基于BP神经网络的烟草叶片质体色素高光谱反演

摘要

为促进高光谱遥感技术在烟叶品质监测中的应用，以云烟87为研究对象，设置不同的光质处理，测定了不同生育时期烤烟叶片的质体色素（叶绿素a、叶绿素b和类胡萝卜素）含量（质量分数）及相应的叶片光谱反射率。采用小波能量系数提取法将原始光谱数据从2151个降维为21个，并分别建立了以质体色素实测值作为输出因子，降维后的光谱数据作为输入因子的BP神经网络预测模型，训练函数采用L-M优化算法函数trainlm，输入层和输出层传递函数分别为S型正切传递函数tansig和线性传递函数purelin，叶绿素a、叶绿素b和类胡萝卜素的神经网络的隐含层节点数分别为27、32和45。结果表明，叶绿素a、叶绿素b和类胡萝卜素各模型决定系数R2分别为0.84、0.86和0.76；均方根误差RMSE分别为0.12、0.14和0.10；相对误差绝对平均值K分别为0.23、0.21和0.15。3种质体色素模型的拟合精确度均较高，误差较小，整体效果良好。%To promote the application of hyperspectral remote sensing technology to tobacco quality monitoring, experiments were conducted using different light quality to examine cv. Yunyan 87, the contents of plastid pigments (chlorophyll a, chlorophyll b and carotenoid) and their corresponding hyperspectral reflectances of flue-cured tobacco leaves at different growing stages. Hyperspectral reflectance data was reduced from 2 151 to 21 by wavelet coefficient extraction, then the reduced inversion models based on back propagation neural network for predicting the contents of chlorophyll a, chlorophyll b, carotenoid in tobacco leaves were established with plastid pigments as output factors and reduced hyperspectral reflectance data as input factors. The training function adopted trainlm of L-M optimization algorithm, while the transfer functions of the input layer and output layer were tansig and purelin, respectively. The number of the hidden layer nodes in BP neural network models for chlorophyll a, chlorophyll b and carotenoid was 27, 32 and 45, respectively. The results showed that the determination coefficients of BP neural network models for chlorophyll a, chlorophyll b and carotenoid were 0.84, 0.86 and 0.76; the root mean square errors were 0.12, 0.14 and 0.10; the absolute mean relative errors were 0.23, 0.21, and 0.15, respectively. The three models well predict the contents of plastid pigments with high precisions and low errors.