Meta-learning for wood identification under imbalanced few-shot conditions

Wood identification in forest production contexts is essential for sustainable forestry management, legal wood trade, and biodiversity conservation. However, real-world forest environments present data constraints such as few-shot samples, class imbalance, and cross-sectional anatomical variability. To address these challenges, the study proposes a meta-learning-based classification framework that integrates a focal loss-enhanced meta-learning strategy with a dual-section weight fusion mechanism. The model, characterized by adapting radial and tangential section images, is typical in logging sites or wood assessment stations. A dataset of 5,597 images from ten wood species was leveraged to evaluate the method under few-shot scenarios. Results demonstrate the proposed approach significantly improves classification accuracy, generalization, and robustness under imbalanced field-like data conditions. This work provides a foundation for intelligent wood identification systems applicable in forest resource monitoring and future on-site deployment.