Skip to Main Content
Super-resolution mapping (SRM) is a promising technique to generate a fine resolution land cover map from coarse fractional images by predicting the spatial locations of different land cover classes at subpixel scale. In most cases, SRM is accomplished by using the spatial dependence principle, which is a simple method to describe the spatial patterns of different land cover classes. However, the spatial dependence principle used in existing SRM models does not fully reflect the real-world situations, making the resultant fine resolution land cover map often have uncertainty. In this paper, an example-based SRM model using support vector regression (SVR_SRM) was proposed. Without directly using an explicit formulation to describe the prior information about the subpixel spatial pattern, SVR_SRM generates a fine resolution land cover map from coarse fractional images, by learning the nonlinear relationships between the coarse fractional pixels and corresponding labeled subpixels from the selected best-match training data. Based on the experiments of two subset images of National Land Cover Database (NLCD) 2001 and a subset of real hyperspectral Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) image, the performance of SVR_SRM was evaluated by comparing with the traditional pixel-based hard classification (HC) and several existing typical SRM algorithms. The results show that SVR_SRM can generate fine resolution land cover maps with more detailed spatial information and higher accuracy at different spatial scales.