Comparative analysis of deep learning architectures for fine-grained classification of dahlia varieties

Fine-grained visual classification (FGVC) of flowers remains a significant challenge due to subtle inter-class variations and large intra-class diversity. In precision agriculture, deploying highly accurate models on resource-constrained edge platforms is severely restricted by computational overhead. To reduce this bottleneck, we introduce a specialized 63-category dahlia dataset comprising 3,279 field-captured images and study a Cross-Architecture Knowledge Distillation (CAKD) framework that transfers soft targets from a compact Vision Transformer (Swin-Tiny) teacher to a compact CNN student (MobileNetV3). Under a strictly identical training protocol for all models, the distilled student reaches a test accuracy of 97.51% ± 0.99% (mean ± standard deviation over three seeds) with only 4.28M parameters and 0.22 GMACs, exceeding every teacher in the pool: ResNet152 (97.01%, 58.27M), Swin-Tiny (96.82%, 27.57M) and EfficientNetV2-M (96.62%, 52.94M). Homogeneous distillation from the twice-larger ResNet152 teacher reaches 97.61%, a difference within seed variance, so the value of CAKD is efficiency rather than a higher accuracy ceiling. The student runs at 17.0 ms per image (58.8 FPS) on a consumer smartphone. We additionally report balanced accuracy, macro-F1 and head/middle/tail metrics, which show that overall accuracy conceals a substantial drop on rare cultivars. The compact student was exported to ONNX and integrated into an edge-cloud WeChat Mini Program that collects hard examples from the field for periodic retraining. The results position the system as a practical accuracy-efficiency trade-off for agricultural phenotyping and variety protection, rather than as a state-of-the-art claim on a newly constructed dataset.