Optimizing variables selection of random forest to predict radial growth of Larix gmelinii var. principis-rupprechtii in temperate regions

Drought induced great tree growth declines and high mortality, leading to high uncertainty in carbon storage estimation. The influences of droughts on tree growth had been extensively explored, however, how to predict tree growth during drought years to reduce uncertainty in carbon storage prediction is still challenging. We utilized a combined approach of random forest importance assessment and the "VSURF" package in R software to optimize the variable selection, and then used the selected variables in random forest and multiple linear regression (MLR) method to predict the tree growth based on 132 monthly climate variables and tree-ring network consisting of 1198 Larix gmelinii var. principis-rupprechtii trees from 48 sites. Forty-three out of 132 variables were selected if the random forest importance assessment was only used to selected climate variable, however, twelve important variables were identified by optimized variable selection, which further improve model efficiency with the least variables. The comparison between random forest and MLR showed that the predictions of the random forest model showed a better fit with the observed tree-ring values than MLR from 1989 to 2018. The predicted growth of L. gmelinii var. principis-rupprechtii is better in dense sites compared to sparse sites. During summer drought years, random forest performs well to predict tree growth in densely distributed sites. Our results highlighted the usage of optimized variable selection method combined with the random forest model to predict drought-year growth over the dense sites. Our study is crucial to predict drought-year carbon sink of planted larch forests under different climate changes scenarios in the future.