UNSUPERVISED STATISTICAL APPROACH FOR TREE-LEVEL SEPARATION OF FOLIAGE AND NON-LEAF COMPONENTS FROM POINT CLOUDS

Abstract

Accurately classifying foliage and non-leaf components in point clouds is essential for remote sensing forest applications. Existing methods rely on radiometric attributes or local geometric features, often requiring time-consuming manual labelling. In this paper, we propose a statistical approach using flexible thresholds determined based on tree-level geometric features. Selected features (local anisotropy, curvature, linearity, first principal component, verticality, and sphericity) have shown robustness in earlier studies. Threshold values are identified as points of inflection in the fitted distributions for each tree. Our method requires only two parameters and was tested with manually labelled Terrestrial Laser Scanning (TLS) data and non-labeled data from an oblique and above canopy setup (Permanent LiDAR scanner setup). We tested two boreal tree species, Scots Pine, and silver birch, with 28 trees in total (14 trees for each species) using two data sources. Compared to two alternative methods (namely, fixed thresholding and CANUPO), our approach consistently outperforms in terms of recall. We achieved an average overall accuracy of 85%, recall of 88.5%, precision of 83%, and f1 score of 85%. Visually assessing oblique and above canopy results, our algorithm effectively captures tree structures. Our statistical approach provides an effective solution for foliage and non-leaf separation, with processing times of less than five minutes for individual tree point clouds containing up to 2 million points and no need for extensive manual labelling or parameter adjustments.