1. Introduction
Despite only covering a relatively small area of the planet’s surface—estimated to cover ~3% of the terrestrial surface of Earth—inland waters have great importance for numerous critical functions, since they provide ecosystem services such as hydroelectricity production, flood control, navigation, water supply, and fisheries [
1,
2,
3]. These water bodies provide services that influence human welfare, directly and indirectly, and, therefore, they emerge as a limiting factor in quantity and quality for human development and ecological stability [
4,
5]. In addition, inland water bodies act as sentinels of the ever-changing environment in their surroundings, reporting the status of phenomena such as climate change, developmental pressure, and land-use and land-cover change [
6]. Reservoirs are a distinct example of inland water bodies and are an easy target for waste disposal. Currently, freshwater ecosystems show an increase in degradation in water quality and ecosystem services due to human activities. Soil occupation, agriculture, urbanization, and industrial affairs comprise actions already described to affect these water bodies [
6]. So, reservoirs are subject to a wide diversity of anthropogenic stressors, making it necessary to evaluate reservoir ecosystem changes and to understand the magnitude and implications of these alterations [
7].
In order to protect and manage aquatic ecosystems, the Water Framework Directive 2000/60/EC (WFD), European legislation in the water policy and quality, requires each European Union (EU) member state to assess the quality of their water bodies (e.g., rivers and reservoirs) [
8]. The WFD classifies reservoirs as heavily modified water bodies (HMWB) due to physical alterations caused by human activity, substantially changing them in character [
8]. Traditionally, the ecological status of water bodies is defined according to their biological, chemical, and physical characteristics in comparison with reference values [
9]. However, for natural water bodies, the reference condition is their high ecological status (HES), while for HMWB it is the maximum ecological potential (MEP) [
10]. According to the WFD, the MEP is defined as the state where “the values of the relevant biological quality elements reflect, as far as possible, those associated with the closest comparable surface water body type, given the physical conditions, which result from the heavily modified characteristics of the body” [
10].
The evaluation of ecological potential in reservoirs is remarkably challenging due to their complex nature, which represents an environment different from lakes and rivers [
9,
11,
12]. Besides, consistent monitoring using the metrics defined in the WFD follows traditional in situ methods (e.g., water sample collection and laboratory analysis) that are often very time and money consuming to estimate the quality of water regularly [
13]. The WFD implementation has required a large expansion in the monitoring of water bodies which made it possible to perceive their ecological status. Indeed, many researchers advocate some change in the WFD-monitoring measures to provide sufficient spatial and temporal resolution, and, in many cases, to make it more cost-effective [
14]. In recent years new monitoring tools have become available, including Earth Observation [
15,
16]. The use of satellite data for monitoring has demonstrated a high potential to better standardize measures across Europe. This approach has been recognized by several authors through enhancing both spatial coverage and frequency of monitoring of variables, such as water color, chlorophyll
a (Chl
a), cyanobacteria, and emergent macrophyte coverage [
16].
