1. Introduction
In recent decades, increases in eutrophic conditions in coastal systems have led to an increased frequency of harmful algal blooms (HAB) [
1,
2,
3,
4,
5,
6]. HABs can cause wide-ranging effects on the ecology of systems, local economies, commercial fisheries, and human health. Commercial fisheries are often hardest hit due to shellfish bed closures and mortality effects on fishery populations [
7,
8]. These effects are the result of the production of toxic species of phytoplankton and cyanobacteria or of secondary effects such as hypoxia. In 2018, the National Marine Fishery Service reported that HAB-related impacts on commercial fisheries totaled over 5.6 billion USD [
8].
The causes of HAB events are being actively researched and may be influenced by the confluence of many environmental factors (biotic and abiotic). It is known that sustained increases in nitrogen and phosphorus inputs have played a role as primary drivers of increased HAB events and their duration [
7,
9,
10]. Climate change has also contributed to these conditions adding to changes in mean water temperatures, salinity, rainfall, and sea-level rise that may also contribute to HAB frequency and duration [
11].
Increased awareness by the public and reported health-related effects on humans and animals have resulted in increased monitoring of HABs and the water quality indicators that may contribute to these events. Developing databases to assess trends in HABs and associated conditions can be both time consuming and expensive [
12]. There are many ongoing monitoring programs focused on the monitoring and prediction of HAB events in freshwater systems including those that supply drinking water [
13]. Similarly, in large coastal estuaries such as the Chesapeake Bay, there are robust water quality monitoring programs. These programs are developing large-scale open-source water quality datasets that are maintained by various state, federal, and research institutions such as the Water Quality Portal (United States Geological Survey, USGS) and AquaSat [
14,
15] (
Table 1). However, there are few monitoring programs or databases developed for HABs in tidal riverine and small coastal systems (e.g.,
https://mywaterquality.ca.gov/habs/, accessed on 8 August 2021).
An alternative approach to HAB field monitoring is the use of high-resolution remote sensing imagery to collect multispectral data across a wide area in a single satellite pass. Multispectral satellite sensors can detect changes in water quality by collecting surface reflectance containing the spectral characteristics of the water column. In open oceans, remote sensing technology has been used for over 50 years and has led to the production of large publicly available datasets combining in situ and remote sensing reflectance over long time periods [
16]. Many algorithms have been developed from these datasets to estimate chlorophyll concentrations [
17]. Early satellite platforms, such as Coastal Zone Color Scanner (CZCS—launched in 1978), Moderate Resolution Imaging Spectroradiometer (MODIS), and Medium Resolution Imaging Spectrometer (MERIS), had the ability to survey large swaths of open ocean but had poor spatial resolution. Although some of these platforms had high-frequency revisit times (e.g., 2-day revisit period for MODIS), their use for small-scale system research was limited due to their poor spatial resolution and, thus, inability to resolve smaller coastal estuaries and tributaries [
18].
Today, state-of-the-art satellite platforms have been launched and are providing improved data at higher spatial resolutions and a broader range of wavelengths. Platforms such as the European Space Agency’s Sentinel-2 multi-spectral imaging system and NASA’s Landsat 7 system have a higher spatial resolution (10 and 30 m, respectively) and more frequent revisit times than earlier platforms. Sentinel-2 comprises a dual satellite system, Sentinel-2A and 2B, each being identical in spatial, temporal, and spectral resolutions. With two operational platforms in orbit, the Sentinel-2 revisit time is only 5 days compared with Landsat 7, which has a revisit time of 16 days. For these reasons, we chose to use Sentinel-2 imagery data for this project.
Satellite imagery can be combined with field and in situ water quality data to develop predictive algorithms to fill in spatial and temporal gaps in field monitoring. Previously, combined databases such as those provided by AquaSat included primarily freshwater data and only limited estuarine data [
15]. AquaSat combines data from Landsat 5, 7, and 8 with data from the National Water Quality Data Portal (NWQP) [
14]. AquaSat consists of a database of matched remote sensing and water quality constituents. The AquaSat database contains over 600,000 matching data points over a temporal period spanning 1984–2019. Parameters incorporated in this database include total suspended solids (TSS), dissolved organic carbon (DOC), water color, chlorophyll a, and Secchi depth (SD). The developers of AquaSat also provide open-source tools and scripts to allow others to utilize their methods.
