Texture segmentation for remote sensing image based on texture-topic model

Textures of land covers provide significant evidences for segmentation and classification. Inspired by resent researches on topic model, we work on a novel texture segmentation method for very high resolution (VHR) remote sensing images based on Latent Dirichlet Allocation (LDA). In order to model spatial relationship between words in LDA, a constraint random variable which is used to control the selection of neighboring features of each specific texture is introduced to the model. The proposed method is evaluated on segmenting remote sensing images by finding the homogeneous regions in texture-topic map. The experimental results show our method has great potential for remote sensing image segmentation.