Comparison of models for forecasting peak abundance of Dendrolimus punctatus larvae
Author of the article:ZHOU Xia-Zhi1** WANG Zhen-Xing2 YU Yan2 LI Shang1 BI Shou-Dong2*** ZHANG Guo-Qing3 FANG Guo-F
Author's Workplace：1. School of Forestry and Landscape Architecture, Anhui Agricultural University, Heifei 230036, China; 2 School of Science, Anhui Agricultural University, Hefei 230036, China; 3. The Forest of Qianshan County, Anhui Province, Qianshan 246300, China; 4. The Forest Disease and Pests Prevention and Control Station，Shenyang 110034, China
Key Words： the peak occurrence quantity of Dendrolimus punctatus larvae, the prediction model of stationary time series, the prediction model of regression, the prediction model of BP neural network, the prediction model of Markov chains, the prediction model of contingency table analysis
Abstract：
[Objectives] To improve the accuracy of forecasting the peak occurrence of Dendrolimus punctatus, and provide a foundation for choosing suitable predictive models. [Methods] Models based on five different methods; stationary time series, regression forecasting, BP neural networks, Markov chains, and contingency table analysis, were used to predict peaks of abundance of first and second generation Dendrolimus punctatus larvae over a 33 year period from 1983 to 2016 in Qianshan County Anhui Province. [Result] The single regression model used peak egg abundance as an independent variable, whereas the multiple regression and stepwise regression model used the actual 0.21 to 0.31 number of larvae per strain. The predicted and actual results of the single regression model differed by 1.06 to 1.58 larvae per strain. The actual results for 2015 and 2016 were identical to the predictions of the stationary time model for those years. If the standard error of the model based on the BP neural network was 1 larvae per strain, the accuracy of the forecast first generation from 1983 to 2014 was 90.32％, and that of the second generation was 100%. Forecasts based on Markov chains in 2015 and 2016 were highly consistent with the actual results. Forecasts based on contingency table analysis in 2015 and 2016 were all completely consistent with the actual results. The accuracy of forecasted peaks of larval abundance from 1983 to 2014 was 90.32％ for the first generation, and 83.87% for the second generation. However, when the peak abundance of second generation larvae was less than 3.5 larvae per strain, then the forecast accuracy was 74.19%. [Conclusion] Choice of independent variable was the key to the accuracy of regression forecasts. Models based on stationary time series were suitable as temporal patterns of larval abundance conformed to stationary time series. Classification standards directly affect the accuracy of predictions based on contingency table analysis and Markov chains. BP neural networks can be used to investigate nonlinear relationships among independent variables and were a relatively ideal forecasting method.