An adaptive hybrid vision architecture for sugarcane leaf disease identification

Identifying sugarcane leaf diseases in natural field environments remains a challenging problem, due to complex weather and farming conditions, which can lead to significant variability of symptoms within a given category. To address this issue, this study proposes EnhancedConvFormer, a hybrid vision architecture that includes an adaptive dynamic token mixer (ADTM). A lightweight, data-driven gating mechanism is also employed to dynamically optimize the balance between local feature extraction (through depthwise separable convolutions) and global context modeling (through multi-head self-attention), each tailored to the distinct visual characteristics of individual input images. A rigorous McNemar test confirms this dynamic fusion approach significantly surpasses static fusion methods during feature extraction from scratch ( P < 0.05). Leveraging convolutional inductive bias also allows the ADTM to function as a structural regularizer, effectively mitigating feature co-adaptation and ensuring robust data efficiency, even with limited training samples. The resulting model achieved an F1-score of 96.30% on complex real-world datasets, demonstrating exceptional performance. Notably, EnhancedConvFormer is highly optimized for agricultural edge deployment, requiring only 3.957 GFLOPs and 24.7 million parameters. The real-time inference throughput reached 51.42 FPS on a single NVIDIA RTX A5000 GPU, with a peak memory usage of ~130 MB. This outcome demonstrates the potential of EnhancedConvFormer as an efficient and practical solution for automated, real-time crop disease diagnosis in precision agriculture.