Model transfer-based filtering for airborne LiDAR data with emphasis on active learning optimization

This paper proposes a novel airborne LiDAR (Light Detection And Ranging) filtering method which combines transfer learning (TL) theory with active learning (AL) to refine the filtering result. This method requires that a reliable training set is available only for classified datasets (source domain) and not for another dataset to be filtered (target domain). Initial training set includes all points in the source domain. Then this training set is optimized by an AL procedure. At each iteration of AL, the most informative samples are selected from target domain and added into training set. After several times of iteration, target LiDAR dataset is filtered. Our method focuses on model transfer and aims at taking advantage of the already available knowledge on the source domain to supervise correct and fast classification of the target domain. Experimental results show that our method is instructive and effective to provide a reliable training set and that combining with AL strategies can achieve higher accuracy (type II errors all lower than 8.28% when batch size reaches 5000) and faster convergence, compared to random sampling.