基于最优特征空间构建的随机森林算法在WorldView-2影像分类中的适用性研究

目前面向对象的分类研究中,对于研究区影像的分割尺度问题多以试验者的多次试验以及主观推断为主,缺乏定量化的评价标准。同时,在对遥感影像分类的算法选择以及在分类过程中,有效特征空间的选取均存在一定程度的主观性。针对遥感影像面向对象分类过程中分割尺度选择盲目及分类空间构造主观性较强的问题,以World View-2遥感影像数据为例,首先利用改进的全局最优分割尺度的方法获取研究区影像的最优分割尺度,在此基础上选取了研究区分割对象的48个特征,利用OOB误分率对各个特征的重要性排序;然后按重要性顺序以5为步长讨论特征数量对分类精度的影响,构建了用于分类的最优特征空间;最后将采用最优特征空间的随机森林算法获得的最佳分类结果,与面向对象的最邻近像元、决策树以及支持向量机分类算法进行了比较。结果表明,用于分类的特征数量与分类精度之间,并不是简单的正相关关系;与面向对象的最邻近像元、决策树以及支持向量机分类算法相比,利用最优特征空间进行随机森林分类的分类精度最高,表明该方法更适合于高分辨率World View-2数据的分类。

The study for the applicability of the random forests classification algorithm in the WorldView-2 image based on the construction of optical feature space

In the study of the current object-oriented classification methods, the choice of the segmentation scale for the remote sensing image is always based on the multiple tests or subjective inference with little quantitative evaluation criteria. Also, the choices of the classification algorithm and effective feature space in the classification process are of some subjectivity. In order to solve these problems, the study of object-oriented random forests classification based on a worldview-2 remote sensing image is carried out, hoping to promote the random forest algorithm in worldview-2 image classification. Firstly, the optimal segmentation scale for the image is calculated based on the improved method calculating global optimal segmentation scale. Then 48 features of the objects are selected for classification using random forest classifier. The importance of the 48 features is sort through OOB (out of bag) misclassification rate. The effect of amount of features on the overall accuracy is discussed at 5 intervals and the optical feature space is constructed according the OOB misclassification rate. The results show that with the number of features increasing, the classification accuracy is not always getting higher. And compared with KNN, Decision Tree and SVM classifier, the random forests algorithm based on the optical feature space performs best, indicating its best applicability in the classification of WorldView-2 image.