Remote sensing image semantic labeling based on conditional random field

Remote sensing images exhibit abundant information and complicated texture, and remote sensing image semantic labeling provides important information and clue for the subsequent object recognition, detection, scene analysis and high-level semantic extraction, which makes it a significant and extremely challenging task in remote sensing image understanding field. To address the task of remote sensing image semantic labeling, we propose to utilize the conditional random field (CRF) framework to model the low-level features and context information in remote sensing images. A texture descriptor 'Texton' is combined with the association potential in CRF framework to capture the texture layout in remote sensing images. 'Texton' feature selection and model parameter learning are carried out by employing the joint Boosting algorithm. And color information in Lab color space is used in the interaction potential in CRF to describe the color context. Then a Graph Cut algorithm is utilized to infer the CRF model to get the automatic semantic labeling results of the images. At the same time, we establish an optical remote sensing image database Google-4 and label all the images by manual annotation. Experimental results on Google-4 show that the CRF modeling scheme combined with `Texton' and color feature can accomplish the semantic labeling task of remote sensing images more accurately compared to the support vector machine (SVM)-based methods.