Urban point sources of nutrients were the leading cause for the historical spread of hypoxia across European lakes

Significance

Using a compilation of data arising from over 1,500 European watersheds, we have identified the relative role of different drivers in initiating hypolimnetic hypoxia, a critical indicator of lake health. In particular, our regional synthesis of laminated lake sediments indicated a significant acceleration in the spread of lacustrine hypoxia in the 1900s, which occurred well before the general use of commercial fertilizers in the mid-20th century and the onset of supraregional climate warming in the 1970s. The spread of hypoxia was best explained by urban expansion and the associated intensification of anthropogenic point sources of phosphorus, whereby changes in lifestyle increased the discharge of nutrients from treated and raw sewage, and ultimately led to enhanced lacustrine biological productivity.

Abstract

Enhanced phosphorus (P) export from land into streams and lakes is a primary factor driving the expansion of deep-water hypoxia in lakes during the Anthropocene. However, the interplay of regional scale environmental stressors and the lack of long-term instrumental data often impede analyses attempting to associate changes in land cover with downstream aquatic responses. Herein, we performed a synthesis of data that link paleolimnological reconstructions of lake bottom-water oxygenation to changes in land cover/use and climate over the past 300 years to evaluate whether the spread of hypoxia in European lakes was primarily associated with enhanced P exports from growing urbanization, intensified agriculture, or climatic change. We showed that hypoxia started spreading in European lakes around CE 1850 and was greatly accelerated after CE 1900. Socioeconomic changes in Europe beginning in CE 1850 resulted in widespread urbanization, as well as a larger and more intensively cultivated surface area. However, our analysis of temporal trends demonstrated that the onset and intensification of lacustrine hypoxia were more strongly related to the growth of urban areas than to changes in agricultural areas and the application of fertilizers. These results suggest that anthropogenically triggered hypoxia in European lakes was primarily caused by enhanced P discharges from urban point sources. To date, there have been no signs of sustained recovery of bottom-water oxygenation in lakes following the enactment of European water legislation in the 1970s to 1980s, and the subsequent decrease in domestic P consumption.

Changes in land cover and land use have been identified as important drivers of phosphorus (P) transfers from terrestrial to aquatic systems, resulting in significant impacts on water resources (1–3). In post-World War II Europe, changes in land cover, land use, and P utilization caused widespread eutrophication of freshwaters (3). Elevated rates of P release from point sources to surface water bodies increased in step with population increases, with the novel use of P in domestic detergents and with enhanced connectivity of households to sewage systems that generated concentrated effluents (4). The intensification of agriculture and drastic increased use of fertilizers from industrial and manure sources resulted in elevated P concentrations in runoff from diffuse sources (4). These trends have now metastasized from Europe and North America to most nations, which explains the almost global development of eutrophication problems in surface waters (1).

Much of our understanding regarding the interactions between changes in land cover/use, climate, and lake eutrophication comes from detailed studies of individual lakes (5), modeling exercises (1), and/or regional-scale syntheses of instrumental data (6, 7); these studies are largely based on relatively short time series (8). Depending on the multitudinous local differences in catchment and lake morphology, river transport capacity, climate, geology, and regional trajectories in socioeconomic development, the responses of lakes to surrounding land changes can differ greatly in intensity, modalities, and kinetics (9–12). Multiple sites need to be investigated to quantify a regional trend, as well as to evaluate local to regional heterogeneities. Only a few studies have interpreted the long-term trajectories of lakes (based on >100-year lake records) in terms of eutrophication on a regional scale by analyzing trends in nutrient and dissolved CO₂ concentrations (13, 14), carbon burial rates (15), cyanobacterial dominance (16), and hypoxia development (17). However, only one of these studies (13) considered the temporal dynamics of land cover and use, and only a few studies (16, 17) considered modern land cover. Our current lack of knowledge of the effects arising from cumulative environmental pressures presents the potential for serious underestimation of the long-term impacts of land use changes and hinders our ability to identify the relative importance of P sources to lake ecosystems (18).

Recent progress in land use science has provided an insightful large-scale perspective spanning centuries to millennia (19–22). Additionally, European high-resolution datasets (23, 24) allow for investigations to be conducted at the scale of individual lake watersheds. The present study relies on existing datasets of changes in land cover at the watershed scale [Historic Land Dynamics Assessment model (HILDA v2; ref. 24)], climate data [University of Delaware Air Temperature & Precipitation (UDEL model; ref. 25)], and a database on the historical onset of hypoxia in lakes (17) to (i) reconstruct the European dynamic of lacustrine hypoxia during the Anthropocene, and (ii) decipher whether P from diffuse sources (agriculture) or point sources (urbanization) is responsible for the spread of lacustrine hypoxia in Europe.

One widely studied response of lakes to eutrophication is bottom-water hypoxia ([O₂] < 2 mg⋅L⁻¹). Bottom-water hypoxia in lakes is detrimental not only for the biota that would normally inhabit oxic aquatic and benthic environments but also facilitates biogeochemical reactions that generate methane and further mobilize pollutants from previously accumulated sediment, including P (26–28). Hypoxia can develop naturally, but is more often the result of (i) cultural eutrophication, which enhances biomass production and, ultimately, its decomposition through microbial oxygen respiration (29–31), and (ii) rising mean temperatures, which decrease oxygen solubility in water (32), stimulate microbial oxygen respiration (30), and/or strengthen thermal stratification (33, 34). Among these forcing mechanisms, recent paleolimnological studies identified excess P availability, and not climate, as the main driver for the onset of lacustrine hypoxia during the Anthropocene (17, 35). These studies used the presence and environmental signals of varved (i.e., annually laminated) sediments in lakes distributed in the French Alps and worldwide to assess the long-term dynamics of hypoxia. Indeed, hypoxic conditions are recorded in lake sediments by virtue of preserved laminations after crossing a critical threshold in bottom-water oxygenation that prevent macrobenthic bioturbation in the deeper parts of basins (35). The onset of sustained lamination (including varves) in modern lake sediments is an unambiguous and independent proxy for the timing of hypoxic, anoxic (i.e., complete absence of oxygen), or even euxinic (i.e., sulfidic) bottom-water conditions on a regional scale. The well-defined geochronology of lacustrine varves provides forensic evidence to quantify the timing, prevalence, and causes of aquatic regime shifts (17).

