter, the automatic optimization of kernel function and so on. In the stage of feature selection, the RF model provide better results compared to the typical F-score method. For the automatic optimization of kernel function, we developed the method of selecting optimal kernel automatically from four common kernels, which are linear kernel, polynomial kernel, RBF kernel and sigmoid kernel, abiding by root mean squared error (RMSE) minimum principle.(4) A novel approach to the estimation of forest AGB from multisource remote sensing images based on the composed model of SVM and random feature selection has been presented. The composed model, on the basis of RF feature selection, trains a group of SVM models by random feature selection technique and aims at exploting the peculiarities of an ensemble of SVM to improve the robustness and accuracy of the estimation process. The defining of optimal parameter and the automatic optimization of kernel function are handled by the single SVM in composed model. Three result combination strategies are adopted which are average-based method, weight-based method and selection-based method. Results show that: the composed model is more effective than regular SVM and other nonparametric methods to estimate forest AGB whether the training data are uncertainty or not, since it has the characteristic of self-adapted to some extent.The innovations in the thesis are as follows:(1) Introduce the basic flow of using SVM to estimate forest AGB and propose a novel approach to estimate forest AGB from multisource remote sensing images by combining SVM and random feature selection technique.The gain of the proposed algorithm is noticeable especially significant when working with very reduced training sets and different noise sources.(2) Implement the optimal estimation model based on KNN for estimating forest AGB from multisource remote sensing images. The optimal model improves the accuracy further by combining the RF feature selection method.(3) Use the RF algorithm to select features in the three kinds of nonparametric estimation method and validate the effectiveness by comparing to the result of typical F-score algorithm.更多还原