基于聚类分析的多特征融合遥感图像场景分类

Scene classification of multi-feature fusion remote sensing image based on clustering analysis

高分辨率遥感图像相比于中低分辨率遥感图像,能够提供更详细的地物信息,但各种场景内物体和空间结构分布较复杂,针对场景分类中不同目标的特征有效性不尽相同、彼此存在互补现象,提出了一种基于聚类有效性分析的分层多特征融合场景分类方法。首先提取图像数据集的Gabor纹理与颜色直方图特征;其次聚类分析训练样本,选取最佳聚类个数,计算对应的各类别的聚类一致性,对于聚类一致性较好的子类训练分类器,并对其余的类别提取局部特征并进行频繁项集挖掘,然后训练基于精简特征的分类器;最后用这两个分类器对测试样本进行分类。实验结果表明,在一个2 100幅图像构成的大型遥感图像数中,提出的算法比仅用单一特征分类方法的最高精度要高;与其他融合方法相比,该方法取得了较高分类精度,达到了97.28%,算法时间复杂度也大为降低。

Compared with the low-resolution remote sensing images, high-resolution remote sensing images can provide more detailed ground object information, but the distribution of objects and spatial structures in various scenes is more complex, so the feature validity of different targets in scene classification is not the same. This paper proposes a hierarchical multi-feature fusion scene classification method based on clustering validity analysis. Firstly, the Gabor texture and color histogram features of the image data set are extracted, then the training samples are analyzed, the optimal number of clustering is selected, and the corresponding clustering consistency is calculated. For the subclass classifier with good clustering consistency, the local feature is extracted and the frequent item sets are mined for the other categories, then the classifier based on the reduced feature is trained. Finally, the two classifiers are used to classify the test samples. The experimental results show that the proposed algorithm is more accurate than the single feature classification method in the number of large-scale remote sensing images composed of 2 100 images, and achieves a higher classification accuracy of 97.28% compared with other fusion methods.