Additive mixed-effect models (AMMs) (36) were used to analyze temporal trends and to depict differences among groups of watersheds in Europe: (i) 51 watersheds with lakes recording recent hypoxia onset, (ii) 97 watersheds with lakes recording natural hypoxia, (iii) 769 benchmark watersheds extracted from the Lake Core Database (37), and (iv) 690 benchmark watersheds from the Global Lakes and Wetlands Database (GLWD) (38). Lakes of the GLWD have been selected randomly in Europe to represent various gradients of human pressure, climate conditions, land cover, and land use.

Results

Our sampling captured the wide ranges of lake morphometric properties, catchment sizes, modern human activities, and climatic conditions that are spread across Europe (Fig. 1A, Fig. S1, and Table S1). General trends in land cover change in Europe during the past 300 years corresponded to increases in the percentages of urban and cultivated areas, albeit some regions were more affected than others (Fig. 1 B and C).

Fig. 1.

Location of the 1,607 study sites and changes in land cover over the past 300 years (CE 1700–2000). (A) Fifty-one recently hypoxic lakes, 97 naturally hypoxic lakes, and 1,459 benchmark watersheds composed of 769 lakes from the Lake-Core Database and 690 randomly selected European lakes from the GLWD database. (B and C) Increases in cultivated areas (%) and urban areas (%) for the past 300 years were observed in all of the watersheds according to an M-K test, where a higher coefficient indicates a stronger increase (69).

Fig. S1.

Distribution of modern land cover, climate, and geomorphological properties in the studied European watersheds: cultivated area, forested area, pastured area, mean annual precipitation, mean annual air temperature, human population density, lake area, lake catchment size, and maximum lake water depth.

Table S1.

Descriptive statistics of lake and catchment properties

	Recently hypoxic lakes (n = 51)			Naturally hypoxic lakes (n = 97)			Benchmark, Lake Core Database (n = 769)			Benchmark, GLWD (n = 690)
Parameters	Mean	Min.	Max.	Mean	Min.	Max.	Mean	Min.	Max.	Mean	Min.	Max.
Altitude, m	511	25	2,445	494	21	2,826	438	0	3,028	317	1	2,550
Lake area, ha	60	0	582	7	0	290	25	0	5,650	21	0	1,856
Maximum depth, m	62	2	310	28	0	410	14	0	449	24	0	514
Catchment, km2	1,620	0	25,972	524	0	22,523	253	0	47,375	1,725	0	556,923
Precipitation, mm⋅a−1	821	471	1,577	819	406	2,002	1,022	458	2,755	742	420	1,853
Urban area, %	8	0	100	5	0	100	3	0	100	2	0	98
Cultivated area, %	12	0	100	9	0	100	10	0	100	15	0	100
Forested area, %	38	0	99	52	0	100	34	0	100	50	0	100
Pastured area, %	27	0	82	24	0	100	42	0	100	20	0	100
Other land, %	6	0	100	7	0	100	8	0	100	2	0	97
Inhabitant per km2	183	0	1,530	77	0	1,504	108	0	8,537	69	1	3,527

Mean, minimum (Min.), and maximum (Max.) are reported for the four lake categories of this study: recently hypoxic lakes, naturally hypoxic lakes, and two sets of benchmark lakes. Air temperatures are mean annual anomalies (°C).

Based on our analyses (Materials and Methods), we found that the fraction of lakes recording hypoxia in Europe increased over the past 300 years, from an initial annual rate of 0.06 ± 0.004% per year (a⁻¹) (Pearson’s test, p < 0.0001) between CE 1850 and 1900 to rates of 0.20 ± 0.01% a⁻¹ between CE 1900 and 2000 (p < 0.0001) (Fig. 2A). In total, we found that 51 lakes shifted to hypoxia during the past 300 years (Table S2). The catchments of these 51 lakes with recent hypoxia onset had higher percentages of both cultivated and urban areas in CE 2000 than the benchmark watersheds (Fig. S2). Furthermore, most of the lakes with recent hypoxia onset were low-elevation sites (48 of 51 were situated between sea level and 1,000 m above sea level). We also found that the patterns of historical change in land cover and land use for these 51 lakes were best described by nonlinear (i.e., additive mixed-effect) models; urbanized areas increased sharply at the end of the 19th century (from 0.02% in CE 1700 to 4.1% in CE 2000), whereas the proportion of cultivated lands has expanded more gradually since the early 18th century (from 7.8% in CE 1700 to 23.4% in CE 2000) and occurred well before the first spread of hypoxia (Fig. 2 B and C). More than half of the 51 lakes shifted to hypoxia before the introduction of fertilizers in Europe in the middle of the 20th century. Climate warming, as well as changes in precipitation, is also an unlikely primary driver for the onset of hypoxia because the main warming signal in the air-temperature record postdates the initial spread of hypoxia (Fig. 2 D and E).

Fig. 2.

Trends in the prevalence of lake hypoxia and urbanization, as well as observed climate change dynamics, during the past 300 years in Europe. (A) Spread of lacustrine bottom-water hypoxia shown as a cumulative number of lakes (blue curve) based the onset of varve deposition in lake sediments from the 51-lake subset and the human population in these watersheds (red dashed curve). Percentages of urban (B) and agriculturally cultivated (C) areas in watersheds of the 51 lakes that shifted to hypoxia. In B and C, temporal trends and 95% confidence intervals were calculated according to centennial land use data and an AMM. The black arrow in B indicates early European water nutrient mitigation legislation in the 1970s and 1980s (70). The dark gray and green-shaded peaks in C indicate the respective nitrogen (N) and phosphate (P₂O₅) fertilizer applications in the European Union since the 1950s (71). European trends in air temperature (D) and April/May/June (AMJ) precipitation (E) are reconstructed from tree rings (72).

Table S2.

