Artificial neural networks as an alternative method to nonlinear mixed-effects models for tree height predictions

Tree heights are one of the most important aspects of forest mensuration, but data are often unavailable due to costly and time-consuming field measurements. Therefore, various types of models have been developed for the imputation of tree heights for unmeasured trees, with mixed-effects models being one of the most commonly applied approaches. The disadvantage here is the need of sufficient sample size per tree species for each plot, which is often not met, especially in mixed forests. To avoid this limitation, we used principal component analysis (PCA) for the grouping of similar plots based on the most relevant site descriptors. Next, we compared mixed-effects models with height-diameter models based on artificial neural networks (ANN). In terms of root mean square error (RMSE), mixed-effects models provided the most accurate tree height predictions at the plot level, especially for tree species with a smaller number of tree height measurements. When plots were grouped using the PCA and the number of observations per category increased, ANN predictions improved and became more accurate than those provided by mixed-effects models. The performance of ANN also increased when the competition index was included as an additional explanatory variable. Our results show that in the pursuit of the most accurate modelling approach for tree height predictions, ANN should be seriously considered, especially when the number of tree measurements and their distribution is sufficient