Anomaly Detection and Machine Learning for Stand-Level Growth and Yield Modeling in Hybrid Eucalypt Plantations

Abstract Accurate prediction of forest growth and yield is essential for strategic planning in intensive plantation management. This study evaluates whether unsupervised anomaly detection can be used as a systematic data-quality layer in stand-level growth and yield modeling, after standard consistency checks have been applied. We used a multi-regional continuous forest inventory of hybrid Eucalyptus urophylla × Eucalyptus grandis plantations in Minas Gerais, Brazil (6,553 measurements from 1,749 permanent plots in three regions). Four unsupervised methods, Isolation Forest, One-Class SVM, Local Outlier Factor and a dense autoencoder, were applied within each region to identify multivariate anomalies in stand age, volume, basal area and dominant height. Using Isolation Forest with a contamination rate of 0.10, approximately 10% of records were flagged as anomalous, but this corresponded to the complete removal of 2–4% of plots, while most plots either retained all measurements or lost only a subset of them. We then calibrated nine nonlinear machine learning models to predict stand volume at the second measurement (V₂) under two scenarios: using the full dataset and using the anomaly-filtered dataset. Gradient Boosting and other tree ensembles achieved the best performance in both cases (R 2 ≈ 0.86–0.88, relative RMSE ≈ 13–15%), and differences in accuracy between full and filtered datasets were small (changes in R 2 < 0.03 and in relative RMSE < 1 percentage point). A sensitivity analysis across contamination levels from 0.05 to 0.20 confirmed that a 10% threshold offers a practical compromise, removing a consistent set of multivariate extremes while preserving plot coverage and predictive performance.