Evaluation of the random forest regression machine learning technique as an alternative to ecoregional based regression taper modelling

• Diameters along the tree stem and trees’ volume differ in dissimilar ecoregions. • RFr models outperformed taper equations in estimating stem volume and diameters. • Model accuracy depends on data range; RFr extrapolation is range-limited. • High data quality is vital for RFr to ensure biologically plausible outputs. • Method choice depends on accuracy needs, expertise, and resource availability. Ecologically oriented management plans are based upon accurate estimation of stand growth and yield under various climatic and growing environmental conditions. Furthermore, the precise calculation of the volume quantities of the wood classes that can be obtained from a tree is of great importance for the production of forest management plans and projections for the future of the forest products industry. Despite their limitations, stem taper regression models are one of the most widely used methods for estimating tree diameters along the bole and estimating the tree stem volume. In this study, the non-parametric ensemble algorithm RFr was evaluated as an alternative machine learning approach for ecoregion-based taper modelling. RFr models were developed to accurately estimate stem diameters along the tree bole for trees from three distinct ecoregions, yielding precise diameter predictions. A comparative analysis between the RFr models and traditional taper regression models demonstrated that both methods are capable of reliably predicting stem diameters and tree volumes. However, the non-parametric nature of the RFr modelling approach, which effectively reduces the variance of individual regression tree learners, allowed it to outperform the conventional taper regression. The RFr models produced more accurate predictions of both stem diameter and volume and exhibited high reliability in their prediction intervals, as confirmed by uncertainty assessments. Overall, the RFr approach showed superior performance on both training and testing datasets.