In situ chl a concentrations and physicochemical parameters were extracted from open-source international, national, and regional water quality databases (Table ). We first obtained data from queries to the Earth System Science Data portal (https://www.earth-system-science-data.net/, last access: 17 June 2024), the Environmental Data Initiative repository (https://edirepository.org/, last access: 17 June 2024), and the Scientific Data portal (https://www-nature-com.sire.ub.edu/sdata/, last access: 17 June 2024). We then conducted a literature search on Web of Science (https://www.webofscience.com/wos/, last access: 17 June 2024) and Scopus (https://www.scopus.com/, last access: 17 June 2024) for further existing long-term chl a and nutrient time series. To do so, we used the following search terms: TITLE or ABSTRACT (oligotrophication, reoligotrophication, chlorophyll, time series); and in TITLE or ABSTRACT (lake, river, estuary, coastal, estuarine); and in EVERYTHING (trend, long term, multi-decadal). When public and accessible, we directly extracted the datasets and proceeded with data harmonization. The database architecture (Fig. ) allows researchers to easily complement it with additional time series in the future. New additions to the database will be eased by a set of scripts available in a dedicated version control GitLab repository (https://gitlab.com/OLIGOTREND/wp1-unify, last access: 7 May 2025), allowing users to reproduce, update, or add more time series from level L0a to higher data levels and products.

We gathered data as raw measurements, i.e. unprocessed or non-aggregated time series, and defined herein these data as level L0a. Extracted variables included chlorophyll a (chl a), water temperature (wtemp), conductivity (cond), pH, dissolved oxygen as concentration (o2) and percentage of saturation (o2sat), dissolved inorganic nitrogen (din), nitrate (no3), nitrate × nitrite (no23), ammonium nitrogen (nh4), Kjeldahl nitrogen (nkjel), total nitrogen (tn), orthophosphate or soluble reactive phosphorus (po4), total phosphorus (tp), dissolved organic carbon (doc), and total suspended solids (tss). The ecosystem types covered in this database included lakes and reservoirs, rivers, estuaries, and coastal bays.

We primarily targeted databases identified with long periods of records without any filter on geographic location (Table ). We discarded chl a datasets obtained with remote sensing techniques to ensure a strict comparability among observations. For stratifying deep lakes, we extracted values either for the euphotic layer or from the upper 10 m if euphotic depth was unavailable to avoid using data from light-limited conditions.

First, L0a time series were individually reformatted into standard units and data matrix headers, forming an ensemble of time series defined here as level L0b. Nutrient concentrations were expressed as mgL-1, except chl a, which remained in µgL-1. Time series were named with a unique identifier (uniquID) per site corresponding to the concatenation of the following data separated by underscores: “ecosystem type”, “basin”, and “station ID”, e.g. “river_loire_04000100”. Basin names were derived from site geographic coordinates and the corresponding watershed according to the FAO dataset (Food and Agriculture Organization of the United Nations and FAO Land and Water Division, 2011). Ecosystem type was either “estuary”, “lake”, or “river”, corresponding to estuary or coastal bay, lake or reservoir, and river, respectively. The “station ID” was the one provided by the original data source. For each sampling site, the geographic coordinates found in the original metadata were used to create a point shapefile labelled with the station unique identifier (uniquID) as explained above. Stations with no geographic coordinates were discarded from the database.

Data quality was assessed and checked for all L0b time series from sampling stations presenting at least 15 years of chl a data. The resulting dataset comprises the OLIGOTREND L1 data level (Fig. ). We did not remove any data in response to data curation (QA/QC) to allow users to design their own quality check procedure. Instead, we flagged potentially anomalous or suspicious observations. Valid observations were indicated with flag = 0. Quality control identified missing values (flag = 1), possible outliers (flag = 2), and abnormally repetitive values (flag = 3). Observations were considered outliers when the corresponding values exceeded 3 times the interquartile range defined by a site. Observations were considered abnormally repetitive when, at a given site and for a given variable, the corresponding value appeared more than 5 % of the time in the time series, not necessarily consecutively. Obvious mistakes in the units found in the original datasets at level L0b were identified and corrected by plotting the density of the distribution of observed concentrations and scatterplots by pairs of variables (e.g. chl a vs. tp, tp vs. po4, etc.) throughout the database.

