Rapid Detection Of Nitrogen Content In Citrus Leaves Based On Hyperspectral Imaging Technology

Nitrogen, as an indispensable element for the growth and development of fruit trees, not only is an important part of the chemical compounds such as protein, chlorophyll and enzyme, but also affects the fruit trees’growth, fruit yield and quality improvement. Therefore, timely and accurate monitoring of the fruit trees’nitrogen level not only is good for rational fertilization, production increase and improvement of fruit quality, but also mitigates the surface and underground water resource pollution caused by excessive nitrogen fertilization. This study selected citrus plant as the experimental fruit crop. Leaves’ nitrogen content were measured by rapid N cube, and based on the hyperspectral images of fresh citrus leaves, this study picked out optimal pretreatment methods and appropriate characteristic wavelengths, and also established several effective prediction models for determining the N content in citrus leaves with chemometrics methods. Furthermore, the texture features of the citrus leaves were extracted to build prediction models, and finally, new and reliable two bands vegetation indices were developed based on conventional ones. The main research conclusions obtained in this study are as follows:(1)Three optimal pretreatment methods were selected from eleven different pretreatment methods including Savitzky-Golay (SG) smoothing, standard normal variate (SNV), multiplicative scatter correction (MSC) and so on. They were SG smoothing, Detrending and SG smoothing-Detrending. After treating raw spectral data with these three selected pretreatment methods, successive projection algorithm (SPA) was used to select characteristic wavelengths, which were then used as the inputs for prediction models established with partial least squares regression (PLS), multiple linear regression (MLR) and back propagation neural network (BPNN) algorithms. As a result, nine models were obtained, among which the two models based on the methods of SG smoothing-Detrending-SPA-BPNN (Rp:0.8513, RMSEP:0.1881) and Detrending-SPA-BPNN (Rp:0.8609, RMSEP:0.1595) achieved the best prediction results.(2)Based on these three optimal pretreatment methods, BPNN models were found to have better prediction results than PLS and MLR models, which indicated that the nonlinear regression calibration methods (such as BPNN) might be more suitable than the linear ones (such as PLS and MLR) for nitrogen prediction modeling of citrus leaves.(3)Based on gray histogram, the BPNN nitrogen prediction model achieved good prediction result (Rp=0.8058, RMSEP=0.1847), showing that the texture feature variables of gray histogram had potential for citrus leaves’ nitrogen prediction. The MLR prediction models based on gray level co-occurrence matrix and both of gray histogram and gray level co-occurrence matrix achieved good correlation coefficient for modeling datasets(Rc>0.9) but showed terrible correlation for prediction datasets(Rp<0.2). This indicated that the nitrogen prediction models based on texture feature variables needed to be improved.(4)These four new developed two bands vegetation indices including (Rλ1-Rλ2)/(Rλ1+Rλ2),(Rλ12-Rλ22)/(Rλ12+Rλ22),(Rλ1-Rλ2)/Rλ2and Rλ1/Rλ2had the best correlation (R>0.8) with nitrogen content with the characteristic bands combination of λ1=856nm and λ2=814nm.(5)Models based on (R856-R814)/(R856+R814),(R8562-R8142)/(R8562+R8142),(R856-R814)/R814and R856/R814had the similar prediction results (Rc>0.8and Rp>0.72). Based on these four vegetation indices, all the quadratic polynomial function models obtained good prediction results (Rc>0.84and Rp>0.79), and the new developed two bands vegetation indices had better nitrogen prediction results than conventional vegetation indices and red edge parameters, suggesting that the new two bands vegetation indices with the characteristic bands combination of λ1=856nm and λ2=814nm had potential for citrus nitrogen status prediction.