Long‐Term Trends in Productivity Across Intermountain West Lakes Provide No Evidence of Widespread Eutrophication

Eutrophication represents a major threat to freshwater systems and climate change is expected to drive further increases in freshwater primary productivity. However, long‐term in situ data is available for very few lakes and makes identifying trends and drivers of eutrophication challenging. Using remote sensing data, we conducted a retrospective analysis of long‐term trends in trophic status among lakes greater than 10 ha across the Intermountain West, a region with understudied water quality trends and limited long‐term data sets. We found that most lakes (55%) were not exhibiting shifts in trophic status from 1984 to 2019. Our results also show that increases in eutrophication were rare (3% of lakes) during this period, and that lakes becoming increasingly oligotrophic were more common (17% of lakes). Lakes that were not trending occupied a wide range of lake and landscape characteristics, whereas lakes that were becoming more oligotrophic tended to have larger residence times and were located in catchments with greater clay content and more development. Our results highlight that while there are well‐established narratives that climate change can lead to more eutrophication of lakes, this is not broadly observed in our data set, where we found more lakes in the Intermountain West becoming more oligotrophic than lakes becoming eutrophic.


Introduction
Widespread eutrophication is a global phenomenon that threatens water quality, recreational industries, and ecosystem function (Amorim & Moura, 2021;Gatz, 2020;Paerl et al., 2001).A common outcome of eutrophication is an increase in the biomass of phytoplankton, both algae and cyanobacteria, in freshwater, transitional, and ocean environments (Anderson et al., 2008;Hudnell, 2010;Wurtsbaugh et al., 2019).In many cases, this rapid and excessive growth can become severe and lead to the development of Harmful Algal Blooms (HABs) (Heisler et al., 2008;Smith, 2003).HABs are of particular concern due to the threats they pose to human health and drinking water sources (Christensen & Khan, 2020;Falconer & Humpage, 2005;Fleming et al., 2002).Thus, the wide-ranging effects that eutrophication and HABs have on aquatic systems and their threat to human health have highlighted the need to understand the factors which drive them.
Generally, eutrophication and algal blooms are attributed to excessive loading of nitrogen (N) and phosphorus (P) as well as high water temperatures (Beaver et al., 2018;Gobler et al., 2016;Paerl & Paul, 2012;Rejmánková & Komárková, 2005).However, in shallow lakes, warmer temperatures and higher light absorption have been found to be more significant drivers of productivity (Kosten et al., 2012).In other words, the combination of factors that drive rapid increases in lake productivity may differ between individual water bodies or geographic regions, hence smaller and more focused state and regional scale studies may be more useful in describing changes in lake productivity dynamics (Oleksy et al., 2022).
Large scale studies have highlighted that water quality trends are context dependent and vary across regions (Beaver et al., 2018).However, some regions with unique landscape features remain understudied regarding lake productivity trends.For example, the Intermountain West region (including the US states Colorado, Idaho, Montana, Utah, and Wyoming) has very different hydrological dynamics and landscape features compared with other regions, yet water quality trends remain mostly undocumented.The region undergoes quick wet-dry seasonal transitions, with most of the streamflow generated by snowmelt (Bales et al., 2006).Higher gradients in temperature and precipitation with elevation make hydrologic processes significantly different compared with lowelevation regions (Bales et al., 2006).Land use in this region also differs, with substantial amounts of grassland pasture and range contributing to increased organic nutrient loading to streams and rivers (Agouridis et al., 2005).
An increase in awareness and reporting of HABs in the Intermountain West suggests that lakes in the region may be becoming more eutrophic, yet our understanding of lake productivity trends is very limited.As nation-wide research and understanding of HABs has grown, so have management and sampling plans, educational materials, and overall public awareness (Hudnell, 2010).However, this increase in awareness and reporting has the potential to create a perception that blooms are increasing in intensity and frequency (Hallegraeff et al., 2021).Recent analyses of lake color trends in the region highlight that lakes are experiencing roughly equal trends of changing from blue to green or changing green to blue, indicating there is not overwhelming evidence that they are becoming more eutrophic (Oleksy et al., 2022).It remains unclear whether recent trends in algal blooms are a result of representative increases in intensity or a result of heightened monitoring.Therefore, retrospective data analyses and long-term monitoring are needed to identify consistent productivity trends (Hudnell, 2008), particularly in understudied regions like the Western US.
Remote sensing and long-term satellite imagery create opportunities to address key research gaps surrounding what factors are driving freshwater productivity across regions.In situ sampling methods are often limited by resources such as time and funding.Therefore, in situ water quality data tends to be focused on relatively large lakes (>20 ha) and long-term records are rare (Stanley et al., 2019).Importantly, leveraging remote sensing data can address water quality dynamics over large spatial and temporal scales where in situ data is lacking (Topp et al., 2020).Remote sensing data with high spatial and temporal coverage are also useful to understand how global change is affecting productivity and bloom dynamics (Harvey et al., 2015;Ho et al., 2017;Seegers et al., 2021).These tools can be used to determine water quality parameters in freshwater systems such as chlorophyll-a (Boucher et al., 2018;Kuhn et al., 2019;Papenfus et al., 2020), suspended sediments (Pavelsky & Smith, 2009), and organic matter (Kutser et al., 2005;Slonecker et al., 2016).
In this study, we address two gaps in our understanding of lake productivity dynamics in the Intermountain West.Specifically, we aimed to identify (a) whether long-term trends in trophic state suggest an increasing trend of eutrophication, and (b) the environmental factors driving trophic state change over the past several decades.We use remote sensing imagery and in situ chlorophyll-a data, covering 1984-2019, to create a model that predicts lake trophic state based solely on satellite imagery.This approach allowed us to document productivity trends in 1,169 lakes over 35 years.By increasing the level of understanding of historical trends in lake productivity and their drivers in this region, our analysis can also shed light on the intensification of algal blooms in lakes and provide important information for water quality management.

