Global predictions of primary soil salinization under changing climate in the 21st century

The study addresses how projected climate change will influence primary (naturally occurring) soil salinization in global drylands. Salinization adversely affects plant growth, soil biota, and ecosystem services, with cascading environmental and socioeconomic impacts including reduced agricultural productivity, land degradation, and health issues. While spatial mapping of salt-affected soils has advanced, quantitative global-scale projections of future salinization remain scarce due to complex processes, data limitations, and uncertainties in climate drivers. The authors aim to develop and validate data-driven models to predict top-soil (0–1 m) electrical conductivity (EC) as a proxy for primary soil salinity under multiple climate scenarios for mid- (2031–2060) and long-term (2071–2100) horizons, focusing on drylands (Aridity Index ≤ 0.65). Secondary salinization processes (irrigation, groundwater rise, sea-water intrusion) are excluded to isolate primary climate-driven effects.

Prior work has mapped salt-affected soils from local to global scales (~8.31–11.73 Mkm² affected), with concentration in drylands where evapotranspiration exceeds precipitation. Studies have described climate change links to salinization but often lack quantitative, spatially explicit projections. Early estimates suggested substantial expansion of salt-affected areas (e.g., North Mediterranean by 2050 per 1 °C warming; Australian drylands at risk rising to 170,000 km² by 2050). Schofield et al. used indicators (relief, moisture flux, flow deficit) to flag potential risk areas but without quantitative EC projections and based on a single older GCM. Many studies focus on secondary salinization (irrigation, coastal intrusion) at local scales. Mechanistic and stochastic models of salinity exist but require detailed inputs (initial salinity, soil/groundwater parameters) not available globally. Machine learning has shown promise in digital soil mapping, including recent global spatio-temporal EC/sodicity prediction for past decades, motivating ML for forward projections.

Scope: Predict primary soil salinity (ECe) in drylands globally (AI ≤ 0.65) at 0.5° resolution, annually from 1904–2100, focusing on 0–1 m depth. Secondary salinization sources (irrigation, groundwater, sea-level rise) are excluded. Data – response: Geo-referenced soil EC profiles from WoSIS (1950–2014). After removing records without sampling dates and excluding cropland samples (identified using GLCC v2.0 for pre-1997 and MODIS MCD12Q1/MCD12C1 for 2000/2006/2014/2018), 44,708 samples from 11,517 profiles remained. Depth information (upper/lower sample depths) retained. Predictors (14 total): - Purely spatial (9): WRB soil classes; soil clay content (%); field capacity (mm); wilting point (mm); effective rooting depth (m); elevation (m); slope (degrees); plus sample upper and lower depths (cm). Spatial layers sourced from ISRIC SoilGrids, IGBP-DIS, Yang et al. rooting depth, Esri World Elevation; processed to common projections and resolutions; gaps filled by neighbourhood averaging. - Spatio-temporal (5; 5-year moving averages): annual precipitation frequency (day−1); precipitation intensity (cm); daily evapotranspiration (cm day−1); sea salt dry and wet deposition rates (mg m−2 day−1). Derived from GCM outputs. Climate drivers and scenarios: Historical and future outputs from 16 GCMs: CMIP5 (RCP 4.5, RCP 8.5) and CMIP6 (SSP2-4.5, SSP5-8.5; forced by RCP4.5/8.5). Only models providing daily precipitation (to compute wet days) and sea-salt deposition were used. All climate fields interpolated to 0.5° regular grid; spatio-temporal predictors computed as 5-year moving averages to smooth interannual variability. Modeling: For each GCM dataset, trained supervised ensemble regression tree models (MATLAB fitrensemble; LSBoost aggregation selected) using the 14 predictors to predict EC. Categorical WRB encoded as integers; others numeric. Hyperparameters tuned via Bayesian optimization (expected-improvement-per-second-plus), with holdout CV during tuning; maximum 100 evaluations; up to 500 learning cycles. Training repeated 30 times per GCM (480 runs) to address optimizer non-reproducibility; best model per GCM selected by lowest RMSE. Overall 16 best-fit models retained. Validation: 10-fold cross-validation used to compute R², RMSE, MAE, NSE with 95% CI via BCa bootstrap (1,000 samples) across 30 runs. Additional depth-wise validation by binning depths (0–20 to 100–200 cm). Country/continent-level aggregation comparisons to measured EC means. External comparisons with HWSD (0–30 cm) and WISE-30 (0–20 cm) surface EC at country/continent level. Projection and aggregation: Global drylands base map (UNEP-WCMC) rasterized to 0.5°; only AI ≤ 0.65 cells retained (24,045 grid cells). For each cell-year, predicted EC at depths 0, 10, 30, 60, 100 cm; averaged to 0–1 m using trapezoidal rule; outliers removed via MAD-based filter. Three periods defined: reference (1961–1990), mid-term (2031–2060), long-term (2071–2100). Relative change computed as (Future mean − Reference mean)/Reference mean. Areas of salt-affected soils computed annually for thresholds EC ≥ 2 dS m−1 (primary) and also EC ≥ 4 dS m−1 (supplementary), aggregated by country/continent using GADM boundaries.

