Abstract
<jats:p>Abstract. Long-term, high-resolution canopy cover data are essential for understanding grassland ecosystem dynamics and informing sustainable management. However, existing products are largely limited to coarse spatial resolutions, constraining their utility for high-precision, large-scale analyses. In this study, we collected over 16,000 drone image tiles (30 m × 30 m) from 2,144 sites across China and developed a machine learning model to produce a spatially seamless, 30 m annual dataset of national grassland canopy cover from 1990 to 2023 by integrating drone and Landsat-series imagery. The model achieves high predictive accuracy (R2 = 0.73, RMSE = 18.4 %) and robust temporal transferability (R2 = 0.68, RMSE = 20.6 %). Comparisons with existing large-scale products demonstrated significantly improved accuracy and reduced residual artifacts, underscoring the robustness of our approach across diverse grassland types and time periods. Spatiotemporal analysis indicated a multi-decadal mean canopy cover of 43.80 ± 18.69 % across China’s grasslands. Over the 34-year period, 41.76 % of grasslands exhibited significant increases, 57.16 % showed nonsignificant change, and 1.08 % experienced significant declines. Climatic factors—including drought, precipitation, and temperature—emerged as the dominant drivers of canopy cover dynamics at the national scale, although their effects exhibited pronounced spatial heterogeneity. In contrast, anthropogenic pressures played a secondary role overall but could override climatic influences at local scales. Collectively, these findings, together with the long-term, high-resolution canopy cover dataset developed in this study, provide an essential basis for advancing the understanding of grassland ecosystem dynamics and for supporting evidence-based conservation and sustainable management strategies, particularly under intensifying climate change and increasing frequency of extreme events. The national grassland canopy cover dataset generated in this study is archived on Zenodo and can be freely downloaded from https://doi.org/10.5281/zenodo.20301123 (Jiang et al., 2026).</jats:p>