The primary goal of our research is to develop a database containing a set of geospatially matched water quality data and remote sensing imagery for given time periods that includes chlorophyll and other ancillary water quality parameters (where available) for small estuaries and tidal rivers of the coterminous United States. This database will provide researchers and environmental managers with matched time-series data of imagery and water quality enabling them to develop new algorithms for chlorophyll estimation and HAB conditions across a range of systems with different optical characteristics.
Although there are many examples of the successful application of remote sensing for chlorophyll mapping in individual estuaries around the globe, most of these are based on limited calibration datasets [
19,
20,
21,
22], and more robust testing is needed of combinations of chlorophyll algorithms and processors for atmospheric correction [
19,
23]. More robust testing has been conducted for the processing of Sentinel-2 and Sentinel-3 images to predict chlorophyll in lakes [
24], but the effectiveness of some algorithms can vary among optical water types (OWTs, [
20,
24,
25]). Historically, most researchers have distinguished between Class I waters, which are dominated by chlorophyll, and Class II waters with more complex spectra contributed by a combination of constituents including chlorophyll, suspended nonalgal particulates, and dissolved organic matter [
26]. More recently, using hyperspectral imagery, 22 OWT classes have been identified, including 13 for freshwater bodies and 9 for marine systems, with only some overlap between the two [
27]. Fewer optical classes have been identified using multispectral imagery using techniques such as fuzzy clustering [
28,
29].
Many challenges are currently limiting the development of robust chlorophyll and HAB estimation using remote sensing at fine scales in estuarine systems. Traditionally, ocean color algorithms (OC2, OC3) relied on absorbance in the blue and green bands which is also influenced by suspended sediment and colored dissolved organic matter (CDOM) which are more prevalent in coastal systems [
30]. In more turbid waters, researchers have had more success with chlorophyll algorithms based on ratios and differences involving red and NIR bands [
19,
23]. Atmospheric correction procedures can also affect the success of chlorophyll algorithms [
31]. Sentinel-2 Level 2C images have been corrected for atmospheric effects using Sen2COR, but Sen2COR is optimized for corrections over land, rather than water [
32], so other atmospheric corrections and cloud and cloud-shadow masking procedures are under investigation for use over water (
https://ioccg.org/group/atm-corr/, accessed on 19 September 2024). Some researchers prefer to focus on spectral-shape indices which sometimes can be applied without atmospheric corrections [
24], but performance without corrections has been inconsistent for estuarine systems in California [
33].
Chlorophyll algorithms are less well developed and tested in estuaries due to the complex optical properties of coastal systems, and some algorithms are optimized only for specific water types. Thus, a secondary goal of our work is to characterize the optical water classes, similar to Le [
34], represented in our matched database so that algorithm performance testing can be accomplished across optical water classes.
4. Discussion
The matched database contains 84,438 Sentinel-2 Level 1C observations and 9761 Sentinel-2 Level 2A observations matching water quality data with Sentinel-2 Level 1C and Level 2A image tiles (
Table 3). As we assumed, we found more matching data when we incorporated Sentinel-2 Level 1C data than Level 2A as non-atmospherically corrected imagery is available for a longer time period from ESA, and our data quality processing excluded additional image tiles. Data and stations were assembled from sources based on found data (
Table 1). These data were a mixture of publicly available data from internet sources, data derived from literature searches, and personal contacts. Our goal was to develop a database that maximized the number of water quality data observations matched with remote sensing imagery, knowing that during intermediate processing, cloud removal, and QA flagging processes, additional data would be removed from the database.
