Lakes have been termed sentinels of climate change because of their wide geographical distribution, high sensitivity to change, and ability to integrate and reflect the dynamics of their catchments and the atmosphere (Pham et al. 2008;Adrian et al. 2009). They can be used to study responses to a range of climate changes, from long-term, gradual warming to extreme events such as droughts and floods (Havens and Jeppesen 2018). However, quantifying the effects of climate change on lakes is complicated because of the many multiscaled disturbances that affect lakes. Additionally, the ecological context, which is the hydrologic, geologic, land use, and climatic setting of lakes, is also structured at multiple spatial scales (i.e., lake, watershed, region) and plays a role in how lakes respond to climate change (Read et al. 2015;McCullough et al. 2019). In fact, lake responses to disturbance can be related to characteristics of the lake (e.g., lake morphometry), watershed (e.g., land use), region (e.g., climate) (Pham et al. 2008;Read et al. 2015;Farrell et al. 2020), as well as cross-scale interactions among them (e.g., Soranno et al. 2014;Taranu et al. 2015). Thus, to understand and *Correspondence: 16xs6@queensu.ca; sunxiny9@msu.edu

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

Data Availability Statement: Local water year data and code have been uploaded to the Environmental Data Initiative (EDI) data portal and are available from: https://portal.edirepository.org/nis/mapbrowse? packageid=edi.1547.2. The methods for creating local water year are described in a preprint that can be found on EarthArXiv https://doi.org/ 10.31223/X5Q121. Data for drought responses and predictors have been uploaded to Zenodo and are available from: https://doi.org/10.5281/ zenodo.11244005.

predict lake responses to climate change, we need macroscale studies that incorporate multi-scale natural and human factors.

Although drought has become more frequent and intensified in many regions globally due to climate change (e.g., Easterling et al. 2000;Prein et al. 2016;Oikonomou et al. 2020), we lack an understanding of its potential effects on lakes at macroscales. Meteorological droughts, usually characterized by the combination of decreased precipitation and increased air temperature (Van Loon et al. 2016), can influence lake ecosystem functioning through decreased water levels and increased evaporation. These changes can lead to concentrated nutrients, increases in pollutant concentration, and alterations to biogeochemical processes (Jeppesen et al. 2015;Vicente-Serrano et al. 2020). Algal biomass (often measured as chlorophyll a) sometimes increases in response to drought, potentially caused by elevated nutrient concentrations and a shift in phytoplankton community composition toward more toxin-producing cyanobacteria that can further degrade water quality (Oliveira et al. 2019). However, these changes are not universal, and multi-scale ecological context can play a crucial role in determining the magnitude and even direction of changes (Mosley 2015). A global review by Mosley (2015) suggested that although most studies reported increases in lake algal biomass during droughts, some observed declines. For instance, phytoplankton biomass decreased in a shallow lake in semi-arid Brazil during a drought despite a higher nutrient concentration, which was due to the sediment resuspension promoted by water-level reduction that decreased light availability (Da Costa, Attayde, and Becker 2016). Moreover, in lakes that receive large nutrient inputs from watersheds, droughts could cause declines in nutrients and algal biomass due to reduced riverine inflows (Jeppesen et al. 2015). Such inconsistencies in responses reveal the importance of considering multi-scaled factors to understand and forecast the impacts of droughts on lakes at macroscales.

One challenge in developing a deeper understanding of macroscale climate change impacts on lakes is the lack of consideration of lake type in past studies. Lake type refers to the origin of the lake, whether it be natural or human-made (i.e., reservoirs), and it is likely to affect lake drought responses (Fergus et al. 2020). Reservoirs are abundant and provide essential ecosystem services for human well-being, but they are also different from natural lakes (Hayes et al. 2017;Rodriguez et al. 2023). For example, reservoirs tend to have higher surface water temperatures (Thornton, Kimmel, and Payne 1990), greater catchment-to-surface area ratio (Doubek and Carey 2017), and shorter water residence time (Beaulieu, Pick, and Gregory-Eaves 2013) than natural lakes. These lake type differences can influence the transport and transformation of energy (e.g., irradiance and heat) and mass (e.g., water, solutes, particles) through the landscape and within waterbodies, resulting in divergent lake chemistry (e.g., nutrient levels) and biota (e.g., productivity) as well as responses to changes (Beaulieu, Pick, and Gregory-Eaves 2013;Hayes et al. 2017;Fergus et al. 2021). Unfortunately, reservoirs are studied far less than natural lakes and we lack a comprehensive understanding of the ecological context of reservoirs (Rodriguez et al. 2023).

There is little consensus on whether and how to integrate lake type into climate change studies. Some existing conceptual frameworks about lake responses to climate change assume that natural lakes and reservoirs share similar responses, thus not considering their differences (e.g., Williamson et al. 2009). In contrast, Hayes et al. (2017) argued that drought effects would differ between lake types as they have distinct catchment and management characteristics. Perhaps because of these potential differences in responses, reservoirs are sometimes excluded from climate change studies (e.g., Adrian et al. 2009). However, since $ 46% of lakes ≥ 4 ha in the conterminous United States are reservoirs (Rodriguez et al. 2023), this approach would ignore large numbers of water bodies at continental scales.