Inventory of sites recording the onset of hypoxia

Site	Country	Latitude, DD	Longitude, DD	Facies top section	Hypoxia onset date, CE	Altitude, m	Lake area, km2	Catchment area, km2	Maximum depth, m	Source	DOI
Alserio	Italy	45.78703	9.21512	Varves	1967	260	1.23	18.3	8.1	(73)	10.1007/s10933-005-6786-2
Ammersee	Germany	48.06470	11.12416	Varves	1958	533	46.60	993.0	81.1	(77)	10.1029/2009WR008360
Annecy	France	45.89803	6.13572	Varves	1952	447	27.60	251.0	41.5	(35)	10.1002/2014GB004932
Arendsee	Germany	52.89035	11.47772	Varves	1965	22.8	5.14	29.8	48.7	(78)	10.1016/S0031-0182(01)00403-5
Baldeggersee	Switzerland	47.19777	8.26249	Varves	1885	463	5.22	68.4	66.0	(55)	10.1007/BF02522361
Białoławki	Poland	53.73538	21.82862	Laminated	1971	116	2.11	12.7	36.1	(58)	10.1007/s10933-013-9741-7
Blelham Tarn	England	54.39593	−2.97734	Laminated	1980	44	0.11	4.3	14.5	(79)	10.1023/A:1024437426878
Bourget	France	45.80111	5.82652	Varves	1933	231	44.50	560.0	145.0	(46)	10.4319/lo.2013.58.4.1395
Bussjösjön	Sweden	55.44559	13.80172	Laminated	1900	40	0.03	0.5	3.0	(80)	10.1023/A:1007967832177
Constance	Germany	47.65811	9.28918	Varves	1890	395	571.50	11,477.0	254.0	(81)	10.1007/BF02538288
Dgał Mały	Poland	54.12157	21.78924	Laminated	1972	120	0.94	33.0	16.8	(58)	10.1007/s10933-013-9741-7
Enonselka Vesijarvi	Finland	61.01707	25.60596	Varves	1960	81	26.00	84.0	33.0	(82)	10.1007/978-94-017-3622-0_42
Frickenhauser See	Germany	50.40286	10.23707	Varves	1870	315	0.11	0.1	14.5	(83)	10.1177/0959683607086762
Gallocanta	Spain	40.97208	−1.50390	Laminated	1960	990	14.14	543.0	2.5	(84)	10.1016/S0037-0738(01)00217-2
Garbas	Poland	53.90390	22.16229	Laminated	1968	129	0.42	8.0	38.0	(58)	10.1007/s10933-013-9741-7
Geneva	Switzerland	46.21825	6.16220	Varves	1950	372	582.00	7,975.0	310.0	(35)	10.1002/2014GB004932
Gennarbyviken	Finland	60.02937	23.31161	Varves	1957	41	10.77	88.2	32.0	(85)	10.1007/978-94-009-7290-2_24
Greifensee	Switzerland	47.35885	8.669117	Varves	1916	435	24.00	147.8	32.0	(86)	10.1023/B:HYDR.0000014038.64403.4d
Iseo	Italy	45.67459	10.05160	Varves	1962	186	60.90	1,736.0	251.0	(87)	10.1016/j.chemosphere.2011.06.037
Jaczno	Poland	54.28322	22.87151	Laminated	1963	163	0.41	9.0	29.6	(58)	10.1007/s10933-013-9741-7
Jyväsjärvi	Finland	62.23918	25.77259	Varves	1905	78	3.40	372.0	27.0	(74)	10.1023/B:JOPL.0000007229.46166.59
Kameduł	Poland	54.26666	22.86666	Laminated	1968	160	0.25	10.8	24.5	(58)	10.1007/s10933-013-9741-7
Kocioł	Poland	53.72016	21.85844	Laminated	1988	116	2.90	27.7	26.4	(58)	10.1007/s10933-013-9741-7
Kolje	Poland	54.27964	22.88721	Laminated	1968	149	0.16	3.8	27.5	(58)	10.1007/s10933-013-9741-7
Kuokkajarvi	Finland	61.68806	30.62095	Laminated	1825	21	2.55	28.5	19.0	(88)	10.1007/s10933-005-2542-x
Lago Albano	Italy	41.74656	12.67117	Varves	1948	293	6.00	3.7	175.0	(89)	10.1007/BF00684032
Lago Grande di Avigliana	Italy	45.06648	7.38725	Varves	1935	353	0.83	10.7	26.0	(90)	10.1007/s10933-006-0002-x
Laitialanselka Vesijärvi	Finland	61.07889	25.40059	Varves	1970	81	21.50	159.0	18.0	(82)	10.1007/978-94-017-3622-0_42
Lavijarvi	Finland	68.52848	22.60507	Varves	1935	6	2.01	74.0	22.0	(88)	10.1007/s10933-005-2542-x
Ławki	Poland	53.91050	21.57785	Laminated	1992	117	0.69	6.8	17.0	(58)	10.1007/s10933-013-9741-7
Lemiet	Poland	54.15833	21.81333	Laminated	1981	116	0.79	4.7	18.3	(58)	10.1007/s10933-013-9741-7
Lucerne Vitznau basin	Switzerland	47.01001	8.44979	Varves	1946	434	113.60	2,124.0	214.0	(91)	10.3406/rga.2003.2229
Lugano	Switzerland	45.98398	8.96945	Laminated	1929	271	48.70	565.0	288.0	(92)	10.1007/BF00878140
Neuchatel	Switzerland	46.94095	6.92148	Laminated	1952	429	215.00	2,672.0	153.0	(93)	10.1023/A:1008005622256
Nylandssjön	Sweden	62.94445	18.28138	Varves	1925	34	0.28	0.9	17.5	(94)	10.1111/j.1365-3091.2012.01343.x
Sacrower See	Germany	52.44589	13.10217	Varves	1873	29	1.07	35.3	38.0	(95)	10.1007/s10933-005-6188-5
San Puoto	Italy	41.28497	13.40856	Varves	1925	2	0.40	0.9	37.0	(96)	10.4081/jlimnol.2002.15
Sarkinen	Finland	64.13443	28.29127	Varves	1925	125	0.45	3.8	12.3	(97)	10.1007/BF00050950
Säynäjälampi	Finland	65.54768	29.62905	Varves	1960	300	263.00	17.9	1.5	(98)	10.1007/BF00028421
Sejwy	Poland	54.21166	23.18667	Laminated	1943	149	0.85	39.9	21.5	(58)	10.1007/s10933-013-9741-7
Siniec Maly	Poland	54.15124	21.51453	Laminated	1902	130	0.11	1.1	19.0	(58)	10.1007/s10933-013-9741-7
St. Moritz	Switzerland	46.49405	9.845466	Varves	1910	1,768	0.78	171.0	44.0	(99)	10.1177/0959683607082555
Starnberger See	Germany	47.90414	11.30610	Varves	1939	584	56.40	315.0	120.0	(100)	10.1023/A:1008098118867
Tiefer Klocksin	Germany	53.59256	12.52857	Varves	1925	65	0.75	5.5	62.5	(101)	10.1007/s10933-013-9745-3
Tiefer Uckermark	Germany	53.23517	13.36204	Varves	1967	73	0.04	2.4	26.8	(101)	10.1007/s10933-013-9745-3
Tovel	Italy	46.26041	10.94891	Varves	1945	1,178	0.38	40.6	38.0	(89)	10.1007/BF00014627
Varese	Italy	45.81048	8.74287	Varves	1958	238	15.00	112.0	26.0	(102)	10.1007/978-94-009-4047-5_41
Vesijärvi	Finland	61.04159	25.58476	Varves	1912	81	110.00	515.0	42.0	(82)	10.1007/978-94-017-3622-0_42
Ziegelsee	Germany	53.65552	11.42552	Laminated	1996	38	3.00	11.0	34.4	(103)	10.1023/A:1022952232495
Zug	Switzerland	47.14571	8.48638	Varves	1850	417	38.00	204.0	201.0	(104)	10.1021/es950895t
Zurichsee	Switzerland	47.22266	8.75189	Varves	1895	406	88.00	1,829.0	143.0	(105)	10.1007/BF02538179

