基于遥感影像提取土地利用信息的决策树方法研究

发布时间：2018-07-05 11:22

本文选题：ALOS多光谱影像 + 决策树　；参考：《南京农业大学》2013年硕士论文

【摘要】：赤水河流域人地矛盾尖锐,土地利用方式不合理严重影响了流域土地资源可持续利用和生态环境保护,及时了解赤水河流域各区域的土地利用现状,对流域土地利用合理性分析、水土流失治理、生态环境保护等工作的开展具有重要的意义。探索一种能在研究范围内达到精度与效率统一的土地利用信息提取方法能为流域综合管理的相关工作提供重要支持。本研究以赤水河流域为研究区,以覆盖该区域的10景ALOS多光谱遥感影像为数据源,根据每景影像的覆盖范围将全流域划分为十个子研究区,选择其中一子研究区为试验区,在充分分析和统计八种典型地物在影像光谱特征、归一化差异植被指数(NDVI)、归一化差异水体指数NDWI、数字高程模型(DEM)及影像波段运算后特征值的数值差异的基础上,确定了区分地物的阈值,探索建立了以阈值为规则的二叉决策树模型,进行土地利用分类。将该思路和方法拓展到其余九个子研究区,并分析了影响不同影像分类精度的因子。主要研究结果如下： (1)构建了具有地域代表性的1号子研究区(试验区)的基于规则的决策树模型,该方法对八种地物分类结果的总体精度为89.05%,Kappa系数为0.8741,总体精度相比最大似然法、支持向量机法分别提高12.39%、10.78%；Kappa系数相比最大似然法和支持向量机法则分别提高0.1412、0.1238。总体来说,该模型在不同程度上减少了对林地、草灌、河流、水库坑塘、梯坪地、水田、建设用地这七种地物分类结果的错分、漏分误差,其中减少幅度以河流、水库坑塘最为明显,最大减少幅度分别为50.25%、46.71%。说明该方法在1号子研究区内具有较好的适用性和可操作性。 (2)将试验成功的方法与思路拓展至全流域范围,10个子研究区中,基于规则的决策树法在8个区域的分类结果总体精度和Kappa系数明显优于最大似然法,其分类结果最高总体精度达到90.59%,Kappa系数为0.8811,二者分别比最大似然法分别高出7.4%和0.0887,说明基于规则的决策树法在赤水河流域具有一定的普适性,可以有效提高地物的分类精度。 (3)根据本研究采集的属于赤水河流域范围内的156个GPS野外实测点,经过差分校正后建立数据库,对全流域提取出的土地利用图进行精度验证。结果表明,156个点位中有133个点位的地物分类正确,野外实测点验证精度为85.26%。 (4)数据源的时相对决策树模型的构建有影响。另外,从统计学方面讲,利用多元线性回归模型探讨得出：本研究共有10幅影像10个子研究区域,其分类结果的Kappa系数和区域总面积成反比,和水田面积成正比,也就是说,在10个不同区域内,某幅影像研究区域的区域总面积越大,分类精度越低；某幅影像研究区域的水田面积越大,分类精度越高。
[Abstract]:The contradiction between man and land in Chishui River Basin is sharp, and the unreasonable land use mode has seriously affected the sustainable utilization of land resources and the protection of ecological environment in the basin, so as to understand the current situation of land use in various areas of the Chishui River Basin in a timely manner. It is of great significance to analyze the rationality of land use, to control soil and water loss, and to protect the ecological environment. To explore a land use information extraction method which can achieve the unity of precision and efficiency within the scope of research can provide important support for the relevant work of integrated watershed management. Taking the Chishui River Basin as the study area and the 10 Alos multispectral remote sensing images covering the region as the data source, the whole basin is divided into ten sub-study areas according to the coverage of each scene image, and one of the sub-study areas is selected as the experimental area. Based on the analysis and statistics of the spectral characteristics, normalized differential vegetation index (NDVI), normalized differential water body index (NDWI), digital elevation model (Dem) and the numerical difference of the eigenvalues of the eight typical ground objects in the image band, The threshold value of distinguishing ground objects was determined and the binary decision tree model was established to classify land use. The method is extended to the other nine sub-study areas, and the factors that affect the classification accuracy of different images are analyzed. The main results are as follows: (1) the rule-based decision tree model of sub-research area No. 1 (experimental area) with geographical representation is constructed. The overall accuracy of this method for the classification of eight ground objects is 89.05 and the Kappa coefficient is 0.8741. Compared with the maximum likelihood method, the support vector machine method increases the maximum likelihood coefficient by 12.39 and the Kappa coefficient from the maximum likelihood method and the support vector machine method by 0.1412 and 0.1238, respectively. In general, the model reduces the misclassification and leakage errors of forest land, grass irrigation, rivers, reservoirs, terraces, paddy fields and construction land to varying degrees. The reservoir pit is the most obvious, the maximum reduction range is 50.25 and 46.71 respectively. The results show that the method has good applicability and maneuverability in the No. 1 sub-research area. (2) the successful test methods and ideas are extended to the whole watershed area and 10 sub-research areas. The total accuracy and Kappa coefficient of the rule-based decision tree method in eight regions are better than that of the maximum likelihood method. The highest overall accuracy of the classification is 90.59 and the Kappa coefficient is 0.8811, which is 7.4% and 0.0887 higher than that of the maximum likelihood method, respectively, which indicates that the rule-based decision tree method is universal in the Chishui River Basin. The classification accuracy of ground objects can be improved effectively. (3) according to 156 GPS field survey points which belong to the Chishui River Basin, the database is established after differential correction. The accuracy of the land use map extracted from the whole basin is verified. The results show that 133 of 156 points are correctly classified, and the accuracy of field measurement is 85.26. (4) the time of data source is relative to the construction of decision tree model. In addition, from the statistical point of view, using the multivariate linear regression model, it is concluded that the Kappa coefficient of 10 images in this study is inversely proportional to the total area of the region and is directly proportional to the area of the paddy field, that is, the Kappa coefficient of the classification results is inversely proportional to the total area of the paddy field. In 10 different regions, the larger the total area of an image is, the lower the classification accuracy is, and the larger the paddy field area is, the higher the classification accuracy is.
【学位授予单位】：南京农业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：P237

【参考文献】