FIT student Stefan Reder in cooperation with Jan-Peter Mund, Nicole Albert, Lilli Waßermann and Luis Miranda utilised UAV-orthophotos and an adaptation of the U-Net architecture for the semantic segmentation and localization of windthrown stems. The network was pre-trained with generic datasets, randomly combining stems and background samples in a copy–paste augmentation, and afterwards trained with a specific dataset of a particular windthrow. The models pre-trained with generic datasets containing 10, 50 and 100 augmentations per annotated windthrown stems achieved F1-scores of 73.9% (S1Mod10), 74.3% (S1Mod50) and 75.6% (S1Mod100), outperforming the baseline model (F1-score 72.6%), which was not pre-trained.

These results emphasize the applicability of the method to correctly identify windthrown trees and suggest the collection of training samples from other tree species and windthrow areas to improve the ability to generalize.

Read the complete publication.