Watersheds for each lake were computed using the Shuttle Elevation Derivatives at Multiple Scales (HydroSHEDS) flow accumulation and flow direction rasters. Lake and sediment properties were compiled from published data: site country, coordinates in degree decimal, type of facies in the topmost section, age of hypoxia onset, altitude above sea level, lake area, catchment area, maximum depth, and the original reference and its DOI are shown. DD, decimal degrees.

Fig. S2.

Box plots describing the contemporary proportions of cultivated and urban areas in lake watersheds. Recently hypoxic sites (n = 51) are shown in red, naturally hypoxic sites (n = 97) are shown in blue, and benchmark watersheds (n = 769) are shown in green. Upper and lower population density limits represent the first and third quartiles.

Our statistical analyses support the conclusion that urban point sources were the leading driver for the onset of hypoxia. Using a general additive mixed model (GAMM), we found that the probability of hypoxia onset in our 51-lake subset increased as the proportion of urban area increased over the past 300 years (p < 0.0001), but was unrelated to the changes in cultivated and pastured land area (p > 0.1) (Fig. 3 and Table 1) (R² = 0.23). A common observation across the lakes with hypoxia developing only recently is the acceleration of urbanization around CE 1900 that coincided with the onset of hypoxia (Fig. S3). However, the timing of hypoxia onset was quite variable across lakes (Fig. 3B). Our records showed no evidence of a sustained return to improved oxygenated conditions, despite many efforts at remediation (Fig. 2A).

Fig. 3.

Probability of hypoxia onset increased as a function of urban area (%) in the 51-lake subset. (A) Logistic GAMM showed that the probability of hypoxia onset in lakes increased as the proportion of urban area increased over the past 300 years. (B) Random smooth logistic GAMM further detected that the vast majority of lakes experienced an increase in the probability of hypoxia as urban land cover increased, but that the timing of the onset varied among lakes.

Table 1.

Results of the multiple regression models

Logistic GAMM	edf	Ref. df	χ2	Probability value	Significance
Random slope
s(urban area)	34.0	46.0	118.6	6.3 × 10−14	***
s(cultivated area)	1.9	2.3	3.5	0.24
s(pastured area)	3.3	4.1	5.2	0.27
Random smooth
s(logurban,lake)	65.4	204	200.2	<2 × 10−16	***

A logistic GAMM showed that the probability of hypoxia onset in lakes increased as the proportion of urban area increased over the past 300 years, but was unrelated to the changes in cultivated and pastured land area [adjusted R²_adj (R²_ad ) = 0.227]. By incorporating a term that accounted for differences in the trajectory of urban land use among lakes, our logistic GAMM was able to explain a much larger proportion of the variation in the timing of hypoxia onset (R²_adj = 0.396). ***p = 0.001. edf, effective degree of freedom; Ref., reference degree of freedom.

Fig. S3.

Three hundred-year trends for land cover in Europe based on an AMM, grouping watersheds according to their history of downstream lake hypoxia or reference (Ref.) source. Land-cover values (63) were extracted from each modeled watershed (Materials and Methods). The increase in urban areas from CE 1700–2000 is significantly larger for watersheds with recently developed hypoxia (red solid line) than for watersheds with natural hypoxia (orange dashed line) and benchmark watersheds (green and blue dashed lines). All model results are statistically significant at p < 0.0001. Gray bands indicate 95% confidence intervals of the predicted means based on the AMM.

