Global lake tipping vulnerability under sustained agricultural exposure

Abstract

Intensive agriculture drives lake eutrophication, degrading water quality and biodiversity through harmful algal blooms, hypoxia, and costly regime shifts. It remains unclear whether agriculture triggers sudden change through expansion or instead accumulates vulnerability under long-standing land cover. Here we combine trophic state and nearshore land cover for 859 lakes worldwide to detect 125 abrupt, persistent tipping events from 2003 to 2022. We then use interpretable models to separate chronic agricultural exposure from rapid expansion, and compare simple driver-based forecasts with classical early warning signals. Near-term tipping risk is dominated by a lake’s baseline trophic state and its long-term surrounding cropland; small increases above that baseline raise the 12-month tipping risk by 16%, whereas early warning signals add little predictive power. These findings show that sustained agricultural pressure narrows the safe operating space for freshwater ecosystems and provide a versatile framework for managing other vulnerable social-ecological systems.

Lakes tip into degraded states mainly because of what already surrounds them, not new developments. Chronic agricultural exposure and baseline water quality globally predict abrupt shifts better than early warning signals.

Introduction

Lakes are sensitive indicators of environmental change^1,2 and repositories of public value^3,4. Globally, lakes supply drinking water⁵, sustain fisheries and recreation⁶, and anchor regional biodiversity^2,7. Their small volume relative to catchments and their strong internal feedbacks make them disproportionately sensitive to human pressures⁸. When external forcing intensifies, lakes can cross critical thresholds and reorganize into persistently eutrophic, hypoxic regimes that resist restoration⁹. Preventing such collapses, or reversing them once they occur, requires knowing both where thresholds lie and how quickly we can approach them without triggering instability.

Here, we distinguish three related forms of abrupt change. A changepoint is a statistical discontinuity in a time series. A regime shift is the ecological interpretation of such a discontinuity: a persistent reorganization of lake trophic state rather than a transient interannual excursion. A tipping point is a narrower concept, referring to a regime shift associated with declining resilience, where a small perturbation can move the system into an alternative state maintained by internal feedbacks¹⁰. This distinction is important since not every abrupt change is evidence of tipping.

Canonical tipping paradigms emphasize bifurcation-induced transitions, in which a control variable, such as nutrient loading, crosses a critical threshold, leading to the loss of stability of the clear-water state. Ecologically, this means that a lake can shift from a clear, macrophyte-stabilized regime to an alternative turbid regime maintained by phytoplankton dominance, vegetation loss, and internal nutrient recycling⁹. These feedbacks can maintain the turbid state even after external loading is reduced; recovery may require substantially stronger reductions in pressure than those that caused the shift, a form of path dependence known as hysteresis. Beyond thresholds, the concept of baseline vulnerability recognizes that systems differ in their sensitivity to perturbations: chronic exposure may elevate risk even below critical levels, making systems more susceptible to tipping when triggers occur. This motivates a management framing in which safe operating spaces are defined jointly by absolute exposure levels and context-dependent vulnerability.

Among diffuse watershed pressures, agriculture is a pervasive driver of eutrophication in freshwaters and coasts^11–13. Fertilizer and manure elevate phosphorus and nitrogen loads, promote harmful cyanobacterial blooms^14–16, and drive oxygen depletion through decomposition^17,18; internal loading and vegetation loss can then reinforce turbid states and slow recovery^11,19. In many regions, livestock production and manure redistribution amplify phosphorus surpluses and legacy loading, coupling animal agriculture to long-run eutrophication risk^20,21. Unlike point sources, runoff is spatially heterogeneous²², temporally pulsed²³, and strongly modulated by climate extremes via event-scale erosion and hydrologic connectivity²⁴. With global agricultural production projected to expand by 14% by 2034²⁵, pressure on freshwater ecosystems need not arise only through further land conversion. Intensification within existing farmland, such as through greater fertilizer and manure inputs, higher livestock densities, irrigation, and drainage, can increase nutrient export even where mapped cropland area remains stable^13,25. Practical guidance on how current exposure, management intensity, and recent land cover change shape lake stability is therefore urgently needed.

Global assessments suggest that abrupt trophic shifts are uncommon but occur worldwide, and that lake trajectories covary with climate and socioeconomic pressure²⁶. However, several limitations in the literature inhibit management. Thresholds are often specified in nutrient units that are difficult to regulate and vary idiosyncratically across lakes. Further, the distinction between chronic agricultural exposure and proximate triggers is poorly resolved, obscuring whether land cover operates as a predisposing factor or an immediate cause. Third, the interaction between agricultural pressure and baseline trophic state has not, to our knowledge, been precisely quantified, leaving it unclear whether effects are uniform or context dependent. Finally, early warning signal (EWS) theory predicts that, as resilience declines near some tipping points, lakes may recover more slowly from perturbations, producing rising variance, autocorrelation, or spectral reddening before a shift; however, the empirical forecasting value of these signals in forced, seasonal systems remains uncertain²⁷.

In this work, we address these gaps by combining monthly trophic state with nearshore land cover for 859 lakes worldwide from 2003 to 2022 (Figs. 1 and 2). Gilarranz et al. provided the first global census of regime shifts and stability in lake productivity using satellite-derived trophic state index (TSI) time series, demonstrating the upward trend in tipping events in recent decades and the dependency on population pressure and temperature²⁶. O’Brien et al. evaluated classical and second-generation EWS on multi-trophic lake data, demonstrating that clear critical transitions are rare and that many indicators have limited predictive skill in practice²⁷. Building on these foundations, we evaluate whether baseline trophic state, agricultural exposure, and EWS improve prediction of near-term abrupt trophic change, using within-between decompositions to distinguish chronic exposure from recent deviations. We interpret these models as associational rather than as definitive causal attribution. We detect abrupt, persistent state shifts using a bootstrap-based changepoint approach that quantifies uncertainty in pre-event variance trends, and then estimate logistic models that decompose within-lake deviations from between-lake baselines in both agriculture and trophic state. Minor deviations above stable, long-standing cropland baselines robustly predict near-term eutrophic tipping, whereas oligotrophication signals are weak and buffer-dependent. Baseline trophic state is the dominant predictor; cross-lake differences in agriculture add at most modest further association once state and deviations are controlled, and EWS provide little additional forecasting value (Table 1). The within-between decomposition provides leverage by separating persistent cross-lake exposure from short-term deviations, but unmeasured time-varying drivers prevent strong causal claims. These findings support treating agriculture as a chronic vulnerability driver and prioritizing lakes with already elevated trophic states.

