Attention-enhanced UNet model for recognition of fractional vortex modes under turbulence with small-size samples

Recognition of fractional optical orbital angular momentum (OAM) modes under turbulence is useful for optical communication and information processing. Existing AI approaches still face challenges in addressing fractional OAM states under turbulence with limited data, and they lack interpretability in phase recovery. In this work, we present an attention-enhanced UNet architecture tailored for phase reconstruction and a physics-informed OAM attention classification network. These networks enable cross-turbulence transfer learning, allowing adaptation to unseen turbulence conditions through few-sample fine-tuning. The model also achieves 90% classification accuracy even under 50% area occlusion, and gradient-weighted class activation mapping visualizations reveal that the classifier focuses on physically meaningful interference features after phase recovery. We demonstrate image reconstruction under various atmospheric conditions in a two-dimensional image transmission system.