AI-Driven volumetric estimation of wood piles via detection YOLO models

Wood pile volume estimation is a key step in the wood supply chain. Several studies have explored the replacement of traditional manual measurements of log diameters with machine learning models trained on wood pile photos. Additional efforts have focused on running these models in mobile and offline scenarios. Here, two models based on YOLOv8—one of the most widely used algorithms for precise object detection—were developed and trained on the public HAWKwood dataset of wood piles to estimate their volume. The first model was trained to count log occurrences, while the second was designed to segment the section area of all logs in the pile. These models were then combined into three different methods and tested on a heterogeneous set of web images as well as two Italian datasets from the Calabria and Trentino regions. The best-performing method, in terms of accuracy and compute cost, combined YOLO object detection with section area approximation based on ellipse inside each log bounding box. This approach achieved high accuracy (98.9% in training, 97.0% in validation, and 99.8% in testing) and fast execution times (less than 0.5 seconds). Such results suggest strong potential for mobile phone implementation and future onsite, real-time evaluations.