Beyond standard views: enhancing quarantine true fruit fly identification with a multi-angle deep learning approach (Diptera, Tephritidae)

Deep learning offers a promising pathway for automated species recognition, yet model generalizability is often constrained by image datasets that lack multi-angle views of specimens. Using four tephritid species as a model system, we quantitatively evaluate how specimen imaging angles affect classification performance. Multi-angle imaging substantially improved model generalization even with fewer specimens, and greater angular diversity required substantially fewer specimens to achieve comparable performance—though this advantage diminished under extreme scarcity. Stronger data augmentation and extended training epochs could not compensate for limited angular diversity, whereas selecting appropriate architectures partially mitigated this limitation: lightweight convolutional neural networks sufficed for small species sets, while transformers became increasingly advantageous as taxonomic diversity grew. These patterns held when scaling to a larger dataset of 26 tephritid species, confirming the generalizability of our findings. Together, this study demonstrates that angular diversity provides irreplaceable information for model generalization, a benefit that in silico alternatives cannot replicate. This work provides a quantitative, resource-efficient framework for artificial intelligence-assisted insect identification, offering practical guidance for dataset construction in quarantine surveillance, biodiversity monitoring and taxonomic research.