一种基于特征选择的面向对象遥感影像分类方法

Object Based Remote Sensing Image Classification Based on Feature Selection Method

针对GF—1多空间分辨率遥感数据空间信息丰富,传统影像分类方法无法满足实际应用需要的问题,提出了一种基于特征选择的面向对象遥感影像分类方法——object-RJMC算法,即在影像分割及特征提取的基础上,运用Relief F算法和J-M(Jeffries-Matusita)距离算法去除无关及冗余特征,筛选出适于各类别分类的特征,然后利用CART算法建立分类规则,完成分类过程。以GF-1号2 m、8 m和16 m空间分辨率的三组影像进行算法验证,并与object-CART和pixel-CART影像分类方法进行对比分析。实验结果显示object-RJMC算法的分类精度均高于object-CART和pixel-CART算法的分类精度;且对高空间分辨率的影像分类效果要优于对中低空间分辨率影像的分类效果。该算法减少了特征选择及规则建立的人工干预,克服了以像素为单位的分类算法中由于缺少空间邻域信息而产生孤立、离散、不连通分类结果的问题,可有效地提高GF-1遥感影像分类精度。

With the development of GF-1 multi spatial resolution satellite data,the traditional image classification method has been unable to meet the needs of practical application. Based on this problem,an object-oriented remote sensing image classification method based on feature selection object-RJMC algorithm is proposed. On the basis of image segmentation and feature extraction,relief F algorithm and Jeffries-Matusita distance algorithm are used to remove irrelevant and redundant features to select the features of each category,and classification rules are established by CART algorithm to complete the classification process. Groups of GF-1 2 m,8 m,16 m three different spatial resolution images are used to verify the algorithm and compare the accuracy and precision with objectCART method and pixel-CART method. These results illustrate that the RJMC algorithm classification method has higher classification accuracy than the object-CART classification method and the pixel-CART classification method. Moreover,the object-based method has better accuracy for high spatial resolution than in middle or low resolution images. The algorithm reduces the artificial interference of feature selection and rule establishment,and overcomes the problem of isolation,discrete and non-connected classification by the lack of spatial neighborhood information in the classification algorithm based on pixels,and could effectively improve the classification accuracy of GF-1 image.