Centennial trends in land cover for the 51 watersheds differed notably from trends in the 97 study watersheds recording natural hypoxia (i.e., sites with sustained laminae for >300 years) (Materials and Methods) and the 1,459 benchmark watersheds (Fig. 4). The rate of expansion of urban areas was significantly higher in watersheds associated with recent hypolimnetic hypoxia than in other European watersheds (Fig. 4 and Table 1). To the contrary, the rates of changes in cultivated and pastured areas are similar in watersheds with recent hypoxic lakes and other European watersheds, although the absolute magnitude of cultivated land was generally higher around the sites with recent hypoxia. Mann–Kendall (M-K) tests indicated that hypoxia onsets were preceded by centennial increases in urban (45% of the sites), cultivated (95% of the sites), and pastured (80% of the sites) areas. However, during the transition toward hypoxia (±20 years centered on the time of the onset), urban areas expanded in 71% of the sites, whereas cultivated and pastured areas were decreasing in 61% and 74% of the sites, respectively (Tables S3 and S4). Collectively, these findings suggest that urban point sources of nutrients were the leading factor explaining the spread of lacustrine hypoxia in Europe over the past few centuries. The prevailing importance of urban point sources of nutrients as the preeminent trigger toward the spread of hypoxia was also validated by M-K and AMM analyses of decadal-scale landscape and climatic reconstructions spanning the period between CE 1900 and 2010 (24) (Fig. 4 and Figs. S3 and S4). Finally, basin-scale analyses of modern characteristics confirmed the prevailing importance of local human activities on the presence of hypoxia (Fig. S5).

Fig. 4.

One-hundred–year trends for land cover in Europe based on an AMM, grouping watersheds according to their history of downstream lake hypoxia or benchmark (Ref.) source. Trends in Europe represent decadal percentages of urban, cultivated, grassland, and forested areas. Note the higher increase in urbanization for the recently hypoxic sites during the past 110 years compared with the benchmark sites. Gray bands indicate 95% confidence intervals of the predicted means based on the AMM. DB, database.

Table S3.

Land cover trends for recently hypoxic, naturally hypoxic, and benchmark sites

	Urban		Cultivated		Pastured		Total
Hypoxicity	τ	n	τ	n	τ	n	N	Period
Recently hypoxic	+0.67	31	−0.29	40	−0.35	47	51	CE 1900–2010
Naturally hypoxic	+0.49	29	−0.18	52	−0.14	79	97
Benchmark	+0.44	543	−0.14	674	−0.27	1,143	1,459
Recently hypoxic	+0.92	24	+0.76	45	+0.22	45	51	CE 1700–2000
Naturally hypoxic	+0.91	31	+0.73	78	+0.42	72	97
Benchmark	+0.91	558	+0.70	1,270	+0.55	1,245	14,59

M-K test results are presented for recent (CE 1900–2010) and long-term (CE 1700–2000) periods. A positive score (red) indicates an increasing trend. Decreasing trends are shown in blue.

Table S4.

Interaction between external drivers and lake hypoxia onset

	Sites recording increasing trends		Sites recording decreasing trends		Steady state	Total
Area	%	n	%	n	n	n	Timing
Urban area	71	27	3	1	10	38	At the moment of hypoxia
Cultivated area	29	11	61	23	4	38
Pastured area	21	8	74	28	2	38
Urban area	43	19	0	0	25	44	Before hypoxia
Cultivated area	95	42	2	1	1	44
Pastured area	80	35	20	9	0	44

Increasing or decreasing trends in land cover are presented for two time windows: at the moment of hypoxia onset (±20 years centered on the moment of hypoxia onset) and before hypoxia onset (200 years preceding hypoxia onset). Numbers are color-coded to indicate when more than 40% of the sites showed an increase (red) or a decrease (blue) in urban, cultivated, or pastured area. The M-K tests presented in Fig. S3 confirm the increasing trends in urban area at the moment of hypoxia onset. Note that the urban area was generally increasing at the moment of hypoxia onset, whereas cultivated and pastured areas tended to be decreasing.

Fig. S4.

Changes in air temperature (t.) during the past 110 years. (A and B) Mean annual temperatures are presented for all recently hypoxic sites, naturally hypoxic sites, and benchmark sites. Temperature anomalies have been extracted annually for each studied lake from the instrumental dataset of the University of Delaware Air Temperature & Precipitation (UDEL) model (global data with a resolution of 0.5° × 0.5°). Color lines are fifth-order polynomial trends, and gray lines are 95% confidence intervals. (A) Trends for each subgroup of lakes have been superimposed. In CE 2000, there is no significant difference in temperature anomalies between recently hypoxic sites, naturally hypoxic sites, and benchmark sites. DB, database.

Fig. S5.

(A) Regression tree of hypoxia onset for paleolimnological data. As a preliminary statistical assessment to describe the distribution of varves in European lakes, the modern land use and climate data for each watershed were used as the forcing variables and the presence/absence of varves in lake sediments was used as the response variable in a regression tree analysis using the rpart and the wrapper function “MVPARTwrap” (69) in R, v3.0.2 (R Foundation for Statistical Computing). Plots of the cross-validation results and pruning of the tree using the 1-SE rule (69) were used to select the best tree and to avoid overfitting. (B) Specifically, we chose the complexity parameter associated with the smallest tree where the estimated error rate was within 1 SE of the minimum error, and pruned the tree at this complexity parameter value. Several cross-validation runs were performed to verify that the final tree was not atypical. We compiled data for numerous potential explanatory variables that have previously been shown to explain significant variation in hypoxia distribution. In particular, the variables considered for the regression tree analysis included modern data on lakes, land use, climate, and human activities, as described in Materials and Methods, Land Use and Climate Data. The distribution of hypoxic lakes was also compared with human population density and with gross domestic product (GDP), serving as a potential factor affecting the P yields to lakes. We used a conservative variation inflation factor (VIF < 5) to isolate independent (noncollinear) explanatory variables and eliminated collinear variables to improve interpretation of the regression trees. Explanatory variables retained for the final regression tree included maximum lake depth, urban area, forest area, pastured area, lake surface area, precipitation, air temperature, human population, GDP, and maximum depth. cp, complexity parameter.

Discussion

This regional-scale analysis of paleolimnological records adds to the growing evidence that modern human activities are a widespread force in shaping the structure and function of inland waters (13, 39–42). Our previous paper (17) demonstrated that the spread of lacustrine hypoxia at the global scale was predominantly the result of nondescript human impact. However, the current study specifically pinpoints urban point sources of nutrients as the main forcing mechanism within Europe. Based on earlier water quality studies and paleoecological data, it is known that algal blooms decreased water transparency for most lakes in Europe starting in the middle to late 19th century (37). The eutrophication phase was often more pronounced, beginning in CE ∼1950 (37, 43), but the development of widespread hypolimnetic hypoxia has largely predated the more visible effects of eutrophication in the epilimnion (this study). As such, the spread of hypolimnetic hypoxia can be considered an early warning of eutrophication, caused by enhanced sediment and organic matter fluxes toward bottom waters. The hypolimnion acts as an integrator of processes taking place over the entire water column.

