Prediction Of The Trend Probability Of Chilo Suppressalis Based On Deep Learning

Chilo suppressalis is one of the most harmful pests to rice.Timely and accurately predicting the occurrence trend of Chilo suppressalis can guide plant protection personnel to plan prevention and control measures in advance.For example,under the situation of predicting the occurrence of pests,rational use of pesticides can ensure the yield and quality of rice.The quality can significantly reduce the impact of pests on the growth of rice and improve the stability of the farmland ecosystem.At present,most of the predictions of Chilo suppressalis are based on traditional statistical analysis methods and machine learning methods,and do not have an advantage in the timeliness of modeling of multi-dimensional large samples and longitudinal data.It is not conducive to guiding the high efficiency and efficiency of Chilo suppressalis in actual production.Precise prevention and treatment.This study comprehensively selects meteorological variables,related pest variables and time variables,and implements accurate prediction of the occurrence trend of Chilo suppressalis based on the deep learning model DeepAR.The main work and results of this research include the following aspects:1.The rice planting areas in five counties of Hongjiang City,Dong’an County,Liling City,Linli County and Yuanjiang City are selected as the study area.Pest data comes from the monitoring and early warning information system of major crop pests and diseases in Hunan Province-Rice pests and diseases daily record table(a total of 18,723 tables).The start and end years of Hongjiang,Yuanjiang,Dong’an and Linli are 2000-2020,and the start and end years of Liling are 2010-2020.The recording time is from mid-March to early October each year.On a daily basis,plant protection staff use light traps to attract rice pests,count the number of pests,and record the weather conditions when the lights are turned on.Pests include 13 rice pests including Chilo suppressalis and Cnaphalocrocis medinalis.The meteorological data comes from the ground daily meteorological data downloaded by the National Meteorological Science Data Center,starting and ending from January 1,2000 to October 30,2020,including19 factors such as temperature and precipitation;the lack of pest data and meteorological data Values are interpolated using the average value of adjacent location data,and outliers are smoothed using exponentially weighted moving average.Extract the time characteristics of the pest data such as year,season,etc.to construct the pest time data,and finally match the pest data,meteorological data and time data according to the corresponding time axis to construct a rice pest-meteorological-time data set.2.First,count the number of pest species in the study area.For example,the population of rice planthoppers in Hongjiang from 2000 to 2020 is 2,415,802,which is much more harmful than other pests.The main pests in the study area are rice planthoppers,rice borers,and rice plants.Cnaphalocrocis medinalis.Analyzed the long-term(2000-2020)and shortterm(2020)correlations of the three main pestsand found that in 2020 Chilo suppressalis and rice planthoppers,Chilo suppressalis and rice leaf rollers,rice planthoppers and rice vertical rollers in 2020.The maximum correlation coefficients of leaf borer were 0.62,0.57 and 0.80,respectively,indicating that there is a strong correlation between different kinds of pests.Secondly,analyzing the long-term(2000-2020)and short-term(2020)occurrences and trends of Chilo suppressalis in the study area,it is found that the long-term cycle is 3-7 years,and the short-term cycle is 6-8 weeks.Thirdly,analyze the occurrence and occurrence trends of Chilo suppressalis in the recent period(2017-2020)under different time scales(year,season and month),and found that the two are positively correlated,that is,the more occurrence,the faster the occurrence trend;at the same time,it is found that Chilo suppressalis broke out in Dongan and Hongjiang in late spring(April and May),in Liling and Yuanjiang in early autumn(September),and in Linli in late summer and early autumn(August).Finally,the correlation analysis between the occurrence of Chilo suppressalis and meteorological variables in the study area,including Pearson correlation coefficient(R)analysis and maximum mutual information coefficient(MIC)analysis,among which the R=0.266,MIC of Chilo suppressalis and the minimum temperature =0.661,R=-0.258 to the highest atmospheric pressure,MIC=0.627,that is,Chilo suppressalis is positively correlated with temperature and negatively correlated