We linked inland sampling stations with the global HydroATLAS database (Lehner et al., 2022; Linke et al., 2019). The HydroATLAS has three distinct datasets: BasinATLAS, RiverATLAS, and LakeATLAS, which represent the sub-basin delineations (polygons), river network (lines), and lake shorelines (polygons), respectively. Although we proceeded with the spatial join between HydroATLAS and OLIGOTREND stations, we acknowledge there may be a potential temporal mismatch between HydroATLAS properties and OLIGOTREND temporal coverage. Yet, we assumed this spatial join would succeed at demonstrating the great variability of watershed and ecosystem properties encountered in the OLIGOTREND database.

First, we linked all OLIGOTREND sampling stations to the BasinATLAS by spatial selection of polygons of sub-basins (Pfafstetter level 12, i.e. the highest hierarchical sub-basin level in the BasinATLAS), overlapping with the point shapefile of L1 OLIGOTREND stations. A selection of watershed properties related to their physiography, climate, land cover, hydrology, and anthropogenic pressures was extracted and linked to each station present in the database at the L1 level and intersecting with one of the BasinATLAS sub-basins. Similarly, the intersection of LakeATLAS lake polygons with L1 stations provided an ensemble of lake characteristics for 61 % of the lake stations (418 out of 687). Finally, OLIGOTREND L1 river stations were linked to the RiverATLAS database by identifying the three nearest river segments using the function joinbynearest() in QGIS 3.26.2. For each possible station–segment match, the distance between the station and each segment was calculated, and the quality of the spatial join was assessed using a flagging system: if the distance to the nearest segment exceeded 500 m, a flag (flag = 1) was raised, indicating that the distance might be too large for the join to be considered valid. If the distance to the second- or third-nearest segment was less than 10 % greater than the distance to the nearest segment, a flag (flag = 2) was raised indicating that several river segments could potentially be selected. In that case, if these segments were associated with multiple sub-basins (HYBAS_L12 in HydroATLAS documentation), a flag value of 2.1 was set. If these segments were linked to multiple drainage basins (MAIN_RIV in HydroRIVERS), a flag value of 2.2 was set. All other associations identified during the spatial join were considered valid, and the flag value was set to flag = 0. Only stations with flag = 0 were considered reliable. Overall, out of 924 river stations, 90 % was considered valid. We found that 6.1 % of stations were more than 500 m away from the closest HydroRIVERS segment, and 3.9 % showed possible multiple associations (flag ≥ 2), sometimes with different sub-basins (1.3 %, flag = 2.1) or drainage basins (0.3 %, flag = 2.2). We acknowledge that there is some uncertainty in the spatial join between OLIGOTREND river stations and HydroRIVERS given the spatial resolution of the HydroSHEDS (15 arcsec). This uncertainty could be reduced using a river network derived from a higher-resolution digital elevation model. Stations with unmatched basin, lake, or river segment from the HydroATLAS database were not removed from the OLIGOTREND database, but we did not account for them in the statistics and description of watershed attributes.

We described the OLIGOTREND time series based on multiple metrics. These included the number of observations by each variable and the extent of the period of record as well as the median, average, and standard deviation of all valid values over the entire time series.

As a first step into the trend analysis, we quantified the proportion of time series showing lower annual averages in the second half of the time series compared to the first one. We chose annual averages over growing season averages to increase robustness in the metric because sampling frequency was sometimes unequally distributed seasonally. This further simplified the question of how to identify the growing season among sites across latitudes. We considered that a lower average value in the second half of the time series indicated decline, regardless of the level of trend complexity found in the time series.

