Citrus is the largest kind of fruit in China. Nitrogen is very important for the growth and development of citrus. Real-time and non-destructive monitoring of the nitrogen status of citrus is of great significance for accurate management of nitrogen nutrients. Nitrogen in plants can be divided into assimilable nitrogen, structural nitrogen and functional nitrogen. The content of each component of different forms of nitrogen in citrus leaves has a certain indicative effect on the physiological and biochemical reactions of leaves. Among them, the content of functional nitrogen is an important indicator of nitrogen nutrition status in citrus. Chunjian orange was used as the experimental material in this study. The reflectance spectra of citrus leaves under different nitrogen treatments were measured by the visible-near infrared spectrometer at the fruit swelling period and fruit coloring period, and the functional nitrogen content in leaves was determined by chemical analysis. The correlation between the original spectrum, first-order differential spectrum and the functional nitrogen content of leaves at the fruit swelling and fruit coloring periods of citrus was analyzed, and the sensitive bands were selected. The non-destructive monitoring model of the functional nitrogen content of leaves at the fruit swelling period and fruit coloring period of citrus was constructed by using the full-band and sensitive bands, combined with the spectral vegetation index method, spectral chemical measurement method and machine learning method, and the effects of various spectral variants and spectral preprocessing methods on the accuracy of the model were compared and analyzed. The results showed that the non-destructive monitoring model of functional nitrogen content in citrus leaves constructed by standard normal variate transformation pretreatment of the full-band original spectrum combined with the backpropagation neural network had high accuracy during thefruit swelling period. The calibration set determination coefficient R-c(2) and validation set determination coefficient R-v(2) were all 0.78, and the RMSEC and RMSEV of the modeling set were all 0.82 g center dot kg(-1). The model accuracy based on the original spectrum of the sensitive band combined with the random forest was also high, with R-c(2) and RMSEC were 0.84 and 0.67 g center dot kg(-1), R-v(2) and RMSEV were 0.74 and 0.83 g center dot kg(-1), respectively. In the fruit coloring period of citrus, the full-band original spectrum was preprocessed by standard normal variate transformation. The accuracy of the non-destructive monitoring model of functional nitrogen content in citrus leaves constructed by BPNN was high, with R-c(2) and RMSEC was 0.77 and 1.04 g center dot kg(-1), R-v(2) and RMSEV were 0.76 and 1.13 g center dot kg(-1), respectively. The study has shown that visible-near infrared spectroscopy can achieve non-destructive monitoring of functional nitrogen content in citrus leaves.