Fig. 1. Global trends in lake trophic state and environmental drivers.

a Monthly median total phosphorus (2003–2022; n = 27,007 observations). The black line shows the monthly median; grey bands indicate ± 1 standard deviation (SD). Red/grey shading indicates deviations above/below eutrophic status. Green bars mark abrupt changes in which lakes contribute data; median shifts therefore reflect sample composition, not within-lake phosphorus trends. b Number of lakes contributing phosphorus measurements. c Detrended global median Trophic State Index (TSI; n = 87,858 measurements). Black line shows monthly median; grey bands indicate ± 1 SD. Blue/grey shading indicates deviations above/below the long-term mean.

Fig. 2. Spatial distribution of lakes in our dataset and net change in TSI over time.

Net long-term change in monthly Trophic State Index (TSI) for each lake from 2003 to 2022; the colour scale captures the combined contribution of gradual deterioration, abrupt regime shifts, and recovery transitions, and is not specific to detected tipping events. Colour scale indicates the magnitude and direction of change (red = increase, blue = decrease). The inset histogram shows the frequency of ΔTSI across lakes. Detected tipping events are shown separately in Fig. 3 and Supplementary Tables 5 and 11.

Table 1.

Global model: predictors of eutrophic tipping events (H = 12 months)

Predictor	OR [95% CI]	p value
AGwithin*	1.157 [1.034–1.296]	0.0113
AGbetween	0.903 [0.651–1.252]	0.5402
TSIbetween***	1.513 [1.285–1.781]	<0.001
TSIwithin****	0.276 [0.225–0.338]	<0.0001
AR(1)	1.052 [0.883–1.254]	0.5703
Skewness	0.992 [0.831–1.184]	0.9269
Spectral	1.147 [0.929–1.417]	0.2026

Subscripts within and between refer to within-lake and between-lake components of the corresponding within-between decomposition.

OR odds ratio, CI confidence interval, AG surrounding cropland fraction (5 km buffer), TSI Trophic State Index, AR(1) lag-1 autocorrelation.

*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001

Results

Abrupt trophic shifts were rare but globally distributed

Across 859 lakes meeting data quality criteria from 2003 to 2022, we detected 125 tipping points distributed across 76 lakes (8.9%). Events had a mean magnitude of 8.5 TSI units for eutrophic transitions (median 8.3, interquartile range 3.9-12.5) and exhibited pre-event variance increases consistent with EWS theory. About half of the events represented eutrophication (51%, 64/125), with the remainder oligotrophication (49%, 61/125). While tipping points were rare, many lakes showed gradual trophic deterioration over the study period (Fig. 2), supporting a distinction between widespread drift and discrete threshold crossings.

Tipping was rare in deep and sparsely monitored lakes

Database characteristics for all 859 qualifying lakes were summarized in Supplementary Table 1. Detected tipping events were not distributed uniformly across lake types. In the subset matched to HydroLAKES²⁸ morphometry (636 of 859 lakes), tipping detections were concentrated in shallow systems (45 of 384 very shallow lakes, 11.7%) and were rare or absent in deep lakes (0 of 32 lakes deeper than 30 m; Supplementary Table 2). Tipping detections also scaled strongly with monitoring intensity: across observation-month quartiles of the 859 qualifying lakes, tipping rates rose from 0% (Q1, n = 224) and 4.6% (Q2, n = 217) to 6.4% (Q3, n = 235) and 27.9% (Q4, n = 183); tipped lakes had a median 119 versus 106 observed months for non-tipped lakes (Cohen’s d = 1.01, p < 0.001; Supplementary Tables 3 and 13). The 76 tipping lakes, therefore, represented a high-confidence detected subset rather than a complete census, reflecting both ecological susceptibility and the greater ability of dense time series to support changepoint detection.

Chronic agricultural exposure shaped eutrophic tipping risk

To separate chronic exposure from short-term change, we decomposed each lake’s monthly agriculture level into a lake-specific long-term mean (between-lake variation, AG_between) and monthly deviations from that mean (within-lake variation, AG_within), and applied the same decomposition to TSI. We then modelled the 12-month tipping hazard using cluster-robust logistic regression with these decomposed agriculture and trophic-state terms, EWS, and additional covariates.

Within-lake agricultural deviations predicted eutrophic tipping risk (odds ratio (OR) 1.157, 95% confidence interval (CI) 1.034–1.296, p = 0.011): a one-standard-deviation increase in within-lake cropland deviation was associated with a 16% increase in 12-month tipping probability in the 5 km model (Table 1). Baseline differences in agriculture showed no detectable association at 5 km (AG_between OR 0.903, 95% CI 0.651–1.252, p = 0.54). Baseline trophic state was the dominant predictor: a one-standard-deviation increase in TSI_between corresponded to a 51% rise in tipping probability (OR 1.513, 95% CI 1.285–1.781, p < 0.001), while TSI_within showed a strong protective effect (OR 0.276, 95% CI 0.225–0.338, p < 0.0001), indicating that higher within-lake TSI deviations were associated with lower odds of subsequent eutrophic tipping. Re-estimating the model with a 3 km buffer strengthened this eutrophication signal: AG_within increased to OR 1.363 (95% CI 1.156–1.607, p < 0.001), with AG_between remaining inverse but not statistically significant (OR 0.767, 95% CI 0.543–1.083, p = 0.13; Supplementary Fig. 1), while TSI effects changed little. Approximating distance from the shoreline using HydroLAKES surface area rather than the lake centroid yielded very similar within-lake agriculture odds ratios at every shore-distance buffer tested (Supplementary Table 7).

