Large-scale assessment of pastoral value and carrying capacity of Alpine grasslands using remote sensing and machine learning

European mountain grasslands are undergoing rapid changes in climate and land use, with direct consequences for forage provision and the management of extensive livestock systems. Pastoral value (PV) is a proxy for forage productivity and quality, and a key parameter for assessing grassland carrying capacity. However, traditional methods for PV calculation are field-based and labor-intensive, limiting its applicability at large spatial scales. In this study, we aimed to develop a spatial modelling framework to predict the PV at large-scale using machine learning (ML) algorithms and remotely sensed data, and to derive the first maps of PV and grassland carrying capacity. We used 390 vegetation plots surveyed in the western Italian Alps and 40 remotely sensed predictors describing vegetation phenology and productivity, climate, and topography. Multiple ML models were tested employing forward feature selection to identify the optimal predictor set. Prediction-oriented cross-validation and area of applicability analysis were used to ensure reliable map accuracy. Forward feature selection based on gradient boosted tree algorithm achieved the best performance (RMSE = 6.62, MAE = 5.10, R2 = 0.45) using 7 out of 40 predictors. Although the model performed well for intermediate PV, it struggled to predict extreme values, likely due to limitations in input data and the scarcity of extreme ground-truth observations. Overall, 89% of the western Italian Alps fell within the area of applicability for spatial prediction. The proposed framework demonstrates that ML and remote sensing can provide accurate, scalable, and reproducible predictions of PV, supporting robust assessments of grassland carrying capacity and the design of sustainable grazing management strategies in mountain regions.