【作者】 芶盛；

【导师】 郭科； 陈聆；

【作者基本信息】 成都理工大学 ， 应用数学， 2011， 硕士

【摘要】 我国经济的可持续高速发展对矿产资源与能源需求极大。但是,当前我国主要矿产资源保障形势非常严峻。遥感技术作为有效的对地观测手段,先后经历了全色摄影,彩色摄影、多光谱遥感,高光谱遥感等不同历史阶段。从20世纪70年代起,遥感技术就被广泛的应用于国土资源调查和监测领域,并取得了丰硕的成果。高光谱遥感极大地增强了遥感对地观测能力和对地物的鉴别能力,提高了遥感技术的定量化水平,使遥感从对地物的鉴别发展到对地物的直接识别阶段,利用高光谱遥感技术探测矿产是遥感技术应用的主要方向之一。目前高光谱遥感影像矿物识别分类方法主要分成两种,一种是基于数学变换进行图像降维的数理统计,如主成分分析法。另外一种则根据矿物的光谱形成物理机制,直接利用高光谱遥感数据光谱分辨率高的特点,通过选择吸收谱段,计算光谱吸收特征等方法实现对岩矿识别,如光谱特征拟合法。将这两类方法有效结合起来实现矿物识别分类,从矿物光谱形成的物理机理出发,研究分析地物的诊断光谱特征,并利用数理统计的相关理论方法实现特征匹配达到矿物识别精细与分类填图成为本文研究重点。多特征匹配决策树矿物识别与分类填图技术是一个知识发现与表达、规则定义、建立和运行决策树的过程。现有各类矿物识别决策树要么仅仅使用单一特征(如光谱主吸收峰位置特征)和相同的分类算法进行矿物识别,要么使用不同的特征却使用相同的算法,但是这两种决策树都在一定程度上限制了识别精度的提高。为改善矿物识别精度,本文根据各类典型蚀变矿物的光谱吸收特征建立了多变量决策树：首先计算了研究区典型蚀变矿物的主要光谱吸收特征,并将这些特征以知识的形式表达出来。然后分别计算了光谱特征影像的信息熵,根据信息熵的大小选择了光谱吸收指数、光谱吸收深度、光谱斜率、左区域面积等类特征并结合矿物本身的主要吸收特征完成了矿物识别规则的定义。最后在规则定义基础上构建和运行了决策树,得到了分类填图结果。该结果与中国地质大学(武汉)黄定华等人和中国国土资源航空遥感中心的填图结果基本一致,尤其是绿帘石和蛇纹石填图结果吻合的较好。而对于白云母、绿泥石、方解石矿物,三类结果稍有差异。这说明了用知识表达与规则定义的方法对模拟数据进行矿物填图是可行的、有效地,在一定程度上它可以实现矿物种类及分布填图,同时也说明本文建立的多特征混合决策树分类识别树分类能力强。为定量评价决策树分类精度,本文将构建的多特征决策树和SAM方法分别应用于根据USGS光谱库随机生成的模拟图像：结果显示多特征决策树识别正确率为89.06%,而SAM方法正确率为79.99%。这说明了决策树在研究区地物先验知识部分缺失的情况下仍然可以有效的保持较好的健壮性。但是,本文的研究成果主要针对常见典型蚀变矿物,不断拓展其它类别的蚀变矿物,提高识别的准确率将是今后研究的重点。更多还原

【Abstract】 The demand for mineral resources and energy has become greater and greater with China’s sustainable and rapid economic development. However, the safeguard system of China’s current major mineral resources is very grim. Remote sensing technology, as an effective means of earth observation, has gone through the full-color photography, color photography, multi-spectral remote sensing, and hyper-spectral remote sensing in different historical stages. Since 1970s, remote sensing technology was widely used in the field of land resources survey and monitoring, and has achieved fruitful results. Hyper-spectral remote sensing has greatly enhanced the ability of earth observation and discrimination of ground objects, has increased the quantitative level of remote sensing technology, has made the discrimination of ground objects develop into directly identify phase. And utilizing hyper-spectral to explore mineral resources is one of the main applications of remote sensing technology.At present, mineral discrimination and classification method of hyper-spectral image is divided into two kinds; one is mathematical statistics method, which is based on the mathematical transformation to reduce the dimension of the image, such as principal component analysis. Another is based on the mineral spectra’s physical formation mechanism, directly use hyper-spectral data’s feature of high spectral resolution, by selecting the absorption spectrum and calculating spectral absorption characteristics and other methods to discriminate the rock and mineral, such as the spectral feature fitting. Combining the two methods to discriminate of mineral classification effectively, starting from the physical formation mechanism of the mineral spectra, studying and analyzing the ground objects’diagnostic spectral characteristics, and using the theory of mathematical statistics methods to match features, identify, classify and mapping minerals precisely has been the focus of this thesis.Multi-feature matching decision tree for mineral identification and classification mapping technology is a process that involves knowledge discovery and expression, rule definition, establishment and operation of the decision tree. The existing different mineral identification trees only uses single features (such as the spectral main absorption peaks) and the same classification algorithm for mineral identification, or use different features while use the same algorithm, but the two decision trees are Limited to improve the accuracy of recognition to some extent. To improve the minerals’recognition accuracy, this thesis establishes a multi-variable decision tree according to the spectral absorption characteristics of various typical altered minerals. Firstly, this thesis calculates the main absorption features of the typical alteration minerals in the study area, and expresses these features in the form of knowledge. Then calculates the information entropy of the spectral features images, chooses spectral absorption index, absorption depth, spectral slope, left regional area, and other features according to the entropy, and realizes the definition of mineral identification rules by combining with minerals’main absorption features. Finally, builds and runs the decision classification tree based on the definition rules and gets the mapping results. The mapping results are basically consistent to Huang Dinghua et al from China University of Geosciences (Wuhan) and that of Remote Sensing Center of China Aviation Land and Resources, especially epidote and serpentine mapping results. As for the muscovite, chlorite, calcite minerals, the three mapping results are slightly different. This shows that the method of knowledge representation and rules definition for mineral mapping on simulated data is feasible and effective, and to a certain extent, it can achieve mapping of the mineral type and distribution. It also illustrates the established mixed-decision classification tree has a strong classification ability. In order to evaluate the classification accuracy quantitatively of the decision tree, this thesis applies the multi-feature decision tree and SAM method on simulated images which is randomly generated on the USGS spectral library the results shows that recognition accuracy of the multi-feature decision tree was 89.06%, while the SAM method is 79.99%. This represents that the decision tree can still be effective and keeps good robustness with the absence of the prior knowledge of the study area.However, the research results are mainly focused on the common typical alteration minerals, so how to expand to research methods on other types of alteration minerals and improve the accuracy of identification will be the focus of future research.更多还原