Abrupt vegetation shifts and their drivers in Arid Central Asia

Dryland vegetation responses to environmental change may be nonlinear or abrupt rather than gradual. However, long-term vegetation studies have largely focused on linear trends, leaving abrupt trajectories and their nonlinear environmental associations in Central Asian drylands insufficiently understood. Using 2000–2023 MODIS NDVI, we distinguished linear, curvilinear, and abrupt trajectories through multi-model trajectory diagnosis and used land-cover transition analysis and XGBoost-SHAP to examine abrupt-trajectory patterns and environmental associations. Negative change trajectories were slightly more prevalent, but abrupt trajectories comprised 65.06% of positive and 25.95% of negative change trajectories. Abrupt trajectories accounted for 12.72% of all trajectories, with positive and negative abrupt trajectories accounting for 8.40% and 4.32% of all trajectories, respectively. Positive abrupt trajectories were concentrated at mid-to-high elevations, whereas negative ones occurred patchily along the western and northern margins and around the Aral Sea. Deserts and grasslands had the highest proportions of positive abrupt trajectories; only in shrublands did negative abrupt trajectories outnumber positive ones. Desert-to-grassland transitions were associated with positive abrupt trajectories, whereas grassland-to-desert transitions were associated with negative abrupt trajectories. In the positive abrupt-shift model, trends in soil moisture and atmospheric CO₂ concentration ranked highest, whereas mean temperature and solar radiation ranked highest in the negative abrupt-shift model; SHAP interaction contributions also varied nonlinearly with environmental conditions, and Human Footprint had relatively low SHAP importance in both. These findings improve understanding of nonlinear vegetation dynamics and contrasting abrupt shifts in Central Asian drylands, providing regional information for vegetation monitoring and spatially differentiated management.