基于生态因子和机器学习的烤烟主要化学成分预测模型构建OA

[Objective]The ecological prediction models for forecasting major chemical components of flue-cured tobacco were constructed based on the data of major chemical components of flue-cured tobacco and ecological influencing factors to provide a scientific basis for improving meteorological service work,high-quality production and quality improvement of flue-cured tobacco.[Method]The correlations between major chemical components of K326,a flue-cured tobacco variety,and ecological environmental influencing factors(meteorology and soil)were analyzed by using gray relationship analysis.The ecological prediction models for forecasting major chemical components of flue-cured tobacco were constructed according to screening model input characteristics based on a machine learning method(SVR algorithm and LightGBM algorithm).The SVR prediction model and LightGBM prediction model were optimized by the grid search cross-validation algorithm and Bayesian algorithm to improve the model precision.The simulation effect of two optimized intelligence algorithm prediction models was compared and analyzed under the same computing environment.[Result]The correlation coefficients between the sunlight duration at vigorous growth stage and total sugar,reducing sugar,total nitrogen and potassium oxide content in flue-cured tobacco leveas all were more than 0.700.The correlation coefficient between the average temperature in July and chlorine content and between average temperature during the field growth period,average temperature during vigorous growth stage and nicotine content reached＞0.900 and＞0.690 respectively.The correlation coefficients between soil organic matter content and nicotine,reducing sugar,total sugar,total nitrogen content in flue-cured tobacco leaves all reached＞0.650.The correlation coefficients between soil total nitrogen and nicotine,total nitrogen,potassium oxide,chlorine both reached＞0.660.The overall performance indicators of the optimized prediction models with stronger generalization performance were significantly improved.The MAE and RMSE of six GSCV-SVR chemical component prediction models for forecasting nicotine,chlorine,potassium oxide,total nitrogen,reducing sugar and total sugar content in flue-cured tobacco leaves were 0.012 4-0.700 8 and 0.022 4-1.134 5 respectively.The MAE and RMSE of six BO-LightGBM chemical component prediction models for forecasting nicotine,chlorine,potassium oxide,total nitrogen,reducing sugar and total sugar content in flue-cured tobacco leaves were 0.004 7-0.710 9 and 0.013 7-1.136 6 respectively.The RMSE of BO-LightGBM prediction models for forecasting total sugar,nicotine,chlorine,potassium oxide and total nitrogen content in flue-cured tobacco leaves decreased by 0.252 2,0.013 6,0.008 7,0.005 3 and 0.005 2 compared with the RMSE of GSCV-SVR prediction models respectively.And the R2 of BO-LightGBM prediction models for forecasting total sugar,nicotine,chlorine,potassium oxide and total nitrogen content in flue-cured tobacco leaves increased by 0.078 7,0.062 2,0.009 8,0.015 9 and 0.024 3 compared with the R2 of GSCV-SVR prediction models respectively.[Conclusion]The BO-LightGBM model can improve the prediction precision for forecasting the content of major chemical components in flue-cured tobacco leaves.