Anthropogenic climate change has driven over 5 million km² of drylands towards desertification

The study addresses the persistent uncertainty in quantifying the roles of anthropogenic climate change (ACC) and land use (LU) in driving dryland degradation and desertification. Global estimates of degraded land vary widely, and while drylands show significant satellite-observed greening since 1980, multiple drivers—rising CO₂ and its fertilization effect, climate variability and change, and land-use practices—contribute in complex, spatially heterogeneous ways. Prior approaches often omit CO₂ fertilization, conflate climate variability with long-term climate change, or struggle to detect rapid ecosystem shifts (breakpoints) from disturbances such as deforestation or extreme fires. This work aims to quantify the extent of desertification (long-term declines in vegetation productivity within drylands) since 1982 and to attribute observed vegetation changes among CO₂ fertilization, climate variability (CV), climate change (CC), and LU, thereby informing risk assessment and policy under UNCCD and IPCC frameworks.

Previous studies report substantial global greening driven largely by CO₂ fertilization when LU is prescribed in models, while satellite-based models that do not explicitly account for CO₂ attribute changes primarily to climate and LU. Reported degraded area estimates vary greatly (<10 to 60 million km²). Some studies suggest anthropogenic forcing has increased aridity and dryland extent. However, many methodologies lack explicit treatment of rapid ecosystem changes (breakpoints) and do not disentangle CC from CV. UNCCD, IPCC, and IPBES have identified the need to separate the roles of CO₂, climate, and LU and to incorporate abrupt changes. The present study builds on these gaps by using an observation-based, per-pixel attribution that includes CO₂ fertilization, separates CC from CV, and detects structural ecosystem changes.

Definitions and domain: Desertification follows UNCCD/IPCC usage: land degradation (long-term reduction in biological productivity) within drylands (arid, semi-arid, and dry sub-humid zones). Non-dryland and hyper-arid regions are masked. Vegetation proxy and change detection: Peak growing season NDVI (NDVImax) from GIMMSv3.1g (1/12°, 1982–2015) was used as a proxy for vegetation growth/NPP. Peak NDVI was the annual maximum; if peak occurred in December, the following January–February values were assigned to the previous year. Vegetation change (Observed, Obs) was computed as the difference between expected values at 2015 and 1982 (non-parametric Theil–Sen slope with Spearman significance), accommodating mid-series breakpoints. Attribution framework: 1) CO₂ fertilization: A theoretical photosynthesis–CO₂ relationship (Franks et al.) was used to estimate relative GPP response to atmospheric CO₂ (ca) changes. NDVI time series were adjusted to remove CO₂ effects (NDVIadj) assuming proportional NPP–NDVI response and no long-term change in GPP:Ra ratio. The CO₂-attributed ΔNDVImax was then derived from trends in (NDVIobs – NDVIadj) using non-parametric tests. 2) Climate and land use via TSS-RESTREND v2.15: Applied to NDVIadj to separate climate-driven variability/change from LU. A per-pixel Vegetation–Climate Relationship (VCR) used optimal precipitation and temperature accumulation/offset windows. LU effects were estimated from trends in the VCR residuals while accounting for detected structural changes (breakpoints) using phenological change detection. 3) Separating CC from CV: Observed climatologies (precipitation and temperature, 1962–2015) were 20-year leading-edge smoothed to isolate CC trends (Theil–Sen), which were used to detrend the data to obtain CV. Using the VCR, NDVICC (from smoothed climate) and NDVICV (from detrended climate) were computed and their changes assessed via non-parametric tests. 4) Anthropogenic climate change (ACC): Defined as the combined effect of CO₂ fertilization and CC (ACC = CO₂ + CC). Ensemble and datasets: A 12-member ensemble combined four precipitation datasets (CRU TS v4, CHIRPS, MSWEP, TerraClimate) with three temperature datasets (CRU TS v4, TerraClimate, NOAA CPC). A third ensemble accounted for mixed C3/C4 vegetation using per-pixel C3/C4 fractions from SYNMAP to weight matched C3-responding and C4-nonresponding runs; p-values were combined via Stouffer’s Z. Significance: For Obs and CO₂ components, Spearman’s rho significance per pixel with Benjamini–Hochberg FDR control (αFDR=0.10). For LU, CV, CC, ensemble p-values combined via Fisher’s method then FDR applied. Additionally, an IPCC-style ensemble agreement protocol required >50% of ensemble members to be significant (αFDR=0.10) and >80% of significant runs agreeing in sign; otherwise, component estimates were masked. Breakpoints required >50% of runs to detect a significant breakpoint with 80% within a 3-year window. Gridding and preprocessing: Climate datasets were remapped to the GIMMS grid using first-order conservative remapping. Climate series requiring pre-1962 extension were bias-corrected with TerraClimate via Delta method to enable the 20-year smoothing. Population exposure used GPWv4; human development context used UNDP HDI. Error thresholds and masking: Sensor error bands (±0.001 NDVI) and field significance criteria were applied; non-dryland/hyper-arid pixels were masked. Validation and uncertainty: The total attributable change (CO₂+LU+CC+CV) differed from observed vegetation change by only ~3% and reproduced spatial greening/browning patterns. Sensitivity to C3/C4 responses was examined via separate 12-member ensembles (reported in Supplementary).

• Extent of change (1982–2015) across 44.5 million km² of drylands: 6% experienced significant desertification (browning), 41% significant greening, 53% no significant change. Estimated desertified area: 2.70 million km².
• Attribution (global area fractions where the driver is the largest absolute contributor): CO₂ fertilization 44.1%, Land Use (LU) 28.2%, Climate Variability (CV) 14.6%, Climate Change (CC) 13.1%.
• Mean per-pixel contributions to ΔNDVImax (area-weighted mean ± SD): CO₂ = +0.021 ± 0.011; CV = +0.006 ± 0.020; CC = −0.002 ± 0.023; LU = +0.005 ± 0.032. Relative global contributions: ~67.8% CO₂, −5.6% climate, 15.5% LU; CV accounts for ~19.4% of observed greening.
• Anthropogenic climate change (ACC = CO₂ + CC): Mean greening effect globally (+0.019 ± 0.027), but with a desertifying signal over 12.55% of drylands (5.43 million km²). Approximately 213.4 million people are exposed to areas where ACC exerts a negative effect; 85% live in developing or newly industrialized countries.
• Desertification drivers: In desertified drylands (2.70 million km²), LU was the primary driver in 79.9% and a contributing factor in 99.0% of areas. Mean component changes (ΔNDVImax ± SD) in desertified areas: LU = −0.040 ± 0.034; CC = −0.004 ± 0.030; CV = −0.002 ± 0.024. Only ~2.27% of desertified area was driven solely by climate.
• Risk hotspots and compounding effects: Regions where negative LU impacts are currently offset by positive ACC total ~12.0 million km² affecting ~507 million people; these, plus areas with negative CC but no significant vegetation change (~7.2% of drylands), are at high risk of future desertification.
• Greening attribution in significantly greening drylands (~18.0 million km²): CO₂ is the largest driver in ~40% of areas, LU in ~38%, CV in ~13%, CC in ~8%. Regional greening is strongly influenced by LU and CV in the Sahel, India, China, and Australia.
• Spatial ACC desertification hotspots: parts of the western United States, eastern Brazil, Iraq, Syria, Jordan, Kazakhstan, Uzbekistan, Mongolia, and Australia.
• Population impacts: Current desertification directly affects ~190 million people; high-risk areas (from unsustainable LU or ACC) encompass ~20% of drylands and ~580 million people, disproportionately in low socioeconomic countries.

By explicitly accounting for CO₂ fertilization, separating climate variability from long-term climate change, and detecting structural ecosystem changes, the study clarifies the roles of ACC and LU in dryland vegetation trends. The results show that while CO₂ fertilization has driven widespread greening, it also masks negative LU and CC signals in many regions, creating latent risk. LU emerges as the dominant driver of observed desertification and a major contributor to greening and browning at regional scales, indicating that management practices can exacerbate or mitigate climate pressures. ACC imposes a significant desertifying effect on over 5 million km², affecting hundreds of millions of people—predominantly in developing economies—highlighting equity and adaptation concerns. The findings resolve apparent contradictions between global greening and reports of increasing aridification by demonstrating that net greening coexists with localized degradation where negative LU and CC impacts compound. This nuanced attribution is critical for projecting vulnerabilities and targeting interventions where positive CO₂ effects may not persist or may be outweighed by intensifying climate stress and unsustainable land use.

This study provides the first observation-based global attribution of dryland vegetation change that integrates CO₂ fertilization, climate variability and change, and both gradual and abrupt land-use impacts. Despite widespread greening, 6% of drylands experienced desertification since 1982, primarily driven by LU, with ACC contributing substantial additional desertification risk over 12.6% of drylands. Many regions remain vulnerable where negative LU effects are currently masked by positive ACC signals. Policy implications include prioritizing sustainable land management, improving drought resilience, and focusing adaptation where ACC-driven declines and LU pressures compound, especially in lower-income countries. Future work should refine CO₂ response estimates across C3/C4 systems, incorporate additional ecosystem structure metrics beyond NDVI, extend analyses with higher-resolution and longer climate records, and better integrate socio-ecological datasets to link biophysical changes with livelihoods.

• Proxy limitations: NDVImax may not capture degradation processes that increase greenness (e.g., shrub encroachment, invasive species, agricultural intensification) or saturation effects in dense canopies; bare soil spectral effects in very low biomass systems can bias NDVI.
• Linearity assumption: The approach assumes a linear NDVI–NPP relationship and constant GPP:Ra ratio; both may vary with temperature and ecosystem state.
• CO₂ response uncertainty: Differential C3/C4 responses to elevated CO₂ remain uncertain; long-term experiments show potential for higher-than-expected C4 responses. Results rely on theoretical scaling and weighted ensembles.
• Climate data biases: Dryland climate is poorly sampled; discrepancies across gridded datasets introduce uncertainty despite the ensemble approach.
• Temporal sensitivity: Trend analyses in drylands are sensitive to start/end years and to extreme events (e.g., ENSO); although breakpoints are detected, some abrupt changes may be missed or misattributed.
• Attribution scope: ACC is defined as CO₂ + CC; direct effects of other greenhouse gases cannot be isolated. Land-use attribution is statistical and may conflate unobserved management or disturbance dynamics.
• Spatial/measurement thresholds: Sensor error thresholds (±0.001 NDVI) and masking decisions may exclude subtle changes; hyper-arid and non-dryland areas are excluded by design, limiting generalizability beyond drylands.