Deep learning for species identification in Kirganelia using digitized herbarium specimens

Accurate species identification of herbarium specimens is fundamental to biodiversity research but remains challenging in taxonomically complex plant groups. Here, we evaluated deep learning for species identification in Kirganelia (Phyllanthaceae) using 6475 cropped leaf images representing 30 phylogenomically validated species. Three convolutional neural network architectures were trained using transfer learning, and multiple strategies were compared to address severe class imbalance. Performance was evaluated using conventional image-level metrics and a hierarchical framework aggregating predictions at specimen and species levels. Model architectures showed different trade-offs between overall classification performance and recognition of underrepresented species: EfficientNetV2-S provided strong aggregate performance, whereas MobileNetV3-L with weighted loss achieved the highest minority-class performance. Hierarchical evaluation showed that 19 of 24 sufficiently sampled species (79.2%) were well recognised, while the remaining five were moderately recognised. Five of six species represented by only one or two specimens were also correctly recognised for all available specimens, although these sparsely sampled taxa were evaluated descriptively. Grad-CAM analysis showed that predictions were primarily associated with leaf tissue rather than herbarium background. These findings support imbalance-aware training and hierarchical evaluation for accurate herbarium species recognition while highlighting the importance of class-specific performance, sampling representation, and expert verification.