Applying a convolutional neural network (CNN) to Virginia’s forests: how forest type and age can impact individual tree segmentation

Abstract

Convolutional neural networks (CNNs), specifically U-Nets, successfully detect and segment individual tree crowns when applied to high spatial resolution remotely sensed data. However, much of the prior work for individual tree detection and segmentation has been in arid climates such as the western United States and the African Sahel. In this study, we used the U-Net-id CNN architecture and National Agriculture Imagery Program (NAIP) imagery to investigate forest-cover-specific training on individual tree crown detection in Virginia, U.S. We trained three models using hand-delineated crowns: one model with mixed/deciduous forests (3,457 trees), one with pine plantations (10,680 trees), and one that combined all the training data (14,137 trees) from both taxonomic groups. The training data was selected using 128 × 128 pixel patches across the study area where one interpreter hand-delineated all the tree crowns in the patch. Accuracy for the models was assessed over the entire model application area, by forest cover, and plantation age groups. Accuracy assessment included comparing the hand delineation and model results for tree counts and crown area as well as precision, recall, and F1 scores calculated. The Combined model performed poorly, achieving an F1 score of 0.4. However, further assessment of the forest cover and plantation age level showed increased F1 scores. The Plantation model’s F1 scores indicated the best-performing age class was the 16–19-year-olds, with a score of 0.74. However, the younger plantations had F1scores of 0.42, 0.28, and 0.07 for the 8–11, 4–7, and 0–3-year-olds, respectively. The best count and crown area relationships came from the Mixed/Deciduous model, which achieved an F1 score of 0.64. In this study, we found that the representative quality of the training data versus a larger quantity of data more heavily impacts this U-Net-id CNN architecture and demonstrates the impact of training data quality on a CNN model performance.

Citation

Ritz, A. L., Wynne, R. H., Thomas, V. A., Wagner, F. H., Green, P. C., Schroeder, T. A., & Saatchi, S. (2025). Applying a convolutional neural network (CNN) to Virginia’s forests: how forest type and age can impact individual tree segmentation. International Journal of Remote Sensing, 47(1), 218–245.