It is generally well accepted that contemporary freshwater eutrophication is predominantly caused by diffuse P sources, principally from agriculture (2), in developed nations (i.e., nations having very high human development in 2014 according to the United Nations Human Development Index). In contrast, the situation in developing nations is mixed and includes diffuse sources of P and domestic point sources (18). However, our analysis of longer term trends in Europe (Fig. 2) provides an important historical perspective, whereby intensive fertilization of agricultural soils and associated diffuse sources of P and nitrogen increased through the middle of the 20th century, largely postdating the initial spread of lake hypoxia (2, 18) (Fig. 2 A and C). As such, diffuse sources of P appear to have had a subordinate role compared with point sources for most of the past 300 years, and were not decisive for the onset of lacustrine hypoxia in most of the studied lakes. However, nutrients arising from agricultural areas likely had some effect, because the long-term M-K tests demonstrated that hypoxia onsets were preceded by increases in cultivated and pastured areas, as well as urban areas. Overall, we suggest that lakes have suffered a slow loss of resilience as a result of both point and diffuse P inputs over time until a disproportionate increase from urban point sources tipped the balance toward hypoxia.

In present-day Europe and North America, domestic sewage and industrial waste water mostly receive an efficient treatment, including P removal, before discharging effluents into lakes (44). However, the situation around the end of the 19th century was quite different, when urban waste waters with increasing P content were directly discharged into waterways (44) and began affecting downstream aquatic ecosystems. The problem was fueled by urban expansion, a growing population, an accelerating economy during the industrial revolution, the rising standard of living, and novel domestic and industrial uses of P (45). The first P-containing detergents were introduced around the end of the 19th century and soon enjoyed wide acceptance (45). All of these developments were synchronous with the rapid spread of lake hypoxia.

Importantly, our study shows that lakes with recent hypoxia shifted abruptly and irreversibly to an alternate stable state. For instance, among the lakes considered in this study, three perialpine lakes (Lakes Geneva, Bourget, and Annecy) (Fig. S6) that were previously oxygenated over the past millennia shifted to hypolimnic hypoxia between CE 1930 and CE 1950 following a slight P increase (i.e., with enrichments of only ∼8–10 μg TP L⁻¹; refs. 35, 46, 47). This finding illustrates that even a small increase in P availability can stimulate enough primary productivity to trigger hypoxia without generating algal blooms (because blooms were only observed after CE 1950). Likewise, the temporal trend of oxygenation in European lakes (Fig. 2A) shows a slowing down of the rate of increase, but no turning off of hypoxia after the 1980s, despite the implementation of restoration programs and successful controls on nutrient influx. The crossing of critical thresholds of nutrient loading appeared to have abruptly and irreversibly shifted lacustrine ecosystems from one state to another (48). Imported P, both from watersheds (external load) and remobilized from lake sediments (internal load), can explain the stability of hypoxia over the past ca. 30–40 years. P loads from watersheds to downstream lakes initially accumulate in lake sediments, but may later be remobilized from sediments into overlying waters under hypoxic conditions. P-rich sediments have been identified as the key factor in sustaining hypoxia (49, 50). For instance, the accumulation of organic matter during eutrophic conditions, and the subsequent diagenetic release of P from near-surface sediments, is known to cause lakes to remain in the eutrophic state even if the external input of P has diminished (1). In addition, a reduced ability of ecosystems to remove nitrogen via denitrification and anaerobic ammonium oxidation may be related to hypoxia and could lead to accelerated eutrophication (49). Finally, an increase in water temperature could also decrease the threshold of P concentrations sustaining hypoxia, with more intense stratification, reduced solubility of oxygen at higher water temperatures, and enhanced metabolic rates in warmer bottom waters (51).

Fig. S6.

Specific details on perialpine Lakes Geneva (Leman), Bourget, and Annecy. Changes in land cover (Right) and human population (Left) are presented for the three sites. (Right) Percentages of urban (blue curves with circles), cultivated (black curves with rectangles), and pastured (black triangles) areas are presented. Periods with the hypoxic condition in lakes are highlighted in green. Urban area and population are growing at the time of the hypoxia onset, whereas cultivated and pastured areas are decreasing.

Unfortunately, the lack of past land cover data at a sufficiently high spatial resolution in other regions prevents us from expanding this work globally. Nonetheless, the observed regime shifts to new stable hypoxic conditions highlights the challenges for developing countries facing persistent diffuse P emissions and growing P demands, together with changes in lifestyle (e.g., diet shifts) and expanding urban areas (including the development of megacities and periurbanization). Moreover, wastewater from sewage and industry is often untreated and may be the primary contributor toward eutrophication (52). For example, only 35% of wastewater in Asia and <1% in Africa were treated in CE 2005 (52). Without implementation of wastewater treatment of P in point and mixed sources, the future conditions of lakes in these regions will likely result in prevalent hypoxic hypolimnetic conditions, degraded water quality, and the necessity for decade-long restoration efforts.

In conclusion, our analyses of laminated sediment records indicate that nutrient point sources from growing urban areas were the leading driver for the onset of hypoxia in the hypolimnion of downstream lakes. Point and diffuse sources have always both contributed to the total supply of nutrient inputs to lakes, but with varying intensities over time and space. Our results show that urban point sources of P were the dominant driver of lake eutrophication in European lowland systems during the Anthropocene. During the past few decades, the relative contribution of diffuse P sources has progressively become a major cause of modern freshwater eutrophication in developed countries, as point sources have been reduced and fertilizer use has increased. The lack of reoxygenation of the hypolimnion evident from our analyses highlights the importance of the history and legacy of past land uses, and the need for long-term strategies to maintain and restore water quality in modern lake ecosystems.

