Climate-driven modeling and future projection of grassland aboveground biomass on the Qinghai-Xizang plateau using an interpretable Transformer framework

Grassland aboveground biomass (AGB) on the Qinghai-Xizang Plateau is highly sensitive to climate change, yet existing studies lack fine-scale characterization of month-to-month climate forcing dynamics and their cumulative cross-month effects on AGB. This study proposes an interpretable Transformer framework that mitigates the opacity of black-box prediction by making the model's internal climate-driving contributions traceable. Monthly AGB data spanning 2003–2023 (9123 sampling sites) were used as training labels, and three interpretability methods including occlusion experiments, integrated gradients (IG), and attention weights were cross-compared to identify dominant drivers and key time windows. Both occlusion experiments and IG identified precipitation as the dominant driver, with April precipitation reaching the highest attribution across the entire variable-month matrix (mean |IG| = 1185 kg DW/ha), refining spring-lag effects from the seasonal scale to the individual-month scale. Cross-comparison revealed that attention weights diverged substantially from both methods, cautioning against their use as proxies for climate driver importance. The final model driven by monthly precipitation and maximum temperature (R 2 = 0.6330, RMSE = 331.38 kg DW/ha) was coupled with CMIP6 ensemble projections to simulate potential AGB trajectories through 2060 under a climate-only driving assumption that excludes grazing, land-use change, and other non-climatic drivers. Plateau-wide annual maximum AGB rose by 6.58%, 9.09%, and 11.65% under SSP126, SSP245, and SSP585, respectively, with the high-emission scenario producing sharply divergent projected outcomes: alpine grassland, alpine desert, and alpine desert grassland all exceeded 30% gains, whereas mountain meadow and marsh declined, reflecting competing warming and moisture-deficit effects across grassland types.