Automatic detection of snow breakage at single tree level using YOLOv5 applied to UAV imagery

Abstract

The assessment of forest abiotic damages such as snow breakage is important to ensure compensation to forest owners. Currently, information on the extent of snow breakage is gathered through time-consuming and potentially biased field surveys. In such situations where field surveys are still common practice, unmanned aerial vehicles (UAVs) are increasingly being used to provide a more cost-efficient and objective methods to answer forest information needs. Further, the advent of sophisticated computer vision techniques such as convolutional neural networks (CNNs) offers new ways to analyze image data more efficiently and accurately. We proposed an object detection method to automatically identify trees and classify them according to the damage by snow based on a YOLO CNN architecture. UAV imagery collected across 89 study areas and over the course of the entire year were manually annotated into a total of >55 K single trees classified as healthy, damaged, or dead. The annotated trees, along with the corresponding UAV imagery were used to train a YOLOv5 object detection model. Furthermore, we tested the effect of seasonality, and varying atmospheric and lighting conditions on the model’s performance. Based on an independent test set of data we found that the general model including all of the data (i.e. any seasons, atmospheric conditions, and time of the day) outperformed all other tested scenarios (i.e. precision = 62 %; recall = 61 %). Furthermore, we found that despite the fact that the snow damaged trees represented a minority class (i.e. 16 % of the annotated trees), they were detected with the largest precision (76 %) and recall (78 %). Finally, the general model transferred well across the variation in seasons, atmospheric and illumination conditions, making it suitable for usage for any new UAV image acquisition.