基于决策树模型的高分影像分类研究

发布时间：2018-01-18 08:33

本文关键词：基于决策树模型的高分影像分类研究　出处：《新疆大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来随着航天遥感卫星技术的不断发展,使得遥感影像的光谱、空间及时间分辨率都得到大大提高。特别是高分影像的出现,使得从遥感影像中提取地物信息成为基础数据获取的重要手段。但如何准确的识别高分影像中的地物信息,是当前遥感影像地物识别分类技术中面临的一个热点问题。传统的基于影像像元的分类方法中,主要依据统计学相关理论对不同地物的光谱特征进行统计分析进而实现分类,若将其应用在高分影像中,会出现“同物异谱,异物同谱”现象,往往难以达到理想的地物识别分类效果。本文研究了基于决策树模型的面向对象影像地物识别分类技术。面向对象的分类是以影像中多个像素的集合为分析目标,突破传统的影像分类中以单个像元为分析单元的局限性。基于像素集合为分析目标可以充分考虑地物对象多方面的特征,例如影像的光谱特征、形状特征及纹理特征等,借助决策树模型统计得到不同地物的分类特征,构建特征空间并运用不同分类器进行影像分类研究。本研究以额敏县阔什比克良种场的多源融合影像作为数据源,借助ENVI、eCognition、ArcGis等数据处理工具进行如下研究:(1)对多源融合影像的多尺度最优分割参数进行了详细的分类对比实验,通过对多尺度分割中分割的尺度、波段权重、形状因子及紧致度等参数的实验对比,最终通过目视效果确定了研究区影像的多尺度最优分割参数。(2)通过决策树模型统计不同地物的分类特征,将不同地物的分类特征构建特征空间集,运用K最近邻分类器与CART决策树分类器进行分类对比实验,实现多尺度下地物对象信息的分类。(3)将面向对象的分类结果与传统的最小距离分类结果分别进行精度评价,得出面向对象的分类精度明显好于传统的分类方法得到的分类结果。
[Abstract]:In recent years, with the development of space remote sensing satellite technology, the spectral, spatial and temporal resolution of remote sensing images have been greatly improved, especially the appearance of high-score images. It makes the extraction of ground object information from remote sensing image become an important means of basic data acquisition. But how to accurately identify the ground object information in high score image. It is a hot issue in the current remote sensing image recognition and classification technology. Traditional classification method based on image pixel. According to the related theory of statistics, the spectral characteristics of different ground objects are statistically analyzed and classified. If they are applied in high score images, there will be the phenomenon of "the same object and the same spectrum". It is difficult to achieve the ideal classification effect of ground object recognition. This paper studies the object oriented image recognition and classification technology based on decision tree model. The object oriented classification is based on the set of multiple pixels in the image. . Break through the limitation of traditional image classification which takes single pixel as the analysis unit. Based on the pixel set as the analysis target, we can fully consider the features of the object, such as spectral features of the image. Shape features and texture features, with the help of decision tree model statistics to obtain the classification of different features. Constructing feature space and using different classifiers for image classification. This study uses multi-source fusion images of good seed farm in Eimin County as the data source, with the help of ENVIZE Cognition. ArcGis and other data processing tools are as follows: 1) the multi-scale optimal segmentation parameters of multi-source fusion images are classified and compared in detail, and the scale of segmentation in multi-scale segmentation is analyzed in detail. The experimental comparison of the parameters such as band weight, shape factor and compactness. Finally, the multi-scale optimal segmentation parameter of the study area image is determined by visual effect. (2) through the decision tree model, the classification features of different ground objects are analyzed, and the feature space sets are constructed. The K-nearest neighbor classifier and CART decision tree classifier are used for classification comparison experiment. The classification of object information in multi-scale is realized. (3) the accuracy of the object oriented classification and the traditional minimum distance classification are evaluated respectively. The accuracy of object-oriented classification is better than that of traditional classification method.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】