Since the 1960s, remote sensing (RS) techniques have been used to monitor aquatic environments, by analyzing ocean color under the assumption that Chl
a—
a proxy for phytoplankton biomass quantification—and surface temperature could be estimated remotely [
17,
18]. Based on this, oceanographers started to remotely monitor the optical properties of waters’ constituents, such as phytoplankton, colored dissolved organic matter (CDOM), and total suspended solids (TSS) [
19,
20]. However, the application of RS techniques to inland waters remains a challenge. In general, less success is recognized when applying RS techniques for monitoring inland water bodies’ quality, due to their distinct shapes and sizes, the comparably more significant impact of their
border effect, and the variable composition of water constituents [
6]. Alternatively, satellite RS techniques have been used as an effective tool for supporting the implementation of the WFD [
21]. Today, there are a plethora of programs that make use of satellite RS technologies, such as the Copernicus Program. This is an EU-led initiative designed to establish a European capacity for the provision and use of operational monitoring information for environmental and security applications [
22]. Within the Copernicus Program, the European Space Agency (ESA) is responsible for the development of the Space Component. The Sentinel missions are Copernicus dedicated Earth Observation missions, composing the essential elements of the Space Component [
22,
23,
24,
25]. The Sentinel-2 mission provides continuity to services relying on multispectral high-spatial-resolution optical observations over global terrestrial surfaces [
22]. Sentinel-2 A and Sentinel-2 B are the two polar-orbiting satellites that comprise the Copernicus Sentinel-2 mission. This mission aims at monitoring variability in land surface conditions and makes use of its wide swath width (290 km) and high revisit time (5 days with the two satellites under cloud-free conditions at the equator, which results in 2–3 days at mid-latitudes) to support monitoring of Earth’s surface changes [
26]. Each satellite is equipped with a multispectral instrument (MSI) that works passively by collecting sunlight reflected from the Earth. This instrument is responsible for measuring Earth’s reflected radiance in 13 spectral bands, with 10 and 20 m spatial resolution and 3 bands at 60 m for atmospheric correction [
26].
Despite the capabilities of today’s satellite RS technologies, their direct products do not represent a sufficiently reliable portrait of the Earth’s surface. Satellites measure the light field emerging at the top-of-atmosphere, and thus an atmospheric correction (AC) needs to be performed as part of the processing of water-body data [
27]. Due to the low reflectivity of water, around 90% of the signal that reaches satellite sensors is affected by the absorption and scattering by different particles present in the atmosphere (e.g., water vapor, ozone, oxygen, carbon dioxide and aerosols) [
13]. The atmospheric path traveled by the generally low radiances at the water’s surface makes the requirements for AC very demanding [
27]. However, AC processors can remove the scattered signal of the atmosphere and retrieve the signal from the water’s surface [
28,
29]. The Case 2 Regional Coast Color (C2RCC) is an AC processor made available through ESA’s Sentinel Toolbox Sentinel Application Platform (SNAP). It relies on a database of radiative transfer simulations, inverted by neural networks. The core is a five-component inherent-optical-properties (IOP) model that was derived from the NASA bio-Optical Marine Algorithm Data set in situ measurements. C2RCC has been validated for the different sensors, with good results for Case 2 waters, as well as possessing special neural nets, such as C2X-Nets, which is trained for extremeIOP ranges [
27].
Ongoing developments in RS and geographical information science massively improve the efficiency in analyzing Earth’s surface features. The increased frequency of image acquisition together with the advances in the ability to process data provides new opportunities to understand the complex inland water systems [
30]. Modabberi et al. (2020) provided the first evaluation of the spatiotemporal variation of Chl
a across the Caspian Sea, as this water body had been subject to increasing pollution and environmental degradation [
31]. The authors made use of Level 3 MODIS-Aqua Chl
a data from January 2003 to December 2017 to discover that this water body had suffered from a growing increase in Chl
a, especially in warmer months [
31]. Modabberi et al. (2020) concluded that these trends reflect the increasing rate of degradation in the Caspian Sea [
31]. Ansper and Alikas (2019) used 89 Estonian lakes in a study that aimed to analyze the suitability of Sentinel-2 MSI data to monitor water quality in inland waters [
13]. The authors concluded that, despite their methods being able to provide complementary information to in-situ data to support WFD monitoring requirements, it is important to note that ACs are sensitive to surrounding land and often fail in narrow and small lakes [
13]. In the Iberian Peninsula, Sòria-Perpinyà et al. (2019) worked on Albufera de València—a hypertrophic lake in Valencia, Spain—that aimed to demonstrate the validity of an algorithm for Chl
a concentration retrieval from Sentinel-2 MSI [
32]. With the results obtained, the authors were able to infer that the temporal evolution of Chl
a concentration variations followed an annual bimodal pattern [
32]. In Portugal, Potes et al. (2018) used the Alqueva reservoir as a study site to assess the use of the Sentinel-2 MSI for water quality monitoring [
33]. Despite the set of algorithms being applied with good results, some tuning of the algorithms used was still required to make use of the full potential of the MSI [
33].
