Tree seeds have become a scarce and valuable resource facing high demands following widespread forest decline due to storms and insects and the need for re-establishment and conversion of forests. This paper describes a novel approach to assessing the availability of seed cones of silver fir (Abies alba Mill.) to support full utilization of harvest potential in seed orchards. The method combines UAV-based acquisition of RGB images with the application of AI to automatically detect cones using YOLO11, a convolutional neural network. In a first step, geo-referenced high-resolution RGB images are taken of each crown from the nadir position including the exact RTK position and all relevant optical parameters. In the presented case study, a preliminary training data set was derived by manual labelling all cones in 387 images. This preliminary sample was enlarged by data augmentation to a final training data set of 937 images which was then used to train the YOLO11 script in different variants of resolution. The quality of the results was assessed based on model precision P as the relation of true-positive predictions to the sum of true- and false-positive predictions. After 200 training epochs, P varied between 0.782 for the YOLO11 nano variant and 0.853 for the YOLO11 extra-large variant. In the last step, the trained models were applied to independent images to detect seed cones based on detection probabilities. Together with the positioning data of the respective images these probabilities can be used to reliably assess the harvest suitability of individual trees and whole stands.
Körner et al. (Thu,) studied this question.