Quantifying Bilberry Counts and Densities: A Comparative Assessment of Segmentation and Object Detection Models from Drone and Camera Imagery

Nordic forest management is increasingly emphasizing multi-functional goals, expanding beyond timber production towards non-wood forest products such as wild berries. Wild berry yield maps are based on sample plot data combined with meteorological, remote sensing, and geoinformation data. Automating sample plot data processing is crucial, as manual collection is labor-intensive, time-consuming, and complicated by short berry seasons and fluctuating yields. This study compares two methods for automatic bilberry detection and counting: a deep learning detector YOLO and a machine learning model using the segment anything model (SAM) followed by a random forest classification (SAM-RF). Both system camera and drone imagery were evaluated as input data. YOLOv8 clearly outperformed SAM–RF in berry detection, achieving an R2 of 0.98 and an RMSE of 3.8 berries when evaluated against annotated system camera images, compared to an R2 of 0.80 for SAM–RF. System camera imagery consistently produced higher accuracy than drone imagery due to higher image clarity and more optimal viewing angles, with YOLOv8 achieving an R2 of 0.95 against field counts, compared to 0.81 for drone images. The results also indicate that the primary error source in berry counting arises from the fact that many berries are not visible in the captured images. The results from the data analysis support the use of the developed technologies in yield modeling and even in implementing future ‘follow-me’ drone berry assistants.