Learning-Based Super-Resolution Land Cover Mapping with Additional Transformed Examples

Abstract

Super-resolution mapping (SRM) aims to generate a fine spatial resolution land cover map from input coarse spatial resolution fraction images. The spatial prior model used to describe the spatial land cover patterns at the fine spatial resolution is crucial to the SRM analysis. At present, the learning-based SRM algorithm has shown its advantage, because more information of spatial land cover patterns can be captured from available training fine spatial resolution land cover maps. In practice, for learning-based SRM, the training fine spatial resolution land cover maps should include various spatial land cover pattern examples as rich as possible. However, gathering training fine spatial resolution land cover maps is always a hard work. In order to overcome this shortcoming, this study proposes an approach to provide additional transformed examples (land cover maps) in the learning-based SRM approach. By transforming the original fine spatial resolution land cover maps with rotation and mirroring operations, the number of available training examples can increase eight times. The proposed SRM algorithm is compared with several popular SRM algorithms using both synthetic and real images. Experimental results indicate that more spatial details can be produced when the tranformed samples are applied in the learning-based SRM algorithm and the result produced by the proposed method has higher accuracies than the SRM results used for comparison.