光谱学与光谱分析, 2016, 36 (10): 3237, 网络出版: 2016-12-30
高光谱图像信息检测玉米籽粒胚水分含量
Measuring the Moisture Content in Maize Kernel Based on Hyperspctral Image of Embryo Region
高光谱成像 玉米籽粒 胚 水分 无损检测 Hyperspectral imaging Maize kernel Embryo Moisture content Nondestructive determination
摘要
通过波段比和阈值相结合的方法, 分别提取了玉米籽粒全表面结构和胚结构区域的1 000~2 500 nm近红外高光谱信息, 研究了玉米籽粒水分含量与胚结构区域光谱关系, 同时采用竞争性自适应重加权变量选择算法(CARS)、 遗传算法(GA) 、 连续投影算法(SPA)筛选特征波段, 建立并比较偏最小二乘回归(PLS)模型对水分含量的预测效果。 结果显示, 玉米籽粒水分含量与胚结构区域光谱关系显著, 随着水分含量的增加, 光谱反射值逐渐降低。 预测模型结果表明, 基于玉米籽粒胚结构区域光谱信息所建立的CARS-PLS, GA-PLS和SPA-PLS回归模型预测相关系数Rp分别为0.931 2, 0.917 6和0.922 7, 预测均方根误差(RMSEP)分别为0.315 3, 0.336 9和0.336 6, 所选取的特征波段数量分别为9, 14和6, 较基于全表面光谱信息所建模型的特征波段数量分别少了49, 12和24个, 且预测效果与采用全表面光谱信息无显著差别, SPA-PLS算法为基于玉米籽粒胚结构光谱信息的水分含量预测最高效模型。 提取胚结构区域所用光谱波段为1 197, 1 322和1 495 nm, 建立SPA-PLS回归模型所用特征波段为1 322, 1 342, 1 367, 1 949, 2 070和2 496 nm。 研究结果表明, 采用近红外高光谱技术进行玉米籽粒水分含量无损检测时, 提取玉米籽粒胚结构的图谱信息较全表面光谱信息更高效。
Abstract
Maize is among the most important economic corps in China while moisture content is a critical parameterin the process of storage and breeding. To measure the moisture content in maize kernel, a near-infrared hyperspectral imaging system has been built to acquire reflectance images from maize kernel samples in the spectral region between 1 000 and 2 500 nm. Near-infrared hyperspectral information of full surface and embryo of maize kernel were firstly extracted based on band ratio coupled with a simple thresholding method and the spectra analysis between moisture content in maize kernel and embryo was performed. The characteristic bands were then selected with the help of Competitive Adaptive Reweighted Sampling (CARS), Genetic Algorithm (GA) and Successive Projection Algorithm (SPA). Finally, these selected variables were used as the inputs to build Partial Least Square (PLS) models for determining the moisture content of maize kernel. In this study, a significant relation, which the spectral reflectance decreases as moisture content increase, between moisture content and spectral of embryo in maize kernel was observed. For the investigated independent test samples, all the proposed regression models, namely CARS-PLS, GA-PLS and SPA-PLS, achieved a good performance by using the information of embryo region. The correlation coefficient (Rp) and Root Mean Squared Error of Prediction (RMSEP) and number of characteristic wavelength for the prediction set were 0.931 2, 0.315 3, 9 and 0.917 6, 0.336 9, 14 and 0.922 7, 0.336 6, 16 for CARS-PLS, GA-PLS and SPA-PLS models, respectively. And, compared with models obtained by full surface spectral information, less characteristic wavelengths is used for development of CARS-PLS, GA-PLS and SPA-PLS models, while similar results were obtained. Comprehensively analyzing to both model accuracy and model complexity, SPA-PLS model by using embryo region information achieved the best result. Wavelengths at 1 197, 1 322 and 1 495 nm were applied to extracted the information of embryo region, and the bands at 1 322, 1 342, 1 367, 1 949, 2 070 and 2 496 nm were used to establish the SPA-PLS model. These results demonstrated that near-infrared hyperspectral information from embryo region is more effective for determination of moisture nondestructive in maize kernel.
田喜, 黄文倩, 李江波, 樊书祥, 张保华. 高光谱图像信息检测玉米籽粒胚水分含量[J]. 光谱学与光谱分析, 2016, 36(10): 3237. TIAN Xi, HUANG Wen-qian, LI Jiang-bo, FAN Shu-xiang, ZHANG Bao-hua. Measuring the Moisture Content in Maize Kernel Based on Hyperspctral Image of Embryo Region[J]. Spectroscopy and Spectral Analysis, 2016, 36(10): 3237.