Super Resolution Mapping with Multiple Point Geostatistics

Abstract

Super-resolution is the process of providing fine scale land cover maps from coarse-scale satellite sensor information. Such a procedure calls for a prior model depicting the spatial structures of the land cover types. When available, an analog of the underlying scene (a training image) may be used for such a model. The snesim (single normal equation simulation) algorithm allows extracting the relevant pattern information from the training image and uses that information to downscale the coarse fraction data into a simulated fine scale land cover scene. First, the coarse fraction observed image is downscaled using block indicator Kriging (BIK), with the fine scale indicator variograms computed from the training image. The resulting downscale fractions at any given location are interpreted as a prior probability of having a specific land cover at that location. That prior probability is then merged with a probability lifted from the training image. That latter probability is made conditional to any available fine scale data and previously simulated class data along the simulation path. Land cover types are drawn from the resulting posterior distribution. A servo-system keeps track of the number of simulated classes inside each coarse fraction and assures exact reproduction of the coarse fraction data. By repeating the process with a new path visiting the simulation grid, one can generate several super resolution maps and explore the space of uncertainty for the fine scale land cover. The proposed snesim super resolution mapping algorithm allows to i) exactly reproduce the coarse fraction, ii) inject the structural model carried by the training image, and iii) condition to any available fine scale ground observations. A case study is provided to illustrate the proposed methodology using Landsat TM data from SE China.