A breakpoint and segmented regression analysis was performed using the R package segmented (Fasola et al., 2018). Whenever the Davies test (Davies, 1987) did not identify any non-constant linear regressions in time series, we conducted a Mann–Kendall trend analysis on annual averages with the R package trend (Pohlert, 2023). When the Mann–Kendall test detected a monotonic trend (p < 0.01), we calculated a Sen's slope over the complete dataset. Whenever the Davies test identified non-constant linear regressions, we fitted a segmented regression to the data with two joined segments, and the position of the temporal breakpoint and the corresponding interval estimation were identified. Sen's slope was then quantified for both sides of the given breakpoint. For each segment, there were three possible trend types: declining, no trend, and rising, noted as “-”, “0”, and “+”, respectively. The combination of two joined segments or a single segment only when no breakpoint was detected provided a total of 12 possible trend types: ”-”,”--”, ”+-”, ”0-”, ”-0”, ”0”,”00”, ”+0”, ”-+”,”0+”,”+”, and “++”. We acknowledge a segmented regression with one breakpoint is unlikely to capture all the variety in trend patterns, but it may provide a comprehensive first assessment for nonlinear and non-monotonic temporal patterns robust enough to provide a first overview on multi-decadal temporal trajectories. Outputs from the trend analysis and above-described statistical descriptors constitute level L2 data.

We collected L0 data from 3718 sampling stations, producing a total of 41 979 time series. Among these, 1894 stations had at least chl a for over 15 years and were selected for quality check and harmonization at level L1 (Fig. ). Following the quality check, the OLIGOTREND database includes 4.3 million observations. Across all variables and time series, 83 807 observations (1.7 % of total observations) were flagged as outliers and 691 000 (13.7 % of total observations) as repetitive observations. The highest proportion of abnormally repetitive observations was found for nh4 and tp (34 % and 21 % of the observations, respectively, Table ), likely related to detection and/or quantification limits above the actual concentrations. For chl a, 13 % of the observations was flagged as repetitive (9.9 %) or extreme outliers (3.4 %). We only included the valid data points for all subsequent analysis and time series descriptions. Most L1 time series were multi-decadal with a median time series length of 33 years (Table ).

The majority of chl a time series included five observations per year (Table ); only 16 % of time series were based on monthly sampling. We counted that 95 % of chl a time series exceeded 15 years and 75 %, 43 %, and 11 % covered 20, 30, and 40 years, respectively. The longest chl a time series covering more than 45 years originated from the Lake PCI dataset (10 lake chl a time series located in Sweden), the UK Harmonized Monitoring Program (42 rivers in England and Wales), and the Sacramento Bay Interagency monitoring (13 stations in estuarine area).

Time series duration and mean observation frequency for all other variables were generally similar to the chl a time series. The median period of record was 32 years for both tp and tn. Median sampling frequency was 6 and 10 observations per year for tp and tn, respectively. A small proportion (2 % and 1.8 %, respectively) of tp and tn time series were shorter than 15 years. For tp, 84 %, 57 %, and 9 % of the time series were longer than 20, 30, and 40 years, respectively. For tn, 83 %, 61 %, and 5 % of the time series were longer than 20, 30, and 40 years, respectively. There were 444 stations with joint chl a, N, and P observations for over 15 years. Among these, 220 corresponded to river stations, 169 to estuary stations, and 55 to lake stations.

Across all time series, the median temporal coverage was 1986 to 2022 (Table  and Fig. ). Yet, OLIGOTREND featured early and long chl a time series, with 19 of them starting before 1970 and an average of 50-year-long time series, most of them found in the Lake PCI dataset. Across all variables, the 2000s and 2010s are the decades with the highest coverage. The 2020s were not as covered as the 2010s were, likely indicating that databases are not systematically updated with the most recent observations.

The OLIGOTREND L1 database contains 13 992 time series originating from 1894 sampling stations spanning across 5 continents (Table , Fig. ). There are 238, 687, and 969 stations located in estuaries or coastal bays, lakes or reservoirs, and rivers, respectively (Table ). The three largest data sources are the French national water quality monitoring portal (775 stations), a global database of water quality measurements in lakes (Naderian et al., 2024; 378 stations), and the United States' Chesapeake Bay Program (199 stations).

