Introduction

Chesapeake Bay, the largest estuary in the USA, has undergone considerable ecosystem changes in the last several decades, including eutrophication, seasonal hypoxia, food web shifts, and the loss of key habitats (e.g., seagrasses, oyster reefs; Kemp et al. 2005). Nutrient inputs from the Bay watershed are a primary driver of change and have been linked to elevated phytoplankton growth that reduces light for submerged aquatic plants (Kemp et al. 2005; Orth et al. 2017) and generates organic material that supports elevated rates of oxygen consumption (Boesch et al. 2001; Hagy et al. 2004; Kemp et al. 2005). Thus, efforts to reduce watershed nutrient (and sediment) loads to restore the ecological condition of Chesapeake Bay have long been a management focus. Recently, nutrient load reductions associated with these efforts have been linked to a reversal of some aspects of eutrophication (Lefcheck et al. 2017; Testa et al. 2018; Zhang et al. 2018).

One important symptom of eutrophication is the reduction of light penetration through water-column via elevated sediment and phytoplankton biomass, and several approaches have been implemented to measure water-column light availability. While the quantity and characteristics of light within water can be measured by a variety of methods, long-term monitoring programs typically measure quantities that are relatively affordable and straightforward. These include vertical profiles of photosynthetically active radiation (PAR) measured in water (with an associated surface measurement) and the classical Secchi disk depth (Preisendorfer 1986). Vertical profiles measure the availability of downwelling PAR and can be fit to Beers Law to estimate a coefficient of light attenuation, commonly referred to as kd. Secchi disk depth (hereafter Secchi depth), by contrast, estimates water clarity using the human eye (a measure of transparency) and has sometimes been converted to kd using a variety of empirical conversion factors (e.g., Harding et al. 2015b; Xu et al. 2005). Although these two measures are inherently different, they historically have been highly correlated to one another, leading to a number of “rule of thumb” conversions between Secchi depth and kd ( Xu et al. 2005). Upon recent analysis of long-term observations, the relationship between kd and Secchi depth has appeared to change over time (Gallegos et al. 2011; Pedersen et al. 2014), raising questions about the utility of the approaches and how changes in the spectral quality of the water cause different temporal trajectories for the two variables.

Among the many water properties that influence light availability, inorganic sediment, organic carbon, colored dissolved organic matter (CDOM), and chlorophyll-a are particularly important to water clarity patterns in the estuary. Specifically, riverine sediments block light penetration in the water-column (typically via scattering) and indirectly affect nutrient cycling through transport of particulate nutrients (Beusen et al. 2005; Cerco et al. 2013; Meybeck et al. 2003; Sanford et al. 2001; Schubel 1968). Organic carbon can absorb light in both its particulate and dissolved (CDOM) forms, and upon its oxidization, nutrients are recycled to support additional phytoplankton growth and light attenuation (Alvarez-Cobelas et al. 2012; Bianchi et al. 2007; Dai et al. 2012; Harrison et al. 2005; Spencer et al. 2013). CDOM concentrations are known to vary considerably over space in estuaries, with high concentrations near freshwater or wetland sources and lower concentrations in more oceanic waters (Rochelle-Newall and Fisher 2002; Tzortziou et al. 2008). Chlorophyll-a can modulate estuarine light environments through physical (e.g., light blocking via absorption) and biological (e.g., grazing release of DOM) processes (Harding et al. 2016; Prasad et al. 2010; Roman et al. 2005), and has the unique property of both responding to and causing light attenuation. The integrated impacts of these properties on light availability in the estuary can in turn affect the growth of submerged aquatic vegetation (SAV), which provides key habitats for shellfish and finfish populations (Cerco et al. 2004; Gallegos et al. 2011; Harding et al. 2014; Kemp et al. 2005; Tango and Batiuk 2013).

Recent observations and analyses have indicated that water transparency, as measured by Secchi depth, has remained low or has continued to decline since the 1980s across much of Chesapeake Bay’s tidal habitats, while measures of kd have been stable (Gallegos et al. 2011). Recent efforts have considered changes in light availability in specific regions of Chesapeake Bay (Gallegos et al. 2011; Harding et al. 2015b), but a spatially comprehensive analysis of the changes in light availability in the system over the available record (~ 30 years) has yet to be undertaken. Thus, there is a need to quantify the relationships between available measures of light availability and the key external and internal drivers of light attenuation across all habitats within Chesapeake Bay.

In the above context, this work aimed to test the hypothesis that light availability in the Chesapeake estuarine system is affected by various external and internal drivers whose influence is location-specific. Although it is widely known from past estuarine research that light availability can be affected by different factors (e.g., CDOM, chlorophyll-a, and total suspended solids [TSS]), there is a need to better understand what factors are most dominant along the salinity gradients of estuaries and how these spatial patterns influence long-term changes in light availability in response to variations in watershed sediment and nutrient loadings. To help fill these gaps, we have (1) conducted a comprehensive statistical analysis of Secchi depth variability in Chesapeake Bay and its tributaries to better understand spatiotemporal variations in the primary controlling variables, and (2) utilized statistical techniques to explore the linkage between Secchi depth patterns and external (i.e., nutrient and sediment loads) and internal (i.e., salinity, TSS, and chlorophyll-a) drivers to identify factors that controlled long-term changes in Secchi depth. For this work, we chose Secchi depth over kd as our measure of water clarity because Secchi depth was more regularly and continually measured in the period of record. This effort builds upon previous studies of controls on light availability in Chesapeake Bay by expanding the spatial and temporal scope of the analysis, considering a variety of both in-water and watershed factors, and applying numerous statistical approaches to examine Secchi depth. The results will inform our understanding of long-term changes in light availability and its controls in complex estuarine environments.