Spatially, our observation data are not evenly distributed throughout ecoregions sampled along the East, West, and Gulf Coasts (
Figure 1). Our station sites are concentrated in specific estuaries and ecoregions with large monitoring programs (
Figure 1). Station and observation densities vary between programs and ecoregions (
Figure 1). This is likely due to the activity of water quality monitoring and estuary programs in (NEPs, NERRs) throughout the USA and is indicative of not only areas of environmental concern but also how programs deploy resources in these areas. High sample densities, not depicted in
Figure 1, were seen in Long Island Sound, Chesapeake Bay, Pamlico Sound, South Coastal Florida (East and West Coasts), Apalachicola Bay, Corpus Christi Bay, San Francisco Bay, and some smaller densities in the Northwest in Oregon and Washington State. In Long Island Sound, for example, the USGS has been monitoring stream water quality for over 43 years [
52]. In Chesapeake Bay, agencies and institutions began intensively monitoring the Bay in 1984. This monitoring program is a collective effort that comprises state agencies in Maryland, Pennsylvania, and Virginia and includes NGOs and research/educational institutions [
53]. High-density and high-frequency monitoring program sites increase the probability of finding a match between Sentinel-2 imagery and a given site, especially with the frequent overpass schedule of Sentinel-2 satellites.
The biggest challenge in developing this database was harmonizing methods across monitoring programs. A careful review of metadata was necessary to resolve parameter name and method differences so that the field coding remained uniform throughout the processing iterations needed to generate the final product. It was also necessary to manually review some full datasets and individual files to account for changes in parameter names and formats that may have occurred throughout the monitoring programs’ lifetimes. This also required R code to be customized for specific datasets that had shifting data formats and naming conventions across years. Some other difficulties we experienced were inconsistent quality control measures and workflows within and between data sources. We discovered it was not uncommon for QA flags to change coding and or meaning throughout a program’s life cycle, particularly in some of the long-term buoy datasets. Part of the matching and harmonization process required us to make decisions regarding the handling of missing data. To a large extent, we found matching water quality observations in continuous or semi-continuous records for most of our sample sites. Some parameters such as CDOM and DOC were not well represented in the water quality time-series records. Also, the units and methods reported for CDOM and DOC were variable and sometimes not documented. Ultimately, due to missing calibration information, we determined that all CDOM records in RFU would not be included in the final database files as processing progressed. They were retained in the unfiltered intermediate database for future review.
Chlorophyll presented some particular challenges not inherent in some other supporting water quality data. Reporting units varied across data sources for chlorophyll, some reporting concentration and others only reporting raw RFU units from in situ sensor systems. The RFU data were not included in our harmonized database as we could locate no supporting calibration information relating chlorophyll sensor response (voltage) with the resultant RFU, which makes conversion to concentration units impossible. In some datasets, phaeophytin was not accounted for in the sample analysis or documentation. In those cases, chlorophyll was retained if reported as Chl-
a (total chlorophyll). Since phaeophytin can interfere with Chl-
a analysis, as its absorption and fluorescence peaks are in the same region as Chl-
a, we only include Chl-
a data reported with phaeophytin corrections [
49,
54].
Other chlorophyll data issues arise due to differences between sampling methods, both in vivo and in vitro. For laboratory methods, the accuracy and precision of chlorophyll values can be affected by sample collection, concentration techniques, storage protocols, choice of extraction solvent and method, bandwidth of spectrophotometers, and chlorophyll algorithm applied [
55,
56]. While earlier literature suggests that spectrophotometric or fluorometric analyses overestimate chlorophyll concentrations in comparison with HPLC measurements [
57], others find that in vitro fluorometric and spectrophotometric measurements compare well with HPLC as long as allowance is made for chlorophyllides and allomers [
58]. The accuracy of in vivo fluorometric methods can be more problematic, with the instrument deployments affected by biofouling over time and observations varying with phytoplankton composition, nutrient status, CDOM concentration, and nonphotochemical (NP) quenching in high-intensity light environments [
56]. The accuracy of individual sensors may also vary; Wet Labs ECO-Triplet fluorometers may produce Chl-
a estimates 2–6 times greater than the extracted concentration using the standard factory sensor calibration [
59]. Methods to correct for sensor biases, instrument drift, and NP quenching do exist but are not routinely applied [
60].