At the 12-month horizon, the models showed strong discrimination and accurate calibration. For eutrophication, the global model including EWS achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.883 with a Brier score of 0.0094. The corresponding model without EWS had AUC 0.866 and Brier 0.0075 (Supplementary Figs. 2 and 3; full coefficients in Supplementary Table 12). Oligotrophication models exhibited slightly lower but still clear discrimination (AUC 0.811 with EWS, 0.802 without), with Brier scores of 0.0087 and 0.0066 on the same matched and unmatched samples, respectively. A nested model comparison on a common sample showed that agriculture added statistically detectable but modest predictive information beyond TSI for both directions (likelihood ratio LR  = 17.7, df = 2, p < 0.001 for eutrophic; LR = 15.9, df = 2, p < 0.001 for oligotrophic; Supplementary Table 6); the gain in discrimination was small (ΔAUC ≈ 0.003–0.005).

Agricultural land cover changed little over the two-decade study period. Within-lake variability accounted for less than 0.1% of total variance in surrounding cropland, indicating that most lakes maintained stable land cover while fluctuating slightly around fixed baselines.

Eutrophic signals were robust; oligotrophic ones were not

The agricultural effect did not interact strongly with the within-lake trophic state (interaction OR 0.944, 95% CI 0.790–1.128, p = 0.52). Directional asymmetry, however, was evident. For eutrophication, the within-agriculture effect was robust and strengthened at smaller buffers (OR 1.157 at 5 km and 1.363 at 3 km). For oligotrophication, by contrast, agricultural coefficients were smaller and sensitive to buffer size: weakly elevated risks at 5 km (e.g., AG_within OR 1.120, 95% CI 1.041–1.206, p = 0.003) shrank and lost significance at 3 km in both the EWS and non-EWS models.

Regional and catchment exposure sensitivity

Continental stratification suggested that the agricultural association was not uniform across regions (Supplementary Table 11). The eutrophic within-agriculture coefficient was positive in Europe (OR 1.376, 95% CI 1.162–1.630, p < 0.001) and especially large in Oceania (OR 8.689, 95% CI 6.107–12.363, p < 0.0001), where the signal was driven primarily by New Zealand lakes with intensive pastoral catchments. The eutrophic within-agriculture coefficient was not distinguishable from unity in North America (OR 1.027, 95% CI 0.928–1.136, p = 0.61). The North American oligotrophic model did not converge, and South America had no detected tipping events, so neither is reported; we instead presented continental cases as illustrative trajectories in the Discussion (Fig. 3). We treated all continental models as exploratory, since event counts, cropland ranges, and covariance structures differed substantially across regions.

Fig. 3. Model-based risk surface and representative lakes from different continents.

a Predicted 12-month relative tipping risk as a function of baseline Trophic State Index (TSI; 12-month mean) and within-lake agriculture deviation (in %), derived from our logistic regression model (2003–2022). The colour scale shows the increased probability as a multiplier. Overlaid points indicate representative lakes from Europe (EU, teal), Oceania (OC, magenta), and North America (NA, orange) at their pre-event positions. Detrended TSI time series for representative lakes: b Europe, Braassemermeer (NLD); c Oceania, Te Waihora/Lake Ellesmere (NZL); d North America, Malheur Lake (USA). Black lines show deseasonalized TSI; red horizontal lines indicate the means of the changepoint segments immediately before and after each detected shift, not of the full pre- and post-event records; vertical dashed lines mark detected tipping points; blue dashed lines show the 12-month rolling variance on the secondary vertical axis. Three representative lakes were chosen from a total of 76 lakes with 125 tipping events.

For eutrophication, the within-agriculture association remained positive under alternative agriculture definitions, but statistical support varied: the cropland-only specification (primary) yielded OR 1.157 (p = 0.011); a lenient cropland-plus-grassland specification yielded OR 1.079 (p = 0.046); and a strict cropland-plus-permanent-pasture specification attenuated the coefficient toward unity (OR 1.026, p = 0.594; Supplementary Table 4). Across buffer radii, within-agriculture effects strengthened at 3 km, and oligotrophication estimates were smaller and sensitive to buffer choice (Supplementary Table 5). We therefore reported the 5 km cropland-only buffer as the primary specification and used alternative radii and definitions as sensitivity analyses (Supplementary Fig. 1, Supplementary Tables 4 and 7).

Early warning signals added limited forecasting value

We computed classical early warning signals (EWS) (lag-1 autocorrelation (AR(1)) trend, skewness, spectral reddening) in 12-month rolling windows on detrended TSI, using bootstrap resampling to quantify uncertainty in variance trends. Bootstrap-mean variance trends and confidence-interval widths separated the 125 tipping events from the 121 candidate regime shifts that did not pass the variance-trend criterion (Fig. 4); the rolling-variance signature was a defining feature of our tipping classification. However, these indicators provided limited predictive power in the models. None of the three classical EWS reached statistical significance once baseline TSI and agriculture were included: AR(1) trend (OR 1.052, 95% CI 0.883–1.254, p = 0.57), spectral reddening (OR 1.147, 95% CI 0.929–1.417, p = 0.20), and skewness (OR 0.992, 95% CI 0.831–1.184, p = 0.93). These results were robust to substituting Kendall’s τ for Spearman’s ρ when summarizing rolling-window EWS trends (Supplementary Table 9), and the eutrophic within-agriculture signal remained positive across abruptness thresholds (Supplementary Table 8). Substituting locally estimated scatterplot smoothing (LOESS) or seasonal-trend decomposition using LOESS (STL) for the per-month linear detrending reduced the count of detected events since flexible smoothers absorb step-like changes (Supplementary Table 10 and Supplementary Fig. 4); we retained per-month linear detrending as the primary method since the goal of the changepoint procedure is to detect abrupt, persistent mean shifts.

Fig. 4. Bootstrap uncertainty in pre-event variance trends distinguishes tipping points from regime shifts.

a Bootstrap-mean variance trend (ρ) per event for regime shifts that are not tipping points (n = 121, teal) and tipping points (n = 125, yellow). The variance trend measures the Spearman correlation between time and the rolling variance of detrended Trophic State Index (TSI) in the pre-event period, serving as an early warning signal (EWS) for tipping transitions. Violin plots show full bootstrap distributions with individual data points overlaid. Black lines indicate medians (thick solid) and interquartile ranges (dashed). b Bootstrap 95% confidence interval (CI) width for variance trend estimates, quantifying uncertainty in EWS detection. Tipping points are classified based on positive and significant variance trends (bootstrap mean ρ > 0 with 95% CI excluding zero), distinguishing them from regime shifts that lack this pre-event early warning signal.

Discussion

Our analyses reinforced the long-standing observation that eutrophic lakes are vulnerable to regime shifts. The within-between decomposition showed that the agricultural signal arose primarily from small deviations around largely stable long-run cropland baselines, consistent with a chronic-vulnerability context rather than large new cropland expansion. The agricultural effect was also directionally asymmetric: cropland exposure robustly predicted eutrophic tipping across buffer radii but only weakly and inconsistently predicted oligotrophic shifts. Once baseline trophic state and agriculture were included, classical EWS added only incremental discrimination (Supplementary Table 6), reinforcing concerns about their operational forecasting value in strongly forced systems. Together, these patterns supported a vulnerability-accumulation reading of lake tipping: chronic agricultural exposure had narrowed a lake’s safe operating space, and the proximate triggers of most detected events were more likely to have been hydrologic shocks, climate extremes, or internal feedbacks rather than further cropland expansion. This complemented rather than replaced the classical bifurcation framing: a control variable must still cross a stability threshold. However, the distance to that threshold is jointly determined by sustained exposure and current trophic state, so that lakes can hover near critical points for years before a relatively small perturbation precipitates a shift.

Across buffer radii, the most robust pattern was that within-lake deviations in cropland around a largely fixed baseline predicted eutrophic transitions: modest excursions above a lake’s long-term exposure increased near-term risk. Baseline differences in cropland coverage, by contrast, had at most a modest inverse association once we conditioned on trophic state and within-agriculture deviations, and we did not interpret this as evidence that agriculture was intrinsically protective. Instead, it reflected statistical adjustment for cross-lake differences in vulnerability that baseline TSI and other covariates already captured. The buffer sensitivity and small magnitude of oligotrophication coefficients further supported viewing agriculture primarily as a chronic vulnerability factor for eutrophic tipping, with only weak and non-robust associations with recovery. At the same time, the minimal within-lake variance in surrounding land cover and the rarity of visible pre-event expansion implied that our signal reflected sensitivity to fluctuations around stable exposure, rather than to large new land conversions. Importantly, stable cropland did not imply stable agricultural pressure: intensification within existing farmland could increase nutrient delivery without producing a detectable change in land cover fraction^13,25. Our land cover metric therefore captured the spatial footprint of agriculture but not temporal changes in its management intensity. This picture aligned with nutrient-loading theory and evidence for legacy phosphorus and internal feedbacks that reduce resilience once catchments have been intensively farmed^9,11,19,29.

Baseline trophic state exerted a much stronger influence on tipping risk than agriculture alone, and its effect was broadly consistent across agricultural gradients. Lakes that were already more eutrophic were more likely to tip, regardless of surrounding cropland, indicating that vulnerability accumulated through both external loading and internal processes such as sediment phosphorus release and vegetation loss. The absence of a strong interaction between agriculture and TSI suggested that cropland elevated risk across the trophic spectrum rather than only within particular regimes. Directional asymmetry in the agricultural effect was compatible with agriculture’s role as a nutrient source: additional loading and hydrologic connectivity more readily push systems toward turbid states than they help them recover. Recovery from eutrophic regimes often depends on deliberate reductions in nutrient export, hydrologic flushing, and internal loading controls rather than on marginal land cover changes alone³⁰.

We use the term tipping point throughout in its operational sense (a regime shift accompanied by a significant pre-event variance increase), as evidence consistent with critical slowing down rather than direct observation of the underlying feedback. Tipping points are infrequent, but they are not benign. Abrupt eutrophication collapses ecosystem services on short timescales, and internal feedbacks can lock lakes into bloom-dominated, hypoxic regimes through hysteresis and legacy loading, narrowing restoration options for decades^9,31. This phenomenon explains why preventing thresholds matters even when most trajectories are gradual: a small subset of systems accounts for a disproportionate share of long-run ecological and economic loss.

Oligotrophication events required a different interpretation. They constituted roughly half of our detected tipping events but were mechanistically heterogeneous. Some likely represented genuine recovery transitions following nutrient reduction, hydrologic flushing, or restoration interventions; others may have reflected changes in optical water quality, monitoring composition, or transient declines in primary production. Their associations with surrounding cropland were weaker and more buffer-sensitive than those for eutrophic transitions (Supplementary Tables 5 and 12). We therefore did not interpret them as evidence for a simple agriculture-driven recovery mechanism, but rather as candidate recovery or water-clarity transitions that warranted lake-specific follow-up.

Geographic heterogeneity in the agricultural effect showed that tipping risk depended on regional context rather than a single global dose-response, and the representative lakes in Fig. 3 illustrate how chronic vulnerability and short-term triggers coexist. In Te Waihora/Lake Ellesmere, elevated pastoral agriculture has produced one of New Zealand’s most eutrophic coastal lakes, with macrophyte loss, high external loads, and internal recycling maintaining a turbid, phytoplankton-dominated regime despite occasional clear phases associated with openings to the sea^32–34. North-west European lowland lakes such as Braassemermeer show long histories of high phosphorus and recurrent cyanobacterial scums despite regional nutrient-management efforts and successful mixing interventions in nearby systems³⁵. Malheur Lake in Oregon, an extremely shallow wildlife refuge lake, has experienced repeated high-water and drought cycles, carp expansion, and strong external nutrient and sediment inputs, with internal resuspension now dominating turbidity and nutrient dynamics and maintaining a shallow, turbid, vegetation-poor state despite limited recent change in mapped cropland³⁶. In all three cases, agricultural deviations around each lake’s long-run baseline were small relative to absolute exposure. Despite this, we still detected pronounced eutrophic tipping events, supporting our inference that lakes under chronic agricultural pressure can hover near thresholds for years or decades, with hydrologic shocks, extreme events, or management interventions providing the push into a new regime.

The morphometric results also clarified the likely mechanisms represented by our detected events. Although hypolimnetic anoxia and sediment phosphorus release are often emphasized in deep, stratified lakes, shallow lakes can also experience recurrent diel or episodic anoxia, including during temporary stratification, which can promote internal phosphorus release; climate warming is expected to increase the frequency and persistence of oxygen depletion^11,37–39. The concentration of detected events in shallow systems therefore pointed to a suite of non-exclusive feedbacks involving oxygen depletion and internal phosphorus release alongside wind-driven sediment resuspension, loss of macrophyte stabilization, and increased turbidity. Given that wind fetch, bathymetry, and dissolved-oxygen dynamics were not fully resolved globally, we treated these mechanisms as plausible interpretations supported by the observed depth pattern rather than as direct process attribution.

