Multiscale adaptive reconstruction of missing information for remotely sensed data using sparse representation

Abstract

ABSTRACTDue to the influence of sensor malfunction and poor atmospheric condition, missing information is inevitable in optical remotely sensed (RS) data, which limits the availability of RS data. To tackle the inverse problem of missing information recovery, a multiscale adaptive patch reconstruction method was proposed in this letter. Multiscale dictionaries were learned from different sizes of exemplars in the known image region, which were later utilized to infer missing information patch-by-patch via sparse representation. Structure sparsity was incorporated to encourage the filling-in of missing patch on image structures and determine the patch size for further inpainting. Neighboring information was employed to restrain the appearance of the estimated patch, to yield semantically consistent inpainting result. In view of these ideas, we formulate the optimization model of adaptive patch inpainting and reconstruct missing information through a multiscale scheme. Experiments are performed on cloud removal, gaps filling and quantitative product reconstruction, which demonstrate that our method can well preserve spatially continuous structures and consistent textures without artifacts.