Comparing the random forest algorithm with other modelling approaches to capture the complex patterns of intra-annual wood formation of Chinese fir with different ages

Abstract

Intra-annual wood formation has been examined to define onset and cessation of xylem formation and calculate rate of wood formation which could be modelled by some specific equations. In this study, three traditional methods (Gompertz function and its first derivative and generalized additive model) and two machine learning methods including random forest (RF) and deep learning (DL) were used for modelling dynamics of intra-annual wood formation of Chinese fir in stands with five different ages, and stands were characterized by a middle-subtropical climate. Results showed that RF was the most effective approach to model dynamics of each intra-annual wood formation phase, showing treble-peaks and bimodal patterns of cambial cell division, a bimodal pattern of cell enlarging and wall-thickening, and a two-phases pattern of mature stage and wood cell. Furthermore, these patterns indicated that only younger trees had a treble-peaks pattern of cambial cell division, and the third peak of cambial cell division delayed cell enlarging and wall-thickening; older trees mainly grew in the early growing season, while younger trees mainly grew in the late growing season. In addition, younger trees had a later cessation of cambial cell division, cell enlarging, wall-thickening and mature stage, which led to a longer duration of intra-annual wood formation. Finally, younger trees had a larger number of intra-annual wood cells. The random forest approach, a new way for modelling the intrinsic complexity of intra-annual wood formation, expands our understanding of the complex patterns of intra-annual wood formation.