Objective The aim was to achieve precise diagnosis of rice nitrogen nutrition status and provide scientific guidance for subsequent fertilizer management.

Methods During the tillering and jointing stages of rice growth, a scanner was used to scan the fully expanded the top 1^st, 2^nd and 3^rd leaves of each tiller stem as the dataset. A ResNet34_CFS deep learning diagnostic model was proposed, aiming to accurately identify rice nitrogen nutrient levels. The ResNet34_CFS model was based on the ResNet34 (Residual Neural Network 34-layer) network, integrating the CBAM (Convolutional Block Attention Module) dual-head attention mechanism, FPN (Feature Pyramid Network) structure for multi-scale feature fusion, and SE (Squeeze-and-Excitation) module to optimize feature channel weights. It was focused on key regions of leaves and local details and global morphological features of leaves, in order to enhance the model's representational capabilities. Additionally, it was incorporated transferred ImageNet pre-trained weights, which significantly improved the model's performance in diagnosing rice nitrogen nutrition.

Results The ResNet34_CFS network achieved model testing accuracy rates of 88.96% and 90.35% during the tillering and jointing stages of rice growth. Compared to the pre-improvement network model, the accuracy improved by 7.70% and 9.29%. Compared with network models such as ResNet50, Swin Transformer, and VGG16, the improved ResNet34_CFS network model performed better in terms of accuracy, precision, recall, and F1 score. Additionally, the improved model had a stable training process and fast convergence speed.

Conclusion In summary, the rice nitrogen nutrient diagnosis model established in this study could efficiently and accurately identify the severity of nitrogen stress in rice during the tillering and jointing stages. These results provided a reliable technical solution for rice nutrient diagnosis in smart agriculture.