Machine Learning-based Prediction Of Bark Thickness And Bark Volume Of Dahurian Larch(Larix Gmelinii)

The bark is both an important part of tree organisms and an essential material for biomass energy.As significant predictor variables in forest management decisions and harvesting models,bark thickness and volume play an increasingly critical role in forest resource research,wood production,and biomass.Accurately predicting bark thickness and bark volume becomes especially relevant as the utilization of bark increases with the increase of biomass energy.In this study,six and seven representative traditional models were selected to predict bark thickness and bark volume in larch(Larix gmelinii)natural forests in the Greater Khingan Mountains,respectively,based on a combination of domestic and international bark thickness and volume models.Four machine learning models including neural network(ANN),support vector regression(SVR),decision tree(CART),and random forest(RF)were constructed separately for bark thickness and bark volume using destructively sampled data of larch.Their prediction performance was compared with traditional models using the coefficient of determination(R~2),root mean square error(RMSE),mean absolute error(MAE),and Akaike Information Criterion(AIC).The results showed that the Model(3-5)using the variables D and rh as independent variables was the optimal traditional bark thickness model among the six traditional bark thickness models,with the R~2 reaching 0.6711,and the RMSE and MAE values were smaller than the other models.Among the seven traditional models for bark volume,Model(3-13)yielded the largest R~2 of 0.8407 and the smallest RMSE and MAE of 0.0337 and0.02194,respectively.Both optimal traditional models for bark thickness and bark volume could eliminate heteroskedasticity in the models by using variance functions such as power functions.In the ANN models,the optimization algorithms of the optimal neural networks were both Adam and the activation functions were both Relu for bark thickness and bark volume.In SVR,the minimum RMSE was generated when the kernel function was RBF.In the CART models,the predicted values of the decision tree were often accompanied by high variance,showing discontinuous curves.In RF,the RMSE decreased rapidly with the increase of the number of decision trees,and then gradually leveled off,and the RMSE of the model was basically stable after the number of decision trees reached 500.It was found by calculating the relative importance of variables that the variables with greater influence on the prediction of bark thickness were rh,D,and H in order,and for bark volume were D and H.From the model prediction results,all machine learning models fitted with better accuracy than the traditional model except the CART model.Among the machine learning models,RF fitted the best,followed by SVR,and CART was the worst with low accuracy.Compared to the optimal traditional model,RF improved the R~2 of the model from 0.6711 to 0.7234 and reduced the RMSE from 0.5791 to 0.5310 in predicting bark thickness.While in the bark volume model,RF improved the R~2 of the model from 0.8407 to 0.8890 and reduced the RMSE from 0.0337 to 0.0291The results showed that the machine learning algorithms RF,SVR,and ANN could improve the prediction accuracy of bark thickness and volume compared to the traditional bark thickness and volume model.Among them,RF was optimal for predicting bark thickness and volume of larch in this region.