Use of advanced modelling methods to estimate radiata pine productivity indices

Abstract Site productivity indices have been widely used to describe age normalised height and volume for a range of forest species. In this study we used a wide range of modelling methods to predict Site Index and 300 Index for Pinus radiata D. Don. Site Index normalises height to a standardised age, while the 300 Index normalises volume measurements to a standardised age, stand density and set of silvicultural conditions. These two indices were derived from a national database of 3,676 plots with predictors extracted from geospatial surfaces describing key landform, topographic, climatic, edaphic and species-specific features (e.g. disease severity). Using these data, our objectives were to (i) compare the accuracy of geospatial, parametric and non-parametric models in predicting Site Index and 300 Index, (ii) determine whether regression kriging could be used to improve the accuracy of these predictions, (iii) identify the most influential predictors of these two indices and (iv) produce maps of both indices across New Zealand. All predictions were made on a test dataset (n = 1,104) that was not used for model fitting. The two non-parametric models eXtreme Gradient Boosting (XGBoost) and random forest provided the most precise predictions of Site Index and 300 Index and markedly outperformed both parametric and geospatial models (ordinary kriging, inverse distance weighting). Random forest provided the most precise predictions of Site Index (R2 = 0.811, RMSE = 2.027 m, RMSE% = 6.73%) while XGBoost most precisely predicted 300 Index (R2 = 0.676, RMSE = 3.462 m3 ha−1 yr−1, RMSE% = 12.63%). The use of regression kriging improved the fit of all but one model through accounting for spatial co-variance in the model error. Gains in precision were most marked for the parametric models, and in particular the regression model. After kriging, the three most precise models for both indices were random forest, followed by XGBoost and the regression model. An ensemble model derived from the mean predictions of these three models provided the most precise predictions, among all tested models, for both Site Index (R2 = 0.818, RMSE = 1.991 m, RMSE% = 6.61%) and 300 Index (R2 = 0.691, RMSE = 3.384 m3 ha−1 yr−1, RMSE% = 12.35%). Fitting a range of models to productivity indices was found to be a useful approach as this allows creation of an ensemble model and provides greater insight into the key determinants of productivity.