Modelling tree diameter of less commonly planted tree species in New Zealand using a machine learning approach

A better understanding of forest growth and dynamics in a changing environment can aid sustainable forest management. Forest growth and dynamics data are typically captured by inventorying a large network of sample plots. Analysing these forest inventory datasets to make precise forecasts on growth can be challenging as they often consist of unbalanced, repeated measures data collected across large geographic areas with corresponding environmental gradients. In addition, such datasets are rarely available for less commonly planted tree species, and are often incomplete and even more unbalanced. Conventional statistical approaches are not able to deal with such datasets and identify the different factors that interactively affect forest growth. Machine learning approaches offer the potential to overcome some of the challenges with modelling complex forest dynamics in response to environmental and climatic factors, even with unbalanced inventory data. In this study, we employed a widely used machine learning algorithm (random forests) to model individual tree diameter at breast height (DBH, 1.4 m) in response to age, stocking, site and climatic factors for the following five less commonly planted tree species groups in New Zealand: Cupressus lusitanica (North Island); Cupressus macrocarpa (South Island); Eucalyptus nitens; Sequoia sempervirens; Podocarpus totara; and Leptospermum scoparium. Data to build machine learning models were extracted and combined from three national level databases, and included stand variables, information about sites and climate features. The random forest models were able to predict tree DBH with high precision for the five-tree species (R-2 > 0.72 and root-mean-square error ranged from 2.79-11.42 cm). Furthermore, the random forest models were interpretable and allowed us to explore the effects of site, environmental and climate factors on forest growth. To our knowledge, this is the first attempt to utilize machine learning approaches to model tree diameter of less common planted forest tree species in New Zealand. This approach can be used to forecast more precise forest growth and carbon sequestration to help us understand how different forest types and species are affected by the changing climate.

Automated summary

This study utilized machine learning approaches to model tree diameter of less commonly planted tree species in New Zealand, and explored the effects of site, environmental, and climate factors on forest growth.