Objective  To rapidly and accurately monitor the nitrogen nutrition status of rice and determine the optimal nitrogen application rate for high yield and quality.

Methods Field experiments were conducted over two years (2022-2023)using two locally dominant rice varieties: the conventional early indica rice 'zhongjiazao17' and the hybrid rice 'changliangyou173'. Four nitrogen fertilizer levels (0, 75, 150, 225 kg·hm⁻², denoted as N0, N1, N2, and N3, respectively) were established. Digital camera (Canon EOS 100D, resolution 72 Pixels Per Inch) was used to acquire rice canopy images and corresponding nitrogen nutirtion data. Nitrogen nutrition monitoring models were constructed based on image features and their derived parameters. ResultsThe percentage of rice pixels (PRP) in the image and its derived features showed high correlations with leaf area index (LAI), above ground biomass (AGB), and nutrition accumulation (PNA), with the best model prediciton performance observed at the jointing stage. Further analysis revealed that polynominal funcions based solely on RPR could effectively predict LAI, AGB, and PNA, with coefficients of determination (R²) of 0.76, 0.74, and 0.79, respectively (P＜0.01), and the root mean square error (RMSE) for the model validation was 0.32 g·m⁻², 22.30 g·m⁻², and 2.54 g·m⁻² , and the relative root mean square error (RRMSE) was 8.25%, 7.61%, and 26.49%, respectively. In contrast, high-order exponential function derived from PRP provided superior predictions for LAI, AGB, and PNA, with R² of 0.89, 0.92, and 0.93, respectively (P＜0.01), RMSE values of 0.16 g·m⁻², 3.71 g·m⁻², 0.57 g·m⁻², and RRMSE values of 4.20%, 1.27%, and 5.98% (all＜10% ), indicating excellent model stability.

Conclusion Overall, the feature derivation strategy effectively improves the prediction accuracy and stability of the models, demonstrating significant application values for monitoring rice nitrogen nutrition.