Methods

Site Description

Chesapeake Bay is a partially mixed estuary located within the mid-Atlantic coastal region of the USA. The central open body of the estuary, known as the “mainstem,” is over 300 km long, has a mean low-water estuarine volume of 74.4 km3, and has a mean depth of 6.5 m (Fig. 1). A 20–30 m deep channel (1–4 km wide) runs the length of the middle region of the Bay, but is flanked by broad shallow (< 10 m) shoals to the east and west. Two-layered circulation occurs for most of the year in the estuary, driven by a mean freshwater input of 2300 m3 s−1 from the Susquehanna River that drives a seaward-flowing surface layer and a landward-flowing bottom layer. The upper mainstem estuary (39° N) and the upper regions of the Bay’s tributary systems are vertically well mixed and characterized by low salinity (< 5‰). The mesohaline and polyhaline regions of the Bay and its tributaries tend to be vertically stratified. The mainstem of the Chesapeake is connected to a large number of tributary estuaries that receive and process large freshwater and nutrient inputs (Fig. 1).

Fig. 1
figure 1

Map of (a) Chesapeake Bay watershed with river input monitoring stations for the nine major tributaries, and (b) Chesapeake Bay estuary with 133 water quality monitoring stations

Full size image

Estuarine Data

Measurements of Secchi depth, the light attenuation coefficient (kd), surface-layer (≤ 1 m) nutrient and TSS concentrations, temperature, salinity, and chlorophyll-a were obtained from the U.S. Environmental Protection Agency Chesapeake Bay Program (CBP) Water Quality database (https://www.chesapeakebay.net/what/data). These variables have been measured at 2–4-week intervals at 133 stations within the main body of Chesapeake Bay and its tributaries since 1985 (Fig. 1b).

We pre-processed the estuarine data to generate a uniform, consistent data set for 1985–2016 and to adjust measurements for changes in laboratories and methodology over time. We first averaged any fortnightly data to monthly values and filled-in missing values at each station using information from nearby stations. The neighboring stations were selected as being within 15 km from each other over water from a particular station to nearby stations. If there were no neighboring stations within 15 km, information from the closest station was used. When the measurement method or laboratory changed, we adjusted the pre-change measurements at that station by a level shift calculated as the difference of medians for the values 1 year before and after the change. If the period before or after the method change was shorter than 1 year, then the median of all available data in that period was calculated. When adjustment factors (ratios) from side-by-side comparisons of alternative methods are not available, this approach of comparing relatively short 1-year periods preserves long-term trends in time series better than comparing whole sub-periods and can be applied in an automatic regime in a variety of settings (see Fig. S1 in Supplementary Material for a schematic illustration of the method).

Climatology and Deviations from Long-term Means

We computed a “climatology” for Secchi depth (i.e., long-term monthly means) at each station to characterize mean seasonal variations and perform comparisons among stations. The deviation from the climatological mean was then computed for each month (e.g., SecchiMay_1985–SecchiMay_climatology) at each station to generate long-term time series of climatological anomalies. Annual anomalies were also computed as deviations of the respective annual means from a long-term mean, to further quantify local long-term changes and to relate long-term patterns in these anomalies to changes in controlling variables.

Nutrient and Sediment Loading Estimates

We used the studies by Zhang et al. (2015), Moyer et al. (2017), and Zhang and Blomquist (2018) to compile riverine discharge observations and loading estimates for sediment (SS), organic carbon (total, TOC; dissolved, DOC), chlorophyll-a (Chl-a), total phosphorus (TP), and total nitrogen (TN) at the nine U.S. Geological Survey River Input Monitoring (RIM, Fig. 1a) stations. The RIM stations monitor the fall-line locations of nine major tributaries to Chesapeake Bay and collectively account for a substantial fraction of total Bay input. The published estimates of daily load and their annual aggregates were obtained using the Weighted Regressions on Time, Discharge, and Season (WRTDS) method (Hirsch et al. 2010). This method can offer generally improved estimates over other regression-based approaches because it does not assume homoscedasticity of model errors, fixed seasonal variations in concentration, or a fixed concentration-discharge relationship (Chanat et al. 2016; Hirsch 2014; Moyer et al. 2012; Zhang et al. 2016a).

We aggregated the nutrient and sediment loads from the tributaries for each tidal monitoring station based on geographical location, and the loads from the rivers entering the Bay to the north and/or upstream were summed for those tidal stations. For example, tidal stations at the very north of the estuary were only related to the loads from Susquehanna River (measured at RIM station 1; Fig. 1a), whereas stations in the mid-bay accumulate the influence of Susquehanna, Potomac, Patuxent, and Choptank rivers (sum of loads measured at RIM stations 1, 2, 8, and 9; Fig. 1a) (see Fig. S2 in Supplementary Material for more details).

