A random forest model for basal area increment predictions from national forest inventory data

Abstract Here, we present one of the first attempts to use a machine learning model for the prediction and interpretation of tree basal area increment (BAI) based on data from the National Forest Inventory (NFI). The developed model is based on the random forest (RF) algorithm, trained with 18 independent variables and 15,580 data points (trees from the Slovenian NFI). The RF model was trained for four individual species and two groups of species and evaluated using 10-fold blocked cross-validation. Squared correlation coefficients calculated for independent data ranged from 0.289 for Scots pine (Pinus Sylvestris) to 0.342 for maple and ash species (Acer sp. and Fraxinus sp.), 0.429 for oak species (Quercus sp.), 0.475 for Norway spruce (Picea abies), 0.486 for common beech (Fagus sylvatica), and 0.565 for silver fir (Abies alba). The most important predictor variables were the basal areas of individual trees and their competition status, expressed as the basal area in larger trees and tree social position. Simulations of selected key variables revealed different ecological traits of the studied species: silver fir and Norway spruce have the highest growth characteristics, while common beech has the strongest competition potential. For valuable broadleaves and silver fir, site specific conditions play an important role in tree growth, while oaks and Scots pine have less site-specific demands and wider ecological amplitudes. Finally, in comparison to BAI models from similar studies, the presented RF model showed similar accuracy and could potentially be used as a tool in forest management practices and for making professionally informed decisions.