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

Soil salinization, the excessive accumulation of soluble salts in soil, poses a major threat to agriculture and ecosystems globally. This problem is expected to worsen due to projected climate change impacts on natural salinization processes (primary salinization). The complexity of these processes, involving rainfall, aeolian deposition, rock weathering, and water transport, has hindered accurate predictions of future salinity levels. This research addresses this challenge by developing data-driven models to predict primary soil salinization in the world's drylands up to 2100 under various climate change scenarios. Understanding these future dynamics is crucial for developing effective strategies to manage soil, water, and vegetation resources and for data-driven policy making. The study focuses on primary salinization, excluding secondary sources like irrigation practices and sea-level rise, as data limitations currently prevent accurate modelling of these factors on a global scale. Accurate prediction of future soil salinity is critical for sustainable land management, informing mitigation and adaptation strategies for land managers and policymakers.

Existing studies on soil salinization vary in scale and methodology. Local-scale studies offer detailed analyses of specific areas, but lack the broad perspective needed to understand global trends. Global-scale studies have attempted to map the current extent of salt-affected soils, but quantitative predictions of future changes are scarce. Previous attempts to project the effects of climate change on soil salinization are often descriptive, using limited datasets or focusing primarily on secondary salinization. Some studies have used simplified indicators or older Global Circulation Models (GCMs), limiting the accuracy and spatial resolution of their predictions. The IPCC has acknowledged the likely impact of climate change on primary salinization mechanisms, including alterations in hydrological balance, sea-salt intrusion, and wind-borne salt deposition, but quantitative, spatially explicit predictions on a global scale remain largely absent. This study aims to address this gap by developing a quantitative, global-scale analysis, providing spatially explicit projections of long-term variations in primary soil salinity under different climate change scenarios.

This study uses machine learning (ML) algorithms to predict long-term primary soil salinity on a global scale. The methodology involved several key steps: 1. **Data Collection:** Geo-referenced soil profiles with measured electrical conductivity (EC) values were obtained from the World Soil Information Service (WoSIS) database. Profiles from croplands were excluded to isolate natural salinization processes. A total of 44,708 samples from 11,517 profiles were used. 2. **Predictor Data:** Fourteen predictors were used, including purely spatial data (soil type, texture, rooting depth, topography, etc.) obtained from ISRIC, SoilGrids, and other sources, and spatio-temporal data (precipitation, evapotranspiration, sea salt deposition rates) derived from multiple Global Circulation Models (GCMs) from CMIP5 and CMIP6 projects under various Representative Concentration Pathways (RCPs) and Shared Socioeconomic Pathways (SSPs). Five-year moving averages of spatio-temporal predictors were used to smooth out intra-annual variability. 3. **Model Training and Validation:** Supervised ML algorithms, specifically an ensemble of regression trees, were used to build predictive models. Hyperparameters were tuned using Bayesian optimization with holdout cross-validation (25% holdout). Model performance was evaluated using metrics such as R-squared (R²), root mean squared error (RMSE), mean absolute error (MAE), and Nash-Sutcliffe efficiency coefficient (NSE). Multiple models (30 runs each) were trained for each of 16 GCM outputs, selecting the best-performing model (lowest RMSE) for each. 4. **Soil Salinity Projection:** The trained models were applied to a global soil base map of drylands (Aridity Index ≤ 0.65), predicting annual soil salinity at five depths (0, 10, 30, 60, and 100 cm) for each grid cell (0.5° resolution) between 1904 and 2100. Averages to 1-meter depth were calculated for each year. Outliers were removed using a robust method. Predictions for the reference period (1961-1990), mid-term future (2031-2060), and long-term future (2071-2100) were compared to assess changes in soil salinity. Analysis was conducted at the grid cell, country, and continental levels.