Another factor that likely affects lake responses to drought is lake surface water connectivity. Both natural lakes and reservoirs are connected to streams and other lakes via a range of different surface water connections (Fergus et al. 2017;Hu et al. 2017), from entirely isolated natural lakes and reservoirs to natural lakes with both inflow and outflow and run-of-theriver reservoirs (Rodriguez et al. 2023). For example, many lakes receive surface inflow from upstream lakes that can serve as persistent sources of freshwater and/or nutrient supply during droughts, which may alleviate drought impacts on water levels, water quality, and productivity (Spence et al. 2019). On the other hand, lakes without surface inflow may be more sensitive to droughts due to a lack of surface water supply (Webster et al. 2000;Cobbaert, Wong, and Bayley 2015). However, few studies have considered how surface water connectivity may affect lake responses to climate change (Fergus et al. 2017).

Large differences in climate conditions across a continent with different seasonalities also poses challenges for macroscale understanding of drought effects on lakes. Traditionally, the effects of droughts on aquatic systems have been analyzed based on a calendar year (i.e., January-December), which cannot fully capture the asynchrony between precipitation and water fluxes (Vasas et al. 2010;Yang et al. 2016;Kamps and Heilman 2018). A timeframe that takes into account both timing of surface flow and lag effects, such as water year (also called hydrological year), may be more appropriate than the calendar year for understanding lake responses to drought. In the United States, the duration of a water year is standardized from October 1 to September 30 (U.S. Geological Survey 2024). A drawback of this standard is that regional variations in flow seasonality are not considered, which hinders the accuracy and relevance of the water year to continental-scale research (Wasko, Nathan, and Peel 2020). To account for regional discrepancies, Wasko, Nathan, and Peel (2020) used a global study to illustrate that the water year period varies both across and within countries, including the United States, and proposed a local definition of the water year (i.e., local water year, LWY) whereby the LWY of each site is defined as the 12-month period starting from the month with the lowest local streamflow.

In our study, we aim to build an understanding of how natural lakes and reservoirs across the conterminous United States respond to drought. We asked two questions: (1) Do natural lakes and reservoirs differ in their lake chlorophyll a responses to drought, and how does surface water connectivity influence the response?; and (2) Which multi-scaled ecological context variables at the lake, watershed, and region scales affect natural lake and reservoir (hereafter, collectively referred to as lake) chlorophyll a drought responses? To answer these questions, we applied a LWY for 62,927 lakes ≥ 4 ha across the United States, used lake chlorophyll a concentration as the measure of lake productivity for the 10 yr of 2009-2018, and quantified the difference in lake chlorophyll a between the maximum drought year and the baseline year to characterize response to drought. To our knowledge, this is the first study that examines the effects of connectivity and lake type on lake chlorophyll a drought responses and quantifies the influences of multi-scale natural and human factors on those responses at the continental scale.

The data used in this study were from the LAGOS-US research-ready, open-access platform (Cheruvelil et al. 2021) that includes lake and ecological context data for 479,950 lakes ≥ 1 ha across the conterminous United States. We obtained lake type (natural lake or reservoir) from the LAGOS-US RESERVOIR module that used a machine vision model (ResNet model) trained on lake shapefiles and manual identifications to predict the classification of 137,465 lakes ≥ 4 ha (Polus et al. 2022;Rodriguez et al. 2023). Although validation procedures were conducted on the lake-type classification (about 81% accuracy), the classification includes 7% of natural lakes with a flag for being less than 50 m from a dam (Rodriguez et al. 2023). Additionally, lake types are likely more complex than binary natural lakes and reservoirs. Therefore, we further refined lakes into sub-categories that take into account the critical role that surface water connectivity plays in shaping lakes' ecological context (e.g., Hu et al. 2017). We refined the lake type classification by combining it with four surface water connectivity classes that range from least to most connected (i.e., isolated, headwater, drainage, and drainage-lake [drainageLK]; Fig. S1) from the LAGOS-US LOCUS module (Smith et al. 2021), excluding terminal and terminal-lake classes due to extremely low sample sizes (1823 and 379, respectively). The isolated lakes lack both inflows and outflows, headwater lakes have an outflow but no inflow, drainage lakes have both inflow and outflow and are not connected with an upstream lake that is greater than 10 ha, and drainageLK lakes have both inflow and outflow and are connected with an upstream lake that is greater than 10 ha.

Satellite-inferred chlorophyll a concentration data were obtained from the LAGOS-US LANDSAT module (Hanly, Webster, andSoranno 2024a, 2024b). The LANDSAT module used random forest regression model to predict chlorophyll a using more than 47 million Landsat 5, 7, and 8 lake and scene combinations from 1984 to 2020 based on 43,755 same-day matchups between in situ sampled chlorophyll a from LAGOS-US LIMNO module (Shuvo et al. 2023) and whole-lake surface reflectance retrievals (Hanly, Webster, andSoranno 2024a, 2024b). We extracted all lake-specific chlorophyll a data from 2009 to 2018, processed it, and used it in two ways (described below): (1) as lake-specific chlorophyll a used for calculating lake responses to drought and (2) as 10-yr average chlorophyll a to characterize each lake's long-term average productivity.

