基于高光谱图像光谱与纹理信息的生菜氮素含量检测

高光谱图像包含丰富的光谱与图像信息,该文基于此试图构建生菜氮素检测模型。利用高光谱图像采集系统获取可见-近红外(390~1 050 nm)范围内的生菜叶片高光谱图像,同时利用凯氏定氮法获取对应叶片的氮素值。将光谱反射值较大波长图像与反射值较小波长图像相除并用阈值化法构建掩膜图像,获取感兴趣区域(ROI,region of interest)。由于高光谱数据量大、且数据间冗余性强,因此如何有效的提取一些特征波长十分重要。该文采用主成分分析(PCA,principal component analysis)对原始高光谱图像进行处理,根据前3个主成分图像(PC1、PC2、PC3)在全波长下的权重系数分布图选出662.9、711.7、735.0、934.6 nm 4个特征波长及对应的光谱特征,并且分别提取4个特征波长图像、主成分图像PC1、PC2、PC3在ROI下的基于灰度共生矩阵的纹理特征,最后利用支持向量机回归(SVR,support vector machine regression)分别建立生菜叶片基于特征波长光谱特征、特征波长图像与主成分图像的纹理特征及光谱纹理融合特征与对应氮素值之间的关系模型。结果表明,在校正性能指标决定系数R2C上,基于光谱特征+特征波长图像纹理特征的模型较好,R2C=0.996,校正集均方根误差RMSEC为0.034;在预测性能指标决定系数R2P上,基于光谱特征的模型较好,R2P=0.86,预测集均方根误差RMSEP为0.22。该研究结果可为农作物氮素的快速、无损检测提供一定的参考价值。

Detection of nitrogen content in lettuce leaves based on spectroscopy and texture using hyperspectral imaging technology

Abstract: In this study, we developed a fast and non-destructive technology for the prediction of the nitrogen content in lettuce leaves based on hyperspectral imaging technology which contains abundant spectral and spatial information in an object. First, hyperspectral images of lettuce leaves in the visible and near infrared (390-1050 nm) regions were acquired by the hyperspectral imaging system, and then the corresponding nitrogen content in the lettuce leaves were obtained by Kjeldahl method successively. The binary mask image was successfully determined by the method of dividing the image of very high reflectance intensity by the image of very low reflectance intensity with a certain threshold, and ROI (Region of Interest) in the sample of lettuce leaf was determined by removing the regions of noise using the acquired binary mask image. As the hyperspectral imaging technology provided much more information including spectral and spatial information for all the samples of lettuce leaves, and in which some information is noisy and redundant. In fact, this leads to the difficulty of meeting the needs for fast and efficient detection of some objects. So it is very hard to be directly used for on-line industrial application in our daily life. Therefore, effective selection of several characteristic wavelengths is necessary for the hyperspectral images. In this paper, the initial investigation was carried out by using a principal component analysis (PCA) to identify a number of potential characteristic wavelengths (662.9 nm, 711.7 nm, 735.0 nm, 934.6 nm) according to the weight coefficient distribution curve of the first three principal component images (PC1, PC2, PC3) under the full wavelengths. Both spectral data and texture data based on a co-occurrence matrix were extracted from the four characteristic wavelength images on the ROI, and the texture data of the first three principal component images were also extracted simultaneously. Then spectral data from four characteristic wavelength images, texture data (from four characteristic wavelength images, from the first three principle component images) and the combined data were utilized to develop different SVR (support vector machine regression) models to predict the nitrogen content in the lettuce leaves respectively. According to the performance of all the SVR models in the calibration set and the prediction set, the experiment results show that, from the calibration performance index, the model based on spectral data combined texture data from four characteristic wavelength images is the best with a coefficient of determination (R2C =0.996) and the root-mean-square error (RMSEC) of 0.034. From the prediction performance index, however the model based merely on spectral data is the best with a coefficient of determination (R2P=0.86) and the root-mean-square error (RMSEP) of 0.22. This study provides valuable information for rapid and non-destructive nitrogen detection in crops.