Upstream land-use change and transboundary river water redistribution: a global mapping

1 Introduction

Global transboundary river basins encompass ~40% of global freshwater flow and sustain over 3 billion people across 150 + nations (Wolf et al., 2010). However, every major treaty governing these shared systems regulates extractive, infrastructure-based water use: dams, diversions, and pumping that physically remove water (Wolf et al., 2010; McCracken and Wolf, 2019). More than 800 international water agreements cataloged in the Transboundary Freshwater Dispute Database (TFDD) (2024) share a common architecture treating water availability at international boundaries as a function of climate and geography not as partially determined by upstream land management decisions. This leaves an entire class of transboundary water transfer ungoverned. When upstream countries expand irrigated agriculture, reforest catchments, or clear forests, they alter precipitation partitioning between evapotranspiration and runoff, reducing downstream river flow without physical diversion structures. We term this silent water redistribution: the unrecognized, unregulated transfer of water availability across sovereign boundaries through upstream territorial decisions (United Nations, 1997; te Wierik et al., 2019; Wolf et al., 2003).

Scholars investigate that LULCC alters basin hydrology at governance-relevant scales. Winkler et al. (2021) showed LULCC affected ~32% of the global land surface between 1960 and 2019. Sterling et al. (2013) demonstrated LULCC reduced global terrestrial evapotranspiration (ET) by ~3,500 km³ yr.⁻¹ and Jaramillo and Destouni (2015) showed that irrigation and flow regulation alone raised global human water consumption by ~3,563 km³ yr.⁻¹. Bosmans et al. (2017) confirmed LULCC-driven ET changes comparable to direct withdrawals, concentrated in tropical transboundary river basins. Veldkamp et al. (2017) showed that human interventions cause water scarcity downstream. Wang-Erlandsson et al. (2022) proposed a planetary boundary specifically for green water and demonstrated it has already been transgressed. Mehta et al. (2024) documented 33 million hectares of irrigation expansion from 2000 to 2015, half in water-stressed regions. Recently, Tang et al. (2025) demonstrated that historical LULCC altered land water availability across ~3.8% of the global land surface, and Ma et al. (2024) showed that deforestation-induced runoff changes are dominated by indirect forest-climate feedback that amplifies the transboundary signal beyond direct biophysical effects alone.

Despite this evidence, a critical gap remains in green water governance, which has been called for conceptually by te Wierik et al. (2019) and Keys et al. (2017) The physical basis of atmospheric moisture connections has been reviewed systematically by te Wierik et al. (2021) and modeled across 26 major basins by Wang-Erlandsson et al. (2018), but no study has translated LULCC-driven hydrological changes into basin-specific, treaty-comparable water volumes across global transboundary basins. Jiang et al. (2025) projected transboundary conflict from water scarcity under climate change but did not isolate LULCC as a governable driver. Posada-Marín et al. (2024) assessed atmospheric moisture risks using static governance indices rather than observed land cover transitions. Hoek van Dijke et al. (2022) modeled tree restoration impacts using scenario-based models without governance analysis. No study has translated LULCC-driven hydrological changes into basin-specific, treaty-comparable water volumes across global transboundary basins.

Research constructs a BCU spatial framework (796 BCUs and 153 countries) classified by hydrological position via HydroSHEDS (Lehner et al., 2008). Across 413 upstream and midstream source BCUs, we detect five LULCC transitions from MODIS (Friedl and Sulla-Menashe, 2022) (2001–2023), quantify ET consequences via space-for-time substitution on MOD16A2GF (Running et al., 2021), separate climate signals via Budyko elasticity decomposition using CHIRPS precipitation (Funk et al., 2015) and ERA5 (Hersbach et al., 2020), and compute SWRI against FAO AQUASTAT (Food and Agriculture Organization of the United Nations, 2021). In the next step, we compare SWRI results against TFDD database records (Transboundary Freshwater Dispute Database (TFDD), 2024) and verify predictions with JPL GRACE Mascon gravity (Watkins et al., 2015; Wiese et al., 2016). This research addresses two questions:

• How much of the observed ET change in transboundary source regions is attributable to LULCC rather than concurrent climate variability? And
• How large is the blue-water redistribution that transmitted to each downstream co-riparian, expressed against its renewable endowment and do existing water treaties contain any provisions applicable to this mechanism?

The findings show that upstream LULCC silently redistributes −31.66 km³ yr.⁻¹ of blue water across transboundary river basins, dominated by irrigation (−36.37 km³ yr.⁻¹) and reforestation (−23.76 km³ yr.⁻¹). GRACE data confirm signals in 62% of the 13 basins in which the index reaches at least 1% of the downstream country’s renewable water resources, the significance threshold applied throughout this study. Systematic comparison with the TFDD treaty records (Transboundary Freshwater Dispute Database (TFDD), 2024) reveals that no agreement in force contains an explicit provision governing upstream LULCC-driven water redistribution, leaving 78.5% of the redistributed flux in these basins under treaties that are silent on the mechanism and leave millions of people without institutional protection.

