Prediction for Nitrogen Content of Rice Leaves in Cold Region Based on Hyperspectrum
Wang Shuwen,Zhao Yue,Wang Lifeng,Wang Runtao,Song Yuzhu,Zhang Changli,Su Zhongbin
DOI: https://doi.org/10.11975/j.issn.1002-6819.2016.20.024
2016-01-01
Abstract:In this paper, in order to realize the quick, non-destructive and accurate diagnosis of rice nutritional status, we use hyperspectral imaging techniques as an approach for nitrogen content prediction of rice leaves in cold region. The experiments were carried out for two years (2014 and 2015) at Fangzheng country, Heilongjiang province, China. Longdao 20 was chosen as the test cultivar. 6 nitrogen fertilization rates were applied in our experiments, i.e., N0 (0 kg/hm2), N1 (60 kg/hm2), N2 (90 kg/hm2), N3 (120 kg/hm2), N4 (150 kg/hm2), and N5 (180 kg/hm2). The hyperspectral reflectance and nitrogen content of rice leaves under different nitrogen levels at jointing stage were separately measured using American Headwall imaging spectrometer and German AA3 analyzer. The hyperspectral images of 240 rice leaf samples in the spectral range of 400-1 000 nm were acquired. Average spectrum was extracted from the region of interest (ROI) of each sample. Several regression analysis (RA) estimate models have been built based on different characteristic spectral parameters using different algorithms which include successive projections algorithm (SPA) and segmented principal components analysis (SPCA) combined with correlation analysis (CA) for testing and screening. The first method, a nitrogen content value estimation model based on multiple stepwise regression analysis (MSRA) in the whole wavelength region of 400~1000nm has been built and been predicted. Wavelengths 899, 890 nm were retained as the model independent variables. The second method, 8 characteristic wavelengths i.e., 454, 460, 475, 504, 525, 685, 700 and 735 nm were chosen by SPA and selected as modeling variables of MSRA. Wavelengths 735, 525 nm were chosen as the model independent variables. The third method, we divided the whole wavelength into 5 parts which are 400-504, 505-670, 671-697, 698-724 and 725-1 000 nm using correlation coefficient matrix method. The principal component analysis (PCA) was carried out on each part, and 7 sensitive bands were selected according to the contribution rate of each component. The correlation between the nitrogen content of rice leaves and the characteristic spectral parameters which consist of sensitive bands was analyzed, and 11 characteristic spectral parameters, i.e., single band index (SI(558), SI(866)), red edge position index (REPI(709)), ratio index (RI(866,670)), difference index (DI(730,715), DI(730,558)), double difference index (DDI(730,715,685), DDI(866,685,558)), normalized difference index (NDI(866,670), NDI(866,685)) and green normalized difference index (GNDI(730,558)) were selected to establish simple regression analysis (SRA) models. 4 kinds of characteristic spectral parameters with the highest coefficients of determination (RC2 and RP2) were selected to establish multiple regression analysis (MRA) models. We predicted all the estimate models so that testing its accuracy and stability. The results indicated that, reflectance of rice leaves decreases in the visible region, and increases in the near infrared region, with the raise of nitrogen level. From the calibration performance index, the whole wavelength model based on MSRA is the best with a coefficient of determination (RC2=0.821) and root mean square error (MRSEC) of 0.079. From the prediction performance index, the multivariate SI(866), DI(730,715), DDI(730,715,685), DDI(866,685,558) based on SPCA-CA combined MRA is the best with a coefficient of determination (RP2=0.869) and the root mean square error (RMSEP) of 0.085. The study provides technical support and theoretical basis for the rapid detection of nitrogen content in rice leaves and the precise fertilization management during rice growth.