【作者】 柳萍萍；

【导师】 林辉；

【作者基本信息】 中南林业科技大学 ， 森林经理学， 2012， 硕士

【摘要】 森林资源是地球上最大的陆地生态系统,开展森林资源调查,了解和掌握森林资源现状和变化信息对于提高林业发展决策水平,科学合理的经营管理森林资源等都具有极其重要的意义。与传统森林资源调查方式相比,遥感影像解译因其宏观、周期短、动态等优点,被广泛应用于森林资源的调查监测中。随着遥感技术的发展,高光谱遥感是通过遥感方法获取更丰富信息的必然发展趋势,也是当代遥感的前沿和热点之一。高光谱遥感影像将成像技术与细分光谱技术结合,使得高光谱影像在分类技术的应用方面拥有巨大的潜力,高光谱数据能够更好地识别各类地物,但是其庞大的数据量和数据的高维度使数据的传输和存储受到了限制,也给高光谱数据处理方法的探讨带来了不小的挑战。本文应用Hyperion高光谱数据,对湖南株洲攸县黄丰桥林场作为研究区进行分类研究。针对Hyperion数据波段多和数据量大的特点,对高光谱数据进行了未定标和受水汽影像波段的剔除、像元值与绝对辐射值的转换、坏线的修复、Smile效应的校正、FLAASH大气纠正、几何校正等遥感数据预处理,再利用特征选择与特征提取相结合的基于分段主成分分析和波段指数的高光谱数据降维处理方法,将高光谱数据从高维空间映射到低维空间,然后分别利用最大似然法和光谱角填图法进行森林类型的识别与分析。主要研究结论如下：(1)对Hyperion数据进行波段剔除、转换以及校正等处理后,高光谱的数据量变小,地物曲线更加趋向于真实的植被光谱特征,使得地物更容易区分。(2)通过相关矩阵可知,利用分段主成分法对Hyperion数据进行数据源划分,每段的平均相关系数都在0.9以上,表明Hyperion数据相邻波段间的相关性很高。(3)利用分段主成分分析法结合波段指数法的高光谱降维方法将Hyperion数据的242个原始波段缩减为13个信息量大且相关性弱的波段组合,一方面,分段主成分分析法有效的抑制了全局变换导致局部重要光谱被滤除的可能,另一方面,波段指数法在进行波段选择时兼顾了自适应分段后段与段之间以及各分段中波段间的相关性,有效降低了高光谱数据的维度。(4)波段选择之前进行子空间划分,本文将Hyperion数据划分为三个子空间,此操作可剔除相关性较大的波段,并能减小数据的计算量,从而达到高维遥感数据优化处理和高效利用的目的。(5)根据波段指数来选取波段的子集时,选取波段指数的极大值而不是最大值作为降维后波段子集,从而弥补原始数据在分段时产生的误差。(6)对Hyperion数据进行基于特征空间的最大似然法和基于光谱空间的光谱角填图法两种方式进行森林类型识别,分类误差矩阵结果表明：最大似然法的分类总体精度为85.23%,光谱角填图法的总体精度达到88.94%,说明对于高光谱而言,光谱角填图法针对性的对各类光谱进行识别优于多光谱的传统分类方法。更多还原

【Abstract】 Forest resources is the world’s largest land ecological system, the important component of the national natural resources, the forestry construction foundation, the necessary basis of human survival and development. Developing forest resources survey to understand and grasp the forest resources situation and change information to improve the scientific and reasonable decision-making level of forestry development, management of forest resources has great significance. Compare with the traditional forest resources survey, RS image interpretation for its macroscopic, short cycle, dynamic and other advantages, is widely used in the investigation of the forest resources monitoring.Along with the development of remote sensing technology, we are able to get and use more and more information, hyperspectral remote sensing through the remote sensing method for more information is the inevitable development trend, as well as the frontier and hotspots. Hyperspectral remote sensing image combined imaging technology and subdividing spectrum technique in one, it make hyperspectral images in the application of classification technology has huge potential, and in imaging process, hyperspectral images can obtain the continuous spectrum of the features of information, and the traditional spectrum than remote sensing data, hyperspectral data is better to recognition and classification of all kinds of features, but the huge amount of data and data of high dimension make the data transmission and storage is limited, also give hyperspectral data processing method discussion bring quite a challenge.This paper applied Hyperion hyperspectral data, to the hunan zhuzhou YouXian HuangFeng bridge as the forest classification research area. For the bands of Hyperion data and the characteristics of large volumes of data, execute several data processes, uncalibration and suffer lunt images bands weed out, the conversion of DN value and the absolute value, bad line repair, calibration of Smile effect, FLAASH atmosphere effect correction, the remote sensing data pretreatment geometry correction:using mothod combinated with feature selection and feature extraction for hyperspectral data to reduction dimension, process will let data from a high dimensional space mapping to a low dimensional space, then respectively applied maximum likelihood and spectral Angle mapping method for the forest types of recognition and analysis. The main research conclusions are as follows:(1) After the band removed, conversion and correction processing of Hyperion data, hyperspectral data quantity is small, surface features curves tend to be more real vegetation spectral features, making the surface features easier to distinguish.(2) The correlation matrix shows the segmented principal component analysis of Hyperion data, the division of data sources, each the average correlation coefficient above0.9, that Hyperion data adjacent to the high correlation between bands.(3) Use the segmented principal component analysis and band index combining dimensionality reduction method, Hyperion data the original band of242is reduced to the large amount of information and the correlation weak13bands, on the one hand, segmented principal component analysis effectively inhibitglobal transformation leading to the possibility of local important spectral filter, on the other hand, the band index during band selection, taking into account the adaptive segmentation between paragraphs and subparagraphs in the band after the correlation between the effective lowerthe dimension of the hyperspectral data.(4) Sub-space division, before the band selection can remove larger band, and can reduce the amount of data calculation, so as to achieve high-dimensional remote sensing data to optimize the processing and efficient use of purpose.(5) Band index to select a subset of the band, select band maxima of the index rather than the maximum value as the drop the Wei Houbo piece set to compensate for the error of the original data in the segment.(6) Hyperion data in two ways based on the feature space maximum likelihood and spectral angle based on the spectral space mapping method for identification of forest types, classification error matrix:the maximum likelihood classification of the overall accuracy of85.23%, the spectral anglefill in the diagram method, the overall accuracy of88.94%for the high spectral, spectral angle mapping method targeted to identify superior to the traditional multi-spectral classification of various types of spectra.更多还原