2 Methods

2.1 Spatial framework

Basin polygons for 313 transboundary river basins were obtained from the TFDD database (Wolf et al., 2010; McCracken and Wolf, 2019; Transboundary Freshwater Dispute Database (TFDD), 2024), intersected with national boundaries from the Natural Earth Countries dataset (Natural Earth, 2022) in Google Earth Engine, yielding 796 BCUs spanning 153 countries (Figure 1). The mean flow accumulation was extracted from HydroSHEDS 15 arc-second grid (Lehner et al., 2008) using zonal statistics. Position thresholds were computed on a per-basin basis: BCUs were ranked by mean flow accumulation and assigned to tertile-based categories per basin. This per-basin normalization ensures that headwater nations in small basins are classified relative to their own basin’s hydrological gradient rather than against an inappropriate global reference. The complete dataset includes 307 upstream, 106 midstream, and 323 downstream BCUs (60 BCUs with insufficient HydroSHEDS coverage above 60° N were excluded; 24 single-country BCUs were excluded). The 413 source BCUs served for LULCC detection and ET attribution and 323 downstream for SWRI computation.

2.2 LULCC detection

The MODIS Land Cover Type product (MCD12Q1 Collection 6.1) (Friedl and Sulla-Menashe, 2022; Sulla-Menashe et al., 2019) at 500 m annual resolution from 2001 to 2023 was used. A three-stage noise suppression pipeline was implemented: 5-year temporal majority filter following Abercrombie and Friedl (2016), early-epoch (2001–2005) and late-epoch (2019–2023) composites, and 3-year persistence rule. Five transition types were defined: deforestation; rainfed agricultural expansion; urbanization; reforestation; and irrigated agricultural expansion (Figure 2). A woody-class cycling diagnostic was applied to reforestation. Initial detection was 313,290 km², but 70.4% consisted of pixels cycling between spectrally similar woody IGBP classes (forest ↔ woody savanna). After excluding false positives, a total of 92,665 km² of reforestation were estimated. A 5% per-BCU hydrological significance threshold, grounded in Bosch and Hewlett (1982) findings, was applied to identify hydro-significant on BCUs.

2.3 Irrigation expansion detection

Irrigated cropland expansion was quantified directly from the Mehta et al. (2024) global Area Equipped for Irrigation (AEI) grids (Siebert et al., 2013). The grids were differenced, and positive AEI changes (hectares per cell) were summed within each source BCU and converted to km². Because the estimates are derived directly from aggregate AEI metrics rather than downscaled to 500-meter cropland pixels, this approach avoids the assumption that individual MODIS cropland pixels are irrigated. Rainfed agricultural expansion was then obtained per BCU as the residual of total MCD12Q1-detected agricultural expansion minus AEI-based irrigated expansion (bounded at zero). This procedure yielded 69,336 km² of irrigated expansion across 254 BCUs, with 233,157 km² (76.5% of total agricultural expansion) remaining as rainfed expansion (Figure 3).

2.4 Evapotranspiration quantification

ET data from MOD16A2GF Version 6.1 (Running et al., 2021; Mu et al., 2011) were used to select the band with scale factor 0.1. Annual ET was computed by summing all 46 8-day composites per calendar year after applying a quality mask (0 ≤ raw value < 32,767). The gap-filled product ensures temporal completeness. A critical methodological decision was the use of class-specific control pixels in the space-for-time substitution: deforestation controls used pixels that remained as forest; rainfed expansion controls used pixels that remained as natural vegetation; reforestation controls used pixels that remained in valid source classes (IGBP 9, 10, 12, 14, and 16). Using global controls produced a sign error in deforestation ΔET (+90.8 mm yr.⁻¹ with global controls versus (vs.) − 30.3 mm yr.⁻¹ with class-specific controls). Each BCU was assigned to a Beck et al. (2018) 1 km resolution Köppen–Geiger climate zone (1991–2020) through centroid sampling. Zone-specific ΔET coefficients were derived from published catchment syntheses (Bosch and Hewlett, 1982; Döll and Siebert, 2002; Jackson et al., 2005); the space-for-time estimate was retained only where |ΔETSFT| < 5 × |ΔETlit|, an outlier guard against unstable small-sample contrasts, and the literature coefficient was used otherwise. Irrigation used literature coefficients throughout, and deforestation ΔET was constrained to ≤ − 10 mm yr.⁻¹, reflecting its physically required sign. The full zone × transition coefficient table is provided in Supplementary Table S1. Irrigated agricultural ET change was assigned using literature-based coefficients differentiated by Köppen zone (+200 to +650 mm yr.⁻¹ across zones, with the maximum in arid Zone B; the unweighted mean across the 413 source BCUs is +468 mm yr.⁻¹ and the mean weighted by irrigated-expansion area, 64% of which lies in tropical Zone A, is +535 mm yr.⁻¹). Volumetric blue-water change per BCU was computed as follows:

