Abstract: In order to solve the problem that the feature dimension is too high in the process of road extraction, a kind of feature selection method based on ReliefF filtering and Wrapper encapsulation was put forward. The particle swarm optimization algorithm (PSO) was used as the search algorithm of Wrapper, and the optimized random forest algorithm (OPRF) was used as the classifier of Wrapper to form the PSO OPRF encapsulated subsets evaluator. The feature subset selected by ReliefF was evaluated, the feature dimension was reduced, and the optimal feature subset was selected. According to the selected features, the urban road network was extracted by multilevel segmentation and classification. Taking GF-2 remote sensing images of some urban areas of Taiyuan city, Shanxi province as data sources, the road extraction quality obtained by using the proposed feature selection method was compared with that obtained by using only ReliefF algorithm, using traditional random forest as a classifier, and using J48 decision tree as a kind of method on classifier feature selection. The results show that the road extraction quality is the best for the proposed method, and the extraction qualitiy of three types of road are up to 0.959, 0.853 and 0.931, respectively.

Key words: road extraction, feature selection, data mining, multi-level segmentation, random forest, canny operator