A Novel Semi-Supervised Method for Obtaining Finer Resolution Urban Extents Exploiting Coarser Resolution Maps

Abstract

In this work, we present a new semi-supervised strategy for obtaining finer spatial resolution urban maps from coarser resolution satellite data. Our method first uses a coarse resolution map as a source of training data. Then, we use semi-supervised learning in order to refine the set of initial (labeled) training samples by the inclusion of additional (reliable) unlabeled samples at the finer resolution level, in fully automatic fashion. The new unlabeled samples are automatically generated by our proposed methodology, which only requires a limited number of initial labeled samples for initialization purposes. Then, we conduct land cover classification (at the finer spatial resolution level) using a probabilistic multinomial logistic regression (MLR) classifier-in both supervised and semi-supervised fashion-by considering different numbers of labeled and unlabeled samples. In order to exploit spatial information, we use a Markov random field (MRF)-based postprocessing strategy to refine the obtained classification results. In order to test our concept, we use a global dataset: the European Space Agency's GlobCover product, as the coarser resolution map (300-m spatial resolution). Our experimental evaluation is further conducted using Landsat data (30-m spatial resolution) collected over three different locations in the city of Sao Paulo, Brazil, and over two different locations in the city of Guangzhou, China. We obtain promising results in the generation of finer resolution urban extent maps using very limited training samples, derived in all cases from the GlobCover product. These experiments suggest the potential of GlobCover to provide reliable training data in order to support mapping of urban areas at a global scale.