Global, multi-scale standing deadwood segmentation in centimeter-scale aerial images

Background

This article examines how increasing tree mortality worldwide creates an urgent need for efficient approaches to map standing deadwood, including both fully dead crowns and partial canopy dieback, as a basis for monitoring forest dieback and informing management across biomes. Existing remote-sensing methods using RGB aerial imagery often work only for specific regions, resolutions, or forest types, limiting their transferability and value for global applications.

Goals and Methods

This study aims to develop a generalizable semantic segmentation model that detects standing deadwood in high-resolution orthophotos from both drones and aircraft, spanning all major forest biomes and image resolutions roughly between 1 and 28 cm. The authors assembled 434 labeled orthophoto sets (10,778 ha) from the crowd-sourced deadtrees.earth database, retaining medium- and high-quality manual deadwood delineations, and preprocessed the data through reprojection, multi-resolution rescaling, tiling, and a four-dimensional sampling strategy that balances biome, resolution, deadwood occurrence, and image source. A SegFormer-B5 encoder combined with a U-Net decoder is trained using Focal Tversky Loss to handle strong class imbalance and evaluated with spatial block cross-validation using precision, recall, and F1-scores across biomes and resolutions. F1 score is a crucial metric that balances precision (how relevant are positive predictions?) and recall (how many actual positives were found?) into a single score, representing the harmonic mean. It ranges from 0 (worst) to 1 (best), providing a robust evaluation, especially for imbalanced datasets, by penalizing models with extreme imbalances between false positives and false negatives.

Conclusions and Takeaways

The model achieves a mean F1-score around 0.61, with best performance in temperate forests (F1 ≈ 0.71, recall up to 0.90) and biome-specific optimal resolutions, often 4–8 cm, rather than uniform gains at the finest scales. Qualitative inspection shows that predictions often exceed heterogeneous crowd labels, implying conservative numeric scores and highlighting label quality as a key constraint. By releasing both a machine-learning-ready, multi-biome dataset and pretrained RGB models, this study provides a practical foundation for operational monitoring of tree mortality and for training coarser-scale satellite products using these labels as reference.