To further investigate how drought responses were influenced by multi-scale ecological contexts, we acquired lake location (e.g., latitude and longitude) and morphometric (e.g., lake surface area and elevation) information from the LAGOS-US LOCUS module (Smith et al. 2021) and natural and human factors (e.g., air temperature, hydrology, soil features, and land use) from the LAGOS-US GEO module (Smith et al. 2022). We also incorporated the 17 National Ecological Observatory Network (NEON) regions, which are relatively large (213,800-770,995 km 2 ) ecoregions in the United States that primarily rely on climate as the basis for classification (Hargrove and Hoffman 1999), as a regional factor. Finally, we used the maximum depth of lakes from the LAGOS-US DEPTH module (Stachelek et al. 2021). We removed highly correlated and redundant variables, resulting in 33 variables (see Table S1 for a full list of variables with descriptions).

We quantified drought for 479,950 study lakes across the United States from 2009 to 2018 using an LWY drought index. The LWY timeframe was obtained from Sun andCheruvelil (2024a, 2024b), which built on the definition proposed by Wasko, Nathan, and Peel (2020) and applied a spatial interpolation method to construct a continental-scale LWY timeframe. We assigned a single LWY to each of the 202 4-digit hydrologic unit (HU4) regions in the conterminous United States that range from 4384 to 134,755 km 2 in area (Seaber, Kapinos, and Knapp 1987), whereby all lakes in each HU4 had the same LWY timeframe. Each LWY is a 12-month period starting from the lowest streamflow month and is named by the calendar year in which it ends (Fig. 1; Sun andCheruvelil 2024a, 2024b).

We quantified annual lake-specific drought conditions. To do so, we used the Standardized Precipitation Index (SPI), which has been widely used to characterize meteorological drought at various timescales (Guttman 1999;Hayes et al. 2011;Ford, Chen, and Schoof 2021). Precipitation data is the only input parameter for the SPI, which is available monthly for each lake at the fine-scale hydrologic areas in the United States (12-digit hydrologic unit, HU12 (Seaber, Kapinos, and Knapp 1987); 54,263 areas that range from 38 to 161 km 2 , with 1-564 lakes in each HU12) in the LAGOS-US GEO module. We used a 12-month time scale to examine the medium to long-term drought effects because the duration of drought ranges from several weeks to consecutive months and even to years (Oikonomou et al. 2020;Wu et al. 2020). Monthly precipitation data from 1970 to 2018 were entered into the "SPEI" package (v 1.8.1, Beguería and Vicente-Serrano 2023) in R (v4.3.1, R Core Team 2023) to compute SPI values at the 12-month time-scale using the equation: SPI = (P À P*)/σ p , where P is the precipitation, P* is the mean precipitation of the reference period, and σ p is the standard deviation (SD) of the precipitation. We set 1970-2000 as the reference period for computation and quantified the precipitation conditions of the latest 10 yr in our dataset (2009( -2018 LWY; LWY;Beguería and Vicente-Serrano 2023). Standardized Precipitation Index ranges from negative (precipitation is below the mean value of the reference period) to positive (precipitation is above the mean value of the reference period) values. Values between À0.5 and 0.5 were classified as the baseline "normal" condition, greater than 0.5 were "wet" conditions (i.e., more precipitation than baseline), and lower than À0.5 were "drought" conditions (Lloyd-Hughes and Saunders 2002).

Next, using these annual lake-specific LWY SPI values, we identified 2 yr for contrast, the "driest" and the "baseline" for each lake. The driest LWY among the 10-yr study period was the year with the lowest annual SPI value and the baseline LWY was the year with annual SPI closest to zero (i.e., closest to the mean historical precipitation amount for that lake). We further filtered the lakes to keep only lakes with driest year SPI values lower than À0.5 (i.e., true drought year; SPI mean AE SD = À1.4 AE 0.6) and baseline year values within À0.5 and 0.5 (i.e., true baseline normal condition; SPI mean AE SD = 0.01 AE 0.2). This filtering process reduced the sample size from 479,950 (all lakes with SPI values) to 406,953 (lakes with true drought and baseline years).

Data processing and analyses were performed in R (R Core Team 2023). The pre-processing of our response variable, chlorophyll a, was as follows (Fig. 2). We started with lake-specific monthly average chlorophyll a data from 2009 to 2018 LWY. On average, study lakes had 9 AE 2 (SD) years of data, with 6 AE 3 months of data per year. Then, to account for variation among lakes, we scaled and standardized the lake-specific chlorophyll a data by converting monthly chlorophyll a values to z-scores using the formula: z = (x À μ)/σ, whereby x is the monthly chlorophyll a, μ is the 10-yr average chlorophyll a of each individual lake, and σ is the SD. Next, we extracted the monthly z-scores for the driest LWY of each lake and the same month in the baseline LWY and, for lakes with multiple observations, calculated the median z-scores. Then, we subtracted the median z-score of the baseline year from that of the driest year to get median z-score difference:

, where Z dry and Z base are the median z-score of the driest and baseline year, respectively, and x dry and x base are the monthly chlorophyll a of each lake in those 2 yr. This step eliminated the 10-yr average chlorophyll a (i.e., μ) from the calculation of the response variable, allowing it to be used as a predictor variable later in the analyses. Finally, to take into account the influence of the magnitude of drought on chlorophyll a responses, we divided ΔZ-score by the difference between the driest and baseline year SPI values (ΔSPI) using the formula: ΔZ-score median = ΔZ-score/ΔSPI. The ΔZscore median was used as the standardized median z-score changes for each lake. The absolute value of ΔZ-score median was used to represent the chlorophyll-based responsiveness of a lake to drought. We further categorized ΔZ-score median values to represent the directions of each lake's chlorophyll a response as: decrease, moderate change, and increase. We used AE 0.46 ΔZ-score median values (AE 0.5 unit of SD for all the lakes; e.g., Zanchettin, Traverso, and Tomasino 2008;Jones et al. 2015;Anderson et al. 2023) as the threshold to define non-moderate changes (increase: above 0.46; decrease: below À0.46).

