Downscaling of satellite remote sensing data: Application to land cover mapping

Many satellite images have a spatial resolution coarser than the extent of land cover patterns on the ground, leading to mixed pixels whose composite spectral response consists of responses from multiple land cover classes. Spectral unmixing procedures only determine the fractions of such classes within a coarse pixel without locating them in space. Downscaling, also known as super-resolution or sub-pixel mapping, turns these proportions into a fine resolution map of class labels.; Sub-pixel mapping is undetermined, in that many different fine resolution maps can lead to an equally good reproduction of the available coarse fractions. Thus, the unknown fine resolution land cover map is regarded as a realization of a random set. Simulated realizations are generated using the geostatistical paradigm of sequential simulation. At any pixel along a path visiting all fine scale pixels, a class label is simulated from a local probability distribution made conditional to: (i) the coarse class fraction data, (ii) any simulated land cover classes at fine pixels previously visited along that path, and (iii) a prior structural model.; Two algorithms using different structural model types are proposed for the sequential simulation. The first method proposed is built on block indicator cokriging which allows evaluating the previous local probability distributions by a form of kriging; the structural model is then a series of class labels indicator variograms. The second method is based on the multiple-point simulation algorithm SNESIM where the local probability distributions are read from a training image; the structural function is then that training image which can be seen as an analog image depicting the patterns deemed present at the fine resolution.; Two case studies derived from Landsat TM imagery demonstrates the two approaches proposed. The resulting alternative downscaled class maps all honor the coarse proportion data, any fine scale data available, and exhibit the spatial patterns called for by the input structural model. When that structural model is incompatible with the sensor data the pattern reproduction is poor. Fine scale data such as water, roads and previously mapped fine scale pixels are shown to be well reproduced in the downscaled maps.