Estimating log yard size from remote sensing imagery to predict sawmill productivity

By Stijn Peeters

Abstract
The European forestry sector lacks regional and local production data, hindering targeted policies and sustainable forest management. This thesis explores the potential of remote sensing and machine learning to estimate log yard productivity as a proxy for sawmill activity. A three-stage methodology was employed, starting with manual segmentation of log stacks and log yard area as a baseline, followed by semi-automatic segmentation using the Segment Anything Model (SAM) and finally automatic segmentation using a U-Net model trained on the SAM-generated data. Results show that while there is a significant correlation between log stack area and sawmill production data, the correlation is weak. This is likely due to the temporal alignment between the production data and the log stack imagery. The study also demonstrates the potential of SAM in significantly reducing the time required for log stack annotation compared to manual methods. The comparison between SAM and U-Net reveals that the U-Net model consistently identified larger log stack areas than SAM, which together with a stronger relationship with the U-Net images and sawmill productivity suggest that the U-Net model was able to learn and generalise the features of log stacks effectively. This research highlights the potential of remote sensing and machine learning in addressing the data gap in the forestry sector, but also underscores the need for further research to improve data quality and address the temporal alignment issue.