Despite challenging, the use of RS technologies may be an essential alternative, opposed to using exclusively traditional field-based methods to monitor water quality, as they offer a comparatively low-cost, high frequency, spatially extensive and practical complement for water-quality assessment and monitoring [
34,
35].
In this work we will focus on the study of the temporal and spatial evolution of Chl a and TSS, between the years 2018 and 2020, in order to show the validity of a proposed tool for Sentinel-2 images and an operative method for the multitemporal study in different reservoirs of Portugal: Aguieira and Alqueva. Specifically, we applied the C2RCC AC processor to Sentinel-2 imagery data aiming to (i) assess the portability of this AC processor between different reservoirs, and (ii) validate its use for a rapid assessment of water quality.
4. Discussion
Inland water RS has faced, and continues to face, many challenges, not only in terms of the science underpinning the retrieval of physical and biogeochemical properties over typically highly optically complex waters, but it has also suffered from lack of funding, infrastructure, and the mechanisms needed to coordinate research efforts across an historically fragmented community [
55]. This has meant that the inland water community has often had to make use of data from satellite sensors designed primarily for land applications. While these sensors have adequate spatial resolutions for some water bodies, their spectral coverage and resolution are not optimal for many applications over inland waters (e.g., CDOM retrieval). The optical complexity of inland waters, AC issues and adjacency effects add additional challenges to inland water RS [
55]. In this section, such challenges are approached, in order to assess their influence on the results and further improve the methods.
Regarding the spatial differences between both reservoirs, data from NRMSE and NMBE metrics for Alqueva showed interesting results. For Chl
a and TSS, NRMSE results were as high as 32% and as low as 8.9%, indicating a slight relative error (slight scatter of observations). As for the NMBE results, their values were even lower than the NRMSE, ranging from 4.2% to 7.2%. This means that systematic errors were very low and that in most observations there was a slight overestimation of in-situ data by the satellite products, except for Chl
a when using the 20 m product which indicated a slight underestimation. The Chl
a variable is commonly used as a proxy for the phytoplankton biomass present in a water body [
17,
18]. Hence, this study showed that, in Alqueva, phytoplankton is a more predominant component of suspended solids and, therefore, contributes more to water turbidity than in Aguieira. In particular, the most interesting results came out of applying C2RCC to Sentinel-2 MSI 60 m products in Alqueva. This is the case for both parameters. Although further research is needed to investigate the reasons behind Aguieira’s higher errors, Alqueva’s results indicate that satellite data can be very useful and reliable for monitoring reservoirs.
Plowey (2019) in a study using Sentinel-3 Ocean and Land Color Instrument (OLCI) imagery of lakes, achieved low errors for Chl
a retrieval (NMBE = −7%, RMSE = 40%, n = 156), but high errors when retrieving TSM by using the standard C2RCC neural network [
56]. Moreover, Kyryliuk and Kratzer (2019) in a study using Sentinel-3 OLCI imagery of the Baltic Sea, demonstrated that Chl
a was retrieved with a relatively low systematic error (NMBE = 10%), but a high relative error (RMSE = 97%, n = 26) [
57]. However, similarly to Plowey (2019), the authors observed a large systematic error (NMBE = 103%) and an even larger relative error (RMSE = 167%), when retrieving TSM [
56,
57]. In contrast to both studies, where problems were reported when retrieving TSM, this factor was not an obstacle in this study. In addition, Kyryliuk and Kratzer (2019) found that their results improved when studying the regional specific IOPs of their study area and using them to better configure C2RCC [
57]. This could constitute a solution to obtain more reliable results when monitoring reservoirs, particularly ones of smaller dimensions such as Aguieira. However, it is important to keep in mind that the metrics used are normalized by the range of observed in-situ data to allow for comparable results.
