Technical Report on the Turba Fertilizer Machine Learning Stack in Morocco

Technical Report on the Turba Fertilizer Machine Learning Stack in Morocco. By Abdelghani Belgaid and five others, from Turba Open Lab. Published on the fifth of October. No code repository is linked to it.

Site-specific fertilizer recommendation systems tailor nutrient advice to a location, its soil properties, the crop, and a production target. Scientific reuse of them is constrained when the recommendation functions are reachable mainly through interactive interfaces, when outputs are not versioned, and when trained approximations cannot be independently loaded or benchmarked. This report presents the Turba fertilizer machine learning stack: a three-layer, open-source implementation for reproducible site-specific fertilizer recommendation in Morocco.

The three layers each have a job. turba-client gives programmatic access to publicly accessible site profiles, to crop-specific spaces of target yields, and to recommendation workflows for nitrogen, for phosphorus pentoxide, and for potassium oxide. turba-data distributes analysis-ready snapshots. turba-models packages crop-specific machine learning surrogates of the recommendation outputs. Together they link upstream retrieval, versioned snapshots, reproducible benchmarking across models, and surrogates that can be loaded offline. Throughout, the design keeps three things distinct: what the recommendation system outputs, what was observed in agricultural data, and what a model predicts.

The first dataset starts from forty-four thousand and ninety-six unique ESA WorldCereal locations. Expanding scenarios across the supported cereal workflows produced crop-location recommendation requests numbering almost exactly three times the locations, all under a medium target-yield setting. Each request pairs one crop with one location, which is why the requests outnumber the sites that way: a single site is asked about more than one cereal.

The resulting dataset has twenty-two variables. It covers ten regions, sixty-six provinces, and one thousand one hundred and forty-nine communes. Nine families of regression model were evaluated under a fixed, deterministic protocol that divides the data eighty to twenty, and the current release packages five of the best-performing crop-specific models.

The stack serves anyone who wants to study or reuse these recommendations without going through an interactive interface: researchers who need to retrieve recommendations programmatically, to work from a fixed snapshot that someone else can reproduce, or to load a trained surrogate and benchmark it against another. One caution matters for them. The machine learning task here is emulating the recommendation function. It is not predicting how real crops respond in the field, so a surrogate is a stand-in for the recommender, not for the soil.

The authors present the stack as a reproducible base for what comes next: validation across space and time, uncertainty estimation, comparison with field trials, and integration with additional data. The change it makes is one of access. A recommendation that was reachable mainly through an interactive interface becomes versioned, loadable, and benchmarkable.