Accurate measurement of log volumes remains one of the most persistent challenges in post-harvest forestry operations. Traditional stack-based scaling methods rely on geometric assumptions and conversion factors that can struggle with real-world conditions such as irregular log placement, occlusion, voids within stacks, and variable log diameters. These limitations directly affect valuation, inventory management, and logistics planning.
Working with the Interpine team as part of a postgraduate research project, Ethan Clapham (University of Exeter) explored whether modern computer vision techniques could offer a practical, scalable alternative for measuring log volumes directly from images captured in operational log yards. Interpine was pleased to support this work by providing real‑world imagery from forestry operations, along with expertise in log scaling assessment and the practical application of AI within the sector.
Project Overview
The aim of the study was to develop an image-based pipeline capable of:
- Detecting individual logs within stacked log yard images
- Estimating log diameters from visible cross-sections
- Calculating aggregated log volumes using calibrated image geometry
Rather than relying on idealised stack assumptions, the approach treats each visible log as an individual object, measured directly from RGB imagery.
Real-World Data and Challenges
The dataset comprised RGB images containing approximately 3,900 logs, all captured under real operational conditions rather than controlled environments. Images included:
- Non-uniform lighting
- Partial occlusion of logs
- Background clutter and noise
- Variable log orientations
Each image included an in-frame reference ruler, enabling pixel-to-metre calibration for real-world measurement.
These factors reflect the true complexity of log yard environments and provided a robust test of how well computer vision methods perform outside the laboratory.
Detection and Measurement Methodology
The project used a small segmentation model, a single-stage, real-time object detection and segmentation framework optimised for low computational cost and field deployment.
Key elements of the pipeline included:
- Separate models trained for log faces and ruler detection
- Evaluation of circular vs square bounding annotations, with square annotations producing more consistent detections
- Data augmentation (rotation, scale, illumination variation) to improve robustness
- Instance segmentation masks used to extract log cross-sections
- Diameter estimation via geometric fitting
- Volume calculation following pixel-to-metre calibration using the detected ruler
This workflow allowed volume estimation to be derived directly from image geometry rather than stack-level assumptions.
Detection Performance
Evaluation showed strong detection accuracy under real-world conditions. Square bounding annotations achieved:
- Intersection over Union (IoU): 0.905 (vs 0.800 for circular labels)
- Precision: 0.994
- Recall: 0.990
- mAP50: 0.994
The system successfully detected log faces even with partial occlusion and irregular stack geometry, demonstrating robustness suitable for operational environments.
Diameter and Volume Accuracy
The study demonstrated promising results for automated image-based log measurement. Diameter estimation achieved a mean absolute error of 2.65 cm, while volume estimates showed a mean absolute error of 0.239 m³. A slight negative bias in volume highlighted the need for improved calibration, as relying on a single visible log face can underestimate true volume.
Key Findings and Future Directions
Overall, the project showed that object detection and geometric modelling can provide a practical alternative to manual log scaling, even under challenging conditions such as variable lighting and irregular log arrangements. Future work will focus on improving calibration, adding operator review tools, and validating the system across a wider range of forest conditions.
This project provides a strong proof-of-concept for how computer vision can support the next generation of digital forestry operations. This is an example of Interpine working with a range of student studies across the globe to progress technology implementation across the industry. Thanks again to Ethan Clapham and the support from University of Exeter for this work.
You can view the full research poster below, including detailed methodology, evaluation metrics, and visual examples from the study.