基于高光谱的寒地水稻叶片氮素含量预测

为快速、无损和准确地诊断水稻营养状况,开展了基于高光谱成像技术的寒地水稻叶片氮素含量预测研究。以不同施氮水平下的水稻叶片为研究对象,利用高光谱成像技术,分析拔节期水稻叶片光谱,采用全波段高光谱数据、连续投影算法及分段主成分分析(segmented principal components analysis,SPCA)与相关分析(correlation analysis,CA)相结合的方法建立多种回归分析模型,并对模型进行检验和筛选。结果表明:随着施氮水平提高,水稻叶片反射率在可见光区域降低,在近红外区域升高。在校正集决定系数上,基于多元逐步回归分析的全波段模型较好,校正集决定系数为0.821,校正集均方根误差RMSEC=0.079;在预测集决定系数上,基于SPCA-CA结合多元回归分析的多变量单波段指数、差值指数、双差值指数模型较好,预测集决定系数为0.869,预测集均方根误差RMSEP=0.085。该研究结果为快速检测水稻叶片氮素含量及水稻生长期间精确施肥管理提供了参考。

Prediction for nitrogen content of rice leaves in cold region based on hyperspectrum

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.