The combination of chronic pressure and stochastic shocks also appears in other complex systems. Coastal dead zones emerge where chronic nutrient enrichment promotes episodic hypoxia during periods of strong stratification and favourable weather anomalies⁴⁰. Coral reefs exposed to sustained fishing pressure and nutrient loading can persist close to degraded thresholds until heatwaves or storms precipitate rapid shifts to algal-dominated states⁴¹. At larger scales, climate tipping elements such as ice sheets and overturning circulations respond to the combination of multi-decadal forcing trajectories and variability in external drivers^10,42. The within-between decomposition and variance-based classification in this study provided a blueprint for quantifying how chronic exposure and transient disturbances jointly influence tipping risk whenever long-term monitoring of state and drivers is available.

The limited predictive value of classical EWS in our models reinforced concerns about their operational use in strongly forced, seasonal systems^27,43,44. Although rising variance was ubiquitous before the events we identified, this signature was also common among lakes that did not tip within the forecasting horizon. Tipping mechanisms differ across lakes, monitoring is often sparse and incomplete, and external shocks such as floods, heatwaves, and wildfire runoff can overwhelm endogenous slowing-down dynamics. Lake-specific spatial EWS that exploit within-lake spatial heterogeneity have shown more promise than purely temporal indicators on a single state variable⁴⁵, but they require denser within-lake sampling than is available in our globally pooled monthly TSI data. Complementary detection frameworks such as threshold generalized additive models (TGAMs) can characterize gradual or nonlinear regime shifts that our changepoint-based pipeline targets less directly²⁷; combining abrupt and gradual detectors on the same lake data is a useful direction for future work. These limitations do not preclude the use of EWS for classifying event types or studying individual trajectories. However, they suggested that driver-based risk models are more promising for operational inference at management horizons.

Several limitations qualified our causal interpretation. The within-between decomposition reduced, but did not eliminate, confounding from time-invariant cross-lake differences; it could not account for unmeasured time-varying factors that covary with both agricultural fluctuations and tipping, such as fertilizer application rates, manure management, conservation programmes, hydrologic extremes, or episodic runoff. Our mechanism was inferred rather than directly observed: we assumed agriculture elevated risk through nutrient loading, but we did not measure phosphorus and nitrogen fluxes for every lake. Mixing remote-sensing and in situ trophic-state estimates introduced measurement heterogeneity; satellite-derived TSI may also reflect chlorophyll a, suspended sediments, or coloured dissolved organic matter rather than chlorophyll alone, so we did not interpret rolling-window spectral changes as ground-truthed mechanistic signatures. Fixed-radius buffers approximated but did not replicate hydrologic connectivity, and shore-distance corrections (Supplementary Table 7) only partially mitigated this; catchment-level exposure modelling using HydroSHEDS or HydroBASINS is a priority for future work. For these reasons, the metrics we report should be interpreted as statistical regularities rather than as direct causal elasticities.

Despite these limitations, the results have clear implications for managing safe operating spaces for lakes. Preventing abrupt eutrophication requires reducing agricultural footprints and nutrient exports in catchments that are already moderately productive, rather than merely halting further expansion. Oligotrophic lakes with growing agricultural encroachment require particular protection: they provide valuable ecosystem services yet appear vulnerable to shifts once chronic exposure and baseline trophic state cross intermediate levels. Since transient disturbances likely govern tipping in chronically stressed systems, resilience strategies should pair land use controls with measures that dampen shocks, such as buffer strips, wetland restoration, and floodplain reconnection, which modulate hydrologic extremes. Embedding driver-based tipping-risk models into lake management and land use planning would allow regulators to prioritize vulnerable systems, evaluate trade-offs in agricultural development (including livestock expansion), and track whether interventions are pushing lakes back toward safer regions of their state space⁴⁶. More broadly, applying within-between decompositions and variance-based tipping diagnostics to other social-ecological systems could help identify where chronic pressure has already eroded resilience and where relatively small shocks are most likely to trigger costly regime shifts. Future work linking explicit policy outcomes or management interventions to lake responses would provide stronger causal leverage.

Methods

Water quality data

We assembled a global monthly dataset of lakes and reservoirs from 2003 to 2022 by integrating international and national monitoring archives from multiple primary sources. Specifically, the Global Environment Monitoring System for Freshwater (GEMS/Water) GEMStat database maintained by UNEP⁴⁷, which provides in situ measurements of total and dissolved nutrients, chlorophyll a, Secchi depth, and other water quality parameters from monitoring stations worldwide, trophic state index (TSI) time series derived from remote sensing data, obtained from the compilation published by Gilarranz et al.²⁶, which provides monthly TSI estimates based on satellite-derived water quality indicators, the US Water Quality Portal jointly maintained by the USGS, USEPA, and the National Water Quality Monitoring Council⁴⁸, providing access to water quality data from monitoring programmes across the United States, the LAWA (Land, Air, Water, Aotearoa) database for New Zealand lakes⁴⁹, which provides monitoring data for chlorophyll a, Secchi depth, total nitrogen, and total phosphorus from regional and national monitoring networks, and, finally, Alberta water quality measurements⁵⁰. Monitoring data from all sources were standardized, including lake identifier, sampling date, coordinates, and covariates of interest. All concentrations were converted to consistent SI units. We applied broad but conservative range checks by analyte to remove clearly implausible values (negative nutrient concentrations and Secchi depths exceeding the maximum depths reported for the lake). For lakes with multiple samples within a calendar month, we aggregated to the lake-month level using the median, which is robust to occasional extreme values and heterogeneous within-month sampling intensity. Months lacking any valid measurement for trophic state variables were treated as missing.

