基于颜色参数和BP神经网络的紫叶李叶片花青素含量估算

目前对植物叶片花青素含量的测定主要是湿化学法和高效液相色谱法(high performance liquid chromatography,HPLC)，为简化测定方法，降低成本和提高精度，提出一种利用数码相机获取照片提取的颜色参数构建模型无损估测植物叶片花青素含量的方法。试验测定166份紫叶李叶片的花青素含量及其RGB特征值，对15种颜色参数进行皮尔逊相关分析，构建逐步多元线性回归（stepwise multiple linear regression,SMLR）、一元线性回归（single linear regression,SLR）和BP神经网络（BP neural network,BPNN）估算模型；同时对模型进行验证和比较。结果表明，1）BP神经网络模型建模集的R²、RMSE和MAE分别为0.883、0.412、0.323，验证集的R²、RMSE和MAE分别为0.796、0.462和0.353，相关系数均达到极显著水平；一元线性回归模型中，参数G-B与花青素含量的线性相关性最强，相关系数为-0.820，达到极显著水平；逐步多元线性回归模型的相关系数均达极显著水平，其中建模集的R²、RMSE和MAE分别为0.724、0.630、0.459，验证集的R²、RMSE和MAE分别为0.643、0.616和0.509。2）颜色参数与花青素含量之间具有明显的相关性，利用数码相机获取的颜色特征值估测紫叶李叶片花青素含量具有可行性；3）3种模型中，BP神经网络模型的估测效果最好，能有效地估测紫叶李叶片花青素含量，其次为逐步多元线性回归，一元线性回归模型的预测效果相对较差。

Estimation of Leaf Anthocyanin Content in Prunus cerasifera Based on Color Indices and BP Neural Network

For the determination of anthocyanins content in plant leaf is now mainly wet chemical method and high performance liquid chromatography (HPLC) method.In order to further simplify the determination method,reduce the cost and improve the accuracy of determination,this study proposed a method of estimating leaf anthocyanin content in Prunus cerasifera by using color indices obtained by digital camera to build models.Totally,166 leaf samples of P.cerasifera were measured.The content of anthocyanin was determined by wet chemical method,in the meantime,the ROI tool of ENVI4.3 was used to extract the RGB mean values of selected areas of the leaves in digital photos.Based on the basic eigenvalues of R,G and B,15 color indices were obtained through simple arithmetic combination and form transformation.And then Pearson correlation analysis was performed on the correlation between the 15 color indices and anthocyanin content.Taking these indices as input values,BP Neural Network (BPNN),the single linear regression (SLR) and stepwise multiple linear regression (SMLR) models were established,where the measured anthocyanin contents were set as output values.The results showed that 1) the estimation effect of BPNN model was better,the values of R²,RMSE and MAE of the modeling set were 0.883,0.412,and 0.323,respectively,and the values of verification set’s R²,RMSE,and MAE were 0.796,0.462,and 0.353,respectively.The correlation coefficients reached the extremely significant level.In the univariate linear regression model,the G-B index had the strongest linear correlation with anthocyanin content and the correlation coefficient was -0.820,which reached the extremely significant level.The correlation coefficients of SMLR models reached the extremely significant level,and the R²,RMSE and MAE of the modeling set were 0.724,0.630,and 0.495,respectively.The values of R²,RMSE and MAE of validation set were 0.643,0.616,and 0.509,respectively.2) Compared with the other two models,BPNN model had the best estimation effect,which could effectively estimate the content of leaf anthocyanins in P.cerasifera,followed by SMLR,and SLR model had relatively poor prediction effect.