Agent‑based models are well suited for representing the spatial heterogeneity and interactive behaviors of agro‑ecological systems, making them valuable for decision support. Turning simulation outputs into actionable advice, however, requires identifying which decisions actually drive system behavior. Classical global sensitivity analysis assumes independent inputs defined over a fixed Cartesian product, which fails for decision‑tree structures where many parameters appear only after higher‑level choices are made, leading to biased results.
We treat the decision structure as a first‑class modeling object and, using advanced surrogate‑modeling techniques, encode the crop‑management itinerary of the high‑resolution agro‑ecological platform MAELIA as a hierarchical design space containing both categorical and numerical variables. For this branched space we propose a dedicated method—random‑forest‑supervised HSIC‑ANOVA—that ranks main effects and interactions directly on the tree and places rarely and always active decisions on an equal footing.
Applied to outputs such as nitrogen leaching, soil organic carbon variation, and yield predictions, the method yields interpretable, decision‑oriented insights. For instance, irrigation strategy dominates nitrogen leaching, while fertilizer dosage interacts with crop rotation to drive soil organic carbon changes. Overall, this hierarchical sensitivity analysis provides model users with a clear, decision‑focused reading of simulation results.
Review