Modeling Stand Growth Of Natural Conifer And Broadleaf Mixed Forests In Jilin Province Based On Machine Learning

Forest growth models can predict forest growth change,which can provide a reliable basis for forest management decision-making.The traditional statistical growth models are often applied under certain statistical assumptions,such as the data are normally distributed,homoscedastic and independent.In addition,convergence is rather difficult when modelling forest growth in nonlinear form.The above requirements are usually difficult to be met for forest data.The rapid developed nonparametric and data-driven machine learning(ML)methods provide a new way for forest growth modeling,and have become a trend in growth and yield models.However,there are few systematic applied studies,especially for natural mixed forests.Taking natural conifer and broadleaf mixed forests in Jilin province as the object,this study intends to answer the following scientific issues:How is the model performance of different ML methods in predicting the growth of natural conifer and broadleaf mixed forests?What are the differences between ML and traditional growth models?What are the main drivers to the growth of natural conifer and broadleaf mixed forests?How to simulate the growth of natural conifer and broadleaf mixed forests?In our study,seven typical ML algorithms(including random forest(RF),boosted regression tree(BRT),multivariate adaptive regression splines(MARS),artificial neural networks(ANN),k-nearest neighbors(k-NN)and support vector machine(SVM))were applied for growth modeling of the whole stand and its compositional tree species,which included the models of average height,average DBH,volume,basal area,biomass,stand density,mortality and recruitment during 5-year interval,of the natural conifer and broadleaf mixed forests in Jilin province.Stand data were from 610 sample plots of the 5th,6th,7th,8th and 9th Chinese national forest inventory,climate factor data were extracted from Climate AP and soil attribute data were from Soil Grids system.The performance of traditional statistical model and ML model in predicting average height,basal area,stand density,stand mortality and recruitment were compared.In addition,the growth trends of three stand growth types(mixed larch-broadleaf forests,mixed Korean pine-broadleaf forests and mixed spruce-fir-broadleaf forests)during2014-2099 were simulated.The main results are as follows.(1)Seven typical ML algorithms were applied for modeling whole stand and its compositional tree species growth models,which were the model of average height,average DBH,volume,basal area,biomass,stand density,mortality and recruitment at the end of 5-year interval.The input variables of these models included initial stand factor,structure,site,thinning,climate and so on.The process of parameter tuning of each model were analyzed.Ten-fold cross-validation results showed that all seven models had good model performance except for mortality and recruitment.The coefficient of determination of ten-fold cross-validation(R~2cv)of average height,average DBH,volume,basal area,biomass and stand density models were 0.4808-0.9789,and the relative root mean square error of ten-fold cross-validation(r RMSEcv)were 0.0228-0.9036.Whereas,R~2cv of mortality and recruitment models were 0.0018-0.4934,and r RMSEcv were larger than 1.Generally,the k-NN model had worse performance than other six models in predicting average height,average DBH,volume,basal area,biomass and stand density at the end of 5-year interval of whole stand and its compositional tree species.Whereas,the predictive accuracy of these six models differed slightly.With the independent variables increasing gradually,the prediction accuracy of RF,BRT,MARS,Cubist,ANN and SVM models increased gradually,while that of k-NN model decreased roughly.The estimated value of stand factors,which were acquired from whole stand models and compositional tree species growth models,were different.Compared with compositional tree species growth models,the mean relative error(RE)of whole stand models for predicting basal area,volume,biomass and stand density decreased by 17.57%-31.37%.(2)The relative importance and partial dependence of independent variables were analyzed based on RF and BRT models.The results showed that initial stand factors were the main drivers to these eight stand variables(the relative importance were 45.20%-86.51%),followed by average age,site,stand structure,thinning and climate factors(the relative importance were0.44%-16.64%).What's more,the partial dependence relationship between independent and dependent variables based on both RF and BRT had reasonable biological interpretation.(3)The ensemble learning model RF(users-friendly,simple hyper-parameter optimization process,less requirement for input data pre-processing)which had good performance,was selected as the basic prediction model to simulate the growth trends of the whole stand and its compositional tree species of three stand growth types(mixed larch-broadleaf forests,mixed Korean pine-broadleaf forests and mixed spruce-fir-broadleaf forests)during 2014-2099.The results showed that the model could describe the growth process and difference of different tree species in mixed forests.The natural development of these forests showed a trend of broad-leaved tree species increase during the simulation period,which revealed the growth process of mixed forests.(4)After systematically comparing the performance of traditional statistical model and ML model in predicting stand average height,basal area,density,mortality and recruitment,we found that both traditional statistical model and ML model had good performances except for mortality and recruitment.ML model had a little bit better performance than the traditional statistical model in estimating average height,basal area,stand density(R~2cv increased by 3.44%-32.57%,r RMSEcv decreased by 4.89%-67.34%).Therefore,the ML model had identical generalization ability and statistical reliability as the traditional statistical models.(5)In the process of constructing ML models,it is necessary to tune hyper-parameters.The performance of ML models can vary drastically depending on the value selected for the hyper-parameters.We concluded that ML models had the potential of predicting forest growth and could be applied for growth prediction and simulation of natural mixed forests.