Data Sources and Processing
Our analysis used various remote sensing, water quality, lake and landscape features, and climate data sets.We opted for a machine-learning approach that uses paired satellite reflectance from Landsat observations and in situ water quality data.We acquired Landsat data and in situ chlorophyll-a samples for model training from the AquaSat data set (Ross et al., 2019).AquaSat joins Landsat Tier 1 surface reflectance to water quality samples from the Water Quality Portal (Read et al., 2017) and LAGOS-NE (Soranno et al., 2017) that occurred ±1 day of a Landsat observation.We filtered AquaSat to only include water quality portal observations over the Intermountain West region and Landsat scenes with less than 50% cloud cover.The resulting data set included 1,340 observations across 249 lakes in the region (Figure S1 in Supporting Information S1).
To capture environmental drivers that might be important for predicting and explaining productivity trends, we utilized the following open-source data sets.We merged lake characteristics and catchment level metrics to our training data set from the LakeCat (Hill et al., 2018), LAGOS-US (Cheruvelil et al., 2021), and HydroLAKES (Messager et al., 2016) data sets (Figure S2 in Supporting Information S1).Initially we joined lakes in the training set to corresponding lake polygons included in NHDPlusV2.The LakeCat, LAGOS-US, and HydroLAKES data sets were then added through common NHD identifiers.We then selected metrics that were derived from these data sets based on their potential to drive trends in lake productivity (Table 1).
Daily surface water temperature and corresponding weather data (wind speed) were also included in our model development.We extracted daily water temperature from Willard et al. (2022), which includes estimated daily surface water temperature for 185,549 lakes across the US.In addition to daily surface temperature, we calculated prior 14-day mean temperatures for all 1,340 observations included in our training set.Then, we joined 14-day mean temperature and meridional wind speed to our training set using common NHD identifiers and the date of observation.
Using the same methods, we built our prediction data set using LimnoSat-US (Topp, Pavelsky, Dugan, et al., 2021).LimnoSat-US includes Landsat Collection 1, Tier 1 surface reflectance for lakes greater than 10 ha in the U.S. spanning 1984-2020.As with the AquaSat data set, we filtered LimnoSat-US based on Landsat scenes with less than 50% cloud cover.Surface reflectance values represent the median surface reflectance of a 120-m buffer of the "deepest point" of a lake polygon.The "deepest point" is defined as the center of the largest circle that can fit within a lake polygon.This was implemented in an effort to reduce processing time while still capturing surface reflectance that is representative of whole-lake averages (Topp, Pavelsky, Dugan, et al., 2021).However, this buffer area does not capture shoreline environments, and thus our limited in capturing locations where algal blooms are relatively small and located in shoreline areas.We joined the lake characteristics, catchment level metrics, and climate data described above to our prediction data set, resulting in 1,264,355 observations across 2,596 lakes in the Intermountain West.The average amount of time between observations was 27 days.Given the differences between the spectral response of Landsat sensors, we used this comprehensive set of surface reflectance observations to standardize surface reflectance across sensors.Recent studies have found that these differences, if not corrected for, can distort time-series analyses and produce misleading downward trends (Maciel et al., 2023).This apparent bias has the potential to have widespread effects on the results of trend analyses of water quality data, where observed improvements in water quality trends (downward trends) is a product of these sensor differences.Given these concerns, methods that account for this bias and standardize surface reflectance products across the Landsat archive have been developed for use in terrestrial and aquatic systems (Roy et al., 2016, Gardner et al., 2021).Reflectance values were standardized following the method outlined in Gardner et al. (2021), where values from Landsat 5 and 8 are corrected to match Landsat 7. Because Landsat 7 overlaps with both Landsat 5 and Landsat 8, we compiled the reflectance values over sites during the two time periods where multiple sensors were active (1999-2012 for Landsat 5 and Landsat 7 and 2013-2018 for Landsat 7 and Landsat 8).Then, we used a second-order polynomial regression of the 1%-99% percentiles to generate correction coefficients for each band pertaining to Landsat 5 and Landsat 8. Correction coefficients were applied to all surface reflectance data used in this analysis and as a result reflect a more harmonious time series (Figure 1).
Lastly, we defined categories for three trophic states based on the following chlorophyll-a thresholds: oligotrophic (0-2.6 μg/L), mesotrophic (2.7-7 μg/L), and eutrophic (>7 μg/L).These thresholds were taken from the criteria outlined in the National Lakes Assessment (U.S. Environmental Protection Agency, 2009).This categorical approach was taken because predicting chlorophyll-a concentrations in freshwater systems with remote sensing has been notably challenging, particularly with Landsat imagery (Salem et al., 2017;Smith et al., 2021).Landsat bands are relatively broad with a low signal-to-noise ratio, often resulting in predictions of chlorophyll-a with high levels of uncertainty (Matthews, 2011).Furthermore, the accurate prediction of chlorophyll-a is affected by complex optical conditions in various waterbodies with higher levels of turbidity (Alvain et al., 2005;Tilstone et al., 2001).These challenges were addressed by focusing on broad, trophic level predictions of chlorophyll-a.Water Resources Research 10.1029/2023WR034997

Model Development
We developed an Extreme Gradient Boosting (XgBoost) model to classify categories of chlorophyll-a.These models build on machine learning concepts such as decision trees and ensemble learning (Chen & Guestrin, 2016).Decision trees represent a supervised learning approach where training features are split into internal nodes and evaluated to form homogeneous groups (terminal nodes) (Kotsiantis, 2011).Decision trees can comprise a single univariate classifier or the combination of multiple classifiers, known as an ensemble classifier.Gradient boosting is a method of ensemble learning where a series of models are built with weights assigned to misclassified observations (cases where predicted classes differ from observed classes).Misclassified observations from the previous model are used as training data for the next, and the result is an ensemble classifier that represents an aggregation of individual classifiers and minimizes overall error (Pal, 2007).
We used a combination of optical and climatic variables to build a predictive model for chlorophyll-a.Specifically, we calculated multiple band ratios that have been shown to explain variation in phytoplankton blooms (Ho et al., 2017).We used a 14-day average of lake surface temperature and daily meridional wind speed as additional predictor variables.We tested the addition of static predictor variables (such as lake elevation or watershed land use) yet refrained from including these in our final model because recent studies have shown that static predictor variables can act as "identifiers" and lead to overfitting and over-optimistic evaluation metrics (Meyer et al., 2018).After verifying that the addition of static variables did indeed result in overfit models, we selected only continuous predictor variables that we would not expect to lead to substantial overfitting (Table 2).
We partitioned our training set to reserve 20% for model testing and evaluation and 80% for model training and parameter tuning.XgBoost models include a wide range of hyperparameters and are one of the main tools used to reduce model variance.Hyperparameters were tuned by first establishing a grid of conservative values (to prevent overfitting) and then extracting the hyperparameters that resulted in the lowest validation loss.After training the final model with these hyperparameters, model performance was evaluated through a confusion matrix which shows the relative accuracy of predictions across different categories.

