Answers are grounded in this paper and cite their sources.
Quantification of woody plant species and their structural traits in sparsely forested grasslands is crucial for accurately estimating carbon stocks within grassland ecosystems. Unlike plantations, natural sparse woodlands and grasslands cover extensive areas and exhibit high heterogeneity in tree distribution. Existing deep-learning models, such as Mask R-CNN, can effectively extract canopy information in relatively simple scenarios, but their complex network structures and redundant parameters reduce accuracy and recall in heterogeneous environments. To address these limitations, this study proposed DeepTree, a lightweight two-stage instance-segmentation framework that optimizes the Mask R-CNN design with two variants: MobileNet–NAS-FPN Mask R-CNN (MNMS R-CNN), including a small variant, MNMS R-CNN (S), for computational efficiency and a large variant, MNMS R-CNN (L), for improved accuracy. The dataset included 8,724 single-tree canopy samples for model training and validation, together with 2,296 samples (166 field-measured and 2,130 image-derived) for evaluation. Results showed that MNMS R-CNN (S) and MNMS R-CNN (L) improved inference speed by 88.2% and 19.6%, respectively, and achieved F1 scores of 78.3% and 81.1%, outperforming the baseline Mask R-CNN. MNMS R-CNN (L) also showed high accuracy in extracting segmentation-derived phenotypic parameters on the independent test set, with R² values of 0.98 for both crown width and crown area and RMSE values of 0.30 m and 2.66 m², respectively. By balancing efficiency and accuracy, DeepTree provides a scalable pathway for single-tree monitoring and phenotypic information extraction in complex temperate savanna environments.
No takes yet. Share an insight, caveat, or question.
Duan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: