Field: Environmental Science › Environmental EngineeringType: An approach that didn't workArea: digital soil mappingTarget: Stacked ensemble of machine-learning soil-property modelsStage: full study
Hypothesis
A stacking ensemble combining several machine-learning base learners would predict soil properties (texture, organic carbon, pH, topsoil depth) better than any single model in a lowland agricultural area of Lombardy.
What failed
Under nested cross-validation, the stacking models did not outperform all of the individual base learners; single nonlinear models performed well, with random forest best.
Broken assumption
Stacking machine-learning models always beats the best single model for digital soil mapping.
Why
The abstract reports only the comparison; it does not give a reason the ensemble failed to add accuracy.
From the published paper: Adeniyi et al. Land, 2023. ↗CC BY