关键词: 枯死木数量, Poisson模型, 负二项模型, 零膨胀模型, Hurdle模型, 随机效应

Abstract:

Objective: This study aimed to develope counting model that can accurately predict the number of dead trees in Larix principis-rupprechtii plantation, explore the main reasons affecting the number of dead trees in L. principis-rupprechtii plantation, and provide decision-making basis for the scientific management of L. principis-rupprechtii plantation in the core area of the Winter Olympic Games. Method: 45 sample plots of L. principis-rupprechtii plantation in the core area of Winter Olympic Games in Chongli District of Zhangjiakou City were chosen as the research material. Poisson regression model, negative binomial regression model, zero-inflated Poisson regression model, zero-inflated negative binomial regression model, hurdle passion regression model, and hurdle negative binomial regression model were used to develop the models of L. principis-rupprechtii plantation stand mortality number, and the optimal model would be selected in terms of the AIC value. Based on the optimal model, the random effect of different levels and the combination of various random parameters lying on the intercepts were taken into consideration, and the optimal mixed effect model of stand dead number was constructed. The best random effect level and the most optimal combination of random parameters would be determined according to the model convergence situation and AIC value, and the optimal mixed effect model of stand mortality number would be developed. Result: Stand average diameter, mean height of dominant trees, stand age, stand basal area, stand diameter Gini coefficient were the stand factors affecting stand mortality numbers, and site factors had little effect on stand death. When too many zeros were not considered, the fitting effect of negative binomial regression model was much better than Poisson regression model. After considering the phenomenon of zero expansion, the zero expansion model and hurdle model were used for simulation. It was found that the fitting effect of hurdle negative binomial regression model(HNB) and zero expansion negative binomial regression model(ZINB) were better than other models. Finally, the order of goodness of fit of several counting models considering lots of zero values was HNB regression model ≈ ZINB regression model > HP regression model > ZIP regression model. HNB model and ZINB model were selected for further mixed effect model construction. When the random effect level was specified at plot, the number of random parameters had only one, and the model cannot converge when acting on other covariates except for intercept. Only when the random parameters acted on the intercept at plot level, the model could get to converge and the fitting accuracy was further improved, and the evaluation error ME of HNB model and ZINB model were 14 and 11 trees·hm^-2 respectively. Conclusion: In the core area of the Winter Olympic Games, the main factors affecting the death of L. principis-rupprechtii plantation were stand factors rather than site factors. HNB regression model and ZINB regression model had more advantages than Poisson regression model and negative binomial regression model when fitting too many zeros. The generalized linear mixed effect model considering random intercept effect can improve the fitting accuracy of the model and reduce fitting error.

Key words: mortality trees number, Poisson model, negative binomial model, zero expansion model, Hurdle model, random effect