近红外光谱预测稻谷水分含量特征谱区选择及其效率分析

对364份水分含量在2.24%～32.66%之间的“冈优916”稻谷样品，经均值中心化、一阶微分、标准归一化和多元散射校正等预处理后，采用分段间隔法、组合分段法、滑动窗口法和反向分段法等进行特征谱区选择，分别使用偏最小二乘法(PLS)和主成分回归(PCR)两种定量分析方法，获得稻谷含水量近红外光谱预测模型最佳的特征谱区。首次给出了分段间隔法、组合分段法、滑动窗口法和反向分段法等传统的特征谱区选择方法的计算复杂度的计算公式，并对比分析了这几种特征谱区选择方法的程序运行效率。结果表明：采用PLS建模对稻谷含水量光谱的预测性能优于PCR建模，但PLS建模的效率低于PCR建模；在PLS建模中，采用反向分段法对稻谷光谱含水量的预测性能最好，其预测集的相关系数R_P为0.995 6，预测均方根偏差RMSEP为0.78%；其次是滑动窗口法，其R_P为0.994 3，RMSEP为0.89%；但这两种特征谱区选择方法的程序运行效率较低，反向分段法的平均运行时间为4.87 h，滑动窗口法的平均运行时间为29.82 h。该研究结果为今后在并行计算或分布式计算上开发近红外光谱预测模型的快速算法提供参考。

Feature Wavelength Selection and Efficiency Analysis for Paddy Moisture Content Prediction by Near Infrared Spectroscopy

In order to obtain the best feature wavelength region for predicting paddy moisture content (PMC) by near infrared spectroscopy (NIR), this research selected 364 paddy samples of “Gangyou 916” which moisture content varied from 32.66% to 2.24%, the pretreatments methods such as mean centering (Mean), standard normalized variate (SNV), Savitzky-Golay derivative (SG1) and multiplicative scatter correction (MSC) were performed, and adopted the feature wavelengths selection methods include interval method (IM), synergy interval method (SIM), moving window method (MWM) and backward interval method (BIM), then used partial least squares (PLS) and principal component regression (PCR) quantitative analysis algorithms for the PMC NIR modeling. The calculation formulas of complexity for wavelengths selection methods with IM, SIM, MWM and BIM are firstly provided in this paper. And this paper also compares and analyzes the program operating efficiency of these methods. The results show that the prediction ability of PLS is better than PCR, but the modeling efficiency of PLS is lower than PCR. The BIM is the optimum prediction model for PMC among the four wavelengths selection methods, which root mean square error of prediction (RMSEP) and correlation coefficient (R_p) in prediction set are 0.995 6 and 0.78%, respectively. The second is MWM, which RMSEP and RP are 0.994 3 and 0.89%, respectively. However, the program running efficiency of these two methods is relatively low. The average running time of BIM is 4.87 h, and average running time of MWM is 29.82 h. This work provided a reference comparison for a fast algorithm of near infrared spectroscopy prediction model in parallel computing and distributed computing.