Geographically, the L1 dataset includes stations from 33 different large watersheds (Fig. , and see Table S1 in the Supplement for a detailed list of these watersheds). The five most represented large watersheds are the Seine (France, 320 stations), the United States North Atlantic coast (266 stations), the Mississippi–Missouri Basin (231 stations), the French west coast (183 stations), and England and Wales (163 stations). In total, 7 large watersheds contain more than 100 stations. Data from the Chesapeake Bay (United States North Atlantic coast watershed) and the Elbe River watershed are particularly remarkable in terms of data contribution, covering hundreds of stations along the main rivers and encompassing both freshwater and estuarine zones.

The OLIGOTREND database covers 1229 sub-basins from the HydroATLAS database, distributing over 257 spatially independent large watersheds with no hydrological connections. OLIGOTREND covers a wide range of eco-physiographic contexts (Table ). It covers medium to large watersheds (10th to 90th percentiles were 142 to 11 416 km2), primarily lowlands. Stations extend to four climate zones, from extremely cold and mesic to hot and dry. The share among land-use types also covers a wide range, from 100 % forest or natural grassland areas to heavily impacted urban areas and croplands. Some of the stations are located in nearly pristine areas, but most of them are in highly populous areas.

Similarly, lakes and rivers represented by the OLIGOTREND database cover a wide range of morphometry, from shallow (e.g. Hickling Broad Lake, England, average water column depth ∼ 0.7 m) to deep and large lakes (e.g. Lake Geneva, France–Switzerland, average depth ∼ 155 m) and from headwater streams (e.g. the Evel River in French Brittany draining a basin of 5 km2) to large rivers (e.g. Mississippi, Danube, Rhine, Loire, Seine, Ebro, and Susquehanna rivers).

For most variables, long-term averages are clustered by ecosystem type (Fig. ). The lowest chl a concentrations were found in rivers (7.8 ± 10.7 µgL-1) followed by estuaries (11.8 ± 9.9 µgL-1) and then lakes (18.0 ± 25.3 µgL-1). This greatly contrasted with most P, N, and oxygen time series: for instance, tp and tn distributions showed the highest ranges in rivers (0.13 ± 0.11 mgPL-1 and 3.1 ± 1.8 mgNL-1) and the lowest in lakes (0.06 ± 0.13 mgPL-1 and 1.9 ± 0.9 mgNL-1). For DOC, most time series remained within a similar range of values regardless of ecosystem type, except for four lakes located in the north-east US (global lake database; Naderian et al., 2024). The highest conductivity values appeared in estuaries, much higher than in rivers or lakes. There were only nine lakes with conductivity time series, which explains the density distribution peaks for this ecosystem type. The warmest waters were also found in estuaries.

Across the entire database, chl a annual averages showed moderate to strong correlation with tp and tn (Fig. ). Chl a was strongly and positively correlated with tp (Pearson, r = 0.39) across all ecosystem types. The positive correlation was the strongest for lakes (r = 0.82), moderate for rivers (r = 0.37), and the weakest for estuaries (r = 0.29). Chl a was positively correlated with tn (Pearson, r = 0.40), which was the highest in lakes (r = 0.75), moderate in rivers (r = 0.49), and the lowest in estuaries (r = 0.30). Variables tp and tn were positively correlated across all ecosystem types (Pearson, r = 0.59), with the strongest correlation found in lakes (r = 0.74), slightly lower in rivers (r = 0.63), and the weakest one in estuaries (r = 0.35). There was a clear cluster outlier for these variables in estuaries characterized by low chl a and tp but rather high tn. These observations corresponded exclusively to the Florida Coastal Everglades.

Comparing the mean value of annual averages between the second and the first halves of time series proved to be a simple but effective way to overview temporal behaviour of time series in the database. Across all variables and ecosystem types, 60 % of time series showed a lower average value in the second half. 63 % of chl a time series showed lower values in the second half (Fig. ). For N and P nutrient time series, 78 % to 87 % showed an average concentration that was lower in the second half (it was 85 %, 87 %, 78 %, 85 %, and 86 % for tp, po4, tn, din, and nh4, respectively). An exception was found for no3 with only 45 % time series having a lower concentration in the second half of the time series. Interestingly, we found that the majority (74 %) of tss time series had a lower concentration in the second half, whereas o2, o2sat, pH, and cond showed no clear differences in the second half of the time series with 49 %, 43 %, 42 %, and 42 %. For wtemp, there was a clear indication of a warming trend with 64 % of time series with higher averages in the second half of the time series.

The breakpoint and trend analysis (Fig. ) revealed 15 % of chl a time series were best represented with a segmented trend component, while 62 % had no trend detected, 18 % presented a monotonic declining trend, and 5 % showed a monotonic rising trend (predominantly found in estuaries; see Fig. ). The predominant segmented trend types were “00” (32 %), “0-” (21 %), “-0” (19 %) and “+-” (7 %).

For tp and po4, 29 %–31 % of the time series had a breakpoint with a segmented trend, and 26 %–32 % had no trend detected, while 35 %–42 % presented a declining monotonic trend, and 1 %–2 % were rising. For tp time series, 72 % of segmented trends had a declining trend type, while it was 65 % for po4 time series. Compared to rivers and estuaries, a lower proportion of declining tp trends were observed in lake time series.

For N species, time series were dominated by the no-trend type (38 %–61 %), and significant trends were contrasted: tn, din, and nh4 showed a large number of declining trends (36 %–42 %) and a small proportion of rising trends (less than 2 %), while no3 and no23 were characterized by a larger proportion of rising trends (7 % for no23 and 17 % for no3) and segmented trends (14 % for no23 and 25 % for no3). For no3, 57 % of segmented trends had a declining trend type on the most recent part of the time series as 34 % were “0-”, and 23 % were “+-”. Other variables were characterized by 50 %–60 % of no-trend time series. Interestingly, among the detected trends, tss showed a significant proportion of declining trend types, while o2, o2sat, pH, and wtemp showed a predominance of rising trends.

For chl a, Sen's slopes in estuaries were smaller in magnitude compared to lakes and rivers regardless of the trend type (Fig. a). Lakes exhibited a median Sen's slope of -0.7 µgL-1yr-1; it was -0.4 µgL-1yr-1 in rivers and -0.3 µgL-1yr-1 in estuaries. The fastest declines (below -4 µgL-1yr-1) were found in the Sacramento Bay in California; the Loire (France); and several shallow lakes in the Mississippi–Missouri Basin, the Denmark–Germany coast, and England and Wales. The largest positive chl a trends were found in rivers, with a median slope of 0.79 µgL-1yr-1 compared to 0.13 and 0.23 µgL-1yr-1 in estuaries and lakes, respectively. The fastest rises (above 4 µgL-1yr-1) were found in the Loire (France).

For tp, the fastest rises and declines were observed in river ecosystems (Fig. b) with median slopes of 4.0 × 10⁻³ and -4.7 × 10⁻³ mgPL-1yr-1, respectively, 1 order of magnitude greater than the slopes observed in lakes and estuaries. The fastest declines (below -0.1 mgPL-1yr-1) were observed in the Rhône and Seine rivers (France).

For tn, although the fastest declines were observed in estuary stations (Florida Coastal Everglades) down to -0.4 mgNL-1yr-1, the median value for declining slopes was overall faster in rivers with median slopes of -0.14 mgNL-1yr-1 (Fig. c). It was -6 × 10⁻³ mgNL-1yr-1 in estuaries and -7 × 10⁻³ mgNL-1yr-1 in lakes. Only 11 stations showed rising tn trends (Fig. ), and among them, 3 were in the Chesapeake Bay (US North Atlantic coast), which contrasted with the 145 other estuarine stations in this basin that either showed declining trends (n=89) or no trends (n=56). Note that only seven lacustrine stations showed rising tn and in rivers and that none of the tn time series showed a rising pattern.

