Training dataset generation is a difficult and expensive task for LiDAR point classification, especially in the case of large area classification. We present a method to automatically extent a small set of training data by label propagation processing. The class labels could be correctly extended to their optimal neighbourhood, and the most informative points are selected and added into the training set. With the final extended training dataset, the overall (OA) classification could be increased by about 2%. We also show that this approach is stable regardless of the number of initial training points, and achieve better improvements especially stating with an extremely small initial training set.