ΔBW = − ΔET × A transition × 10 − 6 ( km 3 yr − 1 ) (1)

where ΔETt is the evapotranspiration change associated with transition class t (mm yr.⁻¹), and At is the area of that transition within the BCU (km²). Equation 1 was evaluated separately for each of the five transition classes, and the BCU net signal is their sum, ΔBWnet = Σt ΔBWt.

2.5 Space-for-time (SFT) contrasts vs. literature coefficients

Each contrast is spatial, not temporal. ΔET_SF T max is the mean 2019–2023 ET of the pixels that underwent transition to minus of the class-specific control pixels of the same BCU. Epoch change enters only through the Budyko attribution (Equations 2–4). Contrasts exist for 306 of 315 deforestation: 280 agricultural expansion, 158 urbanization, and 281 of 282 reforestation BCUs; otherwise, the literature coefficient applies. The 5 × screen replaced 31 contrasts, all rainfed expansion and 3% of its area: 20 BCUs where new cropland evapotranspires 108–444 mm yr.⁻¹ less than its natural-vegetation controls and 11 where it evapotranspires 159–469 mm yr.⁻¹ more, consistent with irrigated fields within MODIS cropland pixels. Because MCD12Q1 does not separate irrigated from rainfed cropland and the AEI grids are not pixel-resolved, no contrast exists for irrigation; its ΔET is the Köppen-zone literature coefficient (Supplementary Table S7). Contrasts exceed the literature coefficient in 47% of rainfed-expansion, 42% of urbanization (median −292 vs. −200 mm yr.⁻¹), and 40% of reforestation BCUs (+222 vs. +150 mm yr.⁻¹), but in only 4% of deforestation BCUs, whose contrasts are far smaller than paired-catchment syntheses (median −14 against −30 to −280 mm yr.⁻¹) and 36% positive; the sign constraint set 141 retained contrasts to −10 mm yr.⁻¹.

2.6 Climate attribution via Budyko elasticity decomposition

Annual precipitation per BCU was derived from CHIRPS pentad composites (Funk et al., 2015) for 2001–2023. Monthly ERA5-Land maximum and minimum 2 m temperatures (Hersbach et al., 2020) were averaged to annual values per BCU (zonal mean at 5 km), giving the annual mean temperature and mean diurnal range. Annual Hargreaves–Samani potential evapotranspiration (Hargreaves and Samani, 1985) was then computed per BCU using annual mean extraterrestrial radiation estimated from the BCU centroid latitude, with a floor of 50 mm yr.⁻¹, and years with missing temperature were excluded. The Budyko aridity index φ = PET/P (Budyko, 1974) was computed per BCU per year. Using the Choudhury-Yang parametric Budyko equation (Choudhury, 1999) with n = 2.0 (central estimate), precipitation and PET elasticities were derived analytically following Roderick and Farquhar (2011). The climate-attributed ΔET was computed as follows:

ΔE T climate = ε P × ΔP P ¯ × ET ¯ + ε PET × ΔPET PET ¯ × ET ¯ (2)

where ε^P and ε^PET are the precipitation and potential evapotranspiration elasticities of ET, ΔP and ΔPET are the changes between the baseline (2001–2005) and evaluation (2019–2023) period means, and overbars denote period means. The LULCC-attributed residual yields the attribution fraction:

ΔE T LULCC = ΔE T total − ΔE T climate , f LULCC = ΔE T LULCC / ΔE T total (3)

where ΔETtotal is the observed evapotranspiration change between the two epochs. fLULCC was computed at BCU level and applied uniformly to all transition classes of that BCU. Where |ΔETtotal| ≤ 1 mm yr.⁻¹, the ratio is unstable and fLULCC was assigned the global mean (0.303); the same value was assigned where climate inputs were insufficient (46 BCUs, reported as unclassified in Figure 4e). Estimates were bounded to [−1, 2] to limit leverage of extreme ratios. Climate-corrected blue-water change per BCU is as follows:

ΔB W LULCC = f LULCC × ΔBW (4)

Sensitivity analysis was performed under n = 1.5, 2.0, and 2.5 to characterize Budyko parameter uncertainty.