Correlating Secchi Depth to Environmental Factors

We completed three types of correlation analysis to relate Secchi depth to watershed inputs and to related water-column properties. First, we computed long-term anomalies (annual mean in a given year minus the long-term mean over 1985–2016) in Secchi depth and controlling variables (e.g., chlorophyll-a, TSS) within each mainstem Bay region to evaluate annual-scale co-variability in these variables. Secondly, we used Spearman’s rank correlation to quantify the relationship between Secchi depth and both external material loading (e.g., TSS, total nitrogen) and estuarine conditions (e.g., salinity, chlorophyll-a, TSS). For this set of correlation analyses, we subtracted the trend-seasonal components from each time series to avoid spurious correlations derived from trends in the data. In particular, smoothed trends were estimated using locally weighted scatterplot smoothing (lowess, Berk 2016; Cleveland 1979; Hastie et al. 2001), and additive seasonality was estimated using corrected long-term averages for each month. Plots of the autocorrelation function with 95% confidence bands were checked randomly to confirm that autocorrelation of transformed series at seasonal lags of 12 and 24 months is not statistically significant. These routine steps allowed us to avoid spurious correlations caused by synchronous seasonal cycles in different variables (see Algorithm S1 and Fig. S7 in Supplementary Material for more details). We used alternating conditional expectations (Breiman and Friedman 1985) to determine if the Spearman correlation coefficients were appropriate for examining relationships between variables (see Fig. S8 in Supplementary Material), and the Benjamini-Hochberg adjustment procedure (Benjamini and Hochberg 1995) for maintaining false discovery rate (FDR) at 5% in each family of tests. Finally, we repeated the correlation approach, but lagged the watershed input variables in monthly increments from 1 month prior to 12 months prior. We represent this analysis with a box plot of correlation coefficients (Fig. 8) from 133 stations; thus, the 95% confidence bounds (horizontal dashed lines) are liberal since they are not adjusted for the large number or tests performed.

Multivariate Analysis

To further assess the relative importance of each variable on changes in Secchi depth on a bay-wide scale, we combined 26 potential variables controlling Secchi depth together using classification and regression trees (CART) and random forest (Breiman 2001; Breiman et al. 1984). The goal of CART and random forest is to develop a sequence of splitting rules to separate the observations of response variable (i.e., Secchi depth) into relatively small homogeneous groups of observations using the associated values of explanatory variables. Explanatory variables included watershed loads of SS, TN, TP, and freshwater input (Flow), as well as local variables that include station latitude and longitude, salinity, water temperature, chlorophyll-a concentration, TSS concentration, estuarine nitrogen and phosphorus concentrations (TN; TP; dissolved inorganic N, DIN; dissolved organic P, DOP; total organic N, TON), and dissolved oxygen concentration (DO). Additional variables included lagged TP, TN, and SS loadings (lags of 1 and 2 months), time (decimal year), month (numeric), and cluster. The cluster variable is a factor with three levels, including “Station belongs to the main cluster in 1985–2000 and in 2001–2016,” “Station belongs to the main cluster only in 1985–2000,” and “Other,” as identified by a spatiotemporal clustering approach (i.e., TRUST procedure; see Supplementary Material).

The CART algorithm searches over all values of all explanatory variables, looking for such a value that would split the observations of response variable into the most distinct subsets. Subsequent splits can be made using the same or a different explanatory variable. Thus, nonlinear relationships between response and explanatory variables and (if subsequent splits are based on different explanatory variables) interaction effects between variables can be accommodated in a tree (Berk 2016). The fitted value for some realization of explanatory variables is an average value of Secchi depth in the corresponding terminal node of the tree. Growing big trees with many splits can provide a good fit to the sample data; however, such trees are unstable (adding or removing few observations may change the tree dramatically) and hard to interpret.

A random forest, in contrast, comprises a large number of CARTs, where each new CART is obtained on a new sample with replacement (i.e., bootstrap sample) from the original observations. The bootstrapping allows obtaining multiple different CARTs while having only one sample dataset. The larger the difference between trees in a random forest, the more stable its results are. To make the trees even more different, only m ≤ p variables are considered when defining a new split in a tree (m variables are sampled randomly without replacement from the original pool of p explanatory variables). We set the number of trees to 500 and select at random m = ⌊\( \sqrt{p} \) ⌋ variables at each tree split (see Table S1 in Supplementary Material for more details). We assess the importance of predictors using the permutation-based approach (Breiman 2001). Note that CART and random forest automatically account for various combinations and interaction effects of the original variables, but like classic regression methods, they do not necessarily indicate causal relationships.

Results

We completed a comprehensive analysis of spatial and temporal patterns in Secchi depth in the Chesapeake Bay estuarine complex and associated estuarine biogeochemical data, and utilized several statistical techniques to identify the key water-column and external variables that drive Secchi depth variability across space and time. We identify salinity, a proxy for CDOM, as a key driver of spatial variability bay-wide, but highlight TSS as an important covariate in low-salinity waters while chlorophyll-a is a key covariate in mesohaline and polyhaline waters.

