Detecting Individual-Tree Crown Regions from Terrestrial Laser Scans with an Anchor-Free Deep Learning Model

Abstract

Detecting individual-tree crowns provides a fundamental analysis unit bridging macro ecological patterns and micro physiological functions. This study adapted an anchor-free deep learning model, CenterNet, to detect individual crown locations and regions from dense 3 D terrestrial laser scans. A total of 1181 crowns from twelve plots were manually delineated as reference, among which eight plots were used for training the CenterNet, and another four independent plots for testing model accuracies characterized as the F1-score of location detection and Intersection over Union (IoU) of bounding box area. The maximum training F1-score and IoU were 0.881 and 0.670 over 40k training iterations, respectively. The result testing F1-score and IoU were 0.754 and 0.583, respectively. Five morphological factors were quantified to investigate the causes of accuracy variation among different plots and species, including crown area, tree height, full-width-at-half-maximum, nearest neighbor crown distance, and overlapping ratio of neighboring crowns. Results show that tree height was most important trait for crown detection. A taller, larger, smoother, less crowded, and less overlapped tree was found easier to detect. Among six species, red pine, Scots pine, and silver birch were successfully detected, and Norway spruce, lodgepole pine, and trembling aspen were more difficult to detect.