ABSTRACT:For extraction of road pixels from combined image and elevation data, Wegner et al. (2015) proposed classification of superpixels into road and non-road, after which a refinement of the classification results using minimum cost paths and non-local optimization methods took place. We believed that the variable set used for classification was to a certain extent suboptimal, because many variables were redundant while several features known as useful in Photogrammetry and Remote Sensing are missed. This motivated us to implement a variable selection approach which builds a model for classification using portions of training data and subsets of features, evaluates this model, updates the feature set, and terminates when a stopping criterion is satisfied. The choice of classifier is flexible; however, we tested the approach with Logistic Regression and Random Forests, and taylored the evaluation module to the chosen classifier. To guarantee a fair comparison, we kept the segment-based approach and most of the variables from the related work, but we extended them by additional, mostly higher-level features. Applying these superior features, removing the redundant ones, as well as using more accurately acquired 3D data allowed to keep stable or even to reduce the misclassification error in a challenging dataset.
Road pixel segmentation in airborne data is an important and challenging task. Recently, a sophisticated and robust approach based on superpixels and minimum cost paths has been published. In order to find out which of the numerous features are most essential, we propose a forward-search wrapper approach for feature selection which was tested with two different classifiers and with both generic and customized features. Path connecting the superpixels with a high probability of being roads are then established, filtered, and included as higher order cliques into the non-local energy minimization module thus enforcing the connectivity of the resulting road network. Two minor contributions are adjustment of a segmentation algorithm for multichannel images and an efficient application of the Dijkstra minimum path method for a sparse set of start and end nodes. The results show that both classifiers yield quite different feature sets, which speaks in favor of wrapper approaches. Also, the customized features were ranked among the most relevant, thus emphasizing the importance of sensor data fusion and higherlevel semantical concepts.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.