Geographic Classification of Social Bees Using Automated Wing Pattern Analysis

Intraspecific population classification represents a fundamental challenge in biodiversity conservation, where traditional morphometric approaches face limitations in scalability, reproducibility, and dependence on specialized expertise. We present a deep learning approach for automated geographic classification of social bee populations based on wing morphology patterns. Our analysis encompassed 1040 wing images collected from seven geographical regions using four deep learning architectures: EfficientNetB0, ResNet34, DenseNet121, and MobileNetV3‐Small. Images were standardized to 224 × 224 pixels and enhanced through data augmentation techniques. We employed stratified sampling for dataset division (70% training, 10% validation, 20% testing). Among the tested architectures, EfficientNetB0 achieved the strongest held‐out test performance (accuracy = 0.95, macro F1‐score = 0.93). Multi‐seed evaluation across five seeds (42, 0, 1, 2, 3) further confirmed the robustness of the full training configuration (accuracy = 0.944 ± 0.028; macro F1‐score = 0.902 ± 0.042). We utilized Gradient‐weighted Class Activation Mapping (Grad‐CAM) to visualize model decision mechanisms, revealing region‐specific anatomical structures including diagonal venation patterns, circular configurations, and oval shapes that distinguished populations. Five‐fold cross‐validation confirmed EfficientNetB0's robust generalization (88.40% average accuracy). Our results demonstrate that automated approaches can overcome fundamental constraints of manual classification systems, providing scalable solutions for biodiversity monitoring programs and enabling non‐specialists to conduct accurate taxonomic assessments without extensive training in morphometric techniques.