|
Remote sensing has become an important tool for estimating crop biochemical components such as nitrogen, chlorophyll and carbon, particularly in recent years. Near infrared spectroscopy (NIRS) technology, which is quick and accurate, has provided an effective tool for monitoring growth information of crop plant. In this study, a series of field experiments with different rice varieties and nitrogen levels were carried out in two years. Time-course near infrared spectra (12 500～4 000 cm-1) were taken by Fourier transform near infrared spectrometer in fresh and powder rice leaves. The objective of the research was to develop estimation models for the contents of nitrogen, pigment and soluble sugar, and the ratio of sugar to nitrogen in rice leaves, using different methods of chemometrics including partial least squares (PLS), principal component regression (PCR), stepwise multiple linear regression (SMLR) and back-propagation artificial neural network (BPNN), which would help to realize quantitative diagnosis of rice growth status with NIR technique.At first, the relationships between NIRS and nitrogen content (NC) in rice leaf and powder were analyzed, using different methods of spectra preprocessing in 12 500～4 000 cm-1,8 000～4 500 cm-1,8 000～6 000 cm-1 and 5 500～4 500 cm-1. At the same time, comparative analysis of PLS, PCR, SMLR and BPNN, four NIRS-based models were inspected for estimating nitrogen content in rice. The results showed that the determination coefficient of calibration (RC2) of models for NC in fresh leaf was 0.940, and the root mean square error of calibration (RMSEC) was 0.226, and the best principle constituent factor was 7. The RC2 of models for NC in leaf powder was 0.977, the RMSEC was 0.136, and the best principle constituent factor was 6. And the accuracy of models was evaluated with independent experiment dataset by the determination coefficients of cross-validation (RCV2) and the root mean square errors of cross-validation (RMSECV), the determination coefficients of external validation (RV2) and the root mean square errors of external validation (RMSEP). The RCV2 of models for NC in fresh leaves was 0.857, with RMSEV as 0.251. The models for NC in leaf powder had the Rv2 of 0.897, with RMSEP of 0.211. The RV2 of the models for NC in fresh leaves was greater than 0.800, with RMSEP less than 0.500. The NIRS-based models for NC in leaf powder gave the Rv2 of 0.916, with RMSEP of 0.225. In general, the performance of the models for leaf powder is more satisfactory than that of the fresh leaves.Then, the above-mentioned method was used for establishing quantitative models for estimating leaf Chla+b, Chla, Chlb and Car concentrations. The results indicated that the best pretreatment of the model was first derivate+NDF, and the best wave number was 8 000～4 000cm-1. The best method for estimating Chla, Chlb, Chla+b and Car was PLS, the RC2 all reached 0.8 except for Chla+b model, with the best principal constituent factor as 8. The cross validation results showed that the RCV2 values of Chla, Chla+b and Car were all above 0.8, and the RCV2 of Chlb was 0.796. The external validation exhibited that the best performance was from Chlb model, with Rv2 of 0.842 and RMSEP of 0.123. The Chla and Chla+b models performed less, with Rv2 of 0.741 and 0.749, respectively. And the Car model expressed worst, with RV2 of only 0.625.Finally, the feasibility of spectral sensing for the soluble sugar content (SSN) and the ratio of sugar to nitrogen (S/N) was investigated. The results showed that the pretreatment of first derivative could obviously improve the models’precision. The SSC model of fresh leaf and powder chose 12 500～4 000cm-1 and 5 500～4 500cm-1, with the RC2 as 0.957 and 0.910, respectively. The S/N model of fresh leaf and powder chose 6 800～5 600cm-1 and 7 500-6 000cm-1, with the RC2 as 0.948 and 0.972, respectively. The best method for estimating SSC and S/N of fresh leaf and powder was PLS, with the best principal constituent factors as 7,4,10 and 8, respectively, and the RC2 greater than 0.9 except for the SSC model of fresh leaf. The cross-validation and external validation results showed that the model of powder had higher accuracy and precision than fresh leaf model, while the model of S/N performed better than SSN model. The NIRS-based models for S/N in leaf powder gave best performace, with Rv2 of 0.863 and RMSEP of 0.562.
|