Estimation of leaf nitrogen concentration at different leaf positions in rice based on UAV RGB images combined with machine-learning algorithms

The use of remote-sensing technology to estimate leaf nitrogen concentration (LNC) is an important means of rapidly diagnosing crop nutrition, but most research has focused on LNC across the entire canopy and has overlooked deviations at different leaf positions and their impact on the accuracy of predictive models. A two-factor split field experiment with two rice varieties and four N application rates was conducted in 2022 and 2023. LNC at different leaf positions and growth stages were investigated, vegetation indices, color parameters, and texture features were extracted from unmanned aerial vehicle (UAV) RGB images, and five machine-learning algorithms were used to establish and evaluate models for monitoring LNC at different leaf positions. The results showed that the LNC under all treatments decreased from the upper to the lower canopy leaves. Compared with no nitrogen application, the LNC with 180 kg ha -1 nitrogen application treatment was increased by 9.95%–21.36%. A total of 48 features were extracted from RGB images. Vegetation indices, such as VARI and WI, and color parameters, such as R, B, and H, were significantly correlated with LNC across all three periods. After optimizing the parameters and machine-learning methods, the best-performing models at the jointing, booting, and heading stages were GPR, ANN, and SVM. The highest R 2 was 0.86. This study achieved the rapid monitoring of LNC at different leaf positions, which allowed for a quick diagnosis of nitrogen nutrition based on nitrogen concentrations and their differences among leaves. This provides technical support for timely and effective fertilization measures.