Patterns of Watershed Loadings

Strong spatial and temporal patterns were found for loading estimates of SS, fine-grained SS, TOC, DOC, and Chl-a from the nine major tributaries in 1984–2016 (Zhang and Blomquist 2018; see Fig. S3 in Supplementary Material). Constituent export from the monitored non-tidal Chesapeake Bay watershed (MNTCBW) was strongly dominated (~ 90%) by the three largest tributaries, namely, Susquehanna, Potomac, and James. Among the nine tributaries, the ranking of export generally follows the order of their watershed sizes. SS export was dominated by fine-grained sediment in all nine tributaries. In particular, the sediment in the Susquehanna consisted of almost entirely fine sediment throughout the long-term record. TOC export was dominated by DOC in all tributaries except the Potomac River. Considering all nine tributaries combined, 90% of sediment is fine-grained sediment, while 60% of TOC is DOC.

To illustrate temporal trends in nutrient, sediment, and organic carbon inputs, flow-normalized loads were derived to remove the complication of interannual flow variability (Fig. 2), but note that these flow-normalized loads were not used in subsequent statistical analysis for Secchi depth patterns. The overall increases in SS, fine-grained SS, and TOC for the MNTCBW are largely attributable to rises in loads from the Susquehanna River—the largest tributary to Chesapeake Bay (Zhang and Blomquist 2018). In turn, the increasing inputs from the Susquehanna River were driven by the diminishing trapping efficiency of the Conowingo Reservoir (Hirsch 2012; Zhang et al. 2013; Zhang et al. 2016a, b). Similarly, TP has risen dramatically since the late 1990s because its dominant particulate fraction is also increasing with reduced trapping efficiency (Fig. 2). By contrast, TN (and its dominant dissolved fraction) from the MNTCBW has declined over the last three decades, particularly in 1980s–1990s (Moyer et al. 2017; Zhang et al. 2015).

Fig. 2
figure 2

Long-term trends in flow-normalized loads of selected water quality constituents from the monitored non-tidal Chesapeake Bay watershed. All loads are scaled by their respective long-term medians to aid comparison. Adapted from Moyer et al. (2017), Zhang et al. (2015), and Zhang and Blomquist (2018)

Full size image

Spatial Patterns of Secchi Depth

Secchi depth was much lower in low-salinity regions of the estuary (typically < 1 m and including the tributaries) relative to mesohaline and polyhaline regions, and in these low-salinity regions, seasonal variability in Secchi depth was low (Fig. 3). In contrast, Secchi depth in mainstem Bay stations displayed clear seasonal cycles, with March-April minima in the upper Bay (salinity < 10‰), April–July minima in the middle Bay (mean salinity 10–16‰), and June–July minima in the lower Bay (mean salinity > 16‰; Fig. 4). It should be noted that these spatial groupings were not determined from our statistical analysis, but chosen based upon prior analyses in this estuary that delineate the Bay into longitudinal spatial groupings (e.g., Smith and Kemp 1995; Testa et al. 2018).

Fig. 3
figure 3

Climatology (i.e., long-term mean monthly values) of Secchi depth over 1985–2016 for three clusters of stations in tidal fresh and oligohaline regions of Chesapeake Bay western tributaries

Full size image
Fig. 4
figure 4

Climatology (i.e., long-term mean monthly values) of Secchi depth over 1985–2016 for three clusters of stations in the mainstem of Chesapeake Bay

Full size image

Deviation of Secchi Depth from Long-term Means

We found clear long-term patterns of change for Secchi depth across different regions of Chesapeake Bay. Over the 1985–2016 period, Secchi depth declined in the upper, middle, and lower regions of the Chesapeake Bay mainstem during summer (Figs. 5 and 6), increased slightly in a subset of tidal fresh tributary regions (data not shown), and was relatively constant throughout the winter-spring in the mainstem Bay (Fig. 6). Long-term declines in Secchi depth were significant in summer months (May–July) and fall (September–October) throughout the mesohaline and polyhaline regions of the mainstem (Fig. 6). This focusing of long-term Secchi depth decline in the May to July period corresponds with the season in which seasonal minima in Secchi depth typically occurred in the middle and lower mainstem (Fig. 4). These long-term changes occurred despite significant interannual variability in Secchi depth on a year-to-year, periodic, and near-decadal timescales (Fig. 5).

Fig. 5
figure 5

Long-term changes in the deviation from mean conditions for Secchi depth (left panels), key controlling variables (second from left), the long-term regional median deviation (i.e., regional aggregation) for Secchi and each controlling variable (second from right), and linear correlations between regional medians of Secchi depth and each controlling variable

Full size image
Fig. 6
figure 6

Space-season plots of linear temporal trends, represented by the Pearson correlation coefficients (r) between year and monthly means of Secchi depth (a) and associated p-values (b) in the mainstem of Chesapeake Bay. In the right panel, black areas correspond to p values of 0.05 or less, which were corrected for the propagation of type I errors by the Benjamini-Hochberg procedure (Benjamini and Hochberg 1995)