2.7 Silent water redistribution index

The index is computed as follows:

SWRI ( % ) = ∑ ΔB W LULCC , upstr e am TARW R downstream × 100 (5)

where ΣΔBW^{LULCC,upstream} is the sum of climate-corrected blue-water change across all upstream and midstream BCUs draining into the downstream BCU, and TARWR^downstream is the FAO AQUASTAT (Food and Agriculture Organization of the United Nations, 2021) total actual renewable water resources of the downstream country. Upstream-to-downstream routing exploited the TFDD basin topology: All BCUs sharing a common basin name are hydrologically connected. Of the 313 basins, 272 (86.9%) contained both source and downstream BCUs, yielding 683 unique routing pairs. The remaining 41 basins (13.1%) lack a downstream-classified BCU after the framework’s exclusions (60 high-latitude BCUs and 24 single-country BCUs) and the per-basin tertile assignment; no basin retained a downstream BCU while lacking a source BCU. In 34 basins with multiple downstream BCUs, the upstream signal was distributed proportionally by downstream territorial area. All routed signals summed to −31.66 km³ yr.⁻¹, identical to the unrouted source total, confirming mass conservation in the routing. SWRI severity tiers were defined as High (|SWRI| ≥ 5%), Moderate (1–5%), Low (0.1–1%), and Negligible (<0.1%). A downstream BCU is classed as SWRI-significant where |SWRI| ≥ 1%, and a basin where at least one of its downstream BCUs meets this criterion (15 BCUs in 13 basins). Significance in this sense is a hydrological attribute alone—the magnitude of the redistributed volume relative to the receiving country’s renewable water resources.

2.8 Treaty comparison

The primary data source was the TFDD (Transboundary Freshwater Dispute Database (TFDD), 2024) dataset, cataloguing 3,812 records covering more than 800 internationally recognized freshwater agreements from 1820 to 2023. Each record encodes up to 156 binary and categorical issue-area variables. Linking treaty records to our SWRI basin dataset required a three-tier BCODE matching procedure: exact string matching (185 matches), fuzzy string matching with ≥ 0.80 similarity (84 matches), and contextual verification. A governance criterion was applied conservatively: A treaty row was accepted as evidence of LULCC governance only if it contained an explicit obligation, allocation formula, or institutional mechanism governing upstream land-use-driven evapotranspiration change as a transboundary water redistribution mechanism. Three false-positive flags (Mekong, La Plata Apá River, and Ili/Yili He records) were removed after manual inspection. Screening combined the TFDD issue-area variables (IWRM, surface-water allocation, irrigation, and environment) with full-text keyword queries of the treaty corpus in four groups: evapotranspiration; consumptive use or consumptive loss; land use, land cover, or vegetation change; and return flow or water depletion. The full coding protocol and per-basin decisions are provided in the Supplementary material, and we explicitly distinguish the absence of provisions naming the evapotranspiration mechanism from the presence of general procedural provisions (e.g., IWRM clauses) that could in principle be applied to it. Each of the 272 basins with downstream BCUs was assigned to one of four categories. Basins below the significance threshold were assigned to No Action Needed, and significant basins without a treaty in force to No Governance. Significant basins with a treaty in force were assigned to Partially Governed where at least one linked record contains a provision bearing indirectly on the mechanism and to Governance Gap where the treaty contains no such provision. Variability-management provisions, present in 61 of the 355 linked records, were not admitted to the Partially Governed criterion because they address seasonal flow variability rather than land-use-driven evapotranspiration change. In the final classification, the Partially Governed category is empty because the IWRM field is uncoded or negative in every record linked to the 13 significant basins, and the only keyword matches were the three false positives described above. The category is retained in the protocol and in the released coding tables. Population exposure was quantified using GPWv4 Revision 11 (2020) (Center for International Earth Science Information Network (CIESIN), Columbia University, 2018) via GEE sum reducer at 5 km scale.

2.9 GRACE/GRACE-FO terrestrial water storage verification

Terrestrial water storage anomaly (TWSA) was implemented in GEE. TWSA data were sourced from the JPL GRACE Mascon RL06 Version 3 product (Watkins et al., 2015; Wiese et al., 2016; Landerer et al., 2020) at native ~0.5° (~55 km) resolution, expressed as liquid water equivalent thickness in centimeters relative to a 2004–2009 baseline mean. Two epoch means were computed: early (April 2002–December 2006) and recent (January 2018–September 2023). The epoch change metric (ΔTWSA = TWSA^recent − TWSA^early) was computed for each downstream BCU using a spatial mean reducer at 55 km scale in GEE. Concordance was assessed by comparing the sign of ΔTWSA against the sign of the BCU-level SWRI: Basins where both are negative (SWRI predicts deficit and GRACE detects water storage loss) or both positive (SWRI predicts surplus and GRACE detects storage gain) are classified as concordant.

