基于机器学习的落叶松毛虫发生面积预测模型

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

落叶松毛虫为我国主要害虫之一,其发生严重影响了我国林木生长和森林资源的安全。因此,及时准确地对落叶松毛虫虫害发生趋势进行预测、预报十分必要。虫害的发生受到多种因素的影响,存在复杂的非线性关系,传统的预测方法大多为基于线性的预测,导致其预测效果不够理想。本研究选取当年3月中旬的总蒸发量、上年7月上旬的平均最低气温、当年3月下旬的极端最低气温以及上年11月上旬的平均风速作为自变量,虫害发生面积作为因变量,利用多层前馈神经网络(MLFN)、广义回归神经网络(GRNN)以及支持向量机(SVM)3种机器学习算法对落叶松毛虫发生面积进行预测,并将3种方法的预测结果与传统多元线性回归预测方法相比较。结果表明,机器学习的预测效果均在很大程度上优于多元线性回归预测,并且在3种机器学习算法中,SVM模型的预测效果最好,在30%容忍度下其预测精度可以达到100%,并且该模型还有较低的RMSE值(0.077)和较短的训练时间(1 s)。这表明,机器学习可以应用于生产实际并有效预测虫害发生面积,尤其是SVM模型可以作为一种很好的虫害发生预测手段。

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 suggestthat 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.