Full size image

We found spatially specific relationships between the non-detrended Secchi depth and water-column properties. Given that riverine inputs of sediment during the summer period are typically reduced in this estuary, phytoplankton biomass (for which chlorophyll-a is a proxy) and (perhaps more importantly) phytoplankton-derived detritus become larger contributors to light attenuation during this season. Interannual variability in chlorophyll-a was significantly related to Secchi depth variability in all regions of the estuary, including corresponding long-term increases in chlorophyll-a and declines in Secchi depth in the upper and middle Bay and co-variability of long-term means in the lower Bay (Fig. S9). A more detailed analysis indicates that long-term deviations in the three regions correlated with different (but related) water-column variables (Fig. 5). In the upper Bay, long-term mean deviations in Secchi depth most strongly correlated with TSS (r = − 0.72, p < 0.001), which is consistent with a general seasonal minimum in Secchi depth in Feb–April in this region (Fig. 4), a season during which TSS inputs typically peak and long-term increases in chlorophyll-a have occurred (Testa et al. 2018). In the middle Bay, long-term deviations in Secchi depth correlated highly with chlorophyll-a (r = − 0.73, p < 0.001), indicating that material associated with phytoplankton is the dominant driver of interannual variability in degraded light conditions in this region (Fig. 5). Finally, the lower Bay stands out in that long-term deviations in Secchi depth most strongly corresponded to changes in the ratio of Chl-a:TSS (r = − 0.59, p < 0.001), suggesting Secchi depth is lower when phytoplankton-derived material is a larger fractional contributor to TSS (Fig. 5).

Relationships between Secchi Depth and Local Variables

The relationship between Secchi depth and environmental variables varied widely across space in Chesapeake Bay. Across the analyzed stations, the Spearman correlation coefficients for Secchi depth and TSS concentrations range from − 0.72 to − 0.01 (individually significant at α > 0.05 if above 0.1 by absolute value), with a median of − 0.44 (Fig. 7a). The strongest monotonic relationships are observed in the tributaries and at the mouth of Susquehanna River in the upper Bay. A more drastic separation of correlations (by salinity regime) is observed for the Secchi relationship to chlorophyll-a concentration (Fig. 7b). Specifically, most of the stations in the mainstem (middle and lower Bay) exhibit moderate negative correlations in the range from − 0.6 to − 0.4, whereas the relationship is opposite in tributaries and at the mouth of Susquehanna River, with weak positive correlations from 0 to 0.2. The correlations between Secchi depth and salinity are mostly positive (i.e., higher Secchi depth with increasing salinity; Fig. 7c), with a median of 0.24 and the first quartile of 0.15, which is above the threshold of individual significance of 0.10 at the 95% confidence level (family-wise statistical significance is denoted by different shapes of the points in Fig. 7).

Fig. 7
figure 7

Maps of the Spearman correlation between detrended and deseasonalized series of Secchi depth and two sets of independent variables. a–c Local water-column conditions. d–h Watershed inputs. Note, FDR = false discovery rate

Full size image

Further exploration of relationships between salinity and controls on Secchi indicate a strong spatial pattern. Correlations between Secchi depth and both Chl-a and TSS were calculated separately for each month (rS{Secchi, Chl-a}and rS{Secchi, TSS}) and plotted against the corresponding monthly salinity to reveal the degree of correlation over space (Fig. S10). The strongest negative correlations rS(SecchiMonth, Chl-aMonth) are observed in spring (March–May) in regions with salinity above 7‰; in lower-salinity regions, the correlations are mostly weak positive all year round (Fig. S10a). Similarly, rS(SecchiMonth, TSSMonth) are more strongly negative in spring months, but the separation is not as strong as for chlorophyll-a (Fig. S10b).

Relationships between Secchi Depth and Watershed Loadings

In general, the correlations between detrended watershed inputs and Secchi depth were weak across the estuary and its tributaries. The relationships between Secchi depth and both total SS and fine-grained SS loadings were similar (Fig. 7d, e), where only about half of the stations exhibit statistically significant, weak negative correlation with Secchi depth. The median Spearman correlation across stations was − 0.10, and the critical value was 0.10. Overall, the correlation coefficients between Secchi depth and SS loading varied from − 0.37 (− 0.36 for fine SS) to 0.03 (essentially, 0). The differences between SS and fine SS were trivial because fine SS loading is the main component of the total SS loading, primarily coming from Susquehanna River (Fig. S3). Stations with the most significant correlations were located in the upper mainstem Chesapeake and tidal fresh regions (e.g., upper Potomac, Rappahannock, and James rivers; see Fig. 7d, e), and a comparable analysis without detrending produced significant correlations between Secchi depth and both river flow and fine SS loading (r < − 0.8) in the upper mainstem Bay. The correlations between Secchi depth and nutrient loadings (TN and TP) generally replicated the spatial pattern of correlations with sediment loadings (weak; more negative in upper mainstem and upper tributaries; Fig. 7f, g). However, Secchi depth is more strongly related to nutrient loadings than to sediment loadings. For example, estimated correlation coefficients with TP range from − 0.42 to 0.06 (median of − 0.13); with TN from − 0.53 to 0.07 (median of − 0.15). Upon considering the lagged correlations between watershed input and Secchi depth, lags of 0–2 months produced the only significant relationships and correlation of coincident observations (i.e., at the lag 0) remained the strongest. There were too few stations with significant correlations at lags 3–4 months and none at lags above 4 months, as indicated by testing the significance of the Spearman correlations with Benjamini-Hochberg adjustments (Fig. 8).

