| Tree mortality is an important process in forest ecosystem dynamics.Due to the complexity of forest ecosystems,the prediction of individual tree mortality and the understanding of mortality mechanisms underlying have long been an important topic and difficult area of forest ecology research.Traditional parametric statistical models are often limited in practical application due to their own limitations,such as the requirements for data distribution and multicollinearity,and cannot be well applied in the analytical modeling studies of complex forest ecosystem dynamics.At the same time,with the establishment of a large number of large-scale long-term forest dynamic monitoring plots in recent years,as well as the enrichment of survey content and methods,the quantity and complexity of forest monitoring data have reached an unprecedented new level,which also brings challenges for data analysis and modeling studies.Due to the big data size,large data scale,high data dimensions and continuity,and complex data structure of these new forest inventory data,traditional statistical analysis and modeling methods are often ineffective in dealing with those data sets.However,for this type of big data,new machine learning approaches tend to excel in analytical efficiency and model performance.In addition,machine learning methods do not suffer from anomalies such as multicollinearity and specific statistical distributions of the data.Machine learning approaches can effectively handle large amounts of complex data,and have become essential tools for research in different disciplines in recent years.In this study,parametric statistical model and nonparametric machine learning approaches were used to analysis 35 years consecutive tree census data from a 50-ha dynamic plot on Barro Colorado Island,Panama,and to develop individual tree mortality prediction models.Combining biotic and abiotic data collected extensively in the literature,106 features were constructed as independent variables and predictors,including tree characteristics,functional traits,interspecific interactions,soil chemistry,topographic characteristics,and climatic characteristics,and tree survival status was used as dependent variables.A logistic regression model(traditional parametric method),random forest model and extreme gradient boosting model(nonparametric machine learning methods)were implemented and trained by using the data,respectively.In addition,model analysis tools such as feature importance ranking,partial dependence plots,accumulated effects plots and SHAP values were employed to interpret the models,in order to investigate the mortality mechanism of tropical rainforest.The results showed that the two machine learning models outperformed the classical logistic regression models,with prediction accuracy of over 90% and AUC of over 0.67.The extreme gradient boosting model performed slightly better than the random forest model.The analysis of the models revealed that tree mortality in BCI plot was mainly influenced by functional traits of tree species,individual interactions,and soil chemistry,which with nonlinear characteristics,while climatic factors had relatively little direct influence on individual mortality.In addition,model interpretation results indicated that intraspecific competition had negative effect on individual survival,supporting and revealing the Janzen-Connell hypothesis.The effect of total competition on survival showed a unimodal distribution,suggesting that both competition and facilitation existed in neighbor plants in BCI plots.The negative correlation between tree size and survival is partly due to the increasing negative effect of neighbor competition with increasing tree size,and partly probably due to the fact that larger trees are more susceptible to lightning strikes in tropical rainforests,while older trees are more likely to die due to accumulated historical damage.A unimodal distribution of soil chemical element content and tree survival suggests that competition is more intense at low nutrient availability,validating the resource ratio hypothesis.Sampling growth rate was negatively correlated with tree survival,suggesting that the trees with resource-acquisition strategy are less invested in defense,less susceptible to pests and diseases,and more vulnerable to negative density-dependent effects of conspecifics,indicating a trade-off between rapid growth and defense.In contrast,average wood specific gravity,specific leaf weight,and average tree height were positively correlated with survival,probably because a resource-acquisition strategy predominates in high-productivity forests,ensuring that trees allocate sufficient resources to produce more leaves and stems,and strengthen the canopy to compete for light.The positive effects of LS factors,slope and elevation on survival are contrary to some studies,probably because topographic differences in the Amazon are not significant and tend to be downplayed by the effects of other factors such as competition and soil nutrients,as differences in elevation are not as extreme as in montane forests.Our study analyzes and builds models to predict the mortality of individual trees in tropical rainforests by means of ecological big data mining and machine learning,revealing ecological hypotheses and theories such as the Janzen-Connell hypothesis,interspecific facilitation,and the trade-off between rapid growth and defense,providing new ideas for mining and analyzing forest inventory data of complex large dynamic plot,and demonstrating the prospect of wide application of big data mining and machine learning methods in ecology. |