2.10 SDG 6.5.2 overlay and compound vulnerability classification

National implementation scores for SDG indicator 6.5.2 (proportion of transboundary basin area covered by operational arrangements for water cooperation) were obtained from the UN-Water SDG 6 Data Portal (UN-Water, 2023) and assigned to each SWRI-significant basin as the score of its primary affected downstream country. Because SDG 6.5.2 is reported nationally, this value serves as a proxy for basin-level institutional capacity rather than a basin-specific measurement; for multi-riparian basins, it necessarily reflects the reporting of a single riparian state. For each of the 13 significant basins, we computed a compound vulnerability score (CVS, 0–100) integrating exposure, capacity deficit, and independent observational support:

CVS = 40 × ∣ SWRI ¯ ∣ ∣ SWRI ∣ max + 0.4 × ( 100 − SD G 6.5.2 ) + 20 × C (6)

where |SWRĪ| is the basin-mean SWRI magnitude, |SWRI|ₘₐₓ = 12.89% is the maximum BCU-level magnitude observed across all basins, and C = 1 where the sign of the basin’s GRACE/GRACE-FO terrestrial water storage change agrees with the sign of its SWRI. Vulnerability classes were assigned by explicit rule, independently of the continuous score. Critical where basin-mean |SWRI| ≥ 2%, SDG 6.5.2 < 45, and the GRACE storage sign is concordant with the SWRI sign; High where |SWRI| ≥ 2% (or ≥ 1% with SDG 6.5.2 < 40) without satisfying all three Critical conditions; Moderate where |SWRI| ≥ 1% and SDG 6.5.2 < 55; and Low Priority otherwise. This yielded two Critical basins (Jordan and Gambia), five High, four Moderate, and two Low Priority. The continuous CVS (Equation 6) provides the ranking reported in Figure 5c, whereas the categorical class encodes which combination of exposure and capacity drives risk; the highest-scoring basin (Dniester, CVS 76.8) is accordingly classed High rather than Critical, its cooperation score (53) lying above the 45-point capacity threshold. Compound vulnerability uses basin-mean SWRI throughout; basin-maximum values are reported in Figures 6, 7.

3 Results

3.1 LULCC-driven water redistribution

Across all upstream and midstream BCUs of transboundary river basins, LULCC between the 2001–2005 and 2019–2023 epochs generated a net blue-water redistribution of −31.66 km³ yr.⁻¹ (Equations 1–4) after Budyko elasticity correction for concurrent climate variability (sensitivity range: −27.86 to −35.46 km³ yr.⁻¹ for Budyko n = 1.5 to 2.5). This signal represents water annually removed from the renewable water budgets of downstream co-riparian states by upstream territorial decisions that carry no recognition in international water law. The climate correction retains 84% of the uncorrected net signal as land-use-attributed (−31.66 of −37.48 km³ yr.⁻¹); attribution fractions vary widely across BCUs (unweighted mean 0.30, |ΔBW|-weighted mean 0.89) and are most consequential for deforestation—where concurrent precipitation decline in some source BCUs had partially masked the land-use signal—and least consequential for irrigated agricultural expansion, which carries a climate fraction of only 0.02, confirming that irrigation-driven water redistribution is an almost entirely anthropogenic signal. Of the five detected transition types, irrigated agricultural expansion dominates the consumptive signal (Figure 8) despite covering only 69,336 km²—less than one-quarter of the 233,157 km² of rainfed agricultural expansion (Table 1)—because supplemental water application raises evapotranspiration by +200 to +650 mm yr.⁻¹ above rainfed baselines, depending on climate zone (+535 mm yr.⁻¹ when weighted by irrigated area; Supplementary Table S1). After climate correction, irrigation contributes −36.37 km³ yr.⁻¹, reforestation −23.76 km³ yr.⁻¹, deforestation +25.65 km³ yr.⁻¹, rainfed expansion +2.50 km³ yr.⁻¹, and urbanization +0.32 km³ yr.⁻¹ (Figure 4). The net signal is dominated by two consumptive transitions—irrigation and reforestation—partially offset by the water-releasing effects of deforestation. Reforestation emerges as the second-largest consumptive driver of the transition-type analysis; the policy implications of this finding are examined in the Discussion section. The influence of the observed space-for-time contrasts on this balance between consumptive and water-releasing transitions is revealed by comparison with an accounting based on literature coefficients alone. With literature coefficients applied to every transition, the accounting yields a climate-corrected deforestation surplus of +56.4 km³ yr.⁻¹ and a net global surplus of +12.1 km³ yr.⁻¹; substituting the observed MODIS contrasts reduces the deforestation surplus to +25.6 km³ yr.⁻¹ and increases the reforestation loss from −18.5 to −23.8 km³ yr.⁻¹, giving the net loss of −31.66 km³ yr.⁻¹. Omitting the 5 × screen or the deforestation sign bound shifts the net signal only to −35.1 and −33.3 km³ yr.⁻¹, respectively; the net loss therefore arises from the observed contrasts themselves and is insensitive to the screening rules. The irrigation term, for which no observed contrast is available, is bounded by the sensitivity analysis in Supplementary Table S7.