The analysis revealed several key findings: 1. **Predictor Importance:** Long-term annual precipitation frequency was the most influential predictor, followed by WRB soil classes and daily evapotranspiration. Soil clay content and water holding capacity showed a negative correlation with soil salinity, consistent with previous research. 2. **Spatial Distribution of Salinity Change:** The study projected significant geographical variability in salinity change. Regions projected to experience the most severe increases in soil salinity by 2100 include South America, southern Australia, Mexico, the southwestern US, and South Africa. Conversely, decreases in salinity are projected for the northwestern US, the Horn of Africa, Eastern Europe, Turkmenistan, and western Kazakhstan. The intensity of these changes is generally more pronounced under higher greenhouse gas emission scenarios (RCP 8.5 and SSP 5-8.5). There were inconsistencies between CMIP5 and CMIP6 model predictions in some regions, highlighting the inherent uncertainty in climate projections. 3. **Ensemble Agreement:** The multi-model ensemble agreement on the sign of salinity change varied geographically. High agreement was observed for regions in the southern hemisphere (Australia, South America, Southern Africa) under SSP 2-4.5 and SSP 5-8.5 scenarios, indicating a higher risk of climate-induced salinity increase in these areas. Lower agreement was found for regions in the Middle East, Russia, and Sahara. 4. **Country-Level Analysis:** At the country level, Brazil, Namibia, South Africa, and Mexico showed the highest relative increase in soil salinity under RCP 8.5 and SSP 5-8.5 scenarios. Botswana showed the highest relative increase under SSP 5-8.5. Conversely, Canada, Somalia, and Ethiopia showed substantial projected decreases. 5. **Changes in Salt-Affected Soil Area:** The total area of dryland soils with EC ≥ 2 dS m⁻¹ is projected to increase in Australia and South America and decrease in Asia and Europe under high emission scenarios (SSP 5-8.5). Brazil, Mexico, and Mongolia show the highest estimated increase at the country level, while Canada, Somalia, and Ethiopia show the largest decreases. Similar trends were observed using a higher salinity threshold (4 dS m⁻¹). Model accuracy was assessed through 10-fold cross-validation, achieving a mean R² of 72.79% and a mean RMSE of 3.6 (normalized RMSE ~6%). Comparisons with existing global soil datasets (HWSD, WISE-30) demonstrated the superior predictive accuracy of the developed models, especially at the country level.

The findings of this study provide valuable insights into the potential impacts of climate change on primary soil salinization, particularly in dryland regions. The projected increases in salinity in several key regions align with climate projections of decreased precipitation and increased evapotranspiration in these areas. The geographical variability highlights the need for region-specific adaptation strategies. The high agreement among models in certain regions increases confidence in the projections, while inconsistencies in other areas underscore the challenges of predicting future climate and soil processes. The superior performance of the machine learning models compared to existing global soil datasets demonstrates the value of integrating climate projections into salinity prediction models. However, the study's limitations, discussed below, need to be considered when interpreting the results.

This study presents novel, spatially explicit predictions of global primary soil salinization under various climate change scenarios. The findings identify regions at high risk of increased salinity, highlighting the need for proactive management strategies. The methodology combining machine learning with multiple climate models provides a robust framework for assessing future soil salinity. Future research should focus on improving the resolution and accuracy of climate projections, incorporating secondary salinization mechanisms, and developing physically based models to validate the ML predictions. Further research is needed to improve data availability, especially in data-scarce regions, to improve model accuracy and reduce uncertainty.

Several limitations of the study should be considered: The reliance on large-scale climate projections introduces uncertainty, particularly at finer spatial scales. The exclusion of secondary salinization processes limits the comprehensiveness of the predictions. The accuracy of the models is dependent on the quality and availability of the input data, with potential biases towards regions with more extensive datasets (North America and Australia). The inherent uncertainties in GCM predictions and inconsistencies between CMIP5 and CMIP6 models highlight the need for cautious interpretation of the results. The coarse spatial resolution (0.5°) may mask fine-scale variations in soil salinity. Future studies should incorporate improved data and models to refine these predictions.