基于Block-Gibbs抽样的无限潜Dirichlet分配模型的高分辨率全色遥感影像非监督分类

Unsupervised Classification of High-resolution Panchromatic Remote Sensing Image Based on Infinite Latent Dirichelt Allocation Using Block-Gibbs Sampling

通过引入文本检索算法中的无限潜Dirichlet分配(infinite Latent Dirichlet Allocation,即iLDA)模型,对遥感影像进行建模以获取地物的统计分布及其共生关系,从而实现遥感影像非监督分类。首先,将遥感影像有重叠地划分成一组大小相等的影像块(文集)。其次,以iLDA为基础,构建"像元"(视觉词)、"影像块"(文档)和"地物类"(主题)之间的条件概率关系,并采用Block-Gibbs抽样的方法来估计模型参数,从而构建基于BlockGibbs抽样的iLDA遥感影像非监督分类模型(Block-Gibbs based iLDA,即BG-iLDA)。最后,通过对BG-iLDA模型的逼近求解实现高分辨率遥感影像的非监督分类。实验结果表明,本文提出的基于BG-iLDA的面向对象非监督分类方法相对传统的K-means等算法精度更高,更能有效区分"同谱异物"的地物。

In this paper,the infinite Latent Dirichlet Allocation (iLDA)model for unsupervised classification of images is introduced.An effective unsupervised classification method using the semantic information and the symbiotic relationship from iLDA is proposed,which is used for high-resolution panchromatic images.Firstly,the image corpus is structured by overlapped segmentation of the image into sub-images.Secondly,the relationship of conditional probability among pixels (visual-words), sub-images (documents)and land objects (topics)is built.By which,the proposed method using Block-Gibbs based iLDA (BG-iLDA)is modeled.And the model parameters are estimated using the Block-Gibbs sampling.Finally,the unsupervised classification of high-resolution panchromatic images is realized by approximate solution of the BG-iLDA.Experimental results show the classification precision of the proposed method is better than the K-means method,and the effect of the different object with the same spectral characteristics is appropriately displayed by the classification result.