We identified 444 stations with joint chl a, P, and N data over 15 years and with more than 6 observations per year. Among these, 100 (or 23 %) chl a time series showed a linear declining trend, 251 (or 57 %) had no trend, and 37 (or 8 %) were rising. Declining chl a time series were also linked to declining trends in N and P (Fig. a). Nearly half of the chl a time series with no trend had corresponding no-trend or declining patterns in nutrient time series (Fig. b). Rising chl a time series predominantly corresponded to no-trend or declining patterns in nutrient time series. Only 18 % of the rising chl a time series also had significant rising trends in N or P.

The development of primary producers is far more complex than a single relationship with nutrient availability, especially if one also considers the differences among ecosystem types. Hydraulic flushing, turbulence, exposition to solar radiation, temperature (e.g. Reynolds, 2006), and light climate (Hilt et al., 2011) are crucial environmental variables in lotic systems. Water residence time, internal loading (Jeppesen et al., 2005; Krishna et al., 2021), stratification regime, and underwater light climate are other crucial factors controlling lentic ecosystems (Donis et al., 2021). Such differences are also reflected in the OLIGOTREND database. For instance, on the one hand, rivers had the highest P and N concentrations followed by estuaries and lakes, and on the other hand, the highest chl a concentrations were found in lakes followed by estuaries and then rivers (Fig. ). Further, only 18 % of the chl a time series showed a linear declining trend, which contrasted greatly with a dominating decreasing trend for most nutrient concentrations (Figs. , , and ). Moreover, although lake time series showed the highest correlation between chl a and nutrients (Fig. ), they were also the ones with the highest proportion of non-significant trends (Fig. ). In this context, we argue that the OLIGOTREND database provides a unique opportunity and foundation to further investigate the ambiguous links existing between chl a and nutrient levels over many contrasted waterbodies located in basins with different environmental and climatic conditions.

The OLIGOTREND database evidenced different responses of the individual ecosystem types to nutrient declines (Figs. , , and ). For instance, compared to estuaries and lakes, rivers showed the highest proportion of declining chl a (Fig. ). The inherent specificities of different ecosystems could partly explain why oligotrophication seems to be ecosystem-specific: (i) the successful P reduction in many rivers worldwide (e.g. Le Moal et al., 2019) has led to more frequent P limitation for phytoplankton (Elser et al., 2007), although N or Si may also be limiting primary production (Paerl et al., 2016). (ii) In lakes, longer water residence time and internal nutrient loading can either delay (Jeppesen et al., 2005) or amplify (i.e. through algal blooms; e.g. Krishna et al., 2021) the ecological response following nutrient declines. (iii) Temporal shifts in phytoplankton assemblages towards taxa better adapted to low P levels, and taxa that are barely controlled by zooplankton grazing (e.g. filamentous cyanobacteria; Selmeczy et al., 2019) can often represent overlooked effects, explaining rising or weak trends in primary producers despite nutrient decline over time (Anneville et al., 2019). (iv) In estuaries, the dynamic of primary producers is also largely affected by marine waters, where coastal phytoplankton, sensitive to N (Elser et al., 2007), or N and P availability meets freshwater phytoplankton that is primarily sensitive to P (Kemp et al., 2005). Future analysis of OLIGOTREND time series together with catchment and waterbody attributes could improve our understanding of how aquatic ecosystems respond to nutrient trends in a wide variety of aquatic ecosystems.

The OLIGOTREND database could be explored to further evidence the extent of gradual changes or abrupt regime shifts in water quality time series. In fact, some of the waterbodies represented in OLIGOTREND are known for shifting their primary producer's structure and function following oligotrophication. This is the case of the Loire (France) and the Ebro rivers (Spain), which are known for their long-term gradual regime shifts from phytoplankton to macrophytes in response to phosphorus decline (Diamond et al., 2021; Ibáñez et al., 2012; Minaudo et al., 2015, 2021). Similarly, phytoplankton of the middle Danube now more frequently contains benthic taxa, predominantly diatoms, potentially indicating a long-term regime shift from pelagic to benthic production in recent decades (Abonyi et al., 2018). Moreover, oligotrophication can result in a shift from heterotrophic conditions to dominantly autotrophic processes with lower pollution, as observed for the Elbe River (Wachholz et al., 2024). OLIGOTREND time series could be further analysed to detect possible temporal changes in variance (as a possible early-warning signal; Dakos et al., 2015), seasonal patterns, and relationships between chl a, nutrients, and ecosystem metabolism. This could enhance our understanding of crucial factors underlying regime shifts in river ecosystems, which are comparatively less well known than in lakes (Gilarranz et al., 2022).

