Machine learning forecasting of teak (Tectona grandis L.f.) growth parameters in Ghanaian plantations: a plot-level grouped-split analysis with ablation and sensitivity analysis

Abstract Ghana lacks a centralized, publicly archived permanent sample plot inventory for teak (Tectona grandis L.f.) plantations, which limits growth and yield models that use current field data. This constraint motivates model-extracted training data built from published allometric and site relationships, paired with machine learning, as a route to forecasting tools that can be updated once field inventory becomes available. This study built and tested a multilayer perceptron (GrowthMLP) that forecasts total height and stem volume of Ghana teak from stand, soil, and topographic covariates. Tree-level records (n = 4,995) were extracted using Ghana-calibrated allometric and yield relationships and merged with soil and topographic covariates from two published chronosequence studies. Several tree records were extracted for the same published plot and therefore share identical or near-identical stand and site covariates. To account for this, all data were split at plot level rather than at row level: every tree from a given plot was assigned to exactly one of training, validation, or test, guaranteeing zero shared plots across partitions. A 21-configuration ablation grid tested feature groups, network depth, dropout rate, normalization, and target set on this plot-level 70/15/15 split, training each configuration for up to 120 epochs. The selected model retained all 26 available predictors, used a three-hidden-layer network (512, 256, 128) with a dropout rate of 0.1, and reached a mean test R² of 0.903 and a mean test root mean squared error (RMSE) of 0.294 across height and volume. Stem volume reached a test R² of 0.991, while total height reached 0.816. Diameter at breast height, dominant height, and stand age ranked as the most influential individual predictors under permutation importance. Soil chemical properties as a group produced the largest swing in height error under coordinated perturbation, consistent with field evidence that soil fertility shapes teak growth. Removing the two soil feature groups reduced mean test R² to 0.898, a small but real cost once plot-level leakage was eliminated. A moderate multilayer perceptron can reproduce Ghana teak growth relationships from stand, soil, and site data with high fidelity for volume and reasonable fidelity for height, when evaluated on plots entirely unseen during training. Field permanent sample plot data remain the definitive next step for testing these forecasts outside the allometric relationships used to build them.