Materials and Methods

Reconstructing the Dynamics of Hypoxia.

The sediment textures of many lakes offer a simple proxy for the oxygenation history of bottom waters (53–55). Indeed, the appearance of laminated sediment on top of homogeneous sediment indicates that annual oxygenation conditions fell below a critical threshold in both duration and concentration (35, 56, 57), hence recording the die-out of macrobenthos and the end of its related bioturbation (54, 58, 59) (Fig. S7). If laminations are proven to reflect annual cycles of sedimentation, they offer the additional advantage that the shift from well-oxygenated to at least seasonal hypoxic hypolimnic conditions can often be precisely dated by counting varves from the sediment/water interface down-core (54). The Varves Working Group (VWG) of Past Global Changes has intensively investigated varved lakes over the past decade (54, 60, 61), enabling the assembly of a large dataset of lake hypoxia (17). In Europe, 148 varved sediment records were referenced in the global compilation of the VWG (17) and indicated that the European dynamics of lacustrine hypoxia encompassed: (i) a period of relatively undisturbed conditions before CE 1850 serving as a preindustrialized baseline reference, (ii) a period of major changes during the early industrialization of Western countries and the following “Great Acceleration” phase of the so-called Anthropocene (42), and (iii) the initiation of European lake restoration programs since the 1970s. Land use changes in watersheds of recently varved lakes have been compared with a set of 97 naturally varved lakes to dismiss any sampling bias related to morphometric properties. Preservation of laminated sediments usually indicates that lakes have strong hypoxia; however, strong seasonal hypoxia may not systematically develop laminations, notably due to the absence of contrasting seasonal sedimentation or as a consequence of wind causing sediment resuspension. Our data matrix does not attempt to include all lakes with hypoxia but, instead, includes a conservative and large selection of well-characterized lakes with laminated sediments to provide a statistically sound and relevant basis for constraining the dynamics of hypoxia in Europe.

Fig. S7.

Examples of hypoxia onset in a subset of lake sediment cores used in this study, illustrating the appearance of laminated sediment on top of homogeneous sediment. The appearance of these sedimentary facies indicates that annual oxygenation conditions fell below a critical threshold in both duration and concentration, hence recording the die-out of macrobenthos and the end of its related bioturbation. Specific details on those individual sites across Europe are presented in SI Text [De Vincent et al. (73); Jenny et al. (46); Renberg et al. (76); Jenny et al. (35); Merilianan et al. (74); Lotter et al. (55)].

Numerical Analysis.

An AMM framework generated using the mgcv library in R (62) was used to describe the general nonlinear trends in land use over the past three centuries. It was anticipated from the study by Jenny et al. (17) that watersheds with recently hypoxic lakes would contain an environmental signal reflecting a more urbanized and agriculturally cultivated landscape compared with watersheds serving as benchmarks, as well as naturally hypoxic lakes. Thus, the relationships between urban, cultivated, pastured, and forested areas were evaluated for the four watershed categories of this study. Confidence intervals were derived using the SEs produced by the predict.gam function in R (63), with type = “response” specified in the model [mgcv library (64)]. Multiple regression analyses and nonparametric M-K tests for monotonic trends were conducted to identify the main drivers of hypoxia onsets (SI Materials and Methods).

SI Materials and Methods

Paleolimnological Data.

A literature search was conducted in April 2014 (17) and updated in June 2015 using the Institute for Scientific Information Web of Science database and Google Scholar with different combinations of the following keywords: “varve” and “lake,” and “lamin” and “lake sediment.” The search yielded 148 relevant European lakes that contain laminated or varved sediments. Descriptions and data on varved sites, sediments, and dating methods are available in a study by Jenny et al. (ref. 17 and the references therein). The original chronologies were expressed in CE calendar years. Laminated lacustrine sites had to satisfy several conditions to be included in this synthesis. Accepted sites contained a varved or well-preserved laminated sedimentary sequence and/or featured a published age-model relying on varve counting and/or radiometric dating, and the lakes’ sediment texture had to be explicitly described or illustrated by pictures outlining the laminated intervals. The timing of the first onset of hypoxia was obtained for each lake by examining all relevant published varve data. Where time intervals could not be dated precisely with the help of published data, corresponding authors were contacted and asked for advice. The water depth for each lake was collected and used to verify that lake level fluctuations were not the cause of changes in preservation conditions of the varves. Descriptions and data for lake sites were compiled in this study (Table S1).

Land Use and Climate Data.

Modern data and temporal changes in land use and climate during the past 300 years were analyzed for 1,607 watersheds. Hydrological basins of each site were calculated using the flow accumulation and flow direction rasters made available from hydrological data and maps based on Shuttle Elevation Derivatives at Multiple Scales (HydroSHEDS), together with lake perimeters and areas using the GLWD from the World Wildlife Foundation (38). The following variables were extracted from modeled areas using the “Arc” geographic information system (ArcGIS): (i) modern site characteristics, (ii) past land use from CE 1900–2010 at decadal steps and with a 1-km² spatial resolution, and (iii) past land use from CE 1700–2000 with centennial resolution. Mean local temperatures; precipitations; population densities (65); changes in urban cultivated, pastured, and forest areas (24); and past human population densities (66) were extracted from modeled areas for each watershed.

Multiple regression analyses were conducted to identify the main drivers of hypoxia onsets. For each recently hypoxic lake (n = 51), we created a binomial time series indicating whether the first hypoxic event had occurred or not at each date of the land cover data (i.e., CE 1700, 1800, 1900, 1910, …, 2010). To test the relative importance of the different land cover types, we then ran a logistic GAMM using the binomial time series as the response variable; the percentage of urban, cultivated, and pastured areas as fixed effect explanatory variables; and lake identification as the random effect (testing a random slope and intercept for each lake). To test further whether the smooth term varied among lakes, we tested a random smooth logistic GAMM, which allowed not only the slope to vary among lakes but also the shape of the nonlinear relationship. All GAMMs were fit using the bam function of the itsadug package in R (67).

