Super-Resolution Land Cover Mapping Based on Multiscale Spatial Regularization

Super-resolution mapping (SRM) is a method for allocating land cover classes at a fine scale according to coarse fraction images. Based on a spatial regularization framework, this paper proposes a new regularization method for SRM that integrates multiscale spatial information from the fine scale as a smooth term and from the coarse scale as a penalty term. The smooth term is considered a homogeneity constraint, and the penalty term is used to characterize the heterogeneity constraint. Specifically, the smooth term depends on the local fine scale spatial consistency, and is used to smooth edges and eliminate speckle points. The penalty term depends on the coarse scale local spatial differences, and suppresses the over-smoothing effect from the fine scale information while preserving more details (e.g., connectivity and aggregation of linear land cover patterns). We validated our method using simulated and synthetic images, and compared the results to four representative SRM algorithms. Our numerical experiments demonstrated that the proposed method can produce more accurate maps, reduce differences in the number of patches, visually preserve smoother edges and more details, reject speckle points, and suppress over-smoothing.