Pre-processing of natural and human factors included applying a generalized logit transformation to percent data and a natural log transformation to numeric data (Table S1). The 10-yr average chlorophyll a (an indication of long-term average lake productivity) was calculated for each lake, and annual mean, minimum, and maximum air temperatures were calculated for each HU12 region. Each ecological context variable was assigned a spatial scale (lake (e.g., 10-yr average chlorophyll a), watershed (e.g., soil erodibility), or region (e.g., annual minimum air temperature)), and watershed variables were further refined into three thematic sub-groups: soil (e.g., percent silt in the soil), water supply (e.g., baseflow index that indicates the contribution of groundwater discharge to streamflow), and land use/land cover (e.g., percent of watershed area classified as Fig. 2. Processing steps taken when calculating the response variable, lake chlorophyll a response to drought (ΔZ-score median ). We started with monthly chlorophyll a data from each lake's 2009-2018 local water year (LWY). ( 1) We scaled and standardized chlorophyll a data into z-scores for each lake and extracted the driest and baseline LWY z-scores for each lake. (2) We extracted the month in the baseline year and the same month in the driest year and, for lake-months with multiple measurements, we calculated the median z-score of each lake for the 2 contrasting years. (3) We calculated the difference between median z-score in the driest and baseline year for each lake (ΔZ-score) and calculated the difference between lake-specific drought values (SPI) in the driest and baseline year for each lake (ΔSPI). Using these, we computed the change in z-score per unit change in SPI values for each lake, which is our response variable (ΔZ-score median ). We also created a second response variable by categorizing ΔZ-score median values, representing the direction of each lake's chlorophyll a response: decrease, moderate change, and increase. (4) We merged these response variables with independent variables (i.e., predictors) for further analyses.

wetlands). The final dataset of complete predictors (excluding lake depth) and ΔZ-score median values contained 62,927 lakes (full dataset; 34,422 natural lakes and 28,505 reservoirs). Owing to the limnological importance but the sparsity of lake depth data and the difficulty associated with estimating lake depths (Taranu and Gregory-Eaves 2008;Oliver et al. 2016;Stachelek, Hanly, and Soranno 2022), we created a subset of the full dataset with maximum lake depth, which included 8994 lakes (depth data subset; 4820 natural lakes and 4174 reservoirs).

The full dataset was used to examine natural lake and reservoir responses to drought. A generalized linear model with the gamma distribution (log link function) was employed to examine relationships between the absolute value of ΔZscore median and lake type and connectivity classes ("car" package, v3.1-2, Fox et al. 2023). The model included two categorical predictors, lake type (natural lake/reservoir) and connectivity class (ordinally ranked by connectivity), and an interaction term between them. A multinomial logistic regression model ("nnet" package, v7.3-19, Ripley and Venables 2023) on the categorical ΔZ-score median response (decrease, moderate change, and increase) was utilized to explore the relationships between the direction of chlorophyll a response to drought and lake type, connectivity class, and their interaction. Post hoc analyses were performed using Tukey's contrasts on generalized linear model and multinomial logistic regression model ("multcomp" package, v1.4-25, Hothorn et al. 2023; "lsmeans" package, v2.30-0, Lenth 2018).

The full dataset (33 independent variables) and the depth data subset (32 independent variables) were each used to investigate the effects of lake characteristics and multi-scaled ecological context factors on lake chlorophyll a responses to drought. We applied random forest ("randomForest" package, v4.7-1.1, Cutler and Wiener 2022) and Boruta feature selection ("Boruta" package, v8.0.0, Kursa and Rudnicki 2022) to four models identifying lake, watershed, and region factors that affect ΔZ-score median with either the ΔZ-score median values (continuous) or direction of response (categorical) as the response variable and either the full dataset or the depth data subset (Fig. S2). First, Boruta feature selections with a maximum of 1000 runs were performed to identify relevant predictor variables. All the predictor variables were selected by Boruta when using the full dataset, and one factor (connectivity fluctuates) was rejected when using the depth data subset. Second, we input data (response variable and Boruta-selected predictor variables [32 predictor variables for both the full dataset and the depth data subset]) to random forest models that used five-fold repeated cross-validation. The mean importance scores from Boruta and the percentage increase in mean squared error (i.e., %IncMSE) and decrease in impurity (i.e., Gini) from random forest were used to assess the importance of variables. The effects of the top variable of each spatial scale and sub-group (lake, soil, water supply, land use/land cover, and region) were assessed through partial dependence plots ("pdp" package, v0.8.1, Greenwell 2022) and two-sample Wilcoxon tests.

Across the United States, we observed spatial patterns in lake drought occurrences for 11 of the 17 NEON regions for the 2009-2018 period, but there was no clear pattern for baseline years (Fig. 3). For example, the driest years in the Northeast were LWY 2015 or 2016. For most lakes in the Mid-Atlantic, Southeast, Atlantic Neotropical, Ozarks Complex, Southern Plains, and southern Central Plains, LWY 2011 was the driest. Local water year 2012 was the driest for the Prairie Peninsula, southern Northern Plains, and northern Central Plains, while LWY 2017 was the driest for lakes in northern Northern Plains. Finally, the driest LWY in the Southern Rockies and Colorado Plateau was 2018, whereas LWY 2014 was the driest in the Pacific Southwest.