In OLIGOTREND, we highlighted a significant number of no-trend or rising chl a time series despite declining nutrient levels (Fig. c). This could be related to climatic effects and long-term changes in ecosystem structure, such as in the Chesapeake Bay (Harding et al., 2019). Future analysis of the OLIGOTREND will provide an invaluable source of data to disentangle the effects of climate change and watershed biogeochemistry on multi-decadal chl a and nutrient trends.

The OLIGOTREND database can help boost water quality research if it is combined with other large-scale or long-term ecological datasets. For instance, it is known that shifting baselines because of temporal changes in different, covarying environmental factors can preclude the return of the primary producer to pre-eutrophication conditions (Carstensen et al., 2011; Duarte et al., 2009). As global change intensifies, leading to novel ecosystems (Hobbs et al., 2009), the temporal extension of most available water quality datasets limits a correct estimation of pre-eutrophication baselines. Only a fraction of the OLIGOTREND database covers chl a and/or nutrients during the eutrophication phase, which renders pre-oligotrophication reference conditions impossible to discern and, hence, makes it difficult to validate nutrient remediation actions (Pinay et al., 2018). In this context, combining palaeolimnological observations with water quality monitoring data could have a potential not fully implemented at large spatial scales and across different aquatic ecosystem types (Bennion et al., 2015; Bhattacharya et al., 2022; Dong et al., 2012).

Recent research has shown that nutrient concentrations link to nutrient loads (point and nonpoint sources) at the catchment scale (Ehrhardt et al., 2021; Jarvie et al., 2012; Murphy et al., 2022). Yet, only a few studies have established a mechanistic link between nutrient input management and the development of the phytoplankton biomass. Data-based approaches that jointly analyse decreasing nutrient loadings over multiple decades and sites with corresponding measurements of chl a and nutrients can help better characterize how successful catchment management and environmental measures can affect reverse eutrophication. OLIGOTREND holds the potential to approach oligotrophication longitudinally at the basin scale, where long-term trajectories can be assessed from small streams, rivers, and lakes/reservoirs towards estuaries/coastal ecosystems along with their hydrologically connected time series.

Remote sensing could further supplement crucial water quality information organized in OLIGOTREND. Remote sensing can provide time series data on water quality for inland and coastal aquatic ecosystems, which, if combined with in situ measurements, can increase chl a data coverage both spatially and temporally (Ross et al., 2019; Spaulding et al., 2024). Moreover, regional and Earth system numerical models will improve further if calibrated or validated by in situ observations (Casquin et al., 2024; Liu et al., 2024). The OLIGOTREND database readily represents a centralized and harmonized dataset open for calibration and validation by remotely-sensed water quality data that is available for training and validating regional and large-scale numerical models.

Finally, there is a growing interest in large-scale observations that integrate new and existing databases to answer key questions in aquatic ecology (Barquín et al., 2015). Long-term observations of community data (e.g. via LTER and eLTER, GBIF, Biofresh) may include key functional groups of aquatic food webs, such as phytoplankton, zooplankton, macroinvertebrates (Welti et al., 2024), and fish (Comte et al., 2021). For a selection of sites, chl a trends can be further analysed jointly with long-term community data to investigate the role that community composition and biodiversity may play in responding to long-term environmental change (Jochimsen et al., 2013). Some of the OLIGOTREND time series are linked to lotic community data (i.e. phytoplankton), which have been seldom explored compared to lakes when testing the biodiversity effect on ecosystem functioning and services (Filstrup et al., 2019; Ptacnik et al., 2008).