Search

For multi-resolution land covering classification, many researches have focused on selecting and integrating appropriate feature information from different spatial resolution data of the same area. However, when extending to large scale problems, it is no surprise that low resolution data has worse performance, and high resolution data with wide coverage area has more limitations. To solve this problem, a novel framework is presented which compounds multiple spatial resolution data at arithmetic level without the limitation of full-scale multi-resolution data. The framework allows integrating conditional random fields (CRFs) with "real" likelihood distribution. Discrete feature-likelihood mapping is proposed to represent multi-to-single spatial correspondence. By considering spatial contextual information between pixels, CRFs based classifier offers a robust and accurate framework. Our experiments show that the proposed method can greatly improve the accuracy for large scale land covering classification applications.