Table 1. Twenty BCUs with the largest LULCC area.

3.2 SWRI distribution: where the governance signal concentrates

Silent Water Redistribution Index (Equation 5) ranges from −12.89 to +5.02% across 323 downstream BCUs, with 189 BCUs (58.5%) experiencing a net deficit and 134 (41.5%) a net surplus (Figure 6). The near-zero median reflects the reality that downstream countries with large TARWR denominators absorb even substantial absolute upstream signals into negligible relative fractions. Governance urgency concentrates in the tail: 15 BCUs (4.6%) exceed the 1% significance threshold and 4 BCUs (1.2%) exceed 5% (Table 2). The Jordan basin produces the two most extreme SWRI values. Israel (SWRI = −12.89%) and the Palestinian Territories (−3.10%) face upstream silent redistributions driven entirely by irrigation expansion in the Jordanian and Syrian headwaters, with their small TARWR denominators of 0.75 and 0.80 km³ yr.⁻¹, respectively, amplifying the relative signal to levels comparable to major single-country allocation shares negotiated in bilateral water treaties. The Dniester basin (Moldova, −12.25%) is the second most severely affected, with reforestation as the dominant driver rather than irrigation—upstream Carpathian afforestation is increasing evapotranspiration and reducing runoff toward a downstream country whose entire national water endowment is 1.3 km³ yr.⁻¹. The La Plata system carries the world’s largest absolute upstream deficit of −19.97 km³ yr.⁻¹, driven by Brazil’s 19,449 km² of new irrigated cropland in the headwaters, yet generates only moderate relative SWRI values for downstream Argentina (−1.64%) and Paraguay (−1.97%) because their large TARWR denominators absorb the signal—demonstrating that governance urgency is determined by the signal-to-TARWR ratio, not absolute volume alone. Two basins exhibit positive SWRI values: the Gambia (+5.02%) and the Mekong (Laos, +2.28%). These surpluses are not governance successes but transient liabilities whose reversal, when upstream land use eventually changes, carries no treaty mechanism to anticipate or manage (Figure 6).

Table 2. SWRI severity classification across all 323 downstream BCUs.

Severity thresholds defined relative to FAO AQUASTAT TARWR of the downstream country. Deficit BCUs experience upstream-imposed water loss; surplus BCUs experience upstream-generated water gain. Representative basins listed per severity class.

3.3 A structural blind spot in international water law

Cross-referencing every SWRI-significant basin against the full treaty corpus—3,812 treaty–basin–country records drawn from more than 800 internationally recognized freshwater agreements—yields the study’s central governance finding: At the level of treaty text, no agreement in force contains a provision that explicitly governs upstream LULCC-driven evapotranspiration change as a mechanism of transboundary water redistribution. Among these 13 basins, 11 are classified as Governance Gap, in which a treaty exists but is wholly silent on the mechanism, and 2 as No Governance, in which no agreement of any kind applies (Table 3 and Figure 7). Together, these basins account for a redistributed flux of −24.85 km³ yr.⁻¹, equivalent to 78.5% of the total flux across the 13 basins.

Table 3. All 13 SWRI-significant basins (|SWRI| ≥ 1%) ranked by absolute SWRI.

SWRI is the mean signed value across all downstream BCUs in the basin. Dominant driver is the upstream LULCC transition type contributing the largest absolute blue-water signal.

The coding protocol includes an intermediate Partially Governed category for significant basins whose treaties, although not governing the mechanism explicitly, contain provisions that could be applied to it—a coded IWRM provision or substantive treaty language on watershed, land-use, vegetation, or consumptive-use change (Methods). Manual inspection identified and removed three spurious LULCC flags—an analyst annotation embedded in the Mekong treaty record, a geographic watershed reference within the La Plata Apá River clause, and a reference to the Joint Committee of Chinese-Kazakh Transboundary River Watershed, recorded in a Sino-Kazakhstan joint declaration linked to the Ili/Yili He basin—after which no basin remained in the Partially Governed category. This category is empty in the final classification because of current treaty content, not the protocol’s design. It remains in the published coding tables to allow for future updates if any treaty texts or basin-organization procedural rules are found to activate the mechanism.