Fig. 8
figure 8

Boxplots of Spearman correlation between monthly Secchi depth and lagged nutrient and sediment loadings across 133 stations, i.e., rS(SecchiStation, t, LoadingStation, t + Lag). Horizontal dashed lines correspond to 95% confidence interval for testing individual null hypotheses of no correlation. Purple lines identify critical levels after applying the Benjamini-Hochberg adjustment procedure for maintaining false discovery rate at 5%, for each boxplot (absence of purple lines means no statistically significant correlation was detected after applying the procedure)

Full size image

Multivariate Analysis of Controlling Factors

The CART analysis including bay-wide Secchi depth and controlling watershed and water-column variables automatically selected splits defined by salinity, TSS, TP, and TON concentrations (Fig. 9). The remaining variables had a lower ability to produce distinct subsets of Secchi depth; thus, they were omitted from the final tree (Fig. 9). The smallest terminal node contains 2875 observations (about 6% of the total), which is still a large number, and the heterogeneity of observations was still high (standard deviation within the node is 0.44 m). Therefore, we proceeded by training a random forest with a large number of saturated trees, and similar to the CART results, the two most important predictors were salinity and TSS concentration (Fig. 10). The visible difference between the two methods is in a relatively high importance of geographic coordinates (latitude and longitude) in the random forest plot, which did not appear in the CART plot (Fig. 9). This results from differences in the two methods, where in CART only the most explanatory of two correlated predictor variables (e.g., salinity and latitude) will be selected, but in random forest, a random subset of m variables is considered at each split (m ≤ p; in our analysis, m = ⌊√p⌋); thus, there are multiple instances when correlated variables do not enter the subset of m together and can be selected for splitting the response without “competing” with each other. Regardless of this technical difference, both methods highlight salinity and TSS as the most important predictor variables.

Fig. 9
figure 9

An inverted regression tree for Secchi depth in 133 stations in Chesapeake Bay. In each node, average Secchi depth (in meters) is shown along with the percentage of cases from the total sample size (n = 48,811); terminal nodes are at the bottom. TN, TP, and TSS are all concentrations within the estuary water-column

Full size image
Fig. 10
figure 10

Permutation-based importance of predictors in a random forest for Secchi depth (response variable) in Chesapeake Bay

Full size image

Discussion

Our analysis of 30 years of Secchi depth observations, watershed inputs, and water-column properties in the Chesapeake Bay estuarine complex illustrates strong spatial and temporal patterns in the relative association of key controlling variables (CDOM, TSS, and chlorophyll-a) on long-term trends in Secchi depth. A clear increase in the flow-normalized delivery of watershed SS, TP, and organic carbon from the nontidal Chesapeake watershed has been documented over recent decades (Fig. 2), but correspondence between the temporal variations between these inputs and Secchi depth was only discernible in low-salinity regions. Conversely, clear long-term changes in phytoplankton biomass have occurred throughout the estuary and its tributaries, and organic material associated with this biomass covaries with Secchi depth in seaward regions. These results are consistent with spatially dependent contributions of different factors to Secchi depth in other estuarine systems (Harvey et al. 2019) and underscore that local conditions (e.g., relative freshwater influence) modulate Secchi depth responses to external forcing in Chesapeake Bay.

The location of a station relative to the salinity gradient appeared to determine the strength of correlation between Secchi depth and potential controlling variables. Secchi depth was much lower in low-salinity regions of the estuary (both mainstem and tributaries) relative to mesohaline and polyhaline regions (Fig. 3), which has been well described across these environments due to elevated levels of TSS and CDOM in low-salinity regions (Fisher et al. 1988, 1998). The primary conclusion of these prior studies was that low-salinity waters were the local recipients of suspended sediments and CDOM from adjacent watersheds or tidal wetlands that were maintained in the water-column to cause high light attenuation. In an apparent contrast, the correlation analysis between Secchi depth and external inputs of SS and freshwater revealed generally weak correlations (Fig. 7), which suggests that sediment inputs do not directly associate with elevated turbidity on the timescale of a month on a Bay-wide scale (i.e., effects of inputs may be episodic). The exception to this pattern is that SS input and TSS concentration are significantly correlated in the upper reaches of tributaries and the estuarine turbidity maximum region of the Chesapeake mainstem, indicating that sediment load and TSS concentration are only temporally correlated where the sediment input is efficiently retained and resuspended by local physical factors over longer timeframes. Observations and numerical model simulations have shown that watershed sediment inputs are typically trapped in low-salinity regions, due to high sinking rates associated with these materials and local circulation features (Palinkas et al. 2014; Sanford et al. 2001). In most other regions, Secchi depth was most significantly correlated with the surface concentration of TSS or the concentration of chlorophyll-a (Fig. 7), consistent with previous statistical analysis in Chesapeake (Xu et al. 2005), suggesting that interannual variation in Secchi depth covaries most significantly with in-water properties, which may be further controlled by tidal mixing, wind-stress, benthic communities, and local nutrient availability (Devlin et al. 2008; Riemann et al. 2015). An additional explanation for the lack of strong correlation between watershed inputs and Secchi depth on monthly timescales may be that the removal of trend and seasonal components during the correlation analysis (Fig. 7) excluded important long-term trends in the data.