Trophic state index (TSI) and inclusion criteria

For each lake, we either used a reported trophic state index (TSI) where provided or computed Carlson-style TSI components⁵¹ from in situ Secchi depth, chlorophyll a, and total phosphorus:

$$\mathrm{TSI(TP)}=14.42ln\left(\mathrm{TP}\right)+4.15,$$ 1

$$\mathrm{TSI(CHLA)}=9.81ln\left(\mathrm{CHLA}\right)+30.6,$$ 2

$$\mathrm{TSI(Secchi)}=60-14.41ln\left(\mathrm{Secchi}\right).$$ 3

Here, TP is the total phosphorus concentration (μg L⁻¹), CHLA is the chlorophyll a concentration (μg L⁻¹), and Secchi is the water transparency (m). When more than one component was available, we used the within-month lake-level median across all available components to define TSI; thus, the index was determined by the components that were most reliably observed. We flagged whether TSI was directly reported (from lake-specific monitoring schemes) or derived from these components, and whether the source was in situ or remote-sensing-based. Lakes with fewer than 24 months with a valid TSI over 2003–2022 were excluded from changepoint detection and subsequent modelling.

Land cover data and agricultural exposure

Annual land cover was derived from the European Space Agency (ESA) Climate Change Initiative Land Cover (CCI-LC) product at 300 m spatial resolution (version 2.1), which provides globally consistent land cover maps for 1992 to the present day^52,53. We reclassified CCI-LC legend classes into ecologically coherent categories, including open water, urban/built, forest, shrubland, wetland, and agriculture. Agriculture in the primary analysis comprised rainfed cropland, irrigated cropland, and mixed cropland mosaics. Pasture and grassland-dominated classes were excluded from the primary cropland metric because their nutrient-export pathways and classification uncertainty differed from cropland in the global land cover product. We evaluated pasture-inclusive definitions (cropland + permanent pasture and cropland + all grassland mosaics) as sensitivity analyses (Supplementary Table 4). The eutrophic within-agriculture signal remained positive across pasture-inclusive definitions but attenuated toward unity when permanent pasture was added strictly (OR 1.026, p = 0.594); we therefore retained cropland-only as the primary specification.

For each lake and year, we constructed circular buffers with radii of 0.5, 1, 3, 5, and 10 km around the lake centroid and computed the fraction of each reclassified land cover class within each buffer. These fractions were calculated at the pixel level on the native CCI-LC grid and normalized by the total buffer area intersecting land. Based on predictive performance, we selected the 5 km buffer as the primary agricultural exposure metric, denoted AG_i,t (the fraction of cropland within 5 km of lake i in year t). Other buffer radii were used in sensitivity analyses but were not central to the models reported here (Supplementary Fig. 1); eutrophication results were qualitatively similar at 3 km and 10 km, with larger within-agriculture effects at 3 km, whereas oligotrophication coefficients were smaller and not robust to buffer choice.

Monthly time series construction and detrending

For each lake, we defined a complete monthly time series from the first to the last year where at least one TSI observation was recorded. Let y_i,t denote the raw monthly TSI for lake i at month index t, where t increases by one for each calendar month. Months without any TSI observations were initially coded as missing. We required at least 24 months with a valid TSI, and lakes with <80% observed months were excluded from the analysis.

To remove seasonality while preserving abrupt mean shifts, we applied a monthly linear detrending to the raw TSI time series. For each lake and calendar month, we regressed TSI on a local time index within that month,

$${\mathrm{TSI}}_{i,t}={\alpha }_{i,m}+{\beta }_{i,m}{\tilde{t}}_{i,t}+{\epsilon }_{i,t},$$ 4

where ${\tilde{t}}_{i,t}$ is the centered month index restricted to observations with calendar month m, and ε_i,t is a residual. The detrended series was defined as the residuals ${\widehat{\epsilon }}_{i,t}$, reassembled in chronological order for each lake. This per-month linear detrending followed standard practice in early warning signal (EWS) analysis and changepoint detection, removing smooth seasonal and low-frequency trends without dampening sharp shifts in mean state^54,55.

Per-month linear detrending was chosen as the primary method since the segmentation procedure that follows is designed to detect abrupt, persistent mean shifts; flexible smoothers (LOESS, STL) absorb step-like changes into their fitted trend and reduce detected event counts (Supplementary Table 10, Supplementary Fig. 4).

Gap filling and time series reconstruction

For changepoint detection, we reconstructed an equally spaced monthly residual series for each lake using a moving-window, residual-resampling bootstrap. For a lake with detrended residual series ${\widehat{\epsilon }}_{i,t}$ and a missing value at month t_m, we defined a symmetric window of width 13 months centered on t_m, and restricted to the available residuals within this window. For each missing month, we generated B = 500 bootstrap replicates by resampling, with replacement, from the observed residuals in that local window. We set the imputed residual to the bootstrap mean across B replicates for that month. All bootstrap resampling was performed independently for each lake, using a lake-specific random seed. We excluded lakes where more than 20% of months within its observation span were gap-filled in this way. The resulting bootstrap-mean residual series for each lake, ${\tilde{y}}_{i,t}$, was a fully populated, detrended monthly TSI series on which all subsequent changepoint detection was performed. To quantify sampling intensity, we tracked whether TSI in each month was directly observed or interpolated: months with at least one original TSI observation were assigned sampling weight equal to 1, and purely interpolated months were assigned weight 0.

Changepoint detection and regime shift criteria

We use changepoint, regime shift, and tipping point as nested operational categories. A changepoint is the statistical output of the segmentation model. A regime shift is a changepoint that passes additional filters for ecological magnitude (abruptness a > 1), statistical significance (Welch’s t test p < 0.05), and persistence (at least two-thirds of the first six post-shift observed months remain on the new side of the pre-shift mean). A tipping point is a regime shift with a significant positive pre-event variance trend, used here as evidence consistent with critical slowing down^43,44,54,55. We treated abrupt trophic changes as mean shifts in the deseasonalized TSI residuals. For each lake, we applied a dynamic-programming segmentation algorithm based on normal mean shifts that minimizes the within-segment sum of squared errors (SSE) across all candidate partitions⁵⁶. We considered models with n = 0, 1, 2, 3 changepoints and selected the optimal n using the Bayesian Information Criterion (BIC)⁵⁷, enforcing a minimum segment length of 12 months and discarding candidate changepoints within the first or last 24 months of each lake’s time series to reduce edge artifacts.

