Evaluation of the most scalable, accurate and cost trade-offs machine learning framework to estimate the Mediterranean wood-pasture yield

Abstract Purpose Sustainable grazing management requires precise knowledge of daily nutritional requirements and the quantity and quality of pasture dry matter. Combining multiple data sources with machine learning models can create accurate predictive systems to optimize feeding, cut expenses, and maintain pasture productivity, helping farms stay economically viable long-term. Methods This study evaluated 69 machine learning models—combinations of three algorithms and 23 datasets including thermo-pluviometric, pasture classification, Sentinel 1 & 2, and soil data—from a one-year study on two Mediterranean wood-pasture fields located in Sardinia (Italy). The models were compared for accuracy, scalability, and application cost to identify the most effective framework for predicting pasture and grazing conditions. Results The most accurate framework used the Ensamble Learner (EL) algorithm with thermo-pluviometric, Sentinel-2 and pasture classification data, achieving RMSE 469.92 kg·ha⁻¹ DW, MAE 402.61 kg·ha⁻¹ DW and R² 0.98, but is impractical at large scale because pasture classification inputs require highly qualified staff and time-consuming on-site surveys. A scalable, zero-cost alternative uses EL with thermo-pluviometric and Sentinel-2, with comparable error metrics. Conclusions Research should focus on the whole Machine Learning workflow, from problem definition and covariate selection to preprocessing and evaluation rather than algorithms alone. Reliable pasture-yield modeling should include at least the thermo-pluviometric and Sentinel-2 multispectral data. Future work will apply models to estimate yield, map management zones, and generate grazing-rotation prescription maps using measured pasture utilization. Graphical Abstract