基于多分类器结合的IKONOS影像城市植被类型识别

采用多分类器结合的方法对城市植被进行分类.首先,以分割获得的城市植被分布斑块为处理基元,在不同特征空间中采集不同的样本,通过ISODATA、马氏距离、最大似然、人工神经网络和专家系统法进行分类,并计算各分类结果的关联程度和各植被类型识别的先验概率;然后利用专家投票的大多数规则对分类结果组合,未分类的对象按照先验识别概率最高的结果归类.精度评价表明:多分类器结合方法显著提高了信息识别的能力;采用多分类器结合的方法比单个分类器获得的最高分类精度提高5.5%,Kappa系数提高7.4%;Z统计值均为负,且均通过95%的置信水平检验.

Urban vegetation categories recognition by multiple classifier system from IKONOS imagery

In order to improve urban vegetation classification results,the method of multi-classifier systems was employed.Firstly,taking the "objects" from segmentation as the classification units,five kinds of classifiers of ISODATA(self-organizing data analysis techniques algorithm),Mahalanobis distance,maximum likelihood,neural network and expert system were used to identify vegetation types using different samples in different feature spaces.Secondly,the Q statistic representing relationships between a pair of classifiers and the pre-probability of all categories of vegetation for the five classifier results were calculated.Thirdly,the plurality voting rule was used to combine the five vegetation maps,and the unclassified "objects" was estimated by the category with the highest pre-probability.The results of accuracy assessment show that the vegetation map by plurality voting rules has higher total accuracy and Kappa coefficient,which are respectively 5.5% and 7.4% higher than the solo classifier with the highest accuracy.Z statistic values also show that the method of plurality voting rules greatly improves the classification results.