with atmospheric pressure.The relationship between the occurrence trend of Chilo suppressalis from 2000 to 2020 and meteorological variables is analyzed to further indicate Chilo suppressalis has a certain correlation with the main meteorological variables(temperature,atmospheric pressure).3.Construct the Single-DeepAR probability prediction model of the occurrence trend of Chilo suppressalis.DeepAR,a time series prediction model based on deep learning,uses only the occurrence of Chilo suppressalis as an endogenous variable,without adding external variables such as meteorological variables,related pest variables and time variables,and constructs a Single-DeepAR probabilistic prediction model to realize the dualization of rice Dynamic prediction of multiple generations in a single year.According to a reasonably verified data set,the Mean Scaled Interval Score(MSI)of Liling,Hongjiang,Dongan,Yuanjiang,and Linli are 62.111,5.110,14.991,14.621,and 52.030,respectively,are better than the reference models ARIMA and Wave Net.And Transformer,most of the other evaluation indicators are better than the reference model.In general,the Single-DeepAR model has the best effect on the dynamic prediction of Chilo suppressalis.4.Construct the Multidimensional-DeepAR probability prediction model of the occurrence trend of Chilo suppressalis.DeepAR,a time series prediction model based on deep learning,uses Pearson correlation coefficient analysis,MIC analysis,and Light GBM variable importance analysis to filter pest variables,meteorological variables,and time variables,and use the selected variables as external variables and the occurrence of Chilo suppressalis.Construct a multidimensional Multidimensional-DeepAR probability prediction model for endogenous variables to realize the dynamic prediction of Chilo suppressalis for multiple generations in a single year.According to a reasonable validation set,the performance comparison of the three variable screening methods(Pearson correlation coefficient analysis,MIC value analysis,Light GBM variable importance analysis)and the MultidimensionalDeepAR model without variable screening is obtained.The results showed that Light GBM variable importance analysis was the best method for variable screening.5.This paper has obtained rice pest data from many districts and counties in Hunan Province over the years and performed descriptive statistical analysis on the obtained rice pest data and digs out the correlation of different types of pests,and found rice planthoppers,rice leaf rollers and Chilo suppressalis.Stem borer is the main pest of rice in Hunan Province,and the main pests are related.Furthermore,the descriptive statistical analysis of the Chilo suppressalis data and meteorological data was carried out,and the meteorological variables that had the most influence on the Chilo suppressalis occurrence were excavated,and it was found that the most relevant meteorological variables for the occurrence of Chilo suppressalis were temperature and atmospheric pressure.Finally,the Single-DeepAR probability prediction model and Multidimensional-DeepAR probability prediction model based on the deep learning model DeepAR are constructed.The Single-DeepAR probabilistic prediction model directly uses the historical data of Chilo suppressalis for prediction.The demand for data is not much and the effect is very good.However,this model cannot describe the relationship between Chilo suppressalis and related pests and meteorological factors.Transplant it to an area where no historical data of Chilo suppressalis is recorded for application.Based on Single-DeepAR,the Multidimensional-DeepAR model uses Chilo suppressalis-related pest variables,meteorological variables,and time variables as covariates.The model has higher practical application value,is more universal,and can better reflect pest dynamics The change process can be transplanted to areas with little or no historical data of Chilo suppressalis for prediction.However,the structure of the model is more complex and relies on reasonable variable screening methods.Therefore,this article compares three variable screening methods and finds that the use Light GBM variable importance analysis filter variables are the best for Multidimensional-DeepAR performance improvement.For all research areas,the Multidimensional-DeepAR probabilistic prediction model based on Light GBM variable importance analysis has the best effect.Therefore,the MultidimensionalDeepAR model has good spatial expansion capabilities and is more universal.The Multidimensional-DeepAR model can be applied to other areas and other pest species.