The optical properties of inland waters are highly variable between, and even within, water bodies. These issues confound the development of algorithms for inland waters and typically limit their applicability to different water bodies [
55]. Johansen et al. (2018) evaluated the performances of 29 algorithms that use satellite-based spectral imager data to derive estimates of Chl
a that, in turn, can be used as an indicator of the general status of algal cell densities and the potential for cyanobacterial harmful algal blooms (CHABs) [
58]. They aimed to identify algorithm-imager combinations that had a high correlation with coincident Chl
a surface observations for two temperate inland lakes, as it suggested portability for regional CHAB monitoring [
58]. Even though the two lakes were different in terms of background water quality, size and shape, and the results obtained support the portability of using a suite of certain algorithms across multiple sensors to detect potential algal blooms using Chl
a as a proxy [
58].
In the same line of thought, we also aim to assess the portability of the application of the C2RCC processing chain between the studied reservoirs. For this purpose, the spatial differences between the reservoirs should be considered. From the PCA using in-situ data from both reservoirs we concluded that, except for sampling site Al5, both reservoirs were similar in terms of physical and chemical parameters (
Figure 3). Hence, their dimensions are the biggest difference between themselves. As a reminder, while Aguieira only occupies a drained area of ≈300,000 ha, Alqueva is over ten times bigger, occupying a drained area of ≈5500.000 ha. In addition, it is also important to consider the geographic location of both reservoirs—Aguieira is located in the center of Portugal and Alqueva is located in southern Portugal. Yet, IOPs vary not only across geographic regions but also within the same water mass [
5]. The complexity of the reservoirs is mainly due to the spatial-temporal variability of the water constituents at the same site. In other words, the dominant constituent in the water column at a study site may not only change spatially across short distances, but also seasons [
59,
60].
Concerning both temporal and spatial dimensions, the methods appear more appropriate to be applied in Alqueva than in Aguieira. Notwithstanding, the application of the methods to Aguieira can allow better results if some fine-tuning is performed. This would require approaching some aspects of RS that constitute great challenges to many researchers in this field.
A first aspect that can raise difficulties in RS studies is the process of AC. Pereira-Sandoval et al. (2019) studied the most appropriate AC processor to be applied to Sentinel-2 MSI Imagery over several types of inland waters in Valencia, Spain, including eight reservoirs and a coastal lagoon [
61]. Statistical linear analysis showed that Polymer and C2RCC were the processors with the highest correlation coefficients and lowest errors when comparing in situ measurements and satellite reflectance [
61]. They concluded that, due to the results obtained for both these AC processors, it was possible to support the applicability of Sentinel-2 MSI for inland water quality estimation [
61]. However, Toming et al. (2017) tested the performance of the standard C2RCC processing chain in retrieving water reflectance, IOPs, and water-quality parameters such as Chl
a concentration, TSM concentration, and CDOM in the Baltic Sea [
62]. The Baltic Sea, just like reservoirs, is an optically complex water body where many ocean color products, performing well in other water bodies, fail [
62]. The authors observed that, although the reflectance spectra produced by the C2RCC are realistic in both shape and magnitude, the IOPs (and consequently the water quality parameters) estimated by C2RCC did not correlate with the in-situ data [
60]. However, the authors also observed that some tested empirical RS algorithms performed well in retrieving Chl
a, TSM, CDOM and Secchi depth from the reflectance produced by the C2RCC [
62]. This suggest that the AC part of the processor performs relatively well, while the IOP-retrieval part needs extensive training with the actual IOP data before it can produce reasonable estimates for a given AOI [
62].
Another issue concerns adjacency effect from neighboring land pixels, named
border effect [
63]. Inland water bodies are surrounded mostly by land, and
border effect is especially more significant in situations of raised, undulating topography around the waterbody [
63]. This not only means