Spatially dependent seasonal variability in Secchi depth did, however, reveal how this aspect of water clarity is driven by external inputs and in-estuary processes (Figs. 4, 5, and 8). In low-salinity regions of Chesapeake Bay, the persistent March–April minima (Fig. 4) corresponds with the seasonal peak in Susquehanna River discharge (Zhang et al. 2013, 2015), indicating that in this particular region, local inputs of SS contribute to light attenuation and consequently, low Secchi depth. The seasonality of Secchi depth in more seaward stations in the mainstem was noticeably different, with a consistent April–July minimum in the middle and lower Bay (Fig. 4). Harding et al. (2015b) found that these two regions had similar long-term variations over a timeframe that included years prior to 1985. In these regions, Secchi depth most significantly correlated with chlorophyll-a (Fig. 5), which suggests that phytoplankton biomass and phytoplankton-derived detritus is a much stronger contributor to light attenuation in these types of waters (Fleming-Lehtinen and Laamanen 2012; Xu and Hood 2006), which are largely disconnected from watershed sediment inputs. In fact, the most dramatic declines in Secchi depth have occurred in the summer period (Figs. 5 and 6), a time where surface chlorophyll concentrations increased over the 30-year time series in the upper and middle Bay regions (Fig. S11).

While the spatiotemporal clustering analysis (TRUST) indicated that the vast majority of mainstem Bay and mesohaline tributary stations varied similarly over the last three decades, there was a clear separation of the low- and high-salinity waters (Figs. 7, S10). The overall strong positive correlation between Secchi depth and salinity highlights the strong role of CDOM (which is high at low-salinity) in attenuating light (Branco and Kremer 2005; Gallegos 2001) and that low-flow periods are associated with clearer waters, with either reduced watershed SS inputs (upper Bay) or lower nutrient inputs that would otherwise stimulate phytoplankton growth in the middle and lower regions. Prior studies have indeed found that freshwater inputs are a strong predictor of Secchi depth in statistical models for Chesapeake Bay (Harding et al. 2015b). Both the CART and random forest approaches indicate that salinity and local TSS concentrations were the primary predictor variables for Secchi depth (Figs. 9 and 10), which is consistent with the idea that high TSS levels influence Secchi depth in low-salinity regions, and high TSS levels associated with chlorophyll-a dominate in high-salinity regions. The fact that TP ranked high in the multivariate analyses along with TSS (Figs. 9 and 10) likely indicates the association of phosphorus with suspended inorganic particles (e.g., Keefe 1994), but could also indicate P stimulation of phytoplankton growth. In the high-salinity regions, however, nutrient concentration levels were the next best predictors of Secchi depth, where Secchi depth was lower with higher nutrient levels that indicated nutrient-supported phytoplankton growth (Fig. 9). While we did not have data for CDOM in our analysis, there is abundant evidence that CDOM concentrations are substantially higher in low-salinity waters (Rochelle-Newall and Fisher 2002) and that the strength of salinity as a predictor variable likely reveals the influence of CDOM for light attenuation in low-salinity waters.

The strong relationship between chlorophyll-a and Secchi depth is compelling because recent analyses have suggested that long-term deviations in the relationship between Secchi depth and kd are the result of increases in small organic particles that have altered the scattering to absorption ratio (Gallegos et al. 2011). Prior studies have documented declining values for the product of Secchi depth and kd (i.e., ZSD*kd; Fig. S12) in recent decades (Gallegos et al. 2011; Harding et al. 2015b) associated with Secchi depth declines between 1990 and the most recent decade (Fig. 5). Although the origin and nature of the small organic particles suggested by Gallegos et al. (2011) is not well understood at this time, increases in such particles could be associated with elevated phytoplankton growth and biomass or altered phytoplankton communities, both of which have occurred in recent decades (Fig. 5; Harding et al. 2015a). If we assume this is true, increases in the relative contribution of phytoplankton-derived materials to SS pools would alter the relationship of kd and Secchi depth and explain the decreasing ZSD*kd. The fact that chlorophyll-a was not a strong predictor in the random forest analysis (Fig. 10) could indicate the masking of chlorophyll-a contributions because of its association (and covariation) with TSS in moderate and higher salinity waters. Recent indications of reduced chlorophyll-a and improved light conditions over the past 5–6 years (e.g., Fig. 5; Orth et al. 2017), when Susquehanna River flow has been consistently moderate, have been associated with an apparent reversal of the Secchi depth decline and an increase in the product of kd and Secchi depth (Fig. S12). This is consistent with the temporal dynamics in the TRUST-derived cluster that represents the majority of the Bay, where Secchi depth declined from 1985 to 2000, followed by fluctuations and a slight rise in 2001–2016 (Fig. S5).

Temporal changes in Secchi depth in the mainstem Chesapeake revealed three distinct regional changes. While it is clear that mean deviations in long-term Secchi depth (i.e., changes from a long-term mean condition) indicate long-term declines in all regions of the mainstem Bay (Fig. 5), the local drivers of change appear to be different. In the upper Bay, long-term mean deviations in Secchi depth most strongly correlate with TSS (Fig. 5), which we might expect to result from the long-term increases in flow-normalized sediment inputs from the Susquehanna (Fig. 2). However, winter-spring chlorophyll-a has increased substantially in this region over the same time-period and is highly correlated with changes in particulate nitrogen and phosphorus (Testa et al. 2018), indicating that algal-associated light attenuation may also be a factor in this region. In the middle (mesohaline) Bay, where riverine SS loading was not correlated to changes in Secchi depth, long-term deviations in Secchi depth correlated highly with chlorophyll-a, indicating that material associated with phytoplankton is the dominant driver of light attenuation in this region. Finally, the lower (generally polyhaline) Bay stands out in that long-term deviations in Secchi depth most strongly correspond to changes in the ratio of Chl-a:TSS (Fig. 5), where Secchi depth is lower when chlorophyll-a (and presumably phytoplankton-derived material) is a larger contributor to TSS pools. These patterns emerge despite the fact that TSS can be an ambiguous variable in that it represents inorganic sediments, allochthonous organic material, and autochthonous phytoplankton-derived material. Spatially dependent patterns in the key driving variables for light availability are consistent with the difficulty in clearly associating Secchi depth changes with other patterns of eutrophication (Harvey et al. 2019; Riemann et al. 2015; Sanden and Hakansson 1996).