Across the United States, natural lakes and reservoirs were differ in surface connectivity and productivity. Although there were more natural lakes than reservoirs in our dataset, both lake types were abundant across the conterminous United States. However, the lake types were not equally common across the surface water connectivity classes: natural lakes were predominant in isolated and headwater classes and reservoirs were predominant in drainage and drainageLK classes (Fig. 4a). Note that although reservoirs without an inflowing stream are not included in many definitions of reservoirs, we included them in our study because Rodriguez et al. (2023) found that human-made systems exist as either isolated (no inflow and outflow) or headwater (only has outflow) waterbodies in the conterminous United States. Background chlorophyll a concentrations in reservoirs were higher than in natural lakes (median concentration = 8.7 and 8.0 μg L À1 , respectively; Wilcoxon test, p < 0.001). For both lake types, isolated and drainage lakes had the highest median chlorophyll a concentration (8.7 μg L À1 ), whereas headwater lakes had the lowest median chlorophyll a (7.1 μg L À1 ). When considering lake classification jointly with connectivity class, we found that the median chlorophyll a was highest in drainage reservoirs (9.1 μg L À1 ) and lowest in headwater natural lakes (6.6 μg L À1 ) (Fig. 4b; Table S2).

Across all lakes, the SPI values of the baseline and driest LWYs were 0.01 AE 0.16 (mean AE SD) and À1.45 AE 0.67, respectively, and the average standardized lake drought response was an increase of 0.3 AE 5.8 μg L À1 chlorophyll a (ΔZ-score median = 0.05 AE 0.92; Fig. S3a). Using the 0.5 unit of SD value as the threshold, 17.8% and 20.5% of lakes showed decreases or increases in chlorophyll a, respectively, during drought (11,224 and 12,934 lakes, respectively), whereas 61.6% showed moderate changes (À 0.46 < ΔZ-score median < 0.46, 38,769 lakes; Fig. 5).

We found that lake type and connectivity individually and interactively affected lake chlorophyll a responses to drought, measured by the absolute value of ΔZ-score median (generalized linear model, p ≤ 0.001; Fig. 6a; Table S3). Overall, natural lakes were more responsive to droughts than reservoirs (generalized linear model, p < 0.001). The magnitudes of change in response to drought were greater in natural lakes (0.59 AE 0.78 ΔZscore median , 4.0 AE 5.3 μg L À1 chlorophyll a changes) than in reservoirs (0.50 AE 0.68 ΔZ-score median , 2.9 AE 3.9 μg L À1 chlorophyll a changes). The differences in the absolute value of ΔZscore median between lake types were particularly pronounced between isolated and the two classes of highly-connected lakes (drainage and drainageLK; Tukey, p < 0.001; Fig. 6a). Specifically, the standardized chlorophyll a changes (per unit change in SPI) were 0.8 μg L À1 (ΔZ-score median = 0.4), 0.6 μg L À1 (0.8), and 1.1 μg L À1 (0.8) chlorophyll a higher in isolated, drainage, and drainageLK natural lakes than reservoirs of those same three connectivity classes, respectively. Moreover, isolated lakes, especially isolated natural lakes, were more responsive to drought (5.0 AE 6.6 μg L À1 chlorophyll a changes) than natural lakes and reservoirs of other connectivity classes (Tukey, p < 0.001).

We also found that the directions of lake response to drought (decrease, moderate change, and increase) differed by lake type, connectivity class, and sometimes interactively (multinomial logistic regression model, p < 0.001; Fig. 6b,c; Table S3). For example, natural lakes were more likely to have increased chlorophyll a in response to drought than reservoirs (multinomial logistic regression model, p < 0.001), but there was no difference in the probability of a decreased chlorophyll a (p = 0.113) between the two lake types. Isolated natural lakes (24%) and drainageLK natural lakes (20%) had the highest probabilities of having increased and decreased chlorophyll a during drought, respectively; and drainageLK reservoirs had the lowest probabilities of having both increased (16%) and decreased (15%) chlorophyll a during drought. With the exception of drainageLK natural lakes, the probabilities of changing chlorophyll a decreased with increasing surface connectivity.

We found that 12% of the out-of-bag variance in ΔZscore median was explained in the random forest using the full dataset (n = 62,927 lakes). The 10-yr average chlorophyll a concentration (long-term average productivity) was the most important variable, followed by soil erodibility and annual minimum air temperature (Figs. 7a, S4). Although the effect of lake type and connectivity classes on drought responses was confirmed by the Boruta feature selection algorithm for both the regression and classification models, they were not among the top explanatory variables.

Our examination of the effects of the top variables within each spatial scale and sub-group (10-yr average lake chlorophyll a, watershed soil erodibility, topographic wetness, and wetlands, regional minimum air temperature) on the continuous lake drought response found that lake drought responses and 10-yr average chlorophyll a had a positive relationship until about 25 μg L À1 chlorophyll a, beyond which the relationship became negative (Fig. 7b). Although initially relatively consistent, when erodibility of watershed soil became higher than about 1.5, there was a strong positive association between soil erodibility and drought responses (Fig. 7c). The partial dependence plot curves were close to positive linear for topographic wetness (Fig. 7d) and percent wetland in the watershed (Fig. 7e). Finally, as regional minimum air temperature increased, the relationships changed from positive (until about À22 C), to negative (until about À19 C; Fig. 7f). However, there was another breakpoint at about À15 C, whereby the relationship between temperature and drought responses switched from positive to negative once the minimum temperature was warmer than À15 C.

