Abstract: The spectrum variation of high infraclass and low interclass in high-resolution remotely sensed imagery has seriously disturbed the process of imagery classification. This phenomenon is similar to the misunderstanding of document semantic information caused by synonym and antonym. To solve this problem, a new unsupervised classification algorithm for high spatial resolution remotely sensed imagery, which combines Gabor texture feature and PLSA model (Probabilistic Latent Semantic Analysis), is presented in this paper. Firstly, we extract homogeneous segments from original imagery through MeanShift segmentation. Secondly, Gabor texture features of every pixel in each region are extracted, and clustered into several visual words. Thus, in our case, the imagery segments correspond to the documents, the visual words used to describe the segments correspond to the words in the documents, and the categories to be discovered for each segment correspond to the topics of the documents. Finally, we use PLSA model to analyze each segment, and achieve the image classification by assigning the most likely category for them. The experimental results have shown that the approach can outperforms the existing algorithms in terms of classification accuracy.