Drylands, crucial ecological and biodiversity hotspots, support significant human populations reliant on pastoralism. However, factors such as cropland expansion, grassland degradation, and climate change threaten these systems, resulting in substantial economic and social consequences and impacting several UN Sustainable Development Goals (SDGs). Understanding the drivers of vegetation change in these regions is critical for effective land management and achieving sustainability. While trend detection studies are common, attributing these trends to specific drivers remains a challenge, often hindered by methodological inconsistencies and limitations in accounting for spatial variations and multiple interacting processes. This study addresses these limitations by developing a novel high-resolution multi-stage, multi-model framework to detect and attribute vegetation changes in Kazakhstan (KZ). KZ presents a compelling case study due to its diverse anthropogenic impacts, political history, and significant environmental consequences stemming from events such as the Virgin Land campaign and the Aral Sea shrinkage. This framework will address the extent and spatial patterns of vegetation degradation, relative contributions of land use, climate change, and climate variability, and the primary social-environmental system (SES) drivers contributing to vegetation degradation in KZ.

Existing research on vegetation changes in Kazakhstan has yielded varied conclusions due to differences in methodologies, indicators, spatial and temporal scales, data used, and study sites. Studies have focused on detecting trends, but attributing them to specific drivers remains inconsistent. Previous attribution studies, often using single models or causal graphs, lack generalizability and fail to consider spatial variability and multiple interdependent processes. The limited representation of complex land-use practices further hinders the accurate assessment of human activities' impact on vegetation degradation. This study builds upon this existing research by employing a more comprehensive approach.

This study utilizes a three-stage, multi-model framework for detecting and attributing vegetation changes in Kazakhstan between 2000 and 2019. **Stage 1: Detection and Decomposition of Observed Trends:** This stage uses the Time Series Segmented Residual Trend (TSS-RESTREND) model to detect vegetation trends (using MODIS NDVI) and decompose them into contributions from land use (LU), climate change (CC), climate variability (CV), and other factors (OF). Theil-Sen slope, ordinary least squares regression, Spearman's Rho, and moving window techniques are employed. The absolute maximum slopes of each factor are calculated to distinguish between climate-driven and land-use-dominated changes. **Stage 2: Causation and Contribution of SES Drivers:** This stage employs pixel-wise Granger Causality (GC) analysis to examine the influence of ten independent SES drivers (five land-use and five climate-related) on vegetation degradation in both LU-dominated and climate-driven areas. The analysis considers concurrent and lagged effects. The Augmented Dickey-Fuller test is used to ensure time series stationarity before applying the GC model. Pixels with p-values greater than 0.05 are masked for further analysis. **Stage 3: Attribution of Dominant SES Drivers:** A pixel-wise random forest (RF) model is fitted separately to LU-dominated and climate-dominated degraded pixels, using the significant pixels from Stage 2. The Shapley value attribution method is then applied to quantify the contribution of each SES driver to vegetation degradation. Model accuracy is assessed using R-squared, RMSE, and MAE. Pixels with R² < 0.5 are excluded. The analysis is performed using R, MATLAB, and ArcGIS software. Datasets used include MODIS NDVI, LST, snow cover; CHELSA precipitation, temperature, and VPD; TerraClimate solar radiation; WorldPop population density; DMSP/VIIRS nighttime lights; gridded GDP; and a gridded livestock density dataset developed by the authors.

Vegetation degradation was observed in 45.71% of Kazakhstan (52.02% of the non-masked area), with hotspots concentrated in the western and southern provinces. Increasing vegetation trends were observed in 42.18% (47.98% of the non-masked area). Climate variability was identified as the most significant contributor to overall vegetation change (41.3%), followed by land use (33.54%). However, in degraded areas, land use was the dominant factor (22.54%), followed by climate variability (21.93%). Granger causality analysis revealed that sheep and goat density had a dominant negative impact on vegetation in land-use-dominated degraded areas (significant in 75.6% of pixels), while snow cover significantly impacted vegetation in climate-driven areas (significant in 96.9% of pixels). Random forest modeling with Shapley value attribution further confirmed these findings, showing sheep and goat density as the dominant land-use driver (11.78% contribution) and snow cover as the dominant climate driver (23.3%) in degraded areas. The model attributed dominant SES drivers to 41.57% of degraded pixels (19.81% of KZ).

The findings highlight the complex interplay between land use and climate in driving vegetation change in Kazakhstan. The dominant role of land use, particularly livestock grazing (sheep and goats), in vegetation degradation in northern provinces supports the impact of intensive grazing practices. The importance of snow cover in climate-driven degradation, especially in southern and western areas, underscores the sensitivity of these regions to climate variability and changing snow patterns. The contrasting impacts of snow cover in different regions (positive in the north, negative in the south) reflect the complex influence of snow on vegetation, influenced by factors like snow melt timing and duration. The study's high-resolution approach reveals spatial patterns not captured by previous studies, contributing to a more nuanced understanding of the drivers of vegetation change in the region.

This study presents a novel high-resolution multi-stage, multi-model framework for detecting and attributing vegetation changes. The application of this framework to Kazakhstan identifies land use change, particularly livestock grazing, and climate variability as primary drivers of vegetation degradation. The identification of spatial hotspots of degradation allows for targeted restoration efforts and informs strategies for achieving land degradation neutrality. Future research could focus on refining the model by incorporating additional SES drivers, exploring non-linear relationships, and further investigating the role of specific grazing management practices. The framework’s applicability to other dryland ecosystems warrants further investigation.

The study acknowledges several limitations. Uncertainty may arise from input datasets, model limitations in attributing all vegetation changes, and spatial autocorrelation in predictor variables. The resampling of data to 1 km resolution for GC and RF analyses may induce uncertainty due to mixed pixels and spatial autocorrelation. Potential bias from using a constant GDP dataset after 2015 is also noted. The assumption that NDVI is a robust proxy for vegetation growth in all cases is another potential limitation.