Prediction Of Tobacco Spodoptera Litura Occurrence Based On Machine Learning

The prevention and control of agricultural pests and diseases is an indispensable work to ensure the orderly development of agricultural production.Tobacco planting,one of China’s main economic crops,also suffer from pests.The prevention and control of traditional chemical pesticides runs counter to the concept of sustainable agricultural development due to the environmental pollution,pesticide resistance,and pesticide residues.Therefore,the development of accurate prediction methods for pest occurrence can provide guidance for the formulation of comprehensive control plans,reduce the application of chemical pesticides,and improve tobacco yield and quality.It also is of great significance to the development of green agricultural pest control methods.The occurrence of tobacco insect pests is characterized by multiple generations of damage,which makes its occurrence not only affected by external factors such as meteorological factors,field management,and geographical pattern,but also has a very strong time after effect,that is,it is affected by the previous insect source.Insect pest prediction models based only on external environmental factors cannot reflect their temporal aftereffects,and their prediction performance is limited.This study takes the occurrence of tobacco Spodoptera litura as an example,introduces geostatistics to mine its temporal aftereffects,and combines meteorological factors to integrate multiple feature selection algorithms and machine learning models to predict its occurrence.The results show that the occurrence of Spodoptera litura has a very obvious time after effect,and its global aging step is 6 days.After comparing multiple models,the SVM-MBO model based on Support Vector Machine(SVM)and multi-round selection has the best prediction accuracy,and its Mean Squared Error(MSE)is only 14.56.Further combining the prediction results of multiple models,the MSE of the integrated model is reduced to 5.13.In addition,this study initially explored the application of deep learning models in the prediction of the occurrence of Prodenia litura.First,two types of deep neural networks based on Fully Connected Network(FCN)and Convolutional Neural Network(CNN)are designed,and the influence of multiple hyperparameter combinations on the model training effect is verified.Compared with the FCN model with artificially extracted features,the prediction performance of CNN has been greatly improved,and the optimal MSE is 3.44.The results show that the CNN model can determine a more suitable feature structure through multi-layer convolution and pooling.This provides strong evidence and reference for the application of deep learning in the prediction of agricultural pests and diseases.