Heterogenous controls on lake color and trends across the high-elevation U.S. Rocky Mountain region

Global change may contribute to ecological changes in high-elevation lakes and reservoirs, but a lack of data makes it difficult to evaluate spatiotemporal patterns. Remote sensing imagery can provide more complete records to evaluate whether consistent changes across a broad geographic region are occurring. We used Landsat surface reflectance data to evaluate spatial patterns of contemporary lake color (2010–2020) in 940 lakes in the U.S. Rocky Mountains, a historically understudied area for lake water quality. Intuitively, we found that most of the lakes in the region are blue (66%) and were found in steep-sided watersheds (>22.5°) or alternatively were relatively deep (>4.5 m) with mean annual air temperature (MAAT) <4.5°C. Most green/brown lakes were found in relatively shallow sloped watersheds with MAAT ⩾4.5°C. We extended the analysis of contemporary lake color to evaluate changes in color from 1984 to 2020 for a subset of lakes with the most complete time series (n = 527). We found limited evidence of lakes shifting from blue to green states, but rather, 55% of the lakes had no trend in lake color. Surprisingly, where lake color was changing, 32% of lakes were trending toward bluer wavelengths, and only 13% shifted toward greener wavelengths. Lakes and reservoirs with the most substantial shifts toward blue wavelengths tended to be in urbanized, human population centers at relatively lower elevations. In contrast, lakes that shifted to greener wavelengths did not relate clearly to any lake or landscape features that we evaluated, though declining winter precipitation and warming summer and fall temperatures may play a role in some systems. Collectively, these results suggest that the interactions between local landscape factors and broader climatic changes can result in heterogeneous, context-dependent changes in lake color.


Introduction
High-elevation lakes and reservoirs form the basis of a critical water supply network for arid and semiarid cities and communities downstream. However, climate change threatens these ecosystems via altered temperature and precipitation regimes (Christianson et al 2020, Maberly et al 2020, lake ice phenology (Benson et al 2012, Preston et al 2016, lake temperature (Sadro et al 2018, Christianson et al 2019, Smits et al 2020 and, in turn, ecosystem function and biological composition. In addition to climate change, increasing nutrient loading presents an additional steady change that can lead to increased algal production (Moser et al 2019, Oleksy et al 2021. Despite these potential threats to high-elevation lakes, examining shifts in freshwater ecosystems at large spatial scales is challenging because of sparse coverage and a strong bias of analyses towards a few well-monitored lakes (Stanley et al 2019). Physiochemical studies (e.g. ice-cover duration, water chemistry, surface temperature) in a number of pristine high-elevation lakes suggests that these shifts are significant (Preston et al 2016, Moser et al 2019. Summer warming in combination with nitrogen deposition is leading to algal assemblage shifts and increasing productivity in lakes along the Colorado Front Range (Oleksy et al 2020a). In addition, snowpack and summer weather conditions are strong controls on water chemistry and algal biomass for mountain lakes (Preston et al 2016, Sadro et al 2018, Oleksy et al 2020b, and these drivers are also changing. While there has been recent research examining regional to continental scale changes in lake nutrients (Stoddard et al 2016, Oliver et al 2017, water clarity (Topp et al 2021), lake color (Kuhn and Butman 2021), and algal blooms (Wilkinson et al 2021) there have been no regional studies, to our knowledge, on high-elevation lake shifts likely due to a lack of in situ water quality monitoring data (Read et al 2017).
While remote sensing can be used to directly estimate water quality parameters (Topp et al 2020), lake water color is relatively easy to infer from satellite and is less prone to prediction errors (Giardino et al 2014). Color is also an intuitive and integrative metric that can serve as an indicator of water quality parameters, including colored dissolved organic matter, which can be used to infer estimates of total organic carbon, dissolved organic carbon (Ouyang et al 2006), chlorophyll-a (proxy for algal productivity; Cao et al 2020), and suspended sediment (Dekker et al 2001).
Here we used satellite-derived lake color to address three core objectives to better understand U.S. Rocky Mountains lakes: (a) We evaluated the contemporary spatial distribution of average summer lake color. (b) We quantified how lake color has changed in the region since the beginning of the Landsat record (1984). (c) We examined which lake, landscape, and climatological features of lakes relate to spatial patterns and temporal trends in lake color.
Through these three objectives we aimed to understand the current patterns of lake color across the U.S. Rocky Mountain region and to assess how climate change and other disturbances might be changing lake ecology and related ecosystem properties.

Lake color
We used remote sensing data from the LimnoSat-US database (Topp et al 2020), a robust collection of Landsat surface reflectance data for 56 792 U.S. lakes. The LimnoSat-US data extracts USGS Tier 1 surface reflectance values over Landsat 5, Landsat 7, and Landsat 8 sensors dating back to 1984 from the point furthest from any shoreline ("deepest point"). All the Landsat imagery has been atmospherically corrected, and then adjusted so each satellite had unbiased data across time and between satellites (Topp et al 2020). We limited the analysis to high elevation lakes in the Rocky Mountain Region, which we define as the parts of Idaho, Colorado, Montana, Wyoming, Utah, and New Mexico above 1400 m. This captures many mountain lakes in the region as well as high-elevation plains lakes and reservoirs. We examined spatiotemporal patterns in the color of the lake, called the dominant wavelength (DWL), which maps directly to the Forel-Ule scale, a water transparency classification scale (Wernand and Van der Woerd 2010). The Forel-Ule system is part of a long-studied approach, dating back to the 1890s, of analyzing color of bodies of water; there is a standard scale of 21 colors that classify gross biological activity and transparency of the water based on how it maps onto a human's perception of lake's color (Wernand and Van der Woerd 2010). DWL is quantified by looking at the human visible spectrum surface reflectance values (red, green, blue) and then converted into chromaticity coordinates (Wang et al 2015). For both colormeasuring approaches, blue lakes (DWL < 530 nm) are generally considered oligotrophic, while change in color from blue to green wavelengths generally corresponds to shifts in trophic state from mesotrophic to eutrophic (DWL > 530 nm). Color changes from green toward brown wavelengths can indicate either a dystrophic system or a eutrophic lake with high suspended sediment in the water column (DWL > 575 nm).

Classification of spatial patterns
To understand broad-scale spatial patterns, we examined the median contemporary (2010-2020) lake color across the U.S. Rocky Mountains. We included data from the summer period (1 July-15 September) to minimize seasonal variation and the impact of snow and ice cover, which can persist into June for some of the highest elevation lakes. We joined the LimnoSat-US lake color data to the National Hydrography Dataset (U.S. Geologic Survey 2021), the Global Lake Area, Climate, and Population dataset (Smith et al 2021), watershedlevel metrics from the LakeCat database (Hill et al 2018), LAGOS-US NETWORKS (King et al 2021) and LAGOS-US Reservoir (Polus et al 2021). We used information about the lake, landscape, lake type (natural lake or reservoir), and connectivity features from these datasets to explain lake color spatial patterns in lakes that spanned a broad range of environmental contexts (table 1).
We divided the population of lakes into two categories: blue (DWL < 530 nm; n = 620) or green/brown (DWL > 530 nm; n = 320; figure 2(A)). To address our first research objective, we used a classification and regression tree (CART; Therneau and Atkinson 1997) to determine which environmental characteristics explained variation in lake color across the region using the rpart package (Therneau and Atkinson, 2019). The training dataset included 80% of the total population (n = 765). We calculated the global model accuracy by predicting lake color groupings for the out-of-sample lakes and assessed model performance with a confusion matrix. We visualized the results with the cvms (Olsen and Zachariae 2021) and ggparty (Borkovec and Madin 2019) R packages for the confusion matrix and decision trees, respectively. All analyses and data visualizations were done in R version 4.0.5 (R Core Team 2021).

Trend analysis
For the trend analysis, we built a separate dataset that included only lakes that had at least three cloudfree summer images for a minimum of 30 consecutive years between 1984 and 2020 for a total of 527 lakes in the analysis. This accounts for approximately a quarter of all lakes in this region that are greater than 10 ha in area and over 1400 m in elevation (figure S1). We calculated the non-parametric Theil-Sen's slope for each lake time series of median summer color using the trend package (Pohlert 2020). We used the Mann-Kendall z-score and compared the p-value from that z-score to α = 0.05. We categorized each lake into one of five possible trend categories: For lakes in the Blue->Greener and Green->Bluer categories, we assessed whether the median lake color in the later part of the record indicated a modal shift in color from predominantly blue to predominantly green/brown, or vice versa, consistent with the spatial color categorization.
We conducted a random forest analysis to explore the drivers of color trends (Breiman 2001). Here, we grouped together all lakes with positive trends (Intensifying Green/Yellow and Blue->Greener) and negative trends (Intensifying Blue and Green->Bluer) into composite categories for a total of three trend categories (Negative, No Trend, Positive). Predictors included all those considered in the spatial CART described above (table 1) as well as changes in seasonal precipitation, temperature, and human population size. We used the prism package (version 0.2.0) to download the daily estimate of temperature and precipitation from the Oregon Parameter-elevation Relationships on Independent Slopes Model (PRISM) project (Hart and Bell 2015). For each lake-year, we calculated the mean winter (December-February), spring (March-May), summer (June-August), and fall (September-November) temperature and precipitation. Then, we calculated the Theil-Sen's slope of temperature and precipitation for each lake and season from 1984 to 2020.
We built the random forest models using the rand_forest function in the parsnip package using the 'ranger' engine (Wright andZiegler 2017, Kuhn andVaughan 2021a). We randomly chose 60% of the data as our training data set and 40% as our test dataset which ensured that at least 25% of the observations in each trend category were set aside for validation. We tuned the two hyperparameters using ten-fold cross-validation. The optimum number of predictors at each node (mtry = 4) and the minimum n to split at any node (min_n = 3) for the final model was selected according to the best receiver operating characteristic curve and overall classification accuracy using the yardstick package (Kuhn and Vaughan 2021b). The final random forest model consisted of 1000 trees and was evaluated on the validation data. We present the top ten predictors based on variable importance (VI), computed as the total decrease in node impurity averaged over all trees.

Spatial patterns
Our dataset included 940 lakes above 1400 m across the six-state Rocky Mountain region (figure 1). Between 2010 and 2020, 66% of the lakes were predominantly blue (n = 620) while 34% of the lakes were predominantly green/brown (n = 320; figure S2). The CART analysis revealed that watershed slope, mean annual air temperature (MAAT), and maximum lake depth were important determinants of lake color (figure 2(C)). Most green/brown lakes were found in relatively shallow sloped watersheds with MAAT ⩾4.5 • C. Lakes situated in watersheds with slope angles ⩾22.5 • were most likely to be classified as blue lakes. Similarly, another set of blue lakes were common in less steep watersheds with MAAT ⩽4.5 • C with maximum depth ⩾4.5 m while shallower lakes in those areas were more likely to be green/brown. Watershed slope is negatively correlated with MAAT (r = −0.50) and other factors such as lake elevation that likely influence spatial patterns of lake color ( figure S3). Overall, the CART model was able to correctly classify 84% of blue lakes and 68% of green/brown lakes in the test dataset ( figure 2(B)).

Cross-lake color trends
In the U.S. Rocky Mountains, we detected no trend in lake color between 1984 and 2020 in 55% of lakes (n = 290, figure 3). However, 32% of lakes were trending bluer (n = 166) and reservoirs showed the largest improvements. Specifically, 71% of the lakes that trended bluer were reservoirs (n = 30), and 75% of the lakes that were intensifying blue were reservoirs (n = 30). Most of the lakes trending from Green->Bluer were in Colorado (71% or n = 72; figure 4), including many Colorado reservoirs that switched from Green/brown to Blue (n = 14, table S1). Median lake color shifted toward greener wavelengths in 13% of the population of lakes (n = 71), with 34 lakes in the Intensifying Green/brown category and 37 lakes in the Blue->Greener category. Of the Blue->Greener lakes, six crossed the 530 nm threshold consistently in recent years such that they were classified as Green/brown in the spatial analysis.
Although our random forest model poorly predicted lake greening (figure S4), a combination of static variables and climatic trends partially explained some trends in lake color (figure 5). The variables with highest importance included total human population in the lake-watershed (VI = 6.07), lake elevation (VI = 5.91), changes in winter precipitation (VI = 5.58), urban landcover (VI = 4.4), and changes of spring temperature (VI = 4.25). The majority of the lakes that were Intensifying Blue or trending Green->Bluer were located in relatively urbanized watersheds with some of the highest human population densities in the region (figures 5(A) and (D)). These lakes also tended to be located at lower elevations relative to lakes not experiencing color shifts or lakes that were greening ( figure 5(B)). Both greening and blueing lakes were associated with decreases in winter precipitation between 1984 and 2020 compared to lakes with no trend (figure 5(C)). Furthermore, blueing lakes tended to be in areas where spring air temperatures were cooling slightly relative to greening lakes or lakes without color changes (figure 5(E)), though notably for both the climatic variables only a small subset of the trends were statistically significant ( figure S5, table S2).
Overall, the most widespread climatic trends in the region were increasing summer and fall temperatures (table S2). Although increasing fall temperatures were widespread in this region, there were no differences among color trend groups analysis of variance (ANOVA) F 2,328 = 2.55, p = 0.08). However, absolute rates of summer warming varied among color groups (Kruskal-Wallis H-test, p < 0.001). Specifically, since 1984 summer temperatures increased on average 0.23 • C more in lakes with no change in color compared to lakes that were trending blue (95% CI: 0.06 • C-0.4 • C). Further, rates of summer warming were 0.34 • C higher in the greening lakes compared to the blueing lakes (95% CI: 0.11 • C-0.57 • C; figure S5). For lakes that shifted from Blue->Greener, nearly every lake experienced statistically significant increases in summer air temperature (table S2). Precipitation shifts were highly variable, and most lakes did not experience substantial shifts in PRISMestimated monthly precipitation (table S2).

Discussion
Our analysis showed that most lakes (55%) included in this study showed no substantial change in lake color between 1984 and 2020. This is consistent with both remote sensing and field studies of regional lake water quality trends in arctic (Kuhn and Butman 2021) and temperate regions (Oliver et al 2017, Paltsev andCreed 2022) that showed a minority of study lakes to be exhibiting changes in lake color. For lakes in the Rocky Mountain region that changed over the past 36 years, most trended bluer (70%), suggesting an overall improvement in summer water quality. While there is a growing concern of widespread declines in water quality, our results build on recent studies that show regional improvements in water quality and a more nuanced understanding of changes in lakes occurring across large spatial scales (Topp et al 2021, Wilkinson et al 2021.

Spatial patterns
Our study revealed several putative controls on spatial patterns in lake color in the U.S. Rocky Mountains. Many blue lakes were in steep, high-elevation watersheds, with little vegetative cover and had colder MAAT than green/brown lakes ( figure S6). Together, these factors likely result in limited terrestrial nutrient subsidies and thus lower productivity and clearer waters (Likens andBormann 1974, Leavitt et al 2009). Heterogeneity in additional factors among these high elevation lakes such as lake morphometry and watershed area may also modify this relationship. For example, some green/brown lakes occurred in cold areas (MAAT < 4.5 • C) if they were shallow (<2.5 m average depth), particularly if they had larger watersheds (>12.5 km 2 ). This is expected since small, shallow lakes tend to be more productive than deep lakes (Duarte and Kalff 1989, Genkai-Kato and Carpenter 2005, Richardson et al 2022. Conversely, in some shallow lakes, the color that satellites detect may be capturing benthic algal growth, which can make up a majority of the lake productivity in systems where photic zone extends to the benthos (Lõugas et al 2020). Overall, these spatial patterns are consistent with studies describing continental scale patterns of lake trophic status and water quality, which indicate that high-elevation western mountain ecoregions are generally oligotrophic, with higher prevalence of green, turbid, or eutrophic lakes in the high plains and agricultural ecoregions (Hollister et

Controls on cross-lake color trends
Lakes and reservoirs shifting toward bluer wavelengths represented 32% of all sites and frequently occurred in developed, relatively lower elevation areas. Reservoir management in the Western U.S. typically employs a variety of approaches (e.g. hypolimnetic oxygenation, diversifying water supplies) to maintain water resources under increasing climate variability (Beutel and Horne 1999, Ray 2003, Page and Dilling 2020 and these practices may be a driver of the water quality improvements we observed. However, these apparent changes in water color that may be attributed to local management actions were difficult to capture in our statistical analyses because we lacked broad-scale databases that summarize management efforts for this region. For instance, increases in reservoir storage, resulting in greater volume of water, may result in an apparent blueing of waters, but our study lacks data on changing lake surface area or volume. In addition, managed movement of water across the landscape could further obscure relationships between watershed characteristics and local water quality trends. For example, we observed clusters of reservoirs with blueing trends in the heavily populated Colorado Front Range, but trans-basin water diversions are common in that area (Wiener et al 2008) making it even more difficult to link management practices to changing water color. Our results suggest that management practices over the same period may have led to improving water quality in ecosystems that are often used for drinking water and irrigation.
A relatively small proportion of lakes (13%) exhibited characteristics indicative of decreasing water quality, either shifting from states of blue to greener or intensifying green/brown. Similarly, recent studies of chlorophyll-a trends in U.S. lakes have shown algal intensification to be occurring in a relatively small proportion of lakes with long-term field data (Wilkinson et al 2021). Lakes that did exhibit trends toward greener waters were diverse in their size, shape, watershed area, land cover, and climatic changes. This level of spatial heterogeneity has also been shown in regard to cyanobacteria bloom frequency, where the Rocky Mountain region represented a region where blooms were isolated rather than spatially clustered (Coffer et al 2021). This result reinforces that interactions between local landscape factors and broader climatic changes can results in heterogeneous, context-dependent responses on freshwater systems (Jackson et al 2016, Birk et al 2020. Notably, the random forest model had a limited capacity to classify lakes trending toward greener wavelengths (positive trends; figure S2). These greening lakes tended to be at some of the highest elevations and were sparsely populated by humans relative to the lakes that were blueing (figure 5). Many of these sites experienced slight increases in winter precipitation and decreases in spring temperature. The six lakes that showed the most substantial changes in lake color (table S1) had very little in common except that they all have experienced increases in mean summer air temperature (1.0 • C-1.95 • C since 1984) and were all shallow (less than 3 m mean depth), suggesting that lake color in these systems includes bottom reflectance and possible benthic blooms (Vadeboncoeur et al 2021). It is possible that shallow lakes are particularly sensitive to changes in water volume via increased evaporation rates due to summer warming, and these reduced water volumes result in an apparent greening. While the slope of the greening trends in color in these 39 lakes were statistically significant, we emphasize that most of the color values were within the range of wavelengths that classify these lakes as 'blue' following the approach we used in our spatial analysis. Nonetheless, these lakes appear to be on a 'greening' trajectory and the underlying cause of that shift warrants further investigation.
Winter precipitation and spring temperatures partially explained temporal trends in summer lake color, but they do not fully capture variability in snowpack regimes (Trujillo and Molotch 2014). In many mountainous areas, winter and spring snowpacks control the length of ice duration (Caldwell et al 2021), thus changes in these climatic variables can have cascading effects on lake chemistry and ecology (e.g. algal phenology), and thus color (Cavaliere et al 2021, Hébert et al 2021. Less snow in combination with warmer summers may interactively stimulate lake production in some lakes (Preston et al 2016, Oleksy et al 2020b, but these same climatic changes can have the opposite effect on lakes in other regions (i.e. lower phytoplankton biomass; Hrycik et al 2021), highlighting the need to understand how multiple stressors can have either synergistic or antagonistic effects across lakes.
Finally, there are a few possible explanations for why we did not detect widespread changes in lake color in the region. First, our dataset only included relatively large lakes that were ⩾10 ha, but most lakes in the Rocky Mountains are <10 ha (85.3%, n = 15 568) and the smallest lakes are more abundant at high elevations ( figure S6). This may partially explain why rates of nitrogen deposition did not appear to have an effect of water color trends, even though excess nitrogen is implicated as a driver of ecological change in high-elevation lakes across the region (Moser et al 2019, Oleksy et al 2020a, Burpee et al 2022. Second, we limited our analysis to median summer color, but it is possible that there are dynamics that have helped create the perception of lake greening, such as episodic algal blooms, which are increasing in some systems (Ho et al 2019, Vadeboncoeur et al 2021, Wilkinson et al 2021. This could create issues where algal blooms really are present but are short and intense and thus not captured by Landsat's 8-or 16-day return sampling interval. As such, algal blooms that are increasing in severity, duration, or magnitude may not be detected by our approach. Conversely, by limiting our analysis to summer months, we may be missing shifts in the phenology of lake color, such as early greening in the spring or a second peak of productivity in the fall (Sommer et al 2012). Future studies related to lake changes may consider changes and variability in the entirety of the ice-free season. Furthermore, our understanding of regional changes in water quality will be greatly enhanced by advances in the remote sensing of small lakes.

Conclusions
Climate change impacts are likely to influence highelevation systems faster than others, making highelevation lakes sentinels of climate change (Adrian et al 2009, Moser et al 2019. While eutrophication could pose a major threat to the ability for these systems to continue to provide their vital services to downstream communities, we found that lake color in most large lakes (>10 ha) in this region were stable over the last 35 years. Where we did observe lake color changing, it was consistently towards bluer waters. However, some of the mechanisms for the observed changes, particularly in greening lakes, remain elusive. Future work in this region should investigate the impact of changing water quantity on lake color and how the slow, press changes from climate change interact with short, intense pulse disturbances like floods and fire to alter the ecology of Rocky Mountain lakes and reservoirs.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).