Data Analysis
To summarize lake trends and capture long-term changes in chlorophyll-a, we analyzed the percent occurrence of trophic state observations.First, lakes included in our trend analysis had to have at least two summer observations (June-September) for each year .More conservative filtering criteria, such as at least five observations per year, was explored yet had negligible effects on overall results and resulted in fewer lakes being included in our analysis.The median number of observations per year was 8, which represents a significant improvement in the average number of observations (n = 3) that have been reported for in situ water quality data sets (Stanley et al., 2019).We specifically focused our analysis on summer observations to limit the effect that snow and ice, which typically have much higher reflectance values than water, may have on our analysis of surface water.As a result, 1,169 lakes were included in our analysis after filtering the initial 2,596 based on these criteria.For each summer, the percent occurrence of each trophic state observation was recorded.Then, the average percent occurrence for each trophic state was recorded across two time periods.We split our data set into a historical  and contemporary (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019) record.This choice was made in an effort to preserve the evenness between two subsets of our data, given that early periods of the Landsat record contain fewer observations.Overall, the historical period contained 40% of all data and the contemporary period contained 60%.Importantly, we did test other ways of splitting our data, such as splitting our data set into decades (i.e., 1990-2000 and 2000-2019) and observed similar results.Lastly, lakes were grouped into the following categories based on the shift (if any) in trophic state during these two time periods: 1.No trend: Change in % oligotrophic, % mesotrophic, and % eutrophic was less than 10% across all three categories (Figure 2a) 2. Increasing in % Eutrophic: Number of eutrophic observations increased by 10% or more while the number of oligotrophic observations decreased by 10% or more (Figure 2b) 3. Increasing in % Oligotrophic: Number of oligotrophic observations increased by 10% or more while the number of eutrophic observations decreased by 10% or more (Figure 2c).
Lastly, trend-specific drivers were examined by determining how lake catchment, hydrologic, and climate metrics explained differences across trends.We calculated variable importance across trend categories by applying a random forest model using the randomForest package in R (Liaw & Wiener, 2002).With this approach, we were able to classify the reduction in accuracy that occurred across all three responses when certain variables were excluded.All data processing, model development, statistical analysis, and visualizations were done in Program R (R Core Team, 2022).

Model Performance
Model performance was evaluated through a confusion matrix as well as various accuracy and error metrics (Table 2, Figure 3).In the range of oligotrophic values (0-2.6 μg/L), observations had a balanced accuracy (mean of model sensitivity and specificity) of 78% and only 7% of these observations were misclassified as eutrophic (Table 3).Mesotrophic observations (2.7-7 μg/L) represented the range of values with the lowest prediction accuracy.Our model reported a balanced accuracy of 69% for mesotrophic classifications (Table 3).The most common misclassification among mesotrophic predictions was with observed classes that were oligotrophic (30%) (Figure 3).Lastly, eutrophic observations (>7 μg/L) represented the class with the highest prediction accuracy (85%) (Table 3).In addition, similar to the rare cases of oligotrophic predictions corresponding with eutrophic observations, we observed relatively low prediction error associated with eutrophic predictions and oligotrophic observations (6%).Overall, our model reported a global accuracy of 70% with a 95% confidence interval of between 63% and 76% (Table S1).
The integration of fine-scale, daily temperature and climate features significantly improved our ability to predict across these trophic states.In terms of feature importance measured by model gain, mean 14-day surface water temperature and meridional wind speed were the second and fourth most important predictor variables, behind the band ratio of blue to green and dominant wavelength (Figure 4).In addition, model scenarios without climate variables reported global accuracies of around 63%, with a 95% confidence interval of between 57%-69%.

