面向对象影像分类中基于最大化互信息的特征选择

FEATURE SELECTION BASED ON MAXIMAL MUTUAL INFORMATION CRITERION IN OBJECT-ORIENTED CLASSIFICATION

高分辨率影像面向对象分割后产生了大量的光谱、形状以及纹理特征,如何抽取出最佳特征子集是遥感影像识别的重要问题。本文利用最大化互信息统计独立准则抽取最优特征子集,提高了面向对象遥感影像分类精度。基本过程包含以下3个方面:首先,利用eCoginition软件对高分辨遥感影像进行对象分割;然后,基于互信息最大关联、最小冗余准则(mRMR)获取优选的特征子集;最后,基于支持向量机分类器完成影像分类。以福建省漳州市QuickBird数据为例的实验表明,该方法能够有效提高遥感影像的分类精度,平均误分率降低了约4%。

It is a key problem to select optimal features from the total set where spectral, geometric, shape, texture features and some other features are extracted by the process of image segmentation in object - oriented classification. In this paper, the authors present a method for selecting good features from object - oriented image segmentation according to the maximal statistical mutual information dependency criterion so as to improve the classification accuracy of high spatial resolution image. The proposed method is a three - step classification routine that involves the integration of （1） image segmentation with eCoginition software, （2） feature selection by mutual information minimum redundancy and maximum relevance criterion, and （3） support vector machine for classification. The experiment was conducted on QucikBird image in Zhangzhou city, Fujian province. Furthermore, the proposed method and the well - known feature selection methods such as Tabu search algorithm and fisher discriminate analysis are evaluated and compared with each other. The result shows that the mean error ratio decreases by 4% with the proposed method and that the proposed method for feature selection outperforms the other methods in terms of McNamara＇ s test.