Abstract:

Objective: Existing tree counting methods based on unmanned aerial vehicle (UAV) imagery typically require a large amount of annotated data. To address the challenge of tree counting under limited annotation conditions, a tree counting method based on few-shot learning is proposed. Method: The publicly available TreeAI dataset was integrated with orthophotos of trees collected from the southern edge of the Horqin Sandy Land in Inner Mongolia Autonomous Region, Huangfengqiao Forest Farm in Hunan Province, and Camellia oleifera plantations in Jiangxi Province. Point and ellipse annotations were jointly employed for data labeling to construct a UAV-based tree counting dataset (UAV-TC) containing tree species such as Cunninghamia lanceolata, Pinus sylvestris var. mongolica, Camellia oleifera. The tree counting task was subsequently formulated as a few-shot regression problem. A small number of representative samples were used to guide the model in learning structural similarity features of tree objects, thereby reducing its dependence on phenotypic characteristics specific to individual tree species. Accordingly, a few-shot learning-based tree counting model, named FSTC-Net, was developed. The proposed model consists of a species-independent multi-scale feature extraction network and a density map prediction module, enabling robust counting of multiple tree species in complex environments. Specifically, the feature extraction network incorporated MixNet-L with a feature pyramid structure to enhance multi-scale feature representation of tree targets. The density map prediction module replaced conventional direct feature map inputs with correlation maps generated from sample and image features, thereby enabling cross-species feature alignment and similarity matching. In addition, a random scale augmentation strategy and an adaptive loss function were introduced to improve model generalization and counting accuracy under few-shot conditions. Result: Experimental results demonstrated that FSTC-Net achieved accurate counting of Cunninghamia lanceolata, Pinus sylvestris var. mongolica, Camellia oleifera in the test set, with a coefficient of determination (R²) of 0.949 9. The corresponding mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean square error (RMSE) values were 3.54%, 26.71 trees, and 37.60 trees, respectively. Ablation experiments showed that after integrating the MixNet-L and RoI Align modules into the FamNet model, the R² of the model counting increased by 0.0308 and 0.0329, respectively, while the MAPE was reduced by 0.64% and 0.78%, respectively. When both MixNet-L and RoI Align modules were integrated into the FamNet model, the best performance in terms of R² and MAPE was achieved. Further comparisons with mainstream models, including T-Rex, T-Rex2, and FamNet, demonstrated that FSTC-Net outperformed these methods in terms of both error control and result stability. Specifically, the R² values of FSTC-Net were 0.0854, 0.0493, and 0.0413 higher than those of T-Rex, T-Rex2, and FamNet, respectively, whereas the corresponding MAPE values were reduced by 3.12%, 2.35%, and 1.73%, respectively. In addition, analysis under different canopy densities revealed that although the counting error of FSTC-Net increased in high-canopy-density mixed coniferous and broad-leaved forests, the error remained within an acceptable range. Conclusion: The experimental result has verified the effectiveness and superiority of FSTC-Net for tree counting tasks, and it can provide reliable technical support for UAV-assisted forest resource monitoring.

Key words: tree counting, UAV, few-shot learning, multiple tree species, forest resource monitoring