【作者】 刘炜；

【导师】 常庆瑞；

【作者基本信息】 西北农林科技大学 ， 土地资源与空间信息技术， 2012， 博士

【摘要】 【目的】以陕西省横山县1986年8月2日和2003年8月17日的两期TM图像作为基础数据源，研究适用于农牧交错带土地利用/覆被专题信息提取的数字图像处理技术和自动分类识别方法，并依据所获取的专题信息分析区域土地利用/覆被动态变化特征。【方法】结合实地调查和相关专题图件对原始图像进行系统的预处理，对比分析多种图像增强和特征变换方法增强目标类别光谱特征的效果，包括小波滤波、LBV变换等，然后利用多特征知识建立研究区各种土地利用/覆被类型的目视解译方法。分析各类别常用专题信息提取方法应用在本区出现的问题。分别在Ln{MNDWI}–NDVI、Albedo-NDVI、TM5-MNDWI特征空间中构建描述研究区水体、沙地和居民用地的专题指数CWI、CSI、CRI。验证面向对象分割的方法识别本区各个地类的适用性，并在对TM图像面向对象分割后对比分析最大似然法、BP神经网络法和支持向量机法的分类精度。在此基础上，结合多层级图像分析和面向对象分类的方法为各地类设计专题信息提取流程，并通过测试样本对两期TM图像专题信息提取结果进行目视评判和定量评价。最后利用两期专题信息提取结果得到土地利用/覆被转移矩阵、土地利用动态度和土地利用/覆被转换图谱，从土地利用/覆被数量变化、速度变化和类型变化3个方面分析研究区17年来土地利用/覆被的时空变化特征。【结果】分别对CWI、CSI、CRI进行阈值分割，可以快速分离出与居民用地、水体、沙地光谱特征相近的各种背景地类信息。在对TM图像进行SVM分类前执行面向对象分割操作，能够有效降低分类结果中的椒盐效应，适用于除沙地、荒草地以外的其它地类的专题信息提取过程。不同地类的专题信息提取方法均能够有效减少对同谱地物的误判，取得较高的制图精度和用户精度。【结论】横山县地处毛乌素沙漠东南部与黄土丘陵沟壑区的交接过渡地带，地形、地貌复杂，地表破碎，直接对TM图像分类得到各地类的提取精度有限，并且分类后的图斑散碎、椒盐效应显著，分类结果无法用于专题图制作和土地利用/覆被动态变化分析。试验依据多层级图像分析处理的思想为各个地类设计专题信息提取流程，实现了对各地类精准分布信息的逐步优化和逼近，应用在本区具有较高的可靠性和适用性。试验取得的主要研究成果如下：(1)选取Coiflet3作为小波母函数，采用软阈值函数对TM图像进行阈值滤波处理，能够在较好保持图像清晰度的同时，有效去除较大图斑内部孤立的像元和碎斑，增强图斑同质性，为比差值型光谱指数(NDVI、MNDWI、NDBI等)设置分割阈值和图像分类创造条件。在提取沙地、荒草地时设置小波分解尺度J=3；提取旱耕地、灌木林地、草地等类别时设置小波分解尺度J=2。(2)专题指数CRI、CWI和CSI图像中目标类别与背景类别具有显著的光谱差异，很多光谱微弱并受到相邻地物干扰的目标类别小斑，如细小水流线、道路线和居民点均能够被准确识别。分别对CRI、CWI和CSI图像进行阈值分割和掩膜运算，能够在完整保留目标类别的同时快速分离出大部分背景干扰信息。之后再对阈值分割结果分类，可有效简化分类过程的复杂性，准确区分同谱地物，提高分类结果的制图精度和用户精度。(3)对TM图像进行LBV变换能够增强水浇地、旱耕地、有林地、灌木林地、草地等类别的光谱特征，平滑较大图斑内部细微的光谱差异，提高图斑同质性、增强边缘特征，为准确选取面向对象分割尺度创造条件。(4) TM图像面向对象分割后的基本单元是由相邻、匀质像元组成的同质对象，对对象分类可以降低选取典型训练样本、设置组内聚类中心的难度，有效抑制椒盐效应。在识别水体、居民用地、水浇地、旱耕地、灌木林地和有林地的过程中，试验通过从全图多个位置选取典型训练样区并进行分割测试，确定分割尺度SC。识别上述类别时SC分别取6.2、9.0、7.0、4.2和5.3。(5)相对于BP神经网络法、最小距离法和最大似然法，支持向量机法分类后的沙地、荒草地、草地、灌木林地、旱耕地的图斑相对完整、连续，得到的制图精度和用户精度更高。对于水体、居民用地和道路，试验组合数学形态学开、闭运算作为一对形态滤波器优化初始提取结果的二值图像，在保持各地类较大图斑形状特征基本完好的同时有效去除了噪声图斑，消除细小孔洞，补平缺损并连接断线，并且操作过程简捷、灵活。(6)横山县1986年草地、未利用土地、耕地、林地、水域、居民用地的面积分别为141185.61hm~2、128043.90hm~2、87037.38hm~2、61474.77hm~2、5687.82hm~2、137.70hm~2，未利用土地中沙地的面积比例为56.08%，耕地中旱耕地的面积比例为81.24%，林地中灌木林地的面积比例为84.90%。横山县2003年草地、林地、耕地、未利用土地、水域、居民用地、公路用地的面积分别为181424.88hm~2、84919.95hm~2、80475.30hm~2、72379.44hm~2、3747.51hm~2、394.38hm~2，225.72hm~2。林地中灌木林地的面积比例为88.15%，耕地中旱耕地的面积比例为72.73%，未利用土地中沙地的面积比例为42.06%。(7)横山县1986-2003年间沙地、荒草地、水域、旱耕地面积显著减少。分别有26.96%、23.56%的沙地净转变为草地和荒草地；41.53%、8.21%的荒草地净转变为草地和灌木林地；28.42%的水域面积转变为水浇地；14.22%、5.74%的旱耕地净转变为草地和灌木林地。横山县1986-2003年间水浇地、灌木林地、有林地、草地、居民用地面积增加。扣除逆转面积后分别有8.02%、7.36%、5.00%的水浇地来自灌木林地、水域和草地；13.18%、6.17%的灌木林地来自草地和荒草地；15.22%、12.58%的有林地来自草地和旱耕地；12.88%、10.67%的草地来自荒草地和沙地。31.36%、10.22%、的居民用地来自水浇地和旱耕地。更多还原

【Abstract】 【Objective】Extracting accurate land use/land cover change informantion based on TMimages acquired in Augest2,1986and August17,2003, chosing Hengshan County located inthe farming-pastoral ecotone of Northern Shaanxi as study area.【Method】Firstly, the remotesensing images were pretreated by field survey and thematic maps. After that, methods ofvisual interpretation for each LUCC type were established by multi-feature knowledge.Secondly, different feature transformation and image enhancement methods were investigated,including principal component analysis, tasseled cap transformtion, LBV transformtion,wavelet filtering etc., then comparative analysis of the results were carried out. Thridly,spectral indices of waters, sandy land and residential area were established and described.Validation of object-oriented image segmentation for each LUCC type was carried out. Theclassification accuracy of4methods were compared, including minimum distance, maximumlikelihood, BP artificial neural net and support vector machine. On this basis, the thematicclassification flows for each LUCC type were designed by combining the hierarchical theoryand object-oriented image segmentation. Subsequently, results were visually inspected andquantitatively evaluated by test samples. Evaluation indices included map accurary, user’saccuracy, Kappa coefficient and overall accuracy rate. Finialy, On the basis of comparison ofclassification results, analysis of LUCC changes in Hengshan County from1986to2003wasrealized from the aspects of quantity change, speed change and land type change【.Result】Thegeneral information of residential area, waters, sandy land and waste grassland extracted fromTM image is more accurate by using threshold segmentation of CWI, CRI and CSI. Exceptfor sandy land and waste grass land, the accuracy of object-oriented image classification ismuch better than that of pixel-based classification, as well as salt and pepper effect hasdecreased substantially. Using method of combining object-oriented image segmentation andhierarchical theory,the LUCC thematic information extracted from TM images is moreaccurate and reliable than that of direct supervised classification schema.【Conclusion】Hengshan County is located in the transition zone of the southeast of the Mu Us Desert andloess hilly and gully region, having complex terrain and topography. As a result, fragmentizedpatches, mixed pixel, shadow, metameric substance of same spectrum, metameric spectrum of same substance are ubiquity in TM images. These factors bring too many difficulties inextracting thematic information using direct supervised classification schema. Theseclassification results can’t be in the production of thematic maps and dynamic monitoring ofLUCC because of obvoius salt and pepper effect and low accuracy. Focused on theseproblems, this paper presents a thematic information extraction method of successiveapproximation. Based on this method the optimum sequence arrangement of hierarchicalextraction schema has been realized. And validation by visual inspection and quantitativeevaluation show that hierarchical extraction procedures is beneficial to decreasingmisclassification error the rate of wrong classification and rate of miss classification, theaccuracy and efficiency of LUCC thematic information extracion is improved substantially.Specific research content and its related innovation are carried out as following:(1) Wavelet soft-thresholding filtering to TM image data using Coiflet (order=3) asmother wavelet can give remarkably good results in removing isolated fine patches and pixelsinset bigger blocks and enhancing homogeneity of patches, while maintains image definitionwell. This set the stage for ratio-difference indices as NDVI, MNDWI, NDBI in settingthreshold or for classification. We choose3as the wavelet decomposition scale for extractingsand land and waste grassland, while use2for extracting dry land, shrubbery, grassland andother LUCC types.(2) Due to the marked differences of spectrum between target class and background inimages of CRI, CWI, and CSI index, the slight linear or area object of target class havingweak spectrum and subject to interference of other adjacent ground objects, such as thin river,road, and tiny residential area, can be able to recognition certainly. Combination of thresholdsegmentation and mask operation on CRI, CWI, and CSI images can retain the entire targetclass while separates background interference information from it quickly. Then furtherclassification to threshold segmentation can make the recognition more simply, distinguishobjects with same spectrum certainly, and improve the mapping accuracy.(3) LBV transformation can enhance spectral features of irrigable land, dry land,woodland, shrubbery, grassland and other vegetation type effectivly. Meanwhie it candecrease subtle spectral difference in large patchess. Homogeneity is improved and edgefeatures is also enhanced. Thus LBV transformation provides conditions for exactly choosingsegmentation scale in object-oriented method.