An Active Learning Method for DEM Extraction From Airborne LiDAR Point Clouds

• Published in: IEEE access Vol. 7; pp. 89366 - 89378
• Main Authors: Hui, Zhenyang, Jin, Shuanggen, Cheng, Penggen, Ziggah, Yao Yevenyo, Wang, Leyang, Wang, Yuqian, Hu, Haiying, Hu, Youjian
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Active learning; airborne LiDAR; Airborne sensing; Filtration; Laser radar; Learning; Lidar; Noise measurement; oracle; Photogrammetry; point cloud filtering; Remote sensing; Support vector machines; Surface morphology; Terrain; Three dimensional models; Three-dimensional displays; Training; training sample
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2019.2926497

Airborne Light Detection and Ranging (LiDAR) is a popular active remote sensing technology that has been developing very rapidly in recent years. To solve the problems of low filtering accuracy of airborne LiDAR point clouds in complex terrain environments and avoiding too much human intervention, this paper proposes a point cloud filtering method based on active learning. In the proposed method, the initial training samples are acquired and marked automatically by multi-scale morphological operations. In so doing, no training samples are selected and labeled manually, i.e., the training samples are added gradually according to the oracle used in active learning. In this paper, the oracle is set to a sigmoid function of residuals from the points to the fitted surface. Subsequently, the training model is revised progressively using the updated training samples. Finally, the classification results are further optimized by a slope-based method. Three datasets with different filtering challenges provided by the International Society for Photogrammetry and Remote Sensing (ISPRS) were used to test the proposed method. Comparing with the other ten famous filtering methods, the proposed method can achieve the smallest average total error (5.51%). Thus, it can be concluded that the proposed method performs very well toward different terrain environments.