基于混合效应模型的大兴安岭地被可燃物含水率模型

Ground Surface Fuel Moisture Content by Mixed Effects Models in Daxing'an Mountains

基于黑龙江省大兴安岭林区南瓮河生态站的落叶松(Larix gmelinii)林、蒙古栎(Quercus mongolica)林、落叶松白桦(Betula platyphylla)混交林3种典型林分的288组可燃物含水率数据,选择基于平衡含水率的可燃物含水率实时变化模型为基础模型,采用非线性混合效应(NLME)模型方法,以林分因子作为随机效应,建立具有混合效应的可燃物含水率的实时变化预测模型,并通过给残差方差增加权重的方法解决异方差性问题。结果表明:考虑随机效应和异方差结构的可燃物含水率实时变化NLME预测模型的拟合效果(M_(AE)=0.716 7,M_(RE)=0.026 6)优于不含随机效应的可燃物含水率实时变化预测模型(M_(AE)=0.815 6,M_(RE)=0.031 2);其中以常数加幂函数作为异方差结构的模型精度最高(AIC=547.72,BIC=581.29,-2LL=527.72)且明显优于未给残差方差增加权重的可燃物含水率实时变化NLME预测模型(AIC=961.65,BIC=988.50,-2LL=945.65)。利用独立样本数据对模型进行检验,检验结果表明,对于可燃物含水率实时变化的预测,考虑随机效应和异方差结构的NLME模型的检验精度(M_(AE)=0.495 8,M_(RE)=0.034 2)比利用最小二乘法拟合的多元非线性回归模型(M_(AE)=0.588 5,M_(RE)=0.588 5)有所提高,说明基于混合效应模型的可燃物含水率实时变化模型可以很好地描述区域尺度上不同林分类型的可燃物含水率实时变化规律。

A real-time fuel moisture prediction model with the mixed effects was built with nonlinear mixed effect model （NLMEM） and stand factors as the stochastic effect. The fuel moisture model was with 288 sets of data related to the fuel moisture which were collected from three typical forests including Larix g melinii, Quercus mongolica Fischer and mixture of Larix gmelinii and Betula platyphylla. This model can solve the problems of heteroscedasticity by adding more weights to the residual variance. The results show that the fitting effect of the real-Time NLME prediction model （MAE = 0.716 7, MRE = 0.026 6） with the random effects and heteroscedasticity are better than that of the model （MAE = 0.815 6, MRE = 0.031 2） without considering the random effects. Moreover, the accuracy of the model （ AIC = 547.72, BIC = 581.29, -2LL = 527. 72） using a constant plus power function as the heteroscedastic structure is the highest, and it is obvious superior to the real-time NLME prediction model （AIC = 961.65, BIC = 988.50, -2LL= 945.65） without adding more weights to the residu- al variance. The results from the model test by using independent samples show that the test precision of the NLME model （MAE = 0.495 8 ,MRE = 0.034 2） with the random effects, and heteroscedasticity structure has some improvements in comparison to the multivariate nonlinear regression model （MAE = 0.588 5, MRE = 0.588 5） fitted by the least square method. The real-time fuel moisture content prediction model based on the mixed effects method can well describe the change laws in fuel moisture content for different types of forest in the regional scale of interest.