Using random forests and dendroclimatology to reveal climatic factors in tree growth – case studies from temperate and tropical regions

In this study, we analyse the suitability of the machine learning algorithm Random Forest in exploring dendroclimatic relationships. Linear methods are commonly used to explain climate related changes in radial growth, but inherently fail to capture thresholds and nonlinear effects contained in the data. Machine learning applications like Random Forest can potentially fill this gap. We use tree ring chronologies from three temperate coniferous tree species Picea abies, Pseudotsuga menziesii and Pinus sylvestris from northern Germany as well as a chronology from Cedrela odorata grown in a tropical forest in Suriname to compare response function correlation and Random Forest variable importance for monthly temperature and precipitation values on tree ring width. We further explore the possibilities of using Accumulated Local Effects to display nonlinear effects on tree ring increment. Our results show that Random Forest importance aligns well with response function analysis for the temperate species and mostly aligns with the results on Cedrela odorata. Accumulated Local Effect plots offer valuable insights into nonlinear climate effects and help in explaining unexpected correlations. While the Random Forest algorithm is not a complete substitute for established methods like response function analysis to analyse climate-growth relations, it is a valuable addition to the available toolset in exploring the hidden information in dendrochronological data. The Random Forest algorithm can provide valuable insights about dendroclimatic relationships, Complementing established methods like response function analysis by identifying nonlinear effects and thresholds