The classification random forest model of the direction of change had 96% accuracy in correctly assigning lakes to response categories (i.e., decrease, moderate change, or increase in chlorophyll a in response to drought). All lakes with decreased-and increased-chlorophyll a responses were correctly classified, while about 4% of lakes with moderate change were misclassified, with there being more lake misclassified as increased-than as decreased-chlorophyll a. The 10-yr average chlorophyll a was again the top explanatory variable, and longitude and latitude (lake location) were the next two most important variables (Fig. 8a).

A range of natural and human factors at multiple scales were related to the categorical lake response to drought (Fig. 8b-f). For example, we found that lakes with increasedchlorophyll a responses to drought had significantly higher 10-yr average chlorophyll a (median = 11.8 μg L À1 ) than those with decreased-chlorophyll a responses (11.3 μg L À1 ) or moderate change (10.8 μg L À1 ) (Wilcoxon test, p < 0.001). This result was consistent with the finding from the regression random forest model, despite that the ΔZ-score median and 10-yr average chlorophyll a were negatively associated when chlorophyll a was higher than 25 μg L À1 (Fig. 7b). In contrast, lakes with increased-chlorophyll a responses to drought had significantly lower percent watershed forest (29.1%) than lakes with In plot (b), each circle or dot represents a lake (including both natural lakes and reservoirs). Blue dots or circles represent lakes with decreased chlorophyll a (i.e., more than 2 and between 0.5 and 2 units of SD decreases in ΔZ-score median , respectively). Red dots or circles represent lakes with increases in chlorophyll a (i.e., more than 2 and between 0.5 and 2 units of SD increases in ΔZ-score median , respectively). NEON, National Ecological Observatory Network; SD, standard deviation.

decreased-chlorophyll a responses (31.2%) and moderate change lakes (31.8%; Wilcoxon test, p < 0.001). Additionally, lakes with increased-chlorophyll a responses to drought had statistically significantly lower mean watershed percent baseflow (42.4%) than those with decreased-chlorophyll a responses (44.5%) and moderate change (43.8%) (Wilcoxon test, p < 0.001). Lakes with either increased-or decreasedchlorophyll a responses to drought had similar watershed median percent silt in soil (37.6% and 37.5% respectively), which were statistically significantly higher than moderatechange lakes (36.1%) (Wilcoxon test, p < 0.001). Finally, lakes with increased-chlorophyll a responses to drought were further west (median longitude = À93.2) than lakes with decreased-chlorophyll a responses (À92.2) and moderatechange lakes (À 91.5) (Wilcoxon test, p < 0.001).

The effect of lake depth on lake drought responses Lake maximum depth passed the Boruta feature selection and was one of the top variables in both models (i.e., continuous and categorical response variables; Fig. S4). When using the depth data subset (n = 8994 lakes), the random forest regression model explained 10% of the out-of-bag variance in lake drought response and the classification random forest model had 87% accuracy in correctly assigning equal-weighted categories. Lake type and connectivity were still included as important variables but had low importance. We observed lower chlorophyll a increases in deeper lakes until the maximum depth reached 3 m, beyond which chlorophyll a decreased in response to drought (negative predicted ΔZ-score median ; Fig. 7g). Lakes with increased-chlorophyll a responses to drought were significantly shallower (max depth = 9.6 m) than those with decreased-chlorophyll a responses (10.2 m) and moderate change lakes (11.6 m; Wilcoxon test, p < 0.001; Fig. 8g).

This research extends our understanding of lake responses to climate change. Using a lake-specific LWY to study 62,927 lakes across the conterminous United States, we found that lake chlorophyll a responds to meteorological drought. However, by studying tens of thousands of lakes, we found that lakes differ in their responses by lake type (natural lakes and reservoirs) and surface water connectivity, as well as according to long-term average lake productivity. In addition, we determined that ecological factors from multiple scales affected lake chlorophyll a responses to drought.

This research advances scientific understanding in two main ways. First, we fill an existing knowledge gap about how lake type and surface water connectivity individually and interactively affect lake responses to drought. Reservoirs in general, and their responses to climate change in particular, have been understudied compared to natural lakes despite their abundance and important role in providing diverse ecosystem services (Ho and Goethals 2020;Rodriguez et al. 2023). Reservoirs in our study had higher baseline productivity (measured by median chlorophyll a) than natural lakes, and we found that their drought responses differed by connectivity Fig. 6. Plots showing the magnitude of lake chlorophyll a response to drought (ΔZ-score median ) in all the lakes (a) and the probability of having increased (b) or decreased (c) chlorophyll a in response to droughts by lake type and connectivity class. Plot (a) only presents the ΔZ-score median from 0 to 1 to aid in visualization. A plot with the full range can be found in the supplementary material (Fig. S3). In plot (a), white dots represent mean values, and the bars are standard errors. In plots (b) and (c), each dot represents the estimated probability of increased or decreased chlorophyll a and the bars are lower and upper 95% confidence intervals. The letters above each violin and bar indicate the results of Tukey's contrasts post hoc tests. In each plot, the groups with the same letter(s) have similar ΔZ-score median values or probabilities (p > 0.05). Lake types are ranked from the least to most connected based on the number of streams and upstream lakes (Fig. S1).

