Active Learning based on Random Forest and Its Application to Terrain Classification

In this paper, a novel active learning technique was proposed for solving multiclass classification problem with random forest classifier. By combining uncertainty, density, and diversity criteria, the most informative samples are selected for manually labeling. The uncertainty criterion is implemented by analyzing the difference between the most votes and second most votes from classifier’s output. Samples in dense regions are thought to be more informative than samples in sparse regions. The average distance of a sample to its k-nearest unlabeled neighbors is computed to describe the sample’s density. The distance between a sample and its nearest labeled sample is used to measure the diversity of the sample. The larger the distance is, the less redundancy the sample is. To assess the effectiveness of the proposed method, it was compared with other techniques like traditional active learning based on random forest and SVM. The results of the experiment on terrain classification have demonstrated the effectiveness of the proposed approach.