Active learning for semantic labelling of airborne LIDAR data

Creating training data for classification of airborne LIDAR data is expensive and time-consuming. To label training data, ground truth data is gathered via field surveys or human photo-interpretation of aerial imagery. To avoid getting poor end results due to insufficient training data, organizations often label more training data than is actually needed – at a large expense. Using a semi-supervised, active learning approach for both the segmentation and classification of human-made objects and vegetation in urban, airborne LIDAR point clouds, as is proposed in this work, allows a minimal training data set to be created, tested, and expanded in key areas, as-needed, in an interactive, iterative process. The active learning iterations for segmentation gather linkage constraints to apply on the hierarchical clustering. Then, the number of segments is estimated using an enhanced L-method. The active learning iterations for classification gather additional training patches in uncertain areas according to the SVM results.