A stacked machine learning approach for mapping grazing pressure on the Qinghai-Tibet Plateau: implications for sustainable pasture management

Livestock grazing is one of the primary pressures on the vegetation ecology of the Qinghai-Tibet Plateau (QTP). However, its spatial distribution is highly heterogeneous due to multiple drivers, leading to localized overuse and ecological degradation. This spatial complexity underscores the urgent need for long-term grazing pressure monitoring to support sustainable pasture management. This study developed a spatiotemporal simulation framework for relative grazing pressure (RGP) using a Stacked Generalization Machine Learning (Stacking-ML) model. By integrating county-level actual livestock carrying capacity (ALCC) data with multisource variables (including terrain, vegetation indices, and human footprint indices), we quantified RGP trends in the QTP region over 20 years and evaluated their implications for grazing sustainability. The results showed that: 1) the Stacking-ML model (R 2 = 0.770, RMSE = 0.0015) outperformed single-algorithm models in RGP simulation, achieving high temporal consistency and spatial resolution; 2) RGP exhibited an east-west gradient, with significantly higher growth rates in the eastern QTP ( p < 0.05); 3) policy interventions (e.g., grazing bans in degraded areas) effectively mitigated pressure hotspots in eastern regions. Our framework provides a scalable tool for evidence-based grassland governance, emphasizing the synergy between machine learning and policy-driven sustainability. • Developed a spatiotemporal simulation framework for relative grazing pressure (RGP). • Stacked Generalization Machine Learning (Stacking-ML) model achieved superior accuracy. • RandomizedSearchCV can effectively improve the accuracy of machine learning (ML) models. • RGP presents a trend of transferring to higher grades on the Qinghai-Tibet Plateau.