Temporal analysis of treaty chronology explains that this governance deficit stems from historical agreements predating awareness of land–water coupling. Contrarily, 6 of 11 “Governance Gap” basins are regulated by instruments ratified post-2000, including the Mekong River Commission agreement (2018) and Tigris–Euphrates framework (2015). The mean treaty age within these gap basins is 15.7 years, significantly lower than the global dataset mean of 31.2 years. This discrepancy proves that contemporary transboundary water diplomacy fails to integrate quantified redistribution mechanisms into its legal architecture. While this omission is specific to codified text, it leaves downstream populations vulnerable. Consequently, an estimated 70.3 million individuals reside within downstream BCUs subject to hydrologic redistribution exceeding the 1% threshold. Although residential proximity does not inherently establish functional hydro-dependence, these populations lack formal, legally binding treaty mechanisms to recognize, negotiate, or mitigate these systemic redistributions.

3.4 Compound vulnerability and independent physical verification

The downstream countries most acutely affected by the governance gap are also those with the least institutional capacity to address it. The average SDG 6.5.2 score among downstream countries of the 13 significant basins is only 40.5%, compared to the global average of 54% (Supplementary Table S2). The three downstream countries with the lowest scores—Turkmenistan (22%, Lake Sarygamesh), Israel (33%, Jordan), and Chad (29%, Lake Chad)—are among the territories most exposed to SWRI, demonstrating a systematic co-occurrence of maximum hydrological exposure and minimal institutional adaptive capacity, termed compound vulnerability. Two basins, namely, Jordan and Gambia, meet all three criteria for critical compound vulnerability (|SWRI| ≥ 2%, SDG 6.5.2 < 45%, and GRACE-concordant) (Figure 5). Lake Sarygamesh satisfies the exposure and capacity criteria, but its GRACE signal is discordant, classifying it as High. Independent satellite gravimetry from the GRACE and GRACE-FO missions aligns directionally with the SWRI-predicted signal in 8 of the 13 significant basins (62%); sign agreement alone is indistinguishable from chance (two-sided binomial p = 0.58); however, across the 57 affected BCUs, the continuous association is significant (ΔTWSA on SWRI: slope +1.53 cm per percentage point, 95% CI 0.36–2.71; Pearson’s r = 0.33). The Dniester basin exhibits a terrestrial water storage anomaly epoch change of −24.48 cm, and the Medjerda basin −22.95 cm between the early (2002–2006) and recent (2018–2023) GRACE periods, consistent with satellite ET-based attribution. The five non-concordant basins are individually explicable by documented precipitation trend confounders: La Plata’s positive GRACE signal reflects the well-documented post-1970s precipitation increase in the Rio de la Plata watershed, while the Sahel-adjacent Senegal and Lake Chad basins show positive GRACE signals consistent with the regional precipitation recovery since the 1980s Sahelian drought. The global mean terrestrial water storage anomaly across all 323 downstream BCUs is −7.60 cm, confirming a broad pattern of water storage declines in the downstream territories of the world’s transboundary river basins (Figure 5).

4 Discussion

The findings show that upstream LULCC silently redistributes −31.66 km³ yr.⁻¹ of blue water across the global transboundary river basins, reframing transboundary water governance from a qualitative concern to a quantified governance gap. Previous study established that LULCC alters basin hydrology comparable to direct water withdrawals (Jaramillo and Destouni, 2015; Bosmans et al., 2017), and international water law was not designed for land management (te Wierik et al., 2019; Wolf et al., 2003). This study provides the first globally consistent, mechanistically attributed, quantitative evidence that these facts constitute a structural gap in transboundary water cooperation, not a peripheral concern for a basin’s water stress, but a systematic blind spot operating across 78.5% of the global LULCC-driven water redistribution.

The reforestation paradox is perhaps the most intellectually uncomfortable finding because it challenges the traditional belief that upstream climate action is hydrologically neutral with respect to downstream water rights (Duveiller et al., 2018; Bryan et al., 2018; Tong et al., 2020). Moldova receives −12.25% SWRI from upstream Carpathian afforestation, Chad −1.56% from Sahelian re-greening (Kou et al., 2021), and Namibia −1.12% from Angolan highland reforestation. All actively incentivized by Paris Agreement’s Article 6 climate finance framework and REDD+ mechanisms, with no requirement to assess downstream hydrological consequences, even when the affected downstream country has a TARWR small enough that a multi-percent SWRI represents a nationally significant water deficit. This creates a direct and currently unresolved collision between SDG 13 (upstream afforestation) and SDG 6 (downstream water security), a collision that can be quantified with the SWRI framework. The governance implication is not that upstream reforestation should be discouraged but that climate finance conditionality must expand to include downstream water impact assessment as a standard due-diligence requirement.

