Objective Individual tree growth can be controlled by many factors, such as climate, competition, stand factor and so on. It is necessary to clarify the dominant factors affecting tree growth from climate and stand variables with appropriate methods. Machine learning methods, such as random forest and so on, provide a new way. Testing the reliability of random forest algorithm in analyzing the effects of stand factors and climate on individual tree growth is necessary. The algorithm is expected to provide a new method for forest growth and yield prediction.

Method Long-term continuous monitoring data of larch-spruce-fir sample plots repeatedly measured for 25 years (1986−2010) in Wangqing Forest Bureau of Jilin Province, northeastern China were used. Random forest algorithm was used to build individual tree radial growth model with 52 candidate independent variables as competition, stand factor and climate. The effects of climate and stand factors on individual tree radial growth were analyzed. More concretely, 52 random forest models were built based on 52 hyperparametric combinations (ntree = 1 000 and mtry = 1, 2, 3, ···, 52). And 10-fold cross validation was used to train and evaluate these models. The relative importance and partial dependence of independent variables affecting individual tree radial growth were analyzed based on the full data set and the optimal random forest model.

Result The random forest model with ntree = 1 000 and mtry = 12 had the best generalization ability among all 52 random forest models. This model had the maximal determination coefficient of cross validation (R²_cv, R²_cv = 0.54), the minimal root mean square error of cross validation (RMSE_cv), mean absolute deviation of cross validation (MAE_cv) and relative root mean square error of cross validation (rRMSE_cv) (RMSE_cv = 0.14 cm, MAE_cv = 0.10 cm and rRMSE_cv = 50%). Individual tree radial growth was affected mostly by stand factors with the relative importance over 80.00%. What’s more, among the 9 stand factors, the sum of basal area larger than the subject tree (BAL) had the highest impact on individual tree radial growth, the number of trees per hectare (N) had the lowest impact, and the impact of other 5 factors were in between. Besides, individual tree radial growth decreased with the increase of BAL, basal area per hectare (BA), N and average DBH (D_g). On the other hand, individual tree radial growth increased with the increase of the ratio of DBH of subject tree to average DBH (RD), DBH at the beginning of the period (D₀) and the ratio of DBH of subject tree to the maximal DBH (DDM). In addition, individual tree radial growth was less affected by climatic factors, with the relative importance was below 20.00%. What’s more, all the 44 climatic variables had little impact on individual tree radial growth (the relative importance < 1%). And the ratio of average precipitation in growing season (April−September) to annual precipitation (Pratio), the total annual solar radiation (Asr), the ratio of average precipitation in growing season (April−September) to the relative humidity in growing season (April−September) (Gspgsrh) and solar radiation duration in growing season (April−September) (Gssr) were top four relatively important climate variables.

Conclusion Random forest model can be used to reasonably analyze the complex relationship between each predicted variables and individual tree radial growth. Individual tree radial growth is mostly affected by stand factors, but less affected by climatic factors. In general, the climatic factors had very limited ability in explaining the variation of individual tree radial growth at local scale, while the stand factors, such as competition and stand factor, are the main drivers to individual tree radial growth. Random forest models have good performance in model generalization and accuracy. Both the variable importance and partial dependence plots have reasonable forestry interpretation.