Disentangling climatic and human drivers of Haloxylon shrubland greening (1982–2021): evidence from RESTREND and synthetic difference-in-differences

Abstract

Dryland greening has been widely reported, yet attributing observed vegetation trends to climate variability versus human interventions remains difficult because climate–vegetation relationships are nonlinear, spatially heterogeneous, and rarely evaluated against an explicit counterfactual. Focusing on Haloxylon shrublands in the western Junggar Basin (northwestern China), we combined long-term AVHRR-based Normalized Difference Vegetation Index (NDVI) observations (1982–2021) with multi-source hydroclimate reanalysis to build a reproducible attribution and counterfactual-comparison workflow. The workflow includes: (i) pixel-wise trend and change-point detection (Theil–Sen/Mann–Kendall with false discovery rate control; Pettitt test with block-bootstrap validation), (ii) cross-model identification of dominant climatic drivers using stability selection, random forest with SHAP, and Model-X knockoffs, (iii) climate removal via segmented RESTREND using a generalized additive model (GAM) baseline driven by growing-season precipitation, temperature, and spring precipitation, and (iv) associational counterfactual comparisons using synthetic difference-in-differences (SDID) with event-study and placebo tests. We identify a spatially constrained greening pattern concentrated along oasis margins and the northern desert edge, with synchronized structural shifts emerging in the late 1990s to early 2000s. Across models, soil moisture is the main constraint on interannual NDVI variability, whereas the acceleration of greening is consistent with human water-related interventions. Under outcome-based timing, SDID gaps are consistent with a positive NDVI response associated with inferred human interventions, primarily via a resource-additive level shift; associated changes in climate sensitivity are secondary and heterogeneous, and in oasis-adjacent areas the benefits saturate and then decline after approximately a decade. Together, these results characterize an input-dependent restoration trajectory and suggest considering hydrological sustainability alongside vegetation metrics, without direct validation from pixel-scale water-use or groundwater measurements. Regional inference prioritizes functional-zone summaries (Oasis and Desert); pixel-level SDID on 67 treated cells is supplementary. Targeted 30 m Landsat diagnostics (1993–2021; Supplementary Section S6) and external timing/zone-level SDID robustness (Supplementary Section S7) support scale and timing interpretation without replacing the AVHRR design.