TreeScopeAI: A lightweight YOLOv11-based system for high-accuracy and real time tropical wood species identification

Rapid and reliable wood species identification is essential for forest conservation,biodiversity monitoring, and enforcement against illegal logging. Traditional anatomicalmethods are slow, subjective, and dependent on specialist expertise, creating majorlimitations for real-time verification in tropical forest regions. There is therefore a criticalneed for accurate, interpretable, and field-deployable identification tools. Wedeveloped TreeScopeAI, a lightweight deep-learning system based on the YOLOv11architecture, to classify 21 tropical hardwood species from high-resolution microscopictransverse sections. A dataset of 43,000 image patches was used for training andevaluation. Model performance was assessed through accuracy metrics, confusionmatrices, cross-site validation, PCA and t-SNE feature-space visualisation, and GradCAM interpretability. A deployment-ready application was also implemented to enablereal-time inference on portable devices. YOLOv11 achieved 99.71% overall accuracy,with thirteen species classified at 100% precision and recall, while all remainingspecies exceeded 98.5% accuracy. Misclassifications were rare and concentratedamong anatomically similar taxa. External validation on an independent dataset from adifferent geographical region demonstrated strong generalisation, yielding 99.5%accuracy across five species. Interpretability analyses confirmed that predictions reliedon biologically meaningful anatomical structures, and feature-space embeddingsshowed clear species-level clustering. TreeScopeAI offers a highly accurate,transparent, and computationally efficient solution for automated microscopic woodidentification. The lightweight (26 MB) model supports real-time classification onsmartphones and portable microscopes, providing a practical tool for ecologicalfieldwork, community-based forest monitoring, and frontline enforcement. Thismethodological framework illustrates the potential of modern deep learning to advancespecies identification workflows and improve conservation and timber-legalityverification.