Nonparametric M-K tests for monotonic trends were used to quantify trends of land use for each of the 1,607 watershed time series within the past 300 years. This analysis was based on the Kendall rank correlation coefficient and was conducted using the Kendall library (68). A positive score shows a monotonically increasing trend, whereas a negative value shows a monotonically decreasing trend (69). For each site, M-K tests were run for two time windows to identify the potential effects of slow and fast land cover changes on the hypoxia onset: We anticipated that fast changes in the land cover would show an effect within a short period (±20 years) centered on the time of the onset to be consistent with the uncertainties of reconstructions, and that slow changes in the land cover would show an effect over a longer period (∼200 years) preceding the onset of hypoxia to be consistent with the long-term history and potential legacy effects of past land changes in Europe.

SI Text

Specific details on selected sites are described in Fig. S6 and Table S2.

Lakes Geneva (Léman), Bourget, and Annecy are located in the French and Swiss perialpine region and were intensively studied in the course of the program: Perturbation Impacts on Lake Food Webs, a Palaeo-Ecological Approach (IPERRETRO) of the French National Research Agency (ANR) (2009–2013) (12, 35). Within this program, volumes of hypoxic waters were reconstructed and analyzed during the Anthropocene using limnological and paleolimnological data (35, 46). The three lakes were selected because they share the same P enrichment history: although they were oligotrophic at the end of the 19th century, all three lakes underwent P enrichment as early as the 1920s, concomitant to the population increase in urban areas (Fig. S6). At this time, these lakes received untreated wastewaters from the growing cities of Geneva, Chambéry, Aix-Les-Bains, and Annecy (Fig. S6). A shift to hypoxic conditions occurred in CE 1933 ± 1, CE 1950 ± 1, and CE 1952 ± 1 [in Bourget, Geneva (Léman) and Annecy, respectively] in response to the unprecedented rise in total P concentrations above 10 ± 5 μg of P L⁻¹. Following this shift, hypoxia never disappeared, despite the fact that the environmental policies implemented succeeded in drastically reducing lake P concentrations: Observational data demonstrate that mean total phosphorus (TP) values measured during winter mixing have been successfully reduced to 6 μg of P L⁻¹ in Annecy, 19 μg of P L⁻¹ in Geneva, and 17 μg of P L⁻¹ in Bourget (47).

Lake Alserio, a dimictic hard-water eutrophic lake, is located in the Brianza district near Lake Como in northern Italy (surface area of 1.23 km² and maximum depth of 5.3 m). Beginning in the 1970s, Lake Alserio was affected by elevated external P loads, which resulted in high P concentrations in the water column (73). The major P sources that affected the water quality and the trophic status of the lake were domestic sewage (fluxes of P to the sediments = 1.96 g⋅m⁻²⋅a⁻¹) and runoff (fluxes of P to the sediments = 0.75 g⋅m⁻²⋅a⁻¹) (73). Despite both the reduction of P in detergents and the construction of a pipe conveying urban wastewater and runoff to a treatment plant, the present external P load to the lake is still high. The change from homogeneous sediment to a laminated sequence appears to be linked to the eutrophication process of Lake Alserio occurring since the 1960s (73).

Baldeggersee is located on the central Swiss Plateau, and is characterized by a maximum depth of 66 m, a surface area of 5.2 km², and a water renewal time of 5.6 years. Owing to strong anthropogenically driven nutrient enrichment of this site, Baldeggersee has become hypertrophic, and complete oxygen depletion was observed at water depths of 60 m at the beginning of this century (55). Long-term changes in the watershed showed that growing urbanization was concomitant with the onset of hypoxia in 1880s (as inferred from the lake sediment record). The introduction of treatment plants since 1967 and 1975 has successfully reduced point sources of nutrients (55). Since 1982, bioturbation has prevented varve formation above 55 m of water depth due to the better oxygen regimes at the water/sediment interface. However, below 55 m in water depth, the mineralization of organic carbon is still causing oxygen depletion that allows the formation and preservation of varves.

Lake Jyväsjärvi is located in central Finland (64°14′N, 25°47′E) and has a surface area of 3.4 km² and a maximum depth of 27 m. The town of Jyväskylä was established on the lakeshore in 1837 on the northern bank of Lake Jyväsjärvi, and the lake received untreated municipal wastewater from the town up until 1977 (74). Based on the biological and chemical properties of the sediment strata, Meriläinen et al. (74) distinguished a phase of early changes in the lake ecosystem from the 1870s to ca. the 1940s. This phase of early changes corresponds well to the timing when varves first appeared (at the end of the 1800s) and when the lake received untreated municipal wastewater. Although the first plans for building a wastewater treatment plant were prepared in the early 1930s, municipal wastewater continued to be discharged into the lake in the untreated form until 1977. In recent years, this effluent loading has been reduced.

With a surface of 61.8 km² and a volume of 7.5 km³, Lake Iseo (also known as Sebino) is the fourth largest Italian lake. The watershed includes 83 municipalities, 21 of which are on the shoreline, with a total population of about 180,000 inhabitants. The sources of P in Lake Iseo are derived mainly from point sources (75). During the 1960s, intensifying eutrophication processes and the corresponding deterioration of water quality in northern Italy were becoming apparent, initially in lakes with low maximum depths (Lake Varese and Lake Endine) and subsequently in the deep subalpine lakes. During the same period, Lake Iseo became hypoxic, as evidenced by the appearance of varved sediments.

Nylandssjön is located in northern Sweden (62°57′N, 18°17′E) and is a dimictic, mesotrophic, circumneutral lake. The lake is characterized by a surface area of 0.28 km² and a maximum depth of 17.5 m. The catchment (0.86 km²) is covered mainly by spruce forest and, to a lesser extent, by arable land. Varves are evident in deeper sediment layers of Nylandssjön, but their permanent formation and preservation started at the beginning of the 20th century because of cultural eutrophication (76). The major population increase associated with wastewater release occurred at the end of the 19th century in the region of Lake Nylandssjön [there was a doubling in the number of inhabitants in the major town between CE 1850 and 1910 (Swedish population censuses, Statistikdatabasen), which has since stabilized].