Geostatistical solutions for super-resolution land cover mapping

Abstract

Super-resolution land cover mapping aims at producing fine spatial resolution maps of land cover classes from a set of coarse-resolution class fractions derived from satellite information via, for example, spectral unmixing procedures. Based on a prior model of spatial structure or texture that encodes the expected patterns of classes at the fine (target) resolution, this paper presents a sequential simulation framework for generating alternative super-resolution maps of class labels that are consistent with the coarse class fractions. Two modes of encapsulating the prior structural information are investigated—one uses a set of indicator variogram models, and the other uses training images. A case study illustrates that both approaches lead to super-resolution class maps that exhibit a variety of spatial patterns ranging from simple to complex. Using four different examples, it is demonstrated that the structural model controls the patterns seen on the super-resolution maps, even for cases where the coarse fraction data are highly constraining.