The potential role of changes in phytoplankton biomass in driving long-term patterns of Secchi depth is complicated by the fact that phytoplankton chlorophyll can be both a cause and consequence of altered light availability. Phytoplankton has been clearly associated with the absorption of light, where higher phytoplankton biomass should lead to a reduction of light penetration into the water-column (e.g., Gallegos et al. 2011). However, low-light conditions can stimulate phytoplankton to add additional chlorophyll to increase their light-harvesting capability in order to maintain a positive growth balance (Geider et al. 1996). While it is outside the scope of our analysis to address these issues directly, there are a few lines of evidence to suggest that chlorophyll-a-Secchi relationships are driven by phytoplankton-derived light attenuation. The upper Bay Secchi depth decline is strongly associated with a long-term increase in chlorophyll-a that is also linked to increases in particulate nitrogen (r2 = 0.91, p < 0.01), which suggests that phytoplankton-derived organic material—not simply cellular chlorophyll content—is increasing (Testa et al. 2018). Furthermore, variations in Secchi depth in the lower Bay were strongly correlated to the contribution of chlorophyll-a to TSS, indicating that as phytoplankton-derived material becomes a larger contributor to the water-column particulate pool, Secchi depth decreases. Recent studies have highlighted long-term changes in the carbon-chlorophyll-a (C:Chl-a) ratio associated with altered nutrient inputs (Jakobsen and Markager 2016) and have revealed that C:Chl-a is elevated under high light, low-nutrient conditions. This may be relevant for meso- and polyhaline regions of the mainstem Chesapeake where nutrient availability has recently declined (Harding et al. 2015b; Testa et al. 2018). While it is outside of the scope of this analysis to definitively associate changes in light availability to specific mechanisms, future investigations that examine direct measures of phytoplankton biomass and particulate matter characterizations in concert with chlorophyll-a would help improve this assessment.

Considering that improved water clarity and light availability to benthic autotrophic communities is a major goal of watershed and estuary restoration, explanations for long-term changes in Secchi depth are valuable for state and federal water managers. One of the key goals of the Chesapeake Bay restoration partnership is to increase the potential habitat for submerged macrophytes, and thus the spatial coverage of these plants (Kemp et al. 2004). Our analysis reveals that long-term declines in Secchi depth throughout the majority of mesohaline and polyhaline waters were likely driven by changes in the concentration of material associated with elevated chlorophyll-a concentrations, the reduction of which is the ultimate target of watershed nutrient and sediment reductions. Thus, successful restoration efforts should result in improved light availability in tidal waters. The long-term decline in Secchi depth in the mesohaline regions, with the recent switch to increasing Secchi depth (since 2011; Fig. 5), is consistent with time series indicating a recent increase in SAV coverage in the middle reaches of Chesapeake Bay (Orth et al. 2017). In contrast, SAV coverage has increased substantially in tidal fresh and oligohaline regions of the Chesapeake Bay where nutrient inputs have been reduced (Lefcheck et al. 2018), but SAV has declined in the lower mainstem Bay where Secchi depth has remained at relatively low levels (Lefcheck et al. 2017). An improved understanding of the controls upon and relationships between Secchi depth and light attenuation is needed in this system by expanding upon the statistical analyses presented here with targeted field and experimental studies that consistently measure water-column conditions, light availability, and bio-optical properties across multiple salinity regimes over several years.

Conclusions

Our analysis of 30 years of Secchi depth observations, watershed inputs, and water-column properties in Chesapeake Bay illustrates strong spatial and temporal patterns in Secchi depth as well as the time- and space-dependence of the controlling factors. Our analysis reinforces previous findings related to the strong influence of CDOM and TSS in low-salinity regions, but a larger role of phytoplankton-associated organic material in meso- and polyhaline regions. Although watershed nutrient, sediment, and freshwater inputs did not correlate significantly with Secchi depth at a monthly timescale outside of very low-salinity regions, water-column variables that represent the consequences of those inputs were important explanatory variables in tidal waters, such as TSS and chlorophyll-a. The apparent inconsistency of these two findings may be resolved by considering that there may be temporal lags between nutrient inputs and their impacts on Secchi depth and that there are local and transient controls on TSS and algal-derived organic matter that are not directly related to watershed input and may not be captured by an analysis of this type. Regardless of the specific mechanism, we were able to associate regionally dependent, long-term changes in Secchi depth with changes in TSS and chlorophyll-a concentrations. These findings reinforce the value of watershed restoration actions designed to limit nutrient and sediment inputs to Chesapeake Bay and its tributaries, but highlight regionally specific sensitivities to different factors.