From parametric models to machine learning: Predicting tree height from national forest inventory data

Using two decades of data from the Danish National Forest Inventory (DNFI), we compare baseline parametric fixed- and mixed-effects regression models for tree height prediction with regression tree ensemble models (XGBoost and random forest) and show that these machine learning models outperform the baseline method on unseen plot data. We suggest that the ability to learn from a more extensive feature set enables these models to generalise to unseen data better than the baseline methods. On unseen plot data, the best performing model was XGBoost (R2 = 0.852, RMSE = 2.85 m). The mixed-effects models achieved a good fit to the training data (with R2 = 0.941, RMSE = 1.84 m), demonstrating their utility when localised estimates are possible within a defined dataset, but their predictive performance dropped sharply when plot-level random effects were unavailable. We employ recent methods for calculating Shapley values to illustrate the importance of different features for model predictions; this analysis demonstrates that these methods offer the potential for interpretability of feature interactions in contrast to the perception of machine learning methods as a “black box”. In application to forest inventory, where it is not always possible to estimate a mixed-effects parameter, the machine learning models may provide better height predictions. Improved predictions of tree height may reduce the need for direct measurements of tree height and will enable more accurate estimation of national forest resources.