Common irrigation drivers of freshwater salinisation in river basins worldwide

The study addresses the growing global problem of freshwater salinisation, which affects surface and groundwater quality, ecosystem and human health, and agricultural productivity. Despite its importance, cross-regional assessments of inland salinity status, trends, and sectoral impacts—particularly on irrigation—are scarce. Coastal drivers (e.g., saltwater intrusion, tides, storm surges) are comparatively well studied, but inland drivers require more attention. Human drivers include road salt, mining, and agriculture; of these, road salt impacts are relatively well quantified, while agricultural contributions lack systematic cross-regional assessments. The research aims to quantify inland salinity levels and trends across diverse hydroclimatic regions and to evaluate how agricultural activities—particularly irrigation withdrawals, return flows, and irrigated area—relate to elevated salinity and salinisation trends.

Prior work highlights multiple drivers of freshwater salinisation: coastal processes (e.g., saltwater intrusion exacerbated by groundwater pumping and sea-level rise), hydroclimatic variability (evaporation, discharge), and human activities such as road salt application, mining, and agriculture. Road salt has been extensively studied across temperate regions with snow/ice management. Agricultural influences have been shown at local to regional scales, including evapoconcentration from irrigation, mobilization of salts through return flows, and inputs from fertilisers, with documented cases in arid/semi-arid areas and specific basins (e.g., San Joaquin Valley). However, systematic cross-regional to global assessments of agricultural drivers on surface-water salinisation are limited. Existing machine-learning studies (e.g., US river salinity) implicate combined effects of agriculture, mining, and groundwater pumping, suggesting the need for broader-scale, long-term evaluations integrating multiple driver categories.

Scope and regions: The authors analyzed 401 river sub-basins across seven regional river basins worldwide: Mississippi (North America), Ebro, Danube, Rhine (Europe), Orange (Africa), Mekong (Asia), and Murray–Darling (Australia). Sub-basins were delineated for each salinity monitoring station that met data criteria. Data and period: Long-term river salinity (electrical conductivity, EC) observations from 1980–2010 were compiled from an open global database, GEMStat, and Mekong River Commission. A total of 417,315 observations were aggregated to monthly averages. Stations required at least 15 years of monthly data within 1980–2010. Sub-basin delineation: Using HydroSHEDS 15 arc-sec flow direction data, stations were snapped to flow accumulation cells; sub-basins <1 km² were excluded. Driver variables: Twenty-six drivers were compiled as sub-basin averages, representing three groups: hydroclimatic (temperature, precipitation, discharge, PET, AET, evaporative ratios PET/P and AET/P), geographic (soil salinity, elevation, distance to coast, relative distance to coast), and human (sectoral water withdrawals and return flows for irrigation and non-irrigation sectors; dams and reservoir metrics; land use including total cropland and irrigated area fraction; fertiliser applications of N and P). Data sources included PCR-GLOBWB 2 (5 arc-min monthly), CRU TS4.03 (0.5°), GRanD v1, HYDE/MIRCA2000, WISE30sec, DEM products, and global mining areas; road salt was estimated for Mississippi sub-basins (1992–2010) as a case study. Salinity impact classification and exceedance: Long-term annual average EC per sub-basin was classified into Low (<700 µS cm⁻¹), Moderate (700–1500 µS cm⁻¹), and High (>1500 µS cm⁻¹) classes, based on international irrigation water-use thresholds. Monthly threshold exceedance frequency was computed as the fraction of months exceeding 700 µS cm⁻¹. Statistical comparisons: Driver contributions across salinity classes were normalized (each sub-basin value divided by the group-average for that driver) and compared using Wilcoxon rank-sum tests between classes (Low–High, Low–Moderate, Moderate–High) at p<0.05. Trend analyses: For each sub-basin, Mann–Kendall tests and Sen’s slope estimated trends for salinity (and drivers where possible) over 1980–2010. Kendall’s tau indicated direction and significance (p<0.05). Sen’s slope provided magnitude of EC change (µS cm⁻¹ yr⁻¹). Machine learning: Two Random Forest regression models were built to predict long-term average salinity using driver long-term means and trends: one for sub-basins with significant increasing trends (N=128) and one with significant decreasing trends (N=96). Models were trained on 80% of data and tested on 20%. Tuning explored ntree (500–7500) and mtry; optimal settings were ntree=5000, mtry=12 (increasing) and ntree=5000, mtry=2 (decreasing). Variable importance was assessed with Conditional Permutation Importance (CPI) to account for correlated predictors (threshold δ=0.85; robustness checked 0.8–1). Recursive feature elimination identified minimal variable sets with best performance. Performance metrics included R², RMSE, MAE, and NMAE.

• Salinity status: 65% (263/401) of sub-basins fall in the Low impact class (<700 µS cm⁻¹). The remaining 35% exceed the irrigation threshold: 108 Moderate (700–1500 µS cm⁻¹) and 31 High (>1500 µS cm⁻¹).
• Threshold exceedance dynamics: Even sub-basins with Low long-term averages experienced monthly exceedances. Across all sub-basins, average exceedance frequency was 33% with high variability (0–58%).
• Trends (1980–2010): 57% (229/401) show increasing salinity trends (32% of total with significant increases, p<0.05); 43% (172/401) show decreasing trends (24% significant). Mississippi and Orange have the largest shares of increasing trends (≈73% and 63% of sub-basins, respectively), while Murray–Darling (78%) and Danube (73%) are dominated by decreasing trends.
• Driver associations by class: Irrigation-specific drivers—irrigation water withdrawals (Irr. ww), irrigation return flows (Irr. rf), and irrigated area fraction—are significantly higher in High vs. Low salinity classes. Moderate vs. Low classes also show higher total cropland and fertiliser use. In some basins (Rhine, Murray–Darling), non-irrigation sector withdrawals/return flows are more elevated in higher salinity classes.
• Hydroclimatic patterns: Evaporative indices (PET/P, AET/P) and temperature are lower in higher salinity classes, contrary to simple evapo-concentration expectations. Analysis indicates irrigation intensity is higher in less arid sub-basins, potentially due to water limitations in arid regions and irrigation–climate feedbacks reducing atmospheric aridity in irrigated areas.
• Alternative grouping (>700 vs. <700 µS cm⁻¹): Results remain consistent—higher irrigation-related drivers and cropland/fertiliser in salinity-affected sub-basins; lower evaporative indices.
• Road salt and mining: No significant contributions detected to High/Moderate classes in the multi-region analysis; Mississippi case study ranks road salt below irrigation variables in importance.
• RF importance: For sub-basins with increasing trends, irrigated area is the top predictor, followed by distance/relative distance to coast, elevation, soil salinity, and irrigation withdrawals/return flows. For decreasing-trend sub-basins, distance to coast is most important; non-irrigation withdrawals/return flows rank higher than irrigation variables. Model performance: R²≈0.62 (increasing) and R²≈0.51 (decreasing).

The study demonstrates that inland freshwater salinisation is widespread and increasing across many regions, yet exhibits strong spatial heterogeneity in levels and trends. Monthly exceedances above irrigation thresholds occur even where long-term averages are low, indicating potential seasonal constraints on irrigation suitability and the need to move beyond annual metrics. A central finding is the consistent association between elevated salinity (and increasing trends) and irrigation-specific activities—withdrawals, return flows, and irrigated area—across diverse hydroclimatic contexts. Notably, significant irrigation contributions also occur in less arid regions, likely reflecting both water availability constraints in arid basins and irrigation–atmosphere feedbacks that can reduce apparent aridity where irrigation is intensive. Machine-learning analyses corroborate statistical tests, identifying irrigated area and irrigation water use as key predictors of salinity levels where salinisation is increasing, while geographic context (distance to coast) and non-irrigation sector impacts are relatively more influential where salinity is decreasing. The findings underscore the importance of incorporating irrigation-specific human drivers into water quality assessments, monitoring design, and predictive models. They also call for seasonal and crop-specific assessments to capture intra-annual risk dynamics and agricultural impacts.

Across 401 sub-basins in seven major river basins worldwide (1980–2010), many systems have low long-term salinity, yet a substantial fraction exceed irrigation thresholds seasonally, and 57% exhibit increasing trends. Irrigation-specific activities—withdrawals, return flows, and irrigated area—consistently contribute to higher salinity levels and increasing salinisation, emerging as top predictors in machine-learning analyses. Geographic context, particularly distance to the coast, remains important across trend categories, while non-irrigation sectors are more influential in areas with decreasing salinity trends. These results highlight the need to explicitly include irrigation drivers in water quality monitoring and modelling, and to manage irrigation practices to mitigate salinisation. Future research should: (i) integrate seasonal and crop-specific thresholds to assess agricultural yield risks; (ii) refine irrigation driver characterisation (techniques, sources, crop rotations); (iii) expand assessments to other human uses and ecological impacts; and (iv) improve global datasets for drivers like road salt and mining where regionally relevant.

Uncertainties arise from heterogeneous monitoring records and spatial/temporal coverage (requirement of ≥15 years within 1980–2010), potential measurement gaps, and the use of global models (PCR-GLOBWB, CRU) and gridded datasets with inherent biases and resolution constraints. The classification relies on general irrigation thresholds that do not capture crop-specific or seasonal tolerances. Driver variables can be correlated; CPI mitigates but does not eliminate interpretability challenges. Road salt and mining contributions were limited by data availability and regional relevance; road salt was assessed only for the Mississippi case. Irrigation-specific details (e.g., irrigation method, water source mix, local management) were not resolved, potentially affecting attribution. Aridity–irrigation interactions may confound simple hydroclimatic relationships. Results pertain to the 1980–2010 period and selected basins and may not generalize to all regions or more recent conditions.