For each detected changepoint at time t, we defined adjacent pre- and post-changepoint windows as the contiguous segments immediately before and after t, subject to the 12-month minimum length. Let ${\mu }_{\mathrm{before}}$, ${\sigma }_{\mathrm{before}}$ and μ_after, σ_after denote the mean and standard deviation of the raw TSI in these windows. We computed an abruptness signal-to-noise ratio,

$$a=\frac{∣{\mu }_{\mathrm{after}}-{\mu }_{\mathrm{before}}∣}{0.5\left({\sigma }_{\mathrm{before}}+{\sigma }_{\mathrm{after}}\right)}.$$ 5

We adopted a > 1 as the primary cutoff since it requires the absolute mean shift to exceed one pooled pre/post standard deviation of background variability, providing a scale-invariant ecological signal-to-noise criterion that is conservative relative to thresholds of a > 0.5 used elsewhere; sensitivity to a ∈ {0.5, 0.75, 1.25, 1.5} is reported in Supplementary Table 8. We then compared the pre- and post-changepoint means using Welch’s unequal-variance t-test⁵⁸ and required p < 0.05 for the mean shift to be considered statistically significant. Changepoints were detected on deseasonalized residuals to avoid identifying seasonal structure as a state change. Abruptness was then computed using raw TSI since the ecological magnitude of a shift is most interpretable in original trophic-state units. The pooled pre/post standard deviation scales this raw shift relative to local background variability.

Finally, we imposed a persistence criterion: among the first six months following a changepoint, at least two-thirds of observed TSI values must remain on the new side of the pre-changepoint mean. Changepoints that passed all criteria were labelled candidate regime shifts and direction-coded by the sign of ΔTSI. Positive ΔTSI denotes an upward (eutrophic) shift in trophic state; negative ΔTSI denotes a downward (oligotrophic) shift.

Variance-based early warning classification

For each candidate regime shift, we quantified uncertainty in EWS classification by bootstrapping variance trends before the changepoint. Working on the detrended residuals, we extracted a pre-changepoint segment (m months). For each of B = 500 bootstrap replicates, we resampled the pre-changepoint residual sequence (with replacement, preserving length m), then computed a centered 12-month rolling variance, requiring at least 10 non-missing months per window. We summarized the temporal trend in this rolling variance using Spearman’s rank correlation coefficient ρ between time and the rolling variance⁵⁵. Across the replicates, we obtained a bootstrap distribution of ρ values. The choice of monotonic rank-correlation summary (Spearman’s ρ vs Kendall’s τ) affected only the trend metric applied to each rolling window, not the underlying variance, AR(1), skewness, or spectral series; under either choice, the tipping classification and downstream odds ratios agreed to two decimal places (Supplementary Table 9).

A candidate regime shift was classified as a tipping point if it was a regime shift, the bootstrap mean of the variance trend was strictly positive, and the 95% bootstrap percentile interval for ρ excluded zero, indicating a robust increase in variance leading up to the event. This variance-based diagnostic was meant to capture critical slowing down before a tipping point in noisy, forced systems^42,54,55. We adopted variance, rather than AR(1), as the defining EWS since variance trends have consistently shown high power in simulation and experimental studies of critical transitions^42,55. AR(1), skewness, and spectral reddening were therefore computed as complementary indicators and used as covariates in the hazard models, but not in the tipping classification rule itself.

Early warning signals as model covariates

We recomputed EWS, with the exception of variance due to its use in tipping point criteria, on the filled, detrended TSI series for use as predictors in the models^54,55. For each lake, we defined an integer month index k over its filled series and computed: Autocorrelation: a 12-month rolling estimate of lag-1 AR(1), computed on detrended TSI within each window, requiring at least 10 points. Skewness: a 12-month rolling sample skewness (unbiased) of detrended TSI, requiring at least 10 points. Spectral reddening: a 24-month rolling estimate of spectral reddening, defined as the log-ratio between average low-frequency and high-frequency power in the discrete Fourier spectrum of the deseasonalized TSI segment, requiring at least 16 usable points.

Given that the satellite-derived TSI series in our aggregate dataset could be modulated by chlorophyll a, suspended sediments, and coloured dissolved organic matter, rolling-window spectral reddening was treated as a covariate rather than as a ground-truthed mechanistic signature; lake-specific validation would be needed to attribute spectral changes to a particular optical pathway.

For each month t, we then summarized the trend in these rolling indicators over the preceding 12 months (or 24 months for spectral reddening) using Spearman’s rank coefficient between the time index k and the indicator values in that window. These Spearman coefficients form month-specific early warning trends: EWS_ar1, EWS_skew, and EWS_spec. To ensure temporal precedence, all early warning predictors were lagged by one month before entering the models, so that predictors at month t were used to explain events that may occur from t + 1 onward. This construction allowed us to ask, in a forward-looking hazard framework, whether classical early warning trends improve prediction once drivers and baseline trophic state are already included.

Agricultural exposure decomposition

We linked the annual agricultural fraction within 5 km, AG_i,t, to the water quality measurements by merging on lake identifier and year. Since we were interested separately in cross-lake differences and within-lake temporal changes in agricultural exposure, we decomposed AG_i,t into between- and within-lake components. For each lake i, we computed the lake-specific mean

$${\overline{\mathrm{AG}}}_i=\frac{1}{T_i}{\sum }_t{\mathrm{AG}}_{i,t},$$ 6

where T_i is the number of observed years for lake i. We then defined the between-lake component as ${\mathrm{AG}}_{\text{between},i}={\overline{\mathrm{AG}}}_i$ and the within-lake deviation as

$${\mathrm{AG}}_{\mathrm{within},it}={\mathrm{AG}}_{i,t}-{\overline{\mathrm{AG}}}_i.$$ 7

We applied an analogous within-between decomposition to the 12-month rolling mean of TSI. This decomposition separates long-term differences in trophic state among lakes from short-term deviations within lakes. Then, to investigate interaction terms and facilitate interpretation, we z-scored predictors within each lake. Specifically, for each lake i, we defined

$${\mathrm{AG}}_{\mathrm{within}\_\mathrm{z},it}=\frac{{\mathrm{AG}}_{\text{within},it}-{\overline{\mathrm{AG}}}_{\text{within},i}}{s_{\mathrm{AG},i}},$$ 8

with s_AG,i the lake-specific standard deviation of AG_within,it. We applied the same transformation to TSI_within,it and EWS_ar1,it. Interaction terms were then formed as

$${\mathrm{AG}\_\mathrm{x}\_\mathrm{EWS}}_{it}={\mathrm{AG}}_{\mathrm{within}\_\mathrm{z},it}{\mathrm{EWS}}_{\mathrm{ar1}\_\mathrm{z},it},$$ 9

$${\mathrm{AG}\_\mathrm{x}\_\mathrm{TSI}}_{it}={\mathrm{AG}}_{\mathrm{within}\_\mathrm{z},it}{\mathrm{TSI}}_{\mathrm{within}\_\mathrm{z},it}.$$ 10

