Studies on prediction models of Dendrolimus superans occurrence area based on machine learning

Dendrolimus superans is one of the major forest pest insects, and its occurrence causes serious reductions in forest growth and significant threats to the safety of forest resources in China. Therefore, it is critical and necessary to predict the D. superans occurrence trend and population dynamics timely and accurately. Many factors affect the occurrence and outbreaks of pests, most likely involved in complex nonlinear systems. Unfortunately, most traditional models were based on linear prediction with very poor forecasting accuracy. In this study, following four variables, evaporation in mid March of current year, the average minimum temperature in early July of previous year, the extreme minimum temperature in late March of current year and the average wind speed in early November of previous year, were chosen as the independent variables, whereas the insect pest occurrence area was selected as the dependent variable. Three machine learning algorithms, i.e. multilayer feed-forward neural networks (MLFN), general regression neural network (GRNN), and support vector machine (SVM) were used to predict the D. superans occurrence areas, and these prediction results were compared with those predicted by the traditional multiple linear regression method. Results showed that the prediction efficacies of the machine learning methods were largely superior to multiple linear regression prediction; with the support vector machine model being the best, reaching 100% prediction accuracy with a permissible error range of 30% tolerance, and with low RMSE value (0.077) and short training time (1 second). These results suggest that machine learning algorithms, especially the support vector machine model, might have a great potential for accurate and effective predictions of insect pest occurrence areas as a reliable prediction tool.