Enhancing Bark-Based Tree Species Recognition through Hierarchical Local and Global Feature Learning

We propose a transfer-learning-based Swin-Tiny model with a customized classification head for hierarchical local and global feature representation of bark imagery.

We systematically evaluate four pretrained transformer architectures, namely DeiT-Tiny, ViT-Base, Swin-Tiny, and BEiT-Base, along with the proposed Swin-Tiny model under a consistent experimental framework for bark-based tree species recognition.

We validate the proposed model using accuracy, precision, recall, F1-score, and Cohen’s kappa and compare its performance with existing methods. The proposed model achieves 97.00\% accuracy, 95.17\% and macro F1-score, outperforming the best existing method in accuracy by approximately 9.58%.