基于Elman神经网络的阿拉善荒漠啮齿动物群落组成物种数量预测研究

Prediction of the Number of Rodent Community Composition Species Based on Elman Neural Network in Alasan Desert

群落的数量变动及预测是生态学研究的重要内容,将神经网络技术应用到啮齿动物群落数量预测中是一种新尝试。Elman神经网络通过在前馈网络中增加延时算子,实现了记忆功能,能够对啮齿动物组成物种数量进行动态模拟和预测。以腾格里沙漠东缘荒漠为试验区,以啮齿动物数量为研究对象,采用标志重捕法,逐月监测2006─2014年每年的4─10月捕获量,建立Elman神经网络预测模型,利用2006─2013年的捕获量建立训练网络,以2014年的数据进行验证与测试,比较3种数据处理方法建立预测模型后的平均误差和拟合度,确立最优模型,预测阿拉善荒漠啮齿动物组成物种数量动态。结果表明:(1)未经归一化处理预测结果的平均误差mse为5.30,最小误差1.52%,拟合度为0.80;(2)经[0,1]归一化处理的预测结果平均误差mse为4.51,最小误差1.54%,拟合度为0.82;(3)经[-1,1]归一化处理预测结果的平均误差mse为5.03,最小误差1.63%,拟合度为0.69;(4)3种归一化处理后Elman神经网络模型差异不显著。通过平均误差和拟合度的比较,文章认为采用[0,1]归一化建立的Elman神经网络能较好的预测荒漠啮齿动物数量的变化规律,应用该网络可以预测阿拉善荒漠啮齿动物组成物种数量变化趋势,对指导当地鼠情监控和防治具有重要意义。

The fluctuation and prediction of population is one of important research contents in ecology, however, it is necessary to explore new approaches. In this study, a novel method was used in rodent population prediction by neural network technology. The neural network of Elmam has the function of memory, which can simulate and forecast the quantity of species in rodent by adding delay-units in feedforward networks. Based on the eastern edge of Tengger desert as the study area, with the rodent population as the research object, by use of mark recapture method to continuously looked into 2006─2014(Apr-Oct) and build Elman neural network forecasting model, data between 2006─2013 were used to build training network, data of 2014 were used for test, the mean error and fitting degree of three processing methods were compared and predicted the number of Alashan desert rodent population dynamics. Results showed that:(1) Average error of the prediction results without normalization mse was 5.30 with Minimum error of 1.52%, fitting degree 0.80.(2) After [0, 1] normalization, Average error of the prediction results mse was 4.51 with Minimum error of 1.54%, fitting degree 0.82.(3) After [-1, 1] normalization, Average error of the prediction results mse was 5.03 with Minimum error of 1.63%, fitting degree 0.69.(4) The difference of Elman neural network model was not significant after three kinds of normalized treatment. While comparison of mean error and fitting degree, the article suggested that by use of [0, 1] normalization to establish Elman neural network can better predicted the animal composition and species dynamic trend. Thus providing theoretical basis for guiding and preventing local rodent infestation