The treaty chronology deserves particular attention because it forecloses the most intuitive explanation for the governance gap. If basins were covered exclusively by nineteenth-century treaty agreements negotiated before the hydrological consequences of LULCC were understood, the governance silence could be attributed to historical ignorance rather than contemporary institutional failure. For example. Six (6) of the eleven (11) basins have treaties signed after 2000, including the technically sophisticated 2018 Mekong agreement. Their silence on upstream land-use-driven evapotranspiration impact reflects a conceptual framework in which water availability across national boundaries is treated as a function of climate and geography rather than jointly determined by upstream decisions. The findings acknowledge that the transitions quantified here—satellite-detectable irrigation expansion and climate-finance-driven afforestation—did not operate at comparable extent or intensity in earlier treaty eras and that contemporary-equivalent institutions did not then exist. In addition, the chronology finding therefore characterizes the conceptual architecture of modern agreements rather than faulting historical ones. This framing is embedded in the 1997 UN Watercourses Convention (United Nations, 1997), which defines watercourse use in terms of physical water extraction rather than evapotranspiration changes that alter the hydrological cycle upstream of the boundary. Until this conceptual separation between water governance and land governance is formally bridged, governance gaps persist regardless of treaty recency.

Several methodological limitations warrant acknowledgment. The Budyko climate attribution assumes stationarity of the precipitation–runoff relationship over the study period, an assumption that becomes less defensible in basins undergoing rapid climate changes. The space-for-time substitution method, while validated at comparable spatiotemporal scales (Blois et al., 2013), cannot fully replicate the temporal dynamics of LULCC transitions, and the land-cover specific control pixel approach may introduce uncertainty about the representativeness of control populations that are most tenuous in heavily disturbed basins. Country-level TARWR as the SWRI denominator cannot capture sub-national water distribution inequalities. Finally, the 62% GRACE concordance rate means that for 5 of the 13 significant basins, the corroboration is masked by other terrestrial water storage processes operating at the ~55 km mascon scale: compensating precipitation trends (the post-1970s Rio de la Plata wetting and the post-1980s Sahel recovery), groundwater abstraction, reservoir operations, and storage responses to transitions other than the dominant one. In the La Plata system, widespread rainfed agricultural expansion has itself raised groundwater levels and increased flooding across the flat interior plains (Houspanossian et al., 2023), a within-basin hydrological response that a downstream BCU storage trend cannot separate from the upstream signal. These mismatches may prove transient: Where masking precipitation trends reverse, the SWRI-predicted deficit and the GRACE signal would emerge together. Population exposure counts residents of affected BCUs and should be read as potentially affected population rather than demonstrated dependence on the impacted flows. The AEI-based irrigation estimates end in 2015, so irrigation equipped after 2015 is not captured because irrigation lacks a space-for-time contrast, and its largest accounting term relies on literature coefficients rather than observed ET. The wide range of these coefficients across climate zones (+200 to +650 mm yr.⁻¹) and our ±25% sensitivity analysis (Supplementary Table S7) quantify this dependence. Consequently, developing an observation-based irrigation ΔET at the resolution of irrigated-field maps remains a priority for future work.

Policy pathways to governance reform do not necessarily require new international institutions or wholesale treaty renegotiation. Instead, three incremental interventions are recommended: (1) Organizations covering the eleven Governance Gap basins could incorporate SWRI monitoring into their existing reporting systems to provide an early-warning mechanism for LULCC-driven water shifts above the significance threshold; (2) Future treaty negotiations and periodic review processes could extend Espoo Convention (United Nations Economic Commission for Europe (UNECE), 1991) prior-notification principles to include upstream LULCC transitions, specifically irrigated agricultural expansion exceeding agreed thresholds within BCU, to capture most cases above the threshold while imposing minimal burden on routine land managements, and finally (3) International climate finance mechanisms funding upstream afforestation could condition disbursements on completion of downstream water impact assessments, using the SWRI framework as the evaluation tool (Goldstein et al., 2020). In summary, land use is no longer for food, urban, industry, and ecological purposes; it is also water management. The institutional architecture of transboundary water governance must now consider LULCC in treaty agreements to reflect the reality.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/ffgc.2026.1951830/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Yaseen T. Mustafa, University of Zakho, Iraq

Reviewed by: Pinglang Kou, Chongqing University of Posts and Telecommunications, China

Paula Torre Zaffaroni, Leibniz-Institute of Freshwater Ecology and Inland Fisheries (IGB), Germany