The new EstuarySAT database provides chlorophyll data, including matched Sentinel-2 sets, across a broad array of estuary classes [
61] and geographic regions of the conterminous United States. Given the range of methods used, both laboratory-based and in vivo sensor-based, and uneven geographic distribution, care must be taken in using the database for a broad-scale assessment of chlorophyll status and trends in estuaries of the US without a consideration of methodological differences. However, the diversity of types and optical classes represented will provide researchers with the opportunity to test the representativeness and robustness of different chlorophyll algorithms developed with more limited calibration datasets and using different atmospheric correction processes. Some chlorophyll algorithms such as spectral-shape indices might be used with the top-of-atmosphere reflectances available from Sentinel-2 Level 1 or with the atmospheric corrections from Sen2COR available in Sentinel-2 Level 2C datasets, but previous researchers have had mixed success [
24,
33], so more testing is needed. Other ratio-based algorithms will require testing in conjunction with various atmospheric correction procedures beyond the Sen2COR corrections available in Sentinel-2 Level 2C products [
23].
Our matched dataset includes samples from a broad range of water quality conditions for testing purposes (clear to turbid, trophic classes 1–4, and oligosaline to hypersaline). Optical water classes described by fuzzy cluster analysis of our database had similar spectral signatures to those described as typical of CDOM-dominated waters but lacked classes characteristic of chlorophyll-dominated waters without optical interferences [
45]. Although overlapping, the water quality of three of the classes appeared to be distinguished from others mainly based on salinity differences, which is often not considered in developing empirical chlorophyll algorithms (
Figure 6a,d). Salinity often covaries with CDOM along an estuarine gradient, so these differences may reflect that cross-correlation. In addition, backscattering coefficients vary with both temperature and salinity, so it is reasonable to detect differences in optical signatures for oligohaline vs. mesohaline samples [
62]. The EstuarySAT database can help meet the challenge of determining the robustness of existing algorithm performance across the salinity and chlorophyll gradients.
The EstuarySAT database provides the opportunity for more robust testing of existing chlorophyll algorithms within estuaries and freshwater tidal rivers at the fine spatial and temporal resolutions available from Sentinel-2. Some of these algorithms are more sensitive to the detection of cyanobacteria blooms [
24,
63,
64], while others are not. Even in cases where algorithms can only be used to better quantify chlorophyll levels in general, it is possible to apply chlorophyll thresholds (10 and 24 µg/L) as warning indicators of potential HAB blooms that warrant targeted testing [
20]. Broad-scale mapping of chlorophyll across estuaries can help in overall assessments of estuarine productivity and potential driving variables for bloom patterns in space and time. In addition, chlorophyll levels from remote sensing can be used in conjunction with other environmental variables and/or hydrobiogeochemical models to predict the likelihood of the occurrence of HABs [
22,
65]. The EstuarySAT database contains data from several estuarine systems with histories of HAB formation: e.g., James River, Puget Sound, Albemarle/Pamlico Sounds, San Francisco Bay St. Johns, St. Lucie, and Caloosahatchee.
In the future, we will be improving and updating the database as new water quality and improved remote sensing imagery become available. For example, the imagery provided in the GEE catalog is Sentinel Level 2A which is atmospherically corrected using ESA’s SNAP toolbox (Sen2COR correction algorithm). The literature has suggested that Sen2COR may not perform well over water [
46,
47]. This may affect the resultant estimated chlorophyll concentrations from the imagery data. We are currently reviewing different methods and software tools to atmospherically correct Sentinel-2 Level 1C imagery to improve any derived chlorophyll data and algorithms. Future research will include applying and examining the robustness of existing algorithms for chlorophyll and cyanobacteria bloom estimation in estuaries and tidal rivers. In the future, after adding atmospherically corrected Sentinel-2 Level 1C data, our surface site observation matching will increase by approximately 90%, with some losses after quality control, cloud, and band QA filters are applied. Our research will continue with database updates and additional observations as they become available, evaluating existing chlorophyll and cyanobacteria algorithms for prediction and examining water quality time series to predict blooms.