class. For example, headwater natural lakes and reservoirs had similar drought responses, whereas less connected (i.e., isolated) and more connected (i.e., drainage and drainageLK) lakes differed in their drought responses by lake type. Although our study cannot get at causation, these differences between natural lakes and reservoirs could be affected by multiple factors. For instance, lake elevation could be an influential factor since isolated natural lakes have higher elevations, on average, than isolated reservoirs (Fig. S5a). A previous study found that high-altitude lakes responded to climate change differently than low-altitude lakes due to differences in snow and ice-cover conditions that are sensitive to changes in temperature and precipitation and can affect lake ecosystems (Thompson, Kamenik, and Schmidt 2005).

At the other end of the connectivity spectrum, drainage and drainageLK reservoirs may experience a moderation of the impacts of drought (either increased or decreased chlorophyll) by riverine inflows and upstream lakes, which provided water and/or nutrient supply (Fig. S5b; Jeppesen et al. 2015;Spence et al. 2019). However, the moderation afforded by these inputs may be less effective or non-existent during extreme droughts when streams dry up. We also observed a greater sensitivity to drought in isolated natural lakes than in headwater natural lakes and reservoirs. Although neither group receives surface inflow, headwater lakes receive higher groundwater inputs compared to isolated lakes (Fig. S5c,d), which can alleviate water level declines during droughts (Cobbaert, Wong, and Bayley 2015) and may be contributing to the differences we found. These results are taken within the context of a study documenting how human regulation of reservoir water levels can greatly affect their drought responses (Fergus et al. 2022) and the conceptual model proposed by Hayes et al. (2017) that natural lakes and reservoirs respond differently to climate change due to impacts from catchment and management features that affect energy and mass transport and transformation within waterbodies and through the landscape. Our findings highlight the necessity of including reservoirs in lake climate change studies, accounting for lake type, including measures of surface water connectivity.

Second, we include a large suite of multi-scaled ecological context variables in our models to determine which natural and human factors are affecting lake chlorophyll a drought responses. Although our random forest model explained a small proportion of variance (12%), the top variables in the random forest included all three scales (lake, watershed, and region), demonstrating the importance of incorporating multi-scaled ecological context into climate change impact assessment and prediction. The most important variable for both chlorophyll a response to drought and the direction of that response was 10-yr average chlorophyll a, which gets at long-term productivity and has been found to be a primary driver of lake response to other measures of climate change (e.g., temperature variations, Nõges et al. 2011;warming temperature, Farrell et al. 2020). Using standardized changes in chlorophyll a allowed us to document the complicated nature of these relationships. Although lakes with greater 10-yr average productivity were, on average, likely to show greater increases in chlorophyll a during droughts, that pattern reversed when 10-yr average chlorophyll a exceeded about 25 μg L À1 , which is on the upper end of eutrophic and approaching hypereutrophic conditions (U.S. EPA 2022). This non-linear relationship could be related to differential watershed nutrient loading, or to an, as-of-yet documented, upper limit on the relationship between chlorophyll a and drought. In oligotrophic and mesotrophic lakes, decreased water levels can concentrate nutrients, which increases nutrient availability and induces algae growth during droughts (Mosley 2015). In contrast, in more eutrophic lakes that rely heavily on watershed inputs, droughts could reduce nutrient loadings to lakes, which decreases algal growth (Vanni et al. 2006). However, if droughts coincide with fewer storms during the growing season, then for stratified lakes, internal nutrient loading could be reduced during drought years, further decreasing nutrient inputs into meso-to eutrophic lakes (Soranno, Carpenter, and Lathrop 1997;Orihel et al. 2017).

Our findings also demonstrated that other multi-scaled ecological context variables influence lake chlorophyll a drought responses. For example, we demonstrated that deeper lakes are less sensitive to droughts than shallower ones. Lake depth has been found to regulate climate change impacts (Woolway et al. 2020), perhaps because shallow lakes may experience a greater proportional water volume loss during droughts, leading to greater water temperature increases and sediment (and associated nutrients) resuspension, which can cause reduced light availability (decreasing productivity) or nutrient enrichment (increasing productivity) (Brasil et al. 2016;Seitz et al. 2020).