Productivity Trends
Most lakes included in this study did not show trends in trophic status (Figure 5).Overall, a total of 651 lakes (55%) did not meet our 10% thresholds for shifts across all three categories.More than half of the lakes that weren't changing from 1985 to 2019 were oligotrophic lakes with most observations classified as oligotrophic.In

Water Resources Research
10.1029/2023WR034997 contrast, 24% of lakes within this category were eutrophic lakes.The remaining lakes (16%) in this trend category likely represent a more complex, mesotrophic lake status.
The second most common finding we observed were lakes that had substantial shifts in trophic status by becoming more oligotrophic.We found that 17% of lakes had trends in increasing oligotrophic observations.Most of these lakes tended to be dominated by eutrophic observations, suggesting that they are eutrophic lakes that are improving in water quality.Few lakes showed evidence of shifts in oligotrophic observations greater than 30%, indicating that extreme shifts were rare (Figure S1 in Supporting Information S1).In other words, shifts in oligotrophic observations within this lake trend was relatively moderate (10%-30%, Figure S3 in Supporting Information S1).
Lastly, a minority (3%) of study lakes were shifting toward becoming more eutrophic.Interestingly, these trends were equally distributed across lakes with high numbers of eutrophic observations and those with high numbers of oligotrophic observations.In other words, lakes that were predominately oligotrophic and were becoming more eutrophic were equally as common as lakes that were eutrophic and were intensifying in this way.The magnitude of change was similar to that of lakes that trended oligotrophic, with little evidence of extreme shifts in eutrophic observations (Figure S3 in Supporting Information S1).
The remaining lakes that did not fit into these rigid categories reflect various levels of trophic state change that do not reflect eutrophication trends.For example, 7% of lakes could be described as becoming more oligotrophic and less mesotrophic by the same thresholds outlined in Figure 2. In contrast, few lakes (1%) were found to be becoming more mesotrophic during this time.
The 12% of lakes that did not fit into these categories displayed slight trends in certain categories (such as becoming more oligotrophic), but did not satisfy thresholds for trends in other categories such that we would be confident of defining clear trends in productivity.

Drivers of Trends
Our random forest model was able to identify partially important variables for explaining trends in productivity.Lake catchment data such as 30 years normal mean temperature, clay content, base flow index, residence time, and mean runoff were more important in explaining overall lake trends (global model) (Figure 6).In addition, we explored variable importance within classes to determine how important each variable was within each class.The top five important variables in the global model (listed above) displayed some of the highest mean decrease in accuracy in the no trend and trending oligotrophic classes (Figure 7).For example, lakes becoming more oligotrophic tended to have longer residence times and were located in catchments that were generally less forested (11% compared to 33% among no trend lakes) (Figure 7).However, our model struggled to explain variable importance within the trending eutrophic class, as shown by the relatively low mean increase in accuracy across the top five variables (Figure 7).Additionally, most of the important variables in this class differed from the globally important variables and included mean wet deposition of nitrate, catchment area, % wetland, and % agricultural land on slopes greater than 20%.Forested land use was the only important global variable which was also important for explaining the trending eutrophic class.Lakes that were becoming more eutrophic also tended to be less forested but were located in smaller catchments and were shallower on average (4.13 m) compared with lakes that were not trending (9.12 m).Lastly, a number of climate and landscape metrics displayed a high level of variation across trophic state trends, however some of these metrics had significant cross correlation with other variables (Figure S4 in Supporting Information S1).

