Estimating location of land cover patch in super-resolution mapping by hopfield neural network

Super-resolution mapping (SRM) aims to locate subpixel class fractions geographically in the area represented by a mixed pixel. The accuracy of small sub-pixel class patches are represented by the popular SRM method is explored. It is shown that the accuracy of predicted patch location from the Hopfield Neural of SRM is a function of patch size. Specifically, the accuracy with which patch location is predicted varies inversely with patch size, with very small patches subject to large mis-location errors. A means to reduce the magnitude of mis-location error through the use of multiple sub-pixel shifted imagery is illustrated and the implications to popular site-specific accuracy assessment discussed. The use of multiple subpixel shifted images was able to reduce the error in patch location by more than half for very small patches and represents a simple but effective enhancement to SRM applications.