At the watershed scale, we found soil erodibility and texture to be important, which may point to eroded particles as a source of nutrient and organic matter inputs and the composition of soil determining soil infiltration rates and capacity of water storage (Cleophas et al. 2022), which can buffer water volume decline during droughts. On the one hand, we found that lakes receiving considerable amounts of water inflows from catchments (i.e., higher topographic wetness index values indicating large catchment areas that can provide higher levels of surface water flow) were more responsive to drought. This result may indicate that these lakes are sinks of dissolved organic matter and nutrients (Guillemette et al. 2017), which can be accentuated during droughts and facilitate algae growth. On the other hand, lakes with higher contributions of groundwater discharge to streamflow (i.e., higher baseflow index indicating groundwater inputs) were less responsive to drought, perhaps because these groundwater inputs act as a vital water sources that can mitigate ecological disruptions caused by drought (Cobbaert, Wong, and Bayley 2015). The fact that these two measures of water supply were both important, but had different relationships with lake drought responses highlights the complex hydrological processes (e.g., surface water and groundwater supply) that influence lake responses to drought. Moreover, we found that variables describing land use/land cover characteristics were important to lake chlorophyll a drought responses. For instance, compared to other drought response groups, the increased-chlorophyll a response group had fewer lakes with high watershed forest cover (about 70% forest cover) and more lakes with extremely low forest cover (about 2% forest cover). Previous studies have shown that land use/land cover factors (e.g., forest and wetland cover and agricultural land use) affect dissolved organic matter and nutrient inputs (Sankar et al. 2020), which could contribute to divergent lake chlorophyll a responses to droughts (Fergus et al. 2022). Finally, regional air temperatures, especially minimum air temperatures, were associated with lake chlorophyll a drought responses. However, the direction of the association changed from positive to negative with an increase of minimum temperatures (until about À10 C), suggesting that there could be other factors (e.g., lake geographical location; Adrian et al. 2009) collectively affecting the responses, particularly in areas with extremely low temperatures. Although lake location (via longitude and latitude) was important in the random forests, obvious spatial patterns of lake chlorophyll a responses to drought did not emerge across the conterminous United States. The complexity of the lake chlorophyll a responses to droughts that we found, as well as the fact that a range of lake, watershed, and regional ecological context variables were important for understanding those relationships, demonstrate the importance of further macroscale studies that take into account a wide range of multi-scaled ecological context variables to account for complex underlying mechanisms.

We used a LWY timeframe (Sun and Cheruvelil 2024b) to study lake chlorophyll a responses to droughts at the continental scale. Doing so allowed us to capture the relationships between precipitation and chlorophyll a on temporally appropriate scales that account for regional variation in hydroclimate. However, due to the broad scale of this study and lack of snow/ice melt-relevant data, we did not specifically identify areas where snow is the dominant water source. According to Tedesche, Dahl, and Giovando (2023), snow-dominated areas occupy a small proportion of the conterminous United States, mainly in the Great Basin and Northern Plains regions (NEON 15 and 9 ecoregions). In our study, snow in late fall and winter has likely been taken into account in the Great Basin region because that area has an LWY end month in summer or early fall. However, since the LWYs in some areas in the Northern Plains region commonly have end months of January and December, the LWYs may not have fully covered the snow period, leading to the drought intensity (i.e., SPI values) being overestimated in this region. Future research could further refine the LWY with snow data and examine lake drought responses in these snow-dominant areas.

Our models and results explained part of the variance among lake chlorophyll a drought responses, implying that additional variables (e.g., water nutrient levels, light availability, or water column stability) and interactions among them and ecological contexts should be considered in future studies to improve our understanding of these responses. We studied algal biomass (chlorophyll a) because those data are available across the conterminous United States for both baseline and drought years. However, it is essential to study a range of biotic and abiotic responses to understand the effects of climate change on lakes. Variables such as phosphorus (Shuvo et al. 2021), salinity (Jeppesen et al. 2015), zooplankton (Meerhoff et al. 2022), and water management (Fergus et al. 2022) have all been found to influence lake responses to drought and ought to be considered in future studies. Additionally, we used the change in median z-score per unit change in SPI as our response variable as we expected that a greater drought magnitude (i.e., greater SPI value) would lead to a greater lake chlorophyll a change. Further research could be conducted to investigate the relationship between drought magnitude and lake responses and the multiscale factors affecting the relationship at the macroscale, which would improve the assessment and prediction of drought impacts on lakes. Moreover, previous research suggests that interactions among ecological context variables across scales can sometimes cause unexpected results (i.e., cross-scale interactions; Peters, Bestelmeyer, and Turner 2007;Soranno et al. 2014;Fischer 2018). For example, Soranno et al. (2014) found that local wetlands were positively associated with lake phosphorus concentrations in regions with low percent agricultural land use, but negatively associated in regions with high percent agricultural land use. Therefore, further investigation is needed to understand these nuanced interactions between multiple factors across scales (e.g., latitude and long-term average chlorophyll a; Fig. S6) and to better disentangle the complex relationships between ecological context variables and lake drought responses.

Our study applied a lake-specific LWY timeframe to examine 62,927 lake chlorophyll a responses to droughts at the continental scale from 2009 to 2018. We examined the individual and interactive effects of lake type (natural lake and reservoir) and surface water connectivity on lake chlorophyll a drought responses, and included a large suite of lake characteristics and multi-scaled ecological context variables in models to determine which natural and human factors drive lake chlorophyll a drought responses. Our findings suggest that headwater natural lakes and reservoirs respond similarly to drought, whereas more and less-connected lakes respond differently according to lake type (i.e., natural lakes are more responsive to drought than reservoirs). We also found that features across spatial scales, such as 10-yr average lake productivity, watershed soil texture, groundwater, and land use/cover, and regional air temperature, all determined lake chlorophyll a responses to drought. These results point to the import of including lake type and surface connectivity, as well as a wide range of lake characteristics and multi-scaled ecological context variables when studying lake responses to droughts. Doing so will improve the prediction and management of lake responses to droughts as well as other anthropogenic and climate change disturbances. Thus, there is an urgent need to integrate multi-scale factors into impact analyses and forecasting to bolster our understanding of how these invaluable ecosystems respond to future changes at regional, continental, and global scales.