Abstract: In order to predict the softness of tobacco leaves by using the appearance parameters of fresh tobacco leaves, a LM-BP neural network prediction model was established by studying the relationship between the field ripening appearance parameters such as lightness and darkness (L), redness value (a), yellowness value (b), color saturation (C), hue angle (H), SPAD value, etc. and the softness of post-roasted leaves of YUNYU 87 with different retention numbers of the upper leaves. The results showed that the appearance characteristic parameters of upper leaves with different numbers of retained leaves were different, and the softness of tobacco leaves after baking was also different, and the value of softness after baking was lower in the number of retained leaves of 19 leaves than that in the number of leaves of 16-18 leaves, which ranged from 5.62 to 13.29 mN; There was a correlation between the parameters of tobacco appearance characteristics and the softness of post-roasted tobacco; stepwise regression analysis screened out the factors with greater influence on the softness of post-roasted tobacco as the number of retained leaves, L, H and SPAD value; The LM algorithm was used to replace the gradient algorithm to create the LM-BP neural network prediction model, and the training results showed that the prediction accuracy R2 was close to 1, the average absolute percentage error MAPE<5%, and the root-mean-square error RMSE<2. Properly retaining more leaves increases the softness of the tobacco after roasting; There was a correlation between the maturation appearance characteristics of tobacco leaves in the field and the softness of the tobacco after roasting; The LM-BP neural network was used to create a prediction model with high accuracy, which can be used for intelligent judgement of tobacco maturity in the field.

Key words: roasted tobacco, color value, tobacco softness, BP neural network

摘要： 为利用鲜烟叶外观参数预测烟叶柔软度，通过研究云烟87不同留叶数上部叶田间成熟外观特征参数明暗度(L)、红度值(a)、黄度值(b)、色彩饱和度(C)、色调角(H)、SPAD值等与烤后烟叶柔软度的关系，建立其LM-BP神经网络预测模型。结果表明，不同留叶数的上部叶外观特征参数不同，烤后烟叶柔软度也不同，留叶数19叶的较16～18叶的烟叶烤后柔软度值低5.62～13.29 mN；烟叶外观特征参数与烤后烟叶柔软度间存在相关性；逐步回归分析筛选出对烤后烟叶柔软度影响较大的因子为留叶数、L、H和SPAD值；采用LM算法替代梯度算法创建LM-BP神经网络预测模型，其训练结果预测精度R2均接近1，平均绝对百分误差MAPE<5%，均方根误差RMSE<2。适当多留叶可增加烤后烟叶的柔软度；烟叶田间成熟外观特征参数与烤后烟叶柔软度之间存在相关性；采用LM-BP神经网络创建预测模型准确率较高，可用于烟叶田间成熟智能化判断。

关键词: 烤烟, 颜色值, 烟叶柔软度, BP神经网络