Event definition for modelling

We defined candidate eutrophication (upward) and oligotrophication (downward) events as the set of tipping points identified by the changepoint and variance-based early warning criteria above, with direction given by the sign of ΔTSI. Let Ω_up and Ω_down denote the sets of upward and downward events described by their lake and month indices (i, t^*). For each month t and lake i on the filled panel, we constructed a binary outcome

$$\begin{array}{c}\hfill y_{it}^{\left(H\right)}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}\mathrm{at}\mathrm{least}\mathrm{one}\mathrm{event}\mathrm{occurs}\mathrm{in}\left(t,t+H\right];\hfill \\ 0,\hfill & \text{otherwise},\hfill \end{array}\right)\end{array}$$ 11

with H = 12 months as the primary prediction horizon. We excluded the H months immediately following each event to avoid double-counting and reduce dependence among closely spaced events.

For each direction (upward and downward), we then generated two separate outcome variables, $y_{it,\text{up}}^{\left(12\right)}$ and $y_{it,\text{down}}^{\left(12\right)}$, and associated risk sets that excluded months in the cool-down windows. EWS and covariates were constructed so that they temporally preceded the hazard period: predictors at month t were used to define the hazard of an event in (t, t + 12].

Generalized linear models

We modelled the near-term tipping probability via discrete-time logistic hazard models. For lake i in month t, let z_it be a vector of the aforementioned time-varying covariates. The primary specification decomposes agricultural exposure into within-lake deviations and between-lake baselines:

$$\mathrm{logit}\left\{P\left(y_{it}^{\left(H\right)}=1\right)\right\}={\beta }_0+{\beta }_1{\mathrm{AG}}_{\mathrm{within},it}+{\beta }_2{\mathrm{AG}}_{\mathrm{between},i}+{\mathbf{\gamma }}^{\top }{\mathbf{z}}_{it},$$ 12

where logit is defined by $\mathrm{logit}\left(p\right)=\log \left\{p/\left(1-p\right)\right\}$. The covariate vector z_it in the primary models comprised TSI_between,i, TSI_within,it, the 12-month TSI variance, and lagged EWS. Prespecified variants added interaction terms and explored specifications with and without early-warning predictors (Supplementary Figs. 2 and 3 and Supplementary Tables 6 and 12).

We estimated the models via maximum likelihood using generalized linear models (GLMs) with a binomial family and a logit link (statsmodels v0.14⁵⁹). All continuous predictors were z-scored globally before fitting. Inference was restricted to months in which the trophic state index was actually observed (sampling-intensity flag w_it = 1); months reconstructed by the gap-filling procedure (w_it = 0) were excluded from the analysis sample but retained in the panel structure used to construct rolling-window EWS and within-between decompositions. Standard errors and confidence intervals were computed using the Liang-Zeger cluster-robust covariance estimator with 12-month hazard blocks within each lake as clusters^60,61, matching the mechanical correlation structure of the discrete-time hazard panel: the binary outcome is identically 1 inside each 12-month hazard window, so adjacent lake-months within a window are not independent draws from the binomial likelihood.

Before model fitting, we pruned predictors exhibiting near-zero variance or extreme collinearity. Within each model, predictors with pairwise absolute correlation ∣r∣ > 0.95 were iteratively pruned, and we enforced a full-rank design matrix by dropping the most collinear predictor where necessary. We then computed variance inflation factors (VIFs) and retained only predictors with VIF  < 5 ⁶². Further, we fitted separate models for eutrophication and oligotrophication (up and down events). To explore geographic heterogeneity, we mapped lake identifiers to ISO3 country codes and then to continents using a standard geopolitical grouping. We fitted continental versions of the same models for any continent with at least 100 observations and at least one detected tipping event in the relevant direction, and report only those fits that converged with finite confidence intervals.

Model performance and diagnostics

For each fitted model, we reported the number of lakes, the total number of data observations, and the number of events. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC)⁶³ and Brier score for probabilistic accuracy⁶⁴. Calibration was examined by plotting predicted versus observed event probabilities in deciles of predicted risk. To check sensitivity to early warning predictors, we compared models with and without EWS_ar1, EWS_skew, and EWS_spec, and models with interaction terms versus main effects only (Supplementary Figs. 2 and 3, Supplementary Table 6). Multicollinearity diagnostics (pairwise correlations and VIFs) were inspected to confirm that all retained predictors met the prespecified thresholds.

Statistical reporting

Statistical significance was defined as p < 0.05, and unless otherwise noted, statistical tests were two-sided. Effect sizes were reported as odds ratios with 95% confidence intervals derived from the cluster-robust covariance estimator described above. All analyses were conducted in Python (version 3.11) using pandas⁶⁵, NumPy⁶⁶, SciPy⁶⁷, and statsmodels⁵⁹, and custom code for changepoint detection, bootstrapping, and early warning calculations.

Author contributions

J.S.: Conceptualization; Methodology; Software; Validation; Formal analysis; Investigation; Data curation; Writing - original draft; Writing - review & editing; Visualization. H.W.: Conceptualization; Methodology; Validation; Supervision; Writing - review & editing; Funding acquisition; Project administration.

Peer review

Peer review information

Nature Communications thanks Mariana Meerhoff, Joydev Chattopadhyay and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This research was mainly supported by a research grant from the Alberta Conservation Association. H.W. was partially supported by the Natural Sciences and Engineering Research Council of Canada (Individual Discovery Grant RGPIN-2025-05734) and a Canada Research Chair.

Data availability

The data generated and analyzed in this study, including source data for all main-text and Supplementary display items, are available in Zenodo under https://doi.org/10.5281/zenodo.20076876⁶⁸.

Code availability

All analysis and modelling code is available in the same Zenodo record⁶⁸.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77445-9.