Modelling temporal change in inventory attributes from a LiDAR-derived inventory for the United Counties of Prescott and Russell, Ontario: A comparison of random forest and linear regression methods

Abstract

This study assessed the feasibility of updating a forest inventory derived from 2014 Light Detection and Ranging (LiDAR) data using ground plot data collected in 2021 to model change in basal area, volume, and average stand height. These attributes were determined for a subset (n=32) of stands from the original 2014 inventory. Both 2nd order polynomial regression and random forest learning methods were used to model annual growth increments for these attributes and results were compared. Except for height, the variance explained using random forest regression was greater than that explained using linear regression. As well, root mean square error was lower using random forest as opposed to linear regression for all three attributes, suggesting random forest produced more accurate results overall. Although the random forest results could not be extrapolated to the landscape with confidence due to limitations associated with that approach. Rather, the quadratic equations from the linear regression models were used to predict 2021 landscape values. The results at the landscape scale were deemed to be reasonable in terms of ecological expectations despite recognized model weaknesses. Increasing sample size to capture a greater diversity of stand types and allow for species-specific modeling would no doubt result in much better predictions.