(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Worldwide moderate-resolution mapping of lake surface chl-a reveals variable responses to global change (1997–2020) [1]

['Benjamin M. Kraemer', 'Igb Leibniz Institute For Freshwater Ecology', 'Inland Fisheries', 'Berlin', 'Karan Kakouei', 'Catalina Munteanu', 'Albert-Ludwigs-University Of Freiburg', 'Freiburg', 'Michael W. Thayne', 'Freie Universität Of Berlin']

Date: 2022-12 

We calculated long-term (1997–2020) chl-a trends using 23 years of remote sensing data for 344 lakes (listed in S1 Table ). Trends were calculated from chl-a anomalies where the effects of the day of the year, latitude, longitude, and remote sensing platform as well as the interactions between these variables had been accounted for and removed from the data. We calculated a single lake-wide trend where chl-a anomalies were pooled across pixels and seasons for each lake. We also calculated separate trends for each pixel (where chl-a anomalies were pooled across all days of the year) and for each day of the year (where chl-a anomalies were pooled across pixels). The size of the pixels was 1-km for lakes in Europe and 4-km for the rest of the globe based on the limitation of the data source used here [ 28 ].

We used 742 million chl-a estimates merged across 6 space-borne spectroradiometers (SeaWiFS, MODIS AQUA, MERIS, OLCI-B, VIIRS NPP, and VIIRS JPSS-1) to assess long-term trends in 344 lakes under ice-free and cloud-free conditions from year 1997 to 2020. Chl-a data were retrieved from the “CHL-OC5” product produced by GlobColour [ 28 ] and made available via the Copernicus Marine Environmental Monitoring Service (CMEMS) website: http://marine.copernicuseu/services-portfolio/access-to-products/ . Only the highest-quality chl-a values were used according to the quality criteria provided as part of the CHL-OC5 data product resulting the exclusion of 76% of data due to partial cloud or ice cover. The algorithm used in the CHL-OC5 data product is a five-channel chlorophyll concentration algorithm which was developed for optically complex “case II waters” [ 26 ] and has been partially validated using global in situ data from marine and inland waters [ 29 – 31 ]. We built on this validation by expanding it to in situ data from 53 lakes as described below. The daily chl-a data reflect lake environments in the near surface layer during ice-free and cloud-free conditions. The seasonal extent and the number of chl-a estimates varied across lakes ranging from 2601 estimates for Lake Mogotoyeyo to 365 million for Lake Ladoga. We downloaded and processed the chl-a values in the R environment for statistical computing [ 32 ] using the “data.table” [ 33 ], “dismo” [ 34 ], “sf” [ 35 ], “gbm” [ 36 ], “zyp” [ 37 ] and “lubridate” [ 38 ] packages. Data visualizations were made using “ggplot2” [ 39 ].

We modelled the difference between in situ values and their remotely sensed matchup values as a function of the raw remotely sensed chl-a value, the in situ data source, and 3 lake characteristics (mean lake depth, surface area, and shoreline development index (a metric of shoreline complexity; the ratio of a lake’s shoreline length to the circumference of a circle with the equivalent lake area)) also using a BRT. These three lake characteristic variables were used because they are associated with lake optical characteristics [ 40 ] and are freely available from the HydroLAKES database [ 41 ]. Thus, the BRT allowed the difference between in situ and remotely sensed chl-a values to vary from lake to lake. We then used the resulting model to estimate the difference between in situ and remotely sensed values for all 742 million chl-a estimates used here. We translated the remotely sensed chl-a values into an “in situ analogue” chl-a value by subtracting the modelled differences from the raw remotely sensed values. The resulting in situ analogue chl-a values were used for all subsequent analyses. The calibration of the algorithm using in situ chl-a data reduced the median absolute error from 3.6 μg L -1 to 1.5 μg L -1 .

We adapted remotely sensed chl-a values specifically for lakes based on a comparison of 20,165 in situ chl-a measurements (chl-a extracted from filtered water samples) from 56 lakes matched to interpolated remote sensing data. To interpolate the remote sensing chl-a values, we used fully deterministic BRTs (bag fraction = 1) which modelled remotely sensed chl-a as a function of the decimal date, day of the year, sensor, latitude, and longitude separately for each lake. We used the resulting boosted regression trees (BRTs) to estimate remotely-sensed chl-a concentrations for all 6 sensors at the time and location of each in situ measurement. These modeled values served as the remotely-sensed matchup value for each in situ measurement. In situ data used for this purpose were downloaded, digitized, and compiled from published sources ( S2 Table ; mean chl-a = 8.7 μg L -1 , median chl-a = 3.9 μg L -1 , range = 0.01 to 579.2 μg L -1 ). We would expect the calibrated chl-a algorithm used here to be less accurate near the tails and outside of the in situ chl-a distribution.

BRTs include a variety of tuning parameters which influence the model performance in cross validation. We selected a combination of tuning parameters (bag fraction = 0.62, tree complexity = 5) which reliably gave good performance in 10-fold cross validation across all lakes (median predicted residual error sum of squares = 0.35 μg L -1 , and median correlation between predicted and observed values = 0.85). We optimized the learning rate separately for each of the 10 BRTs for each lake by iteratively running the model with smaller and smaller learning rates (from 0.8, 0.4, 0.2, 0.1, 0.05, to 0.025) until the number of trees in the BRT which minimized the predicted deviance was greater than 1000 as suggested in previous literature [ 43 ].

To estimate lake surface chl-a anomalies, we accounted for and removed variation in each lake’s in situ analogue chl-a data which could be attributed to the day of the year, latitude, longitude, and sensor using BRTs. Thus, the chl-a anomalies were defined as the residuals from a BRT predicting in situ analog chl-a as a function of four predictor variables—the day of the year, latitude, longitude, and sensor. This approach allowed us to account for patchiness in the chl-a data [ 42 ] and bypass the need for gap-filling the time series. This allowed for the calculation of trends over time in chl-a anomalies such that the trends would be less influenced by the seasonal or spatial coverage of the data, the sensor timespan or the interactions among these factors. For the BRTs, we used a tree complexity of 5 to allow for high levels of interactions among variables (e.g. sensor type could have different effects on in situ analog chl-a at specific grid cells within lakes and at specific times of the year). To prevent model overfitting, the BRTs were fit separately for each lake 10 times using a randomly selected 50% of the data. Predicted values were generated from each of the 10 models for all chl-a estimates and averaged to generate more robust estimates. The residuals (i.e. difference between the in situ analog chl-a values and the average of the 10 model predictions) were termed, “chl-a anomalies.” We confirmed that this process successfully removed the variation attributable to the day of the year, latitude, longitude, and sensor using visual diagnostic plotting for all lakes (see example for Lake Erie in S6 Fig ).

We calculated a single lake-wide trend where chl-a anomalies were pooled across pixels and seasons for each lake. We also calculated separate trends for each pixel (where chl-a anomalies were pooled across all days of the year) and for each day of the year (where chl-a anomalies were pooled across pixels). Chl-a percent change was calculated from Theil-Sen nonparametric regression in chl-a anomalies after the effects of the day of the year, latitude, longitude, and sensor as well as the interactions between these variables had been accounted for and removed from the data. Theil-Sen nonparametric regression results were translated into a proportion change by taking the difference between the Theil-Sen modelled chl-a anomalies in 2020 and 1997 as a proportion of the lake’s median chl-a in situ analog value. This proportion was translated into a percentage by multiplying by 100. All stationary variation in chl-a due to space and season was removed using machine learning (BRT) prior to the calculation of trends, thus we would not expect strong bias in the chl-a trends related to ice/cloud cover.

We used a bootstrapped error propagation technique [ 27 ] to estimate chl-a trends and the uncertainty in each trend. The residual errors from the BRT used for adapting remotely sensed chl-a into in situ analogue values were propagated into the estimate of chl-a trend uncertainty. We propagated the errors by adding a residual error (the product of a randomly selected % error residual from the BRT’s error distribution and the original in situ analogue chl-a value) to each chl-a anomaly. Distinct residual errors were iteratively added to each chl-a anomaly value with 100 repetitions. For each repetition, Theil-Sen slopes and intercepts were calculated based on mean annual chl-a anomalies. We calculated p-values of a Spearman rank correlation test relating mean chl-a anomalies to year. The trends and their associated p-values were calculated as the average across all 100 repetitions of the Theil-Sen slope and Spearman correlation calculations.

Results and discussion

We found that when chl-a trends were calculated as lake-wide averages, the median chl-a trend across all lakes was +0.04 μg chl-a decade-1 (total range = -11.98 to +9.64 μg chl-a decade-1, interquartile range = -0.03 to 0.27 μg chl-a decade-1, median % change = +3.2; Fig 1A). Lake-wide average chl-a increased in 65% of lakes (225 out of 344; Fig 2) and decreased in the remaining 35%. Increasing trends in 142 lakes and decreasing trends in 68 lakes had p-values less than 0.1. Thus, more lakes had trends with p-values less than 0.1 than what would be expected based on chance alone and this was true regardless of which arbitrary threshold was used (we tested p-value cutoffs of α = 0.1, 0.05, 0.01). These general findings are comparable to previous work based on Landsat imagery which found that 68% of large lakes had increasing chl-a concentrations since the 1980’s and 39% of lakes had trends with p-values less than 0.1 [2].

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Chl-a percent change (1997–2020) based on lake-wide averages. Map showing lake-wide trends for 344 globally distributed lakes where the sizes of the dots are proportional to the size of the lake and the transparency of the dot is proportional to the p-value (A). Time series show the chl-a anomalies for the 6 lakes with the largest negative trends (B:G) in lake-wide chl-a and the 6 lakes with the largest positive trends (H:M) in lake-wide chl-a when trends were weighted by each lake’s volume. The black lines (B:M) are the ordinary least squares regression line and the size of the dots represents the number of chl-a estimates in each year (B:M). An inset map of Europe can be found in S1 Fig. Continent boundary map data come from Natural Earth. (http://www.naturalearthdata.com/about/terms-of-use/; public domain). https://doi.org/10.1371/journal.pwat.0000051.g001

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Distribution of global chl-a percent change (1997–2020). Percent change at the lake-wide average scale (A, B), and at the lake area scale based on pixel-level trends (C, D). Blue indicates decreasing trends and green indicates increasing trends. The solid portion of each bar represents the proportion of trends with p-values less than 0.1. https://doi.org/10.1371/journal.pwat.0000051.g002

The lake-wide average trends agreed well with published literature on phytoplankton, chl-a, and other phytoplankton proxies based on in situ data sources. For example, in situ phytoplankton biomass proxies decreased in Lake Tanganyika as enhanced thermal stratification due to climate change has reduced the entrainment of deep-water nutrients and primary production [11, 12, 44] (Fig 1). Conversely, chl-a variation in Lake Kivu shown here (Fig 1) matched a strong cycle in phytoplankton biomass observed in situ due to background climate variation [45]. Remotely-sensed chl-a also increased in Lake Erie which has experienced a well-documented nutrient pollution trend for decades due to agricultural expansion and intensification [46] (Fig 1).

While 63% of lakes had positive chl-a trends at the lake-wide average scale, only 44% of the total lake area experienced increases in chl-a at the pixel scale. This difference arose because chl-a tended to decrease in the largest lakes (Fig 1A). The global median chl-a trend was -0.01 μg chl-a decade-1 when pixel-level trends were weighted by pixel area (pixel area varies across latitude when a consistent equal angle grid is applied). In total, we investigated chl-a trends at the pixel scale for 1,383,893 km2 of cumulative lake area. Chl-a increased for 612,014 km2 of the cumulative lake area and 17% (104,871 km2) had a p-value less than 0.1. Chl-a decreased across 771,879 km2 of the cumulative lake area and 19% (148,175 km2) of the lake area with decreasing trends had p-values less than 0.1. Thus, more lake area had trends at the pixel scale with p-values less than 0.1 than what would be expected based on chance alone which was also true using a cutoff of 0.05, and 0.01. This finding (Fig 2D) challenges the broadly held assumption that global change has caused widespread increases in lake chl-a [1].

Decreases in chl-a observed here have been attributed to water management efforts in some lakes (e.g. Ladoga [47]), and to invasive filter feeding mussels in others (e.g. Ontario, Huron, Michigan [21]), but these drivers alone cannot explain decreases across all lakes. Decreases in chl-a as a result of climate change through changes to terrestrial inputs [48], higher heterotrophic consumption of algae [49, 50], or climate change mediated reductions in deep water nutrient entrainment to the well-lit surface waters [4, 51–53] may be an underrecognized response to global change. While this result is most relevant for Earth’s large lakes, widespread bluing trends have been observed in small lakes as well [48, 54, 55]. We recommend future emphasis on how global-change may reduce chl-a, especially because chl-a reductions can heavily affect lake ecosystems and the benefits that humans derive from them [20–22].

Long-term trends based on lake-wide averages hid the complexity of long-term change in chl-a across lake surfaces. Some lakes which had weak overall chl-a trends at the lake-wide scale, had strong simultaneous increases and decreases in chl-a at different locations within the lake (Fig 3). For example, the Caspian Sea had a weak overall trend at the lake-wide scale but chl-a increased substantially in the northern shallows (Fig 3). Increasing trends in the Caspian Sea primarily occurred near the inflowing Ural, Volga, and Terek rivers which deliver large nutrient loads from the surrounding landscape [56] (S2 Fig). At the lake-wide scale, this increase in chl-a in the northern shallows was offset by the statistically weaker but widespread decrease in chl-a in the deeper offshore areas (Fig 3).

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Spatial variation within and across lakes in chl-a trends (1997–2020). The relative positions of the lakes shown here do not reflect geo-spatial location but lake sizes share a common scale except where indicated otherwise in the boxed insets. Lakes shown are those with the largest number of chl-a estimates. The opacity of the color scale reflects the p-values associated with each pixel-level trend. The p-values associated with each grid cell can be found in S3 Fig. Lake boundary map data come from the HydroLAKES database v1.0 which is licensed under a Creative Commons Attribution (CC-BY) 4.0 International License. (https://www.hydrosheds.org/products/hydrolakes). https://doi.org/10.1371/journal.pwat.0000051.g003

In addition to the Caspian Sea, many other lakes experienced local increases in chl-a in the shallower areas closer to the shore near inflowing rivers that diverged from lake-wide average trends (Fig 3). For example, Lake Huron experienced very localized increases in chl-a in the shallow Saginaw Bay where the Saginaw River enters the lake bringing with it a variety of agricultural and industrial pollutants [57] (S2 Fig). At the same time, chl-a decreased in the offshore deeper areas of Lake Huron where the combined effect of nutrient mitigation and invasive species expansion has caused reductions in chl-a [21]. Lake Titicaca experienced very localized increases in chl-a in the smaller south east basin where inflowing rivers drain an increasingly urbanized catchment near El Alto/La Paz, Bolivia [58] (S2 Fig). A strong reduction in chl-a concentrations in Lake Ladoga (Fig 3) suggests that continued efforts at reducing nutrient pollution have improved water quality there. However, patchy greening areas in the shallower parts of Lake Ladoga may indicate the potential for internal re-suspension of past phosphorus loads resting in the sediment [59]. Lake Nasser/Nubia, one of the largest manmade lakes in the world, has a pronounced greening trend near the inflowing Nile River which transitions into a bluing trend near the outflow of the reservoir. This transitioning from chl-a increases to decreases along the axis of water flow was common in large reservoirs presumably because nutrient loads from incoming rivers get diluted as they pass into larger volumes of water with greater thermal stratification. Thus, simultaneous increases and decreases in chl-a for specific locations within lakes can reflect differences in the location and strength of various anthropogenic stressors affecting chl-a. Chl-a changes at the local scale within lakes may diverge widely from the lake-wide average.

Long-term trends based on lake-wide averages also failed to capture the complexity of long-term change in lake surface chl-a across days of the year for individual lakes. For example, in Lakes Michigan, Huron, and Ontario, introduced mussels have reduced chl-a primarily in the less stratified times of the year when the impact of filter feeding on phytoplankton is strongest [19] (Fig 4). Chl-a decreases in Lake Tanganyika were most pronounced during the dry season when climate change mediated changes to stratification patterns have dampened seasonal mixing [11, 12]. Several lakes had serial increases and decreases in chl-a across the seasonal spectrum (Fig 4). For instance, earlier seasonal stratification in Lake Superior due to climate change may have increased chl-a in early spring when phytoplankton benefit from being retained in the well-lit stratified surface layer where they can better photosynthesize. But the increased chl-a in early spring in Lake Superior is combined with a summer decrease in chl-a as prolonged summer stratification reduces the amount of nutrients available for primary producers [60].

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. Seasonal variation in lake chl-a trends (1997–2020). Panels (A:J) show seasonal trends in chl-a for the 10 largest lakes with broad seasonal data coverage. Lines represent the LOESS-smoothed (span of 1) daily trend in chl-a weighted by the number of chl-a estimates included in the chl-a trend calculation for each day of the year. The blue to green color reinforces values on the y-axes and is consistent across panels. The shaded area represents the 95% confidence interval for the smoothed line. https://doi.org/10.1371/journal.pwat.0000051.g004

This within lake complexity (Fig 3, S4 Fig) and seasonal complexity (Fig 4) of chl-a trends discounts the tendency in lake management to emphasize the dichotomous view that lakes are either undergoing increases or decreases in chl-a. 19% of lakes exhibited positive and negative trends in chl-a at different locations or at different times of the year (i.e. more than 10% of positive trends and more than 10% of the negative trends had p-values less than 0.1). The lakes with both positive and negative trends represented 63% of the lake area. This combination of positive and negative trends for individual lakes highlights the heterogeneity of lake responses to global change. Whole-lake chl-a trends can mask spatiotemporal trend variability leading to lakes with weak trends at the lake-wide scale but strongly contrasting increases and decreases in chl-a at finer seasonal and spatial resolution (e.g. Caspian Sea in Figs 3 and 4). Policy makers and lake managers should recognize the potential for local changes at specific times of the year to vary widely from lake-wide averages. Major water management policies such as the Water Framework Directive (EU) and the Clean Water Act (USA) have established management targets based on lake-wide and annual averages which fail to capture the spatiotemporal complexity shown here. Instead, management decisions or policies should be based on finer metrics (i.e. not lake-wide or annual averages) which more fully recognize spatiotemporal variation in water quality and water quality trends.

Overall, we provide a global view of trends in near surface chl-a over the past 23 years for large lakes. Our analysis of chl-a in lakes demonstrates the promise of spatially-explicit long-term satellite observations for tracking chl-a conditions and disentangling the multiple overlapping drivers of change. The approach used here augments the geographically and temporally limited in situ chl-a monitoring efforts where data is often not shared readily nor in near real time. This analysis applies a novel statistical approach to merge chl-a data from multiple sensors to produce a validated data set that documents global lake surface chl-a dynamics with new levels of spatial detail and accuracy. Global datasets documenting the aquatic concentration of chlorophyll-a (chl-a) with remote sensing can serve as a key indicator of lake responses to human activity but so far have been underused. The accuracy of remote sensing based chl-a estimates has been questioned at the local scale due to the presence of surface algae scums, submerged vegetation, sediment, and high concentrations of organic matter [61–63]. However, the cross validation and the performance of the lake-specific chl-a algorithm developed here (S5 Fig) suggests that our main conclusions are robust to such effects. Nonetheless, we caution against the overinterpretation of specific trend estimates reported here and suggest corroborating observed trends in chl-a with in situ measurements wherever possible. In the frequent absence of such in situ measurements, remote sensing data of this type may provide the best approximation of global scale trends yet available.

[END]
---
[1] Url: https://journals.plos.org/water/article?id=10.1371/journal.pwat.0000051

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/