- Predictor importance: Among 14 predictors, 5-year average annual precipitation frequency was most influential (~14% importance across 16 models). WRB soil classes (13.07%) and daily evapotranspiration (9.26%) were next. Partial dependence indicated: EC generally lower with higher elevation, steeper slopes, and higher precipitation frequency; coarse-textured soils associated with higher EC than fine-textured; effective rooting depth influences EC to ~4 m. - Spatial projections (multi-model ensembles): By 2071–2100 relative to 1961–1990, hotspots of increasing primary soil salinity appear in drylands of South America, southern and western Australia, Mexico, southwest United States, South Africa, and to a lesser extent Spain, Morocco, northern Algeria, parts of western/southern Sahara, central India, southeast Mongolia, and northern China. Decreases or stability projected in northwest United States, the Horn of Africa, Eastern Europe, Turkmenistan, and west Kazakhstan. - Ensemble agreement: Higher model agreement (especially CMIP6 SSP2-4.5 and SSP5-8.5) in southern/eastern Australia, South America, and southern Africa; lower agreement/greater uncertainty in Middle East, Russia, and Sahara. CMIP5 vs CMIP6 sometimes disagree (e.g., Russian drylands sign of change). - Country-level relative EC change (grid-cell means, long-term): • RCP 8.5 (2071–2100 vs 1961–1990): Brazil +15.1% (95% CI 13.25–16.95), Namibia +13.57% (12.1–15.04), South Africa +11.2% (9.41–13), Mexico +6.38% (4.96–7.8), Australia +3.31% (2.88–3.73). • SSP5–8.5: Botswana +24.94% (22.71–27.16), South Africa +21.35% (19.84–22.85), Namibia +17.69% (16.14–19.24), Brazil +16.21% (14.77–17.66). - Continental changes in area of EC ≥ 2 dS m−1 (long-term vs 1904–1999 average): Under SSP5–8.5, Africa +1.45%, Asia −0.28%, Australia +3.38%, North America −2.45%, Europe −0.92%, South America +6.70%. Under RCP8.5, Australia +0.60% and South America +4.88% (long-term) with Asia and Europe decreasing. - Country-level area expansion (EC ≥ 2 dS m−1, 2071–2100 vs 1904–1999, SSP5–8.5): Brazil +43%, Mexico +14.5%, Mongolia +8% largest increases among top-30 countries by grid-cell count; largest decreases: Canada −10%, Somalia −8.5%, Ethiopia −5%. - Model accuracy: Average 10-fold CV R² ≈ 72.79%; RMSE ≈ 3.6 dS m−1 (~6% normalized to observed EC range). Depth-wise R² averages ranged ~63.6% (0–20 cm) to ~79.6% (80–100 cm). Country-level aggregation R² between predicted and measured means: 80.41%; continental level: 99.64%. Compared to HWSD and WISE-30 at surface, present models achieved higher R² at both country and continental levels.

The findings quantitatively project where and how primary soil salinity in drylands may change under climate forcing, addressing the longstanding lack of global-scale, spatially explicit predictions. Increases in salinity correlate with projected decreases in precipitation frequency/intensity and increases in evapotranspiration, consistent with salt mass-balance theory. Independent climate studies report drying trends and elevated evapotranspiration in identified hotspots (e.g., southern/western Australia, Mexico, parts of South America), supporting the plausibility of the projections. The study’s ensemble agreement maps highlight regions of higher confidence (southern hemisphere drylands) and areas of greater uncertainty (Middle East, Sahara, parts of Russia). Discrepancies with earlier indicator-based global studies (e.g., Schofield et al.) likely stem from methodological differences (use of a single older GCM and empirical PET vs multi-GCM physically based AET and ML here). Evidence from remote sensing case studies shows that substantial salinity changes can occur over decadal timescales, lending support to the feasibility of projected changes. Overall, the results inform prioritization of monitoring and adaptation in vulnerable drylands and suggest where mitigation (e.g., enhancing leaching efficiency, vegetation management) may be most needed.

This work develops and applies a machine-learning framework to generate annual, gridded (0.5°) global projections of primary soil salinity (0–1 m EC) across drylands from 1904 to 2100 under multiple CMIP5/CMIP6 scenarios. It identifies key climatic and environmental drivers (notably precipitation frequency and evapotranspiration), maps future hotspots (South America, southern/western Australia, Mexico, southwest US, South Africa), and projects areas likely to experience decreases (northwest US, Horn of Africa, Eastern Europe, Turkmenistan, west Kazakhstan). It quantifies country- and continent-level changes in both relative EC and area exceeding EC thresholds (≥2 and ≥4 dS m−1). These projections provide actionable insights for land and water managers to recognize salinization hotspots and plan mitigation/adaptation. Future research should: integrate mechanistic, physically based root-zone salt-budget models where data allow; improve coverage and temporal consistency of soil profiles; develop regional models for underrepresented areas; incorporate reanalysis for historical bias correction; expand GCM ensembles with sea-salt aerosol projections; increase spatial resolution; and produce spatially explicit uncertainty estimates for predictions.

- Scope limited to primary salinization; excludes effects from irrigation, shallow saline groundwater rise, and sea-level induced intrusion due to lack of global projections and data. - Training data biases: WoSIS profiles unevenly distributed (overrepresentation in North America and Australia), fewer/less accurate older samples; potential bias toward recent conditions. - Predictor uncertainties: Purely spatial predictors (field capacity, wilting point, rooting depth) and sea-salt deposition have higher uncertainty in data-scarce deserts; GCM outputs inherently uncertain; CMIP5 vs CMIP6 sometimes disagree on sign. - Coarse projection resolution (0.5°), constrained by GCM grid scales (~1–3°); finer-scale heterogeneity not captured. - No spatially explicit quantification of output uncertainty (e.g., Monte Carlo) due to missing predictor uncertainty maps and computational cost; reliance on cross-validation metrics instead. - External validation limited: comparisons to HWSD/WISE-30 are to datasets without temporal variability and not fully independent of training regions; independent global salinity datasets are lacking. - Ensemble coverage limited by availability of sea-salt deposition fields across GCMs.