Accurate and timely nitrogen status diagnosis is crucial for the nitrogen application management and yield prediction of winter rapeseed. Traditional, destructive manual measurement cannot fully meet the large-scale production in recent years, due to the tedious, time-consuming and laborious. Fortunately, the non-destructive and efficient hyperspectral technology can serve as the necessary means in modern precision agriculture. This study aims to realize the rapid and accurate diagnosis for the prediction of the nitrogen nutrition status of winter rapeseed. Data sources were collected from 90 plant samples (above-ground biomass and plant nitrogen concentration) and hyperspectral measured data at the bud stage of winter oilseed rape under different mulching and nitrogen application treatments for two consecutive years (2022-2023). Eight typical spectral indices (16 in total) were first calculated using the correlation coefficients between the original and the first-order differential (FD) spectrum and the nitrogen nutrition index (NNI). Secondly, the correlation matrix was used to extract the best spectral combination. The optimal spectral index was then selected from the correlation coefficient with NNI. Finally, the optimal spectral index was divided into three groups of input variables for the model (combination 1: the five original spectral indices; combination 2: the five first-order differential spectral indices; combination 3: the five spectral indices, with the highest correlation coefficient with NNI). Support vector machine (SVM), random forest (RF), extreme learning machine (ELM), and back propagation neural network (BPNN) were used to construct the NNI estimation model of winter oilseed rape. Determination coefficient (R²), root mean square error (RMSE), and mean relative error (MRE) were used to evaluate the accuracy of the model. The results show that the correlation coefficient between the first-order differential spectral index and NNI was greater than that of the original, indicating a better NNI prediction using the first-order differential spectral index. There were high correlation coefficients between the spectral indices selected by the three combinations and NNI. The wavelength combinations were located in the red edge (670-760 nm), indicating the NNI variation representing the characteristic information in the red edge. The FDSAVI spectral index shared the highest correlation coefficient (0.674) with NNI. Among them, the wavelength combination was located at 712 and 678 nm. An optimal combination of the RF model and combination 2 were achieved to combine the optimal input variables of the NNI estimation model. The R² value, RMSE, and MRE on the validation set were 0.823, 0.079, and 7.513%, respectively, indicating the high accuracy of the optimal model. The finding can provide a technical basis for the potential application of remote sensing for the monitoring and diagnosis of plant nitrogen nutrition.

Leaf Nitrogen Concentration (LNC) can play a crucial role in the nutrient activity of soybeans, ultimately leading to the photosynthetic efficiency and yield formation. UAV-based remote sensing technology has been commonly used to estimate crop phenotypic traits. However, the traditional remote sensing monitoring of crop nitrogen can usually rely on the spectral data from a single sensor. It is mostly limited to the canopy-level assessment. In this study, a systematic analysis of the vertical heterogeneity of nitrogen was carried out to integrate the UAV-derived multispectral and texture features in the flowering period of soybeans. UAV multi-spectral technology was also combined with two-year (2021–2022) field experiments. The LNC data of each leaf layer were collected with its UAV multi-spectral image data. Firstly, a preliminary analysis was performed on the soybean LNC data. It was found that the LNC concentration in the canopy was higher than that in the middle layer, while the LNC concentration in the bottom layer was the lowest. Secondly, 10 vegetation indices were selected and then calculated. 48 texture features were extracted from 8 second-order probability matrices of each band. Subsequently, 6 texture indices were constructed after random combinations. There were some correlations between the parameters and the LNC of each leaf layer of soybeans. The correlation coefficients of the parameters reached a significant level (P<0.05). Among them, 4 combinations were constructed to serve as the input variables for the model: Combination 1 was the vegetation indices; Combination 2 was the texture features; Combination 3 was the texture indices; and Combination 4 was the group of the vegetation indices, texture features, and texture indices. Three machine-learning models were used to simulate the LNC in each leaf layer of the soybeans, namely Random Forest (RF), Back Propagation Neural Network (BPNN), and Extreme Gradient Boosting (XGBoost). The results showed that the correlation coefficients of the most vegetation indices with the LNC of each leaf layer of soybeans were higher than those of the texture features, thereby reaching a significant correlation level (P<0.05). The correlations with the LNC of each leaf layer decreased in the order of the canopy, middle layer, and bottom layer. The texture indices after random combinations shared the highest correlation coefficients with the LNC in each leaf layer of the soybeans. Among them, the LNC of the canopy, middle, and bottom layers of soybeans had the highest correlation coefficients with the constructed Additive Texture Index (ATI), which were 0.774, 0.726, and 0.650, respectively. Once the single-source data (Combination 1, Combination 2, and Combination 3) was used as the input variables to construct the prediction models, all three machine-learning models showed that the performance of Combination 3 (texture indices) was better than that of Combination 1 (vegetation indices), while Combination 2 (texture features) had the worst performance. The best performance was achieved in the prediction models with the LNC of each leaf layer of soybeans that was constructed by the XGBoost model, when the input variables were Combination 4 (vegetation indices, texture features, and texture indices). The canopy-layer LNC prediction model demonstrated that the coefficient of determination (R²) of the validation set was 0.853, the Root Mean Square Error (RMSE) was 0.321%, and the Mean Relative Error (MRE) was 7.120%; The middle-layer LNC prediction model indicated that the R² of the validation set was 0.822, the RMSE was 0.349%, and the MRE was 7.448%; The bottom-layer LNC prediction model showed that the R² of the validation set was 0.809, the RMSE was 0.340%, and the MRE was 8.042%. The finding can provide a practical technical path to monitor the physiological growth of the crops using an unmanned multispectral platform. Meanwhile, a reliable technical basis can also offer to dynamically monitor the vertical distribution of nitrogen in the precise fertilization of modern agriculture.