Super-resolution land cover classification using the two-point histogram

A geostatistical optimization algorithm is proposed for super-resolution land cover classification from remotely sensed imagery. The algorithm requires as input, a soft classification of land cover obtained from a remotely sensed image. A super-resolution (sub-pixel scale) grid is defined. The soft land cover proportions (pixel scale) are then transformed into a hard classification (subpixel scale) by allocating hard classes randomly to the sub-pixels. The number allocated per pixel is determined in proportion to the original land cover proportion per pixel. The algorithm optimizes the match between a target and current realization of the two-point histogram by swapping sub-pixel classes within pixels such that the original class proportions defined per pixel are maintained. The algorithm is demonstrated for two simple simulated images. The advantages of the approach are its ability to recreate any target spatial distribution and to work with features that are both large and small relative to the pixel size, in combination.