Automated aerial detection of spruce tree crowns through YOLOv5 and watershed segmentation

Abstract

The spruce tree, a key species in Canada, is crucial to industries like timber and pulp. Effective management of these resources relies on monitoring forest parameters to ensure long-term forest health and productivity. Tree crown dimensions contribute valuable insights into these parameters. This study investigates an automated approach for detecting and measuring spruce tree crowns using the YOLOv5 machine learning model combined with watershed segmentation. The method enhances the accuracy of crown measurements from aerial drone images. Over 2,000 spruce trees in a limited forested area of Saskatchewan, Canada, are analysed using top-view images captured by a DJI Mavic 3 Classic drone, which is sufficient for this project as the model trained well. The YOLOv5 model is initially employed to detect trees, following by watershed segmentation to refine the tree crown boundaries. In the forest measuring area, some regions contain closely spaced trees with overlapping crowns that cause challenges for accurate recognition of tree boundaries. Although feeding watershed segmentation with YOLO-detected individual trees addresses this issue, tuning the IoU threshold in the NMS stage, applying data augmentation, and utilizing high-resolution images further enhance detection accuracy. One issue in predicting tree crowns is the underestimation of diameters, often resulting from systematic errors in image capture, measurement methods, environmental conditions, and limitations in image resolution. To address this issue, a linear regression model is applied to adjust the predicted crown diameters, aligning them more closely with the actual field measurements. This approach demonstrates an acceptable accuracy of 89.1% compared to prior research and existing methodologies, which reported accuracies ranging from 67.72% to 95.4%, particularly in complex forest environments. Although the primary focus of this research is on measuring tree crowns, the findings have broader implications for forestry management activities, such as biomass estimation and forest health monitoring. Future research could explore further improvements to the model for real-time forest management applications, as well as expand its use to detect and measure other tree species in mixed forests and agriculture plants.