Bridging CNNs and vision transformers for efficient tea leaf phytopathogen diagnosis: GL-MobFormer

Accurate identification of visible disease symptoms is essential for the sustainable management of tea (Camellia sinensis) cultivation. However, balancing high diagnostic accuracy with the computational efficiency required for deployment on agricultural edge devices remains a significant challenge.

We propose GL-MobFormer, a lightweight hybrid deep learning framework. This architecture integrates the local feature extraction capabilities of MobileNetV3 with the global contextual modeling of a Transformer Encoder. To improve model robustness in unstructured field environments, we applied the CutMix data augmentation strategy. The framework was evaluated on a dataset comprising 5,278 tea leaf images across seven phytosanitary categories.

Empirical evaluations demonstrate that GL-MobFormer achieved a classification accuracy of 95.13% and a Matthews Correlation Coefficient (MCC) of 0.9417. Crucially, this performance was maintained with a low computational footprint of merely 0.33 G FLOPs(Floating Point Operations). Importantly, an occlusion-based sensitivity protocol was implemented to provide quantitative grounding for model interpretability. Results revealed that systematically masking only the top 5% of critical activation regions led to an average reduction of 70.61% in classification confidence, empirically confirming that the model's diagnostic logic is faithfully anchored on pathologically relevant lesion features rather than background noise.

GL-MobFormer achieves an optimal trade-off between diagnostic precision and computational overhead. It provides a practical and highly efficient solution for on-site, real-time phytosanitary monitoring in precision agriculture.