The Studies On Prediction Models Of Occurrence Area Of Dendrolimus Spp.by Using Machine Learning

The damage of forest insect pests and diseases has been a serious problem that restricting the development of forestry in China.Among these issues,the Dendrolimus spp.is responsible for the most damage,as one of the major species of forest pests in our country,its occurrence causing serious reductions in forest growth and posing significant threats to safety of forest resources.Thus,it is necessary to predict the occurrence trend and population dynamics of Dendrolimus spp.timely and accurately.Many factors affect the occurrence and outbreaks of forest pests,most of them involved in complex nonlinear systems.However,most of the previous studies on pest forecasting were based on linear models with low forecasting accuracies,the prediction effect is not precise enough.Therefore,it is necessary to search for more scientific and advanced predict methods to improve the accuracy of prediction,in order to control the potential disaster and minimize losses caused by the pests effectively.In this study,we selected the historical pests data of Dendrolimus punctatus(Walker)and Dendrolimus superans as the study samples,developed three machine learning algorithms:multilayer feed-forward neural networks(MLFN),general regression neural network(GRNN),and support vector machine(SVM)to predict the occurrence areas of D.punctatus and D.superans,,these prediction results were compared with the prediction of the traditional multiple linear regression method.In the end,evaluating the capacity of each models by using RMSE,prediction accuracy and training time.The aim is to determining the feasibility of machine learning in pest forecasting,moreover,to find an effective way to predict the occurrence of the Dendrolimus spp..Results revealed that:(1)For the prediction of D,punctatus,RMSE of the GRNN is 1.92 which is higher than that of multiple linear regression(RMSE=0.494),MLFN(RMSE=0.36)and SVM(RMSE=0.33)models,it indicated that the prediction result of GRNN has relatively large errors,the prediction effect is not satisfactory.MLFN and SVM results are familiar,both of them have lower RMSE than that of multiple linear regression.Under the 30%tolerance,the accuracy rate of MLFN and SVM are 60%and 80%respectively,while the training times are 41s and Is respectively.It demonstrated that SVM has lower training time and higher accuracy rate than that of MLFN.Thus,the SVM model is more suitable for the prediction of the occurrence trend of D.punctatus.(2)For the prediction of D.superans,by comparing RMSE values of each models,the results showed that three machine learning models MLFN(RMSE=0.4002),GRNN(RMSE=0.2565)and SVM(RMSE=0.077)performed better than multiple linear regression(RMSE=0.7474).Under the 30%tolerance,the prediction accuracy of SVM model was 100%,which is higher than that of MLFN(33.33%)and GRNN(66.67%).Besides,in training time,SVM model(Is)has more advantages than that of MLFN model(56s).Therefore,the SVM model is more suitable for the prediction of the occurrence area of D.superans.(3)Using the SVM models which we established in this research,as well as the fundamental principles and the optimization parameter of SVM to developed a software for predicting the occurrence area of Dendrolimus spp..This software was based on MATLAB GUI,in order to realize the visualization of these prediction models and it is convenient for users to input data directly.The pests prediction software can import and process the data quickly,the user interface of it seems briefness and friendly which is simple operation for users,realized human-computer interaction.Moreover,this software has been packaged as an executable file,in order to make the software portable and user-friendly.According to the comparison of the four kinds of prediction models,the results showed that machine learning can be applied to the practical and it has the effective prediction of insect pest occurrence area,especially the SVM model can be used as a good prediction method.