(4) After object-oriented image segmentation, basic unit of remot sensing image is notthe pixel, but homogeneity object formed by adjacent and the same class type pixels.Classificating for object can make the selection of the training samples and setting clustering centers easier. In addition, it is favorable to decrease salt and pepper effect in classificationresults. Thus the rationality and applicability about image classify are improved. In theprocess of identifying waters, residential area, irrigable land, dry land, shrubbery, andwoodland, segmentation scale (SC) is confirmed by choosing the classic training samplingareas form multiple positions of the whole imagery, and the segmentation scales are6.2,9.0,7.0,4.2and5.3respectively.(5) The overall accuracy of support vector machine classification is much better than thatof minimum distance, maximum likelihood and BP artificial neural net, as well as betterintegrity and continuity of patches. Morphological bandpass filter is constructed by usingmorphological open-close operation, and it has a good applicability on optimizing binaryimage of classification results. In post classification phase, open-close operation ofmathematical morphology can filter noise speckles, connect broken lines, fill the holes inorder to improve precision of image classify, meanwhile keeping original shape of largpatches. The operating process of open-close operation has high calculating efficiency,especially for road, residential area and waters extraction.(6) In1986, grassland, unusable land, cultivated land, forest land, waters, residential areain Hengshan County covered141185.61hm~2,128043.90hm~2,87037.38hm~2,61474.77hm~2,5687.82hm~2,137.70hm~2respectively, which were account for33.33%,30.23%,20.55%,14.51%,1.34%,0.03%of the whole area of the county.56.08%of unusable landis sandy land;81.24%of cultivated land is dry land;84.90%of forest land is shrubbery. In2003, grassland, forest land, cultivated land, unusable land, waters, residential area, roadHengshan County covered181424.88hm~2,84919.95hm~2,80475.30hm~2,72379.44hm~2,3747.51hm~2,394.38hm~2,225.72hm~2respectively, which were account for42.83%,20.05%,19.00%,17.09%,0.89%,0.09%,0.05%of the whole area of the county.88.15%offorest land is shrubbery;72.73%of cultivated land is dry land;42.06%of unusable land issandy land.(7) In the period of1986to2003, the coverage area proportion of sandy land, wastegrassland, wters, dry land had certain reduction. Transfer matrix calculations indicated that:the decreased sandy land converted mainly to grassland, waster grassland, which wereaccount for26.96%,23.56%of sandy land area in1986respectively; the decreased wastergrassland converted mainly to grassland, shrubbery, which were account for41.53%,8.21%of waster grassland area in1986respectively; the decreased waters converted mainly toirrigable land, which were account for28.42%of waters area in1986; the decreased dry landconverted mainly to grassland, shrubbery, which were account for14.22%,5.74%of dry land area in1986respectively. From1986to2003, the coverage area proportion of irrigableland, shrubbery, woodland, residential area had certain increase. Transfer matrix calculationsindicate that: the increased irrigable land is converted mainly by shrubbery, waters, grassland,which were account for8.02%,7.36%,5.00%of the irrigable land area in2003respectively;the increased shrubbery is converted mainly by grassland, waster grassland, which wereaccount for13.18%,6.17%of the shurbbery area in2003respectively; the increasedwoodland is converted mainly by grassland, dry land, which were account for15.22%、12.58%of the woodland area in2003respectively; the increased grassland is converted mainly bywaster grassland, sandy land, which were account for12.88%,10.67%of the grassland areain2003respectively; the increased residential area is converted mainly by irrigable land, dryland, which were account for31.36%,10.22%of the residential area in2003respectively.更多还原