Urban areas detection based on conditional random field and multiscale textural features

Due to the complexity of building composition and imaging condition, urban areas show complicated structural information in remotely sensed images. On the other hand, the structural information of each site in urban area depends on that of neighboring sites. In this paper, a discriminative model, conditional random field (CRF), is introduced to learn the dependencies and fuse the multi-scale textural information to detect urban areas. In addition, because of the redundancy in structural information, a feature selection method is employed to reduce the dimension of them before they are put into CRF model, decreasing time consumed in model learning and inferring. By using images of high spatial resolution as input, experiments are performed, indicating that CRF model outperforms SVM in urban areas detection in terms of accuracy, and that, through feature selection it can decrease time consumed in model learning and inferring and obtain competitive result with original data.