基于高光谱图像的龙眼叶片叶绿素含量分布模型

Distribution Model of Chlorophyll Content for Longan Leaves Based on Hyperspectral Imaging Technology

针对传统高光谱单点法检测叶绿素含量效率低、精度不足等问题,提出一种基于高光谱图像和卷积神经网络(CNN)多特征融合的深度学习龙眼叶片叶绿素含量分布预测模型。首先进行Savitzky-Golay光谱去噪,然后通过奇异值分解(SVD)和独立成分分析(ICA)提取特征光谱,再对特征光谱图像提取灰度共生矩阵(GLCM)和CNN纹理特征,最后建立粒子群优化(PSO)支持向量回归(SVR)、深度神经网络(DNNs)分布模型。结果表明,基于特征光谱建模的PSO-SVR预测效果最佳,全期的校正集和验证集模型决定系数R2为0.822 0和0.815 2。对比多种主流模型,基于特征光谱、GLCM纹理、CNN纹理特征的ICA-DNNs模型预测精度最高,校正集和验证集R2分别为0.835 8和0.821 0。试验结果表明,高光谱图像可快速无损地对龙眼叶片叶绿素含量分布进行检测,可为龙眼树实时营养监测和病害早期防治提供理论依据。

Traditional hyperspectral single point detection methods of obtaining chlorophyll content of longan leaves are inefficiency,low accuracy and time consuming. Combined with the state-of-the-art deep learning technology,a distribution model of chlorophyll content for longan leaves based on convolution neural networks（CNN） and deep neural networks（DNNs） was proposed. Firstly,the spectral noise was reduced by Savitzky-Golay filter. The initial features extraction was carried out by using a principle component analysis（PCA） to identity a number of potential characteristic wavelengths（483 nm,518 nm,625 nm,631 nm,642 nm and 675 nm） according to the weight coefficient distribution curve of the first three principle component images（PC1,PC2 and PC3） under the full wavelengths. For the characteristic spectral images and the principal component images,the texture based on the gray level cooccurrence matrix was extracted from those images,and the structure information of those images was also extracted based on CNN simultaneously. Among the 300 samples,there were total of 1 800 spectral images and 900 principal component images,in which a sample corresponding to six characteristic spectrum images and three PCA images for a sample. Gray-level co-occurrence matrix（GLCM） was utilized to extract texture features. The hyperspectral wavelength feature data,texture data,images structure data and the combined data were utilized to develop particle swarm optimization-support vector regression（PSO-SVR） and independent component analysis-deep neural networks（ICA-DNNs）,respectively. The particle swarm optimization（PSO） was introduced to intelligently optimize the parameters（γ and c） in the SVR model to find the optimum. Some main conclusions ware obtained： the performance of PSO-SVR model based on characteristic spectrum was the best,and the coefficient of determination（R2） of calibration set and validation set of the entire growth state were up to 0. 822 0 and0. 815 2, respectively. Multi-source data fusion performance of ICA-DNNs was the best, and the precision of ICA-DNNs model was improved based on feature spectrum,texture feature and images structural feature,the R2 of calibration set and validation set were 0. 835 8 and 0. 821 0,respectively.Compared with traditional methods of SVR,DNNs was more robust for lager data set. The longan leaf textures and CNN characteristics were less relevant to longan chlorophyll content distribution. Chlorophyll content distribution region for tender,pale and dark green longan leaves were： mesophyll of the leaf-root,part of mesophyll of lateral veins and the whole mesophyll. Finally,hyperspectral technology could obtain accurate chlorophyll content of longan leaves rapidly,quantitatively and non-destructively. The research result can provide a theoretical basis for nutrition surveillance of longan growth and longan disease such as leaf spot,brown spot and leaf blight.