Discussion
Eutrophication and the development of algal blooms are global phenomena that threaten aquatic systems.Given the effects of global change and the expected increasing intensity of these disturbances, there has been a substantial level of interest in investigating recent productivity trends in lakes and reservoirs.Our analysis found that most lakes in the Intermountain West region have remained relatively static in terms of their productivity over the last 35 years.In addition, we found that a greater percentage of lakes were improving with regards to productivity, as opposed to becoming more eutrophic.

Modeling Approach
Our research focused on leveraging long-term remote sensing and environmental data sets that would supplement the ongoing debate regarding recent trends in phytoplankton blooms.While the application of remote sensing for inland water quality monitoring has grown over the past decade (Topp et al., 2020), the retrieval of certain optical properties such as chlorophyll-a has remained a challenge (Matthews, 2011).However, by incorporating daily surface temperature and meridional wind speed from data sets leveraging modern deep learning techniques we were able to show substantial improvements in model accuracy.The incorporation of fine-scale lake  climate data over the 35-year time span of this study was instrumental to our ability to document trophic state changes and add evidence to the ongoing debate regarding the recent trends in increasing eutrophication and HABs.
Most notably, surface water temperature was the second most important predictor variable of our trophic state model and could be important for a wide range of remote sensing based water quality models.Water temperature has proven to be an important predictor of chlorophyll-a across inland lakes (Karcher et al., 2020;Liu et al., 2019) as well as oceans (Dunstan et al., 2018).However, applied remote sensing models that predict chlorophyll-a are often limited to strictly optical predictors such as band-ratio (blue-green) models.These models work well in waterbodies where other parameters such as colored dissolved organic matter co-vary with chlorophyll-a (O' Reilly et al., 1998).However, in optically complex waterbodies with higher levels of turbidity and dissolved organic matter band-ratio models struggle to accurately retrieve chlorophyll-a concentrations (Tzortziou et al., 2007;Witter et al., 2009;Zheng & DiGiacomo, 2017).Thus, relying on surface reflectance for predictive models has resulted in a lack of generalizability across a wide range of waterbodies.However, the incorporation of surface water temperature seems to have supplemented existing band-ratio features to better predict across a wide range of lake types.
Wind speed was another climate predictor variable that was substantially important in predicting trophic state.
Correlations between wind speed and chlorophyll have been shown using remote sensing at global scales (Kahru et al., 2010).In addition, wind speed has been documented as an important driver of cyanobacterial bloom development with blooms favoring warm, calm weather (Kanoshina et al., 2003).Overall, the integration of daily, fine-scale weather data greatly improved our ability to predict trophic state and is likely to have a positive impact on similar approaches that leverage remote sensing data.

Productivity Trends
The majority of lakes included in this analysis showed no evidence of substantial changes in trophic state and supplement other regional-scale analyses of in situ chlorophyll-a data.This is consistent with previous analyses demonstrating that magnitude, severity, and duration of algal blooms are not intensifying in US lakes (Wilkinson et al., 2022).Similarly, long-term trends of Florida lakes have indicated that a majority (73%) have not shown evidence of changes in chlorophyll-a and trophic state (Canfield et al., 2018).While there is a growing concern of eutrophication and HABs becoming pervasive in the Intermountain West, our results build on recent studies that suggest no indication of widespread intensification in algal blooms.However, this does not diminish the legacy of eutrophication and algal blooms in some systems.For example, 24% of lakes categorized as no trend were historically eutrophic lakes that remained eutrophic.This potentially points to a longstanding legacy of eutrophication which our approach cannot capture.Rather, the large percentage of lakes not trending combined with the presence of algal blooms across the region suggest a historical baseline of eutrophication and that blooms could have predated the 1980s.
Our analysis revealed that, in fact, the smallest percentage (3%) of lakes were trending eutrophic.Global analyses of long-term phytoplankton blooms have shown a substantial (68%) number of lakes to be increasing in bloom intensity (Ho et al., 2019).However, only 5% of U.S. lakes have been shown to be increasing in the same metric over the past 40 years (Wilkinson et al., 2022).In addition, a recent analysis of lake color found a minority of lakes (13%) in the Rocky Mountain region have shown to be shifting from blue to greener wavelengths during this time (Oleksy et al., 2022).With our analysis, we show that concerns regarding the widespread intensification of algal blooms are not captured in our analysis of trophic state.
Our analysis of lakes that were trending eutrophic revealed several hydrologic and climate factors associated with eutrophication.However, the mean decrease in accuracy of the top five variables for eutrophic lakes were substantially lower (compared to no trend and trending oligotrophic lakes), and suggest a limited capacity to explain these trends.30-year normal mean temperatures tended to be higher among lakes trending eutrophic (median of 7°C compared to 5°C among no trend lakes) and an important variable for explaining overall trends.Mean wet deposition of nitrate was shown to be an important variable for lakes trending eutrophic and has been shown to be an important driving factor (as well as warming) for water quality in high elevation lakes (Burpee et al., 2022;Oleksy et al., 2020).In addition, variables such as residence time suggested that lakes trending eutrophic tended to be shorter.In addition, other variables such as lake depth and lake area, although not shown to be important, suggested lakes trending eutrophic were smaller and shallower than other lakes.Small, shallow lakes often have short residence times and are often more productive than deeper lakes because of the effects that lake morphology can have on ecosystem structure (Henderson et al., 2021;Richardson et al., 2022).Shallow lakes have also been shown to be more sensitive to climate conditions (Mooij et al., 2007) and could explain the interaction between climate and residence time driving these trends.
In contrast, 19% of study lakes were found to be improving by trending oligotrophic.Lake-specific characteristics reveal that lakes improving in water quality were in more developed and less forested catchments with greater clay content, had greater residence times, and tended to be at lower elevations.Soils with high amounts of clay can contribute to increases in runoff and intensify nonpoint source pollution into waterbodies (Waterman et al., 2022).These results are consistent with studies on water clarity (Topp, Pavelsky, Stanley, et al., 2021), lake color (Oleksy et al., 2022), and chlorophyll-a (Wilkinson et al., 2022), that highlight improvements in water quality metrics over the same time period.These trends have been hypothesized to be the result of management actions or restoration projects (Wilkinson et al., 2022), although we lacked the information to make conclusions about the mechanisms of these trends.However, concentrations of nutrients across urban watersheds have significantly decreased over the past 20 years and have been directly attributed to the Clean Water Act (Stets et al., 2020).Given the greater variable importance of developed land use across lakes becoming more oligotrophic (3.9 compared to 1.6 among no trend lakes) and the importance of clay content and its relationship with runoff, it is possible that water quality implementation projects have had a positive effect on mitigating eutrophication in the region.
Despite the 35-year study period and wide range of lakes involved, the remote sensing data used in this study may not capture various spatial and temporal characteristics of eutrophication or algal blooms.Algal blooms tend to have high temporal and spatial variance in the short term, as wind dynamics drive the spatial distribution of phytoplankton blooms (Bosse et al., 2019).Furthermore, the "deepest point" that was used to extract surface reflectance for the LimnoSat data set does not capture shoreline environments.Our approach is thus limited in cases where algal blooms are relatively small and located in shoreline areas.Our approach is also limited by the size of lakes that were included in our analysis.The smallest lakes included in our data set were 10 ha in size.However, a large proportion of lakes in the Intermountain West are less than 10 ha (85%), with a median size of 2.9 ha across the region.Therefore, our analysis is limited to larger lakes and does not capture eutrophication trends in small, high-elevation lakes.This is particularly significant given that small, high elevation mountain lakes have undergone dramatic ecological change (Oleksy et al., 2020).In addition, the 16-day return period for Landsat observations may not capture short-term peaks in eutrophication.Furthermore, some images can be unusable due to extensive cloud cover and may extend the period between observations up to months at a time.However, given that our analysis includes 35 years of data across 1,169 lakes, we would expect to capture widespread eutrophication and the spatial clustering of eutrophication trends if it were present.
Additionally, Landsat's long-term record restricted us to a categorical framework and a coarse analyses of trophic state.This categorical framework and our decision to examine changes across historical and contemporary periods is only one way to supplement in situ data regarding trends in lake productivity.However, we took extra steps to test alternate historical and contemporary periods and explore how this affected results.For example, splitting our data into historical and contemporary decades (i.e., 1990-2000 and 2000-2019) resulted in similar results, where approximately half of lakes remain static, 18% are trending oligotrophic, and 4% are trending eutrophic.In addition, we established trend categories to prevent potentially more observations in the contemporary period solely contributing to certain trends.Given the fluctuations of observations in a given year due to factors such as cloud cover, we tied increases in eutrophic observations with decreases in oligotrophic observations (and vice-versa), in an effort to account for these uncertainties.Furthermore, our analysis does not capture cyanobacteria dynamics or those of cyanotoxins directly.Satellites with spectral resolution that estimate cyanobacteria abundance, such as MERIS and Sentinel-3, have lacked the data availability for long-term, retrospective analyses (Coffer et al., 2021).However, future studies that are able to capture trends in cyanobacteria blooms specifically will help provide further context regarding the concerns of bloom intensification.Furthermore, future methods that account for lake heterogeneity and shoreline areas would greatly improve the ability to provide new information in areas where algal blooms tend to be more frequent.

Conclusions
With increases in global lake temperatures (Maberly et al., 2020), lakes globally are expected to become more eutrophic as a response to climate change (Yang et al., 2022).Yet, there have been conflicting results thus far regarding intensifying eutrophication and algal blooms in U.S. and global lakes (Ho et al., 2019;Wilkinson et al., 2022).While increasing eutrophication is a major threat to freshwaters, our analysis found that lakes in the Intermountain West region have not undergone widespread change.Rather, we found that most lakes were not changing, and a substantial number of lakes were becoming less eutrophic and more oligotrophic over this time period.In addition, the number of eutrophic lakes that have not undergone substantial change over this time period suggests algal blooms have been present in the region since at least the early 1980s.These results highlight the complex nature of observing changes in freshwater lakes across large scales.However, our results suggest that despite the processes that drive eutrophication (warmer temperatures, nutrient accumulation, etc.) which have increased over the past several decades, we haven't yet observed a concurrent increase in eutrophication from a large, unbiased sample of 1,169 lakes in the Intermountain West.This suggests controls on eutrophication in this region are complex and need additional study.

Figure 1 .
Figure 1.Corrected surface reflectance over a 35-year timeseries across three different lakes, highlighting a harmonious timeseries.

Figure 2 .
Figure 2.Examples of three possible trend categories based on the trends in % occurrence of oligotrophic, mesotrophic, and eutrophic observations.Each panel included in this plot represents the trends observed across three different lakes.The median number of total observations per year for each panel (a, b, and c) was 10, 14, and 11, respectively.In addition, the minimum number of total observations used to generate this figure was 4, 5, and 3 for each respective panel (a, b, and c).

Figure 3 .
Figure 3. Confusion matrix illustrating the frequency and accuracy of predictions across all three trophic states.The most common misclassification was among mesotrophic predictions that had observed classes of oligotrophic (middle panel, far left).Overall, our model had a global accuracy of 70% with a 95% confidence interval of 63%-76%.

Figure 4 .
Figure 4. Feature importance, measured as model gain, for the predictor variables included in model development.

Figure 5 .
Figure 5. Spatial distribution of trophic state trends across the five states included in this analysis.

Figure 6 .
Figure 6.Boxplots across trend categories of the top five most important variables based on the decrease in accuracy from the overall (global) random forest model.

Figure 7 .
Figure 7. Mean decrease in accuracy of the top five variables used to explain each trend category in the random forest model.The mean decrease in accuracy describes variable importance by quantifying how much accuracy is lost by excluding that particular variable.

Table 1
Metrics Used in Random Forest Model to Explain Variation Across Trend Categories SILLEN ET AL.

Table 2
Predictor Variables Used for Model Training

Table 3
Model Evaluation Metrics for Each Predicted Class