Abstract:
Objective To verify the feasibility of diagnosing nitrogen stress from images of individual rice (Oryza sativa L.) plant canopies during the tillering stage, and explore the performance advantages of the improved MobileViT model in nitrogen stress level identification.
Method Late rice ‘Huanghuazhan’ was used as experimental material, and six nitrogen application gradients were set to construct a single-plant canopy image dataset at the tillering stage. To address challenges such as subtle phenotypic differences, susceptibility to background interference, and the need for lightweight deployment, a individual plant canopy nitrogen stress diagnostic method based on improved MobileViT was proposed. Based on the hierarchical coding design inherited from MobileViT, the model underwent three structural improvements: Replacing ReLU6 with SiLU to enhance nonlinear expression capability; At Stage 5, a re-parameterized MobileOne block replaced the original MobileViT module to reduce inference latency; At the end of Stage 4, a lightweight ECA channel attention module was integrated to enhance key feature perception and suppress complex background interference. During training, ImageNet pre-trained weights were loaded for transfer learning, and the above structural improvements worked synergistically to enhance the model’s representative capability and efficiency.
Result After improvement, the model’s accuracy, precision, and recall reached 85.67%, 85.93%, and 85.67%, respectively, with accuracy increasing by 3.17 percentage points compared to the baseline model and the model size compressed to 2.31 MB. Compared with five mainstream models including ResNet18 and MobileNet V2, this model achieved the highest accuracy with the smallest parameter count. Ablation experiments showed that the model integrating ECA modules outperformed models integrating SE, CA, and CBAM. Grad-CAM visualization results showed that the model could precisely focus on key phenotypes such as yellowing leaves. Under normal, dim, and intense lighting conditions, the model’s F1 scores were 85.80%, 83.29%, and 80.76%, respectively, all higher than those of the original model, indicating strong robustness.
Conclusion The three structural improvements proposed in this study, combined with ImageNet transfer learning, effectively achieve synergistic enhancements in recognition accuracy, lightweight deployment, and complex light robustness. The improved model can efficiently and accurately diagnose the nitrogen stress level in the canopy of individual rice plant canopies during the tillering stage, providing technical support for precise nutrition diagnosis of rice in smart agriculture.