【作者】 朱清苗；

【导师】 彭世揆；

【作者基本信息】 南京林业大学 ， 森林经理， 2004， 硕士

【摘要】 本文是以中山陵风景园林景区为研究区域，在原始数据TM图像进行光谱特征分析和几何校正的基础上，对遥感图像进行了一系列常规处理，如反差变换、比值增强、主成分变换、缨帽变换、假彩色合成等，然后分析各波段的方差、相关系数，计算各波段的信息熵，以及不同组合的最佳指数（OIF），综合选取了PCI、NDVI、TM4组合作为图像分类的基础波段。 在此基础上，进行了无监分类和有监分类，主要方法选用了最大似然法、最小距离法和神经元网络法。重点对神经元网络法进行了研究，先是用BP网络对图像进行学习训练并分类，然后又用动量方法对BP算法改进，得到较为理想的分类结果。通过以上一系列的分析处理，本文得到了以下结论： TM7波段信息量最为丰富，各地类间的差异也最明显。TM4波段的独立性最强，而且与生物量关系最为密切。在TM的六个波段中经过比值处理后，从影像上看各个地类之间的差异加大，效果比较明显的是归一化植被指数和比值植被指数。通过主成分分析表明，主成分综合了各个波段的信息，且相互之间完全正交不相关，其中PCI达到90.97％的贡献率。 最大似然法的内部算法缺陷少、可靠性好，采用高斯分布模型描述样本的概率分布，对实际上并非高斯分布的样本进行分析时会产生较高的误判率，但对本问而言，所用TM4波段的灰度直方图表明，样本的概率分布服从高斯分布。所得精度结果稍低于神经网络方法。 用动量方法对原始的BP算法进行改进，在节省时间和训练结果上都取得了较好的效果。使收敛总误差（Training RMS Error）在更短的时间达到所期望的值（0.1）。 通过本文的研究神经网络分类方法，无论是总体精度还是K_hat统计值都大于最大似然法和最小距离法，为96.43％和95.33％。各种精度评价指标综合体现出神经网络的优势。分析其原因，神经网络具有良好的自组织和自学习功能，大大放松了传统模式识别方法所需的约束条件，即使模式空间的分布出现锯齿状情况，神经网络也能对模式集进行正确的分类，这也是神经网络分类精度优于最大似然法的主要原因。 为全面比较神经元网络方法与传统的分类方法以体现其优势所在，在精度评价上本文采用了多个评价指标，以期得到比较客观的评价。更多还原

【Abstract】 The traditional remote sensing image classification mthods are Maximum-likelihood, Min-distance classification,etc. In this paper, the artificial neural network(ANN) model was introduced to classify five typical kinds of terrain in Zhong-Shan Mausoleum, such as forest, water, farmland and grassland,building,and others. Based on the principle, the paper compared the distinctness of the three classification methods with the attempt to bring out the trend of remote sensing image classification being used in Zhong-Shan Mausoleum.On the basis of geometric correction for remote sensing images data, detailed character analysis was conducted for the TM images. Then several image transformations such as interactive contrast stretching , ratio enhancement, principal components transformation, tasseled-cap transformation, image color conbine, etc, were implemented.Two indexes was calculated to estimate the best bands union for color combination,one is Optimum Index Factor(OIF), the other is the determinant of the co-variance matrix. It can be seen from the result that for color combination the original optimal bands were TM2, TM4, TM7, the best mixed images were PC1, NDVI, TM4.The main drawback of traditional remote sensing image classification methods is its low precision. The neural networl-based remote sensing image classification technique has been presented. The result demonstrated that the neural network classification system could be used in remote sensing image classification, and its classification precision was superior to that, of the Maximum-likelihood and Min-distance, achieving an accuracy of 96.43%.In this paper,the author applied the gray-level values of three bands union as the network model’s input variables, and chose five kinds terrain as the model’s output variables. First the original Back Propagation Network was used to classify terrain types, the result of the classification demonstrated that Training RMS Error couldn’t descend the expected value. Then Momentum mothed tried to improve the original Backpropagation algorithm, the experimental results show that network converged speedily achieving the expected value(0.1).Accuracy analysis is a necessary work in the classification of remote sensing data. In this paper, the evaluating indexes included overall accuracy, Kappa coefficient, product accuracy, user accuracy, etc. In a word, ANN classification is superior to traditional methods, and has wide prospects in forestry remote sensing.更多还原