Stem crook detection in terrestrial laser scanning data using semi-empirical simulations

Abstract

Abstract Stem crooks are critical stem defects that reduce timber quality and value. This study proposes a simulation pipeline for generating synthetic crooks in Terrestrial Laser Scanning (TLS) point clouds and evaluates the performance of a Convolutional Neural Network (CNN) trained on this dataset for detecting crooks in real-world TLS data. We introduced parameterized crook deformations into TLS-derived stems, producing 14 000 synthetic examples based on Scots pine (Pinus sylvestris L.) trees. The simulated crooks were represented in two patterns: single-and double-directional bends. Each stem was rasterized into 2D images from multiple vertical sections and viewpoints. We trained a YOLOv8s model exclusively on these synthetic images and tested its performance on an independently collected real-world TLS dataset containing 310 stems and 65 visually annotated crooks. The detector achieved F1-scores above 0.6 for low Intersection over Union (IoU) thresholds (<0.3), demonstrating its ability to identify crooks despite domain differences between simulated and real point clouds. However, localization accuracy declined under stricter IoU criteria, with a mean Average Precision (mAP@50–95) of 0.24. Error patterns revealed more commission errors in upper stem regions and more omissions near the butt logs, reflecting branch complexity, point cloud sparsity, and limitations in lower-stem simulations. These results show that synthetic crook data can provide valuable training material for defect detection while also highlighting the challenges associated with using such synthetic data for model training.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.