随机森林算法在高分辨遥感影像土地覆被分类中的实现和应用

发布时间：2018-10-23 07:48

【摘要】：研究土地覆被变化已经成为研究全球变化以及研究现代地学的核心和热点,在土地覆被变化的研究中,土地覆被分类是最为基础和关键的一个环节。土地覆被分类需要强大的分类方法,随机森林是一个功能强大的机器学习分类器,是一种与传统的决策树相比基于非参数回归算法的集成学习方法。目前,随着众多高分辨率卫星陆续发射,高分影像数据量大,混合像元减少、地物信息增多,然而空间分辨率的提高并没有使得分类精度的提高,相反,会出现分类精度的降低。传统基于像元光谱信息分类的方法应用在这些高分影像分类时所得到的精确度远远达不到生产所需的要求。影像中地物信息增多体现在地物的纹理结构随着空间分辨率的提高而变得非常清晰,提取稳定的鉴别性强的纹理特征对提高影像分类精度有十分重要的意义。本文以石台县的两个乡镇为研究区域,选用Rapideye高分影像作为数据源。研究区景观空间异质性程度高,阴影覆盖区域面积大,实地调查中部分类别在影像上显示的颜色可分性不强。为此,本研究采用已广泛地应用于遥感影像的纹理提取的地质统计学方法提取纹理特征,对原始波段进行Band Math运算提取植被指数,应用随机森林计算特征重要性筛选特征,对改变树的数量、特征组合与分类精度间的关系进行试验,并与最大似然法分类结果进行比较。结果表明:(1)随机森林泛化误差随着树的数量(N)增加而收敛于定值,分类精度随着N值增大而无显著增加,并且计算机运算效率降低。本研究中最终选取N=500,既能满足分类精度要又能保证运算效率;(2)应用随机森林和最大似然法对加入纹理特征组合进行分类精度评价,Kappa系数分别为0.7134和0.6315,高于加入植被指数特征组合和对纹理和植被指数特征组合。应用随机森林算法分类精度明显高于应用最大似然法所得的分类精度。此外,纹理信息可以在一定程度上提高分类的精度,对于耕地、建筑等几何形状明显且规律性强的地类分类精度有显著提高,对光谱值接近的不同地类之间的区分提供了有效的依据。随机森林算法综合表现更优,它在保证分类精度的前提下,也能保证运算的效率,更适宜实际生产应用,有一定的应用价值。同时,操作方便,由计算机计算筛选所需特征个数及所需的特征,还能预测满足分类精度要求的N值。最大似然法操作步骤亦不繁琐,但是精度相对较低。
[Abstract]:The study of land cover change has become the core and hotspot of global change and modern geoscience. In the study of land cover change, land cover classification is the most basic and key link. Land cover classification needs powerful classification method. Stochastic forest is a powerful machine learning classifier, which is an integrated learning method based on non-parametric regression algorithm compared with traditional decision tree. At present, with the launch of many high-resolution satellites one after another, the large amount of high-resolution images, the reduction of mixed pixels and the increase of ground object information, however, the improvement of spatial resolution does not improve the classification accuracy. There will be a reduction in classification accuracy. The traditional methods based on pixel spectral information classification can not meet the requirements of production when they are applied to the classification of high-score images. The increase of ground object information is reflected in the fact that the texture structure of the ground object becomes very clear with the improvement of spatial resolution. It is very important to extract stable and discriminative texture features to improve the classification accuracy of the image. In this paper, two villages and towns in Shitai County were selected as data sources. The spatial heterogeneity of landscape in the study area is high, the area of shadow coverage is large, and the color separability of some categories in the field survey is not strong. Therefore, in this study, the geostatistics method, which has been widely used in the texture extraction of remote sensing images, is used to extract the texture features, the vegetation index is extracted by the Band Math operation of the original wave band, and the importance of the feature is screened by the random forest calculation. The relationship between the number of changing trees, the combination of features and the classification accuracy is tested and compared with the results of maximum likelihood method. The results show that: (1) the generalization error of random forest converges to the fixed value with the increase of the number of trees (N), the classification accuracy increases with the increase of N value, and the computer operation efficiency decreases. In this study, the final selection of NX500 can not only meet the classification accuracy but also ensure the operational efficiency. (2) applying the random forest and maximum likelihood method to evaluate the classification accuracy of texture feature combination, the Kappa coefficients are 0.7134 and 0.6315 respectively, which is higher than that of adding planting. By exponential feature combination and combination of texture and vegetation index features. The classification accuracy of the stochastic forest algorithm is obviously higher than that of the maximum likelihood method. In addition, texture information can improve the classification accuracy to a certain extent. It provides an effective basis for the differentiation of different ground types with close spectral values. The stochastic forest algorithm has a better comprehensive performance. It can also guarantee the efficiency of the operation and is more suitable for practical production and application under the premise of ensuring the classification accuracy. It has certain application value. At the same time, the operation is convenient, the number of the required features and the required features can be calculated by the computer, and the N value that meets the classification accuracy can also be predicted. The maximum likelihood method is not complicated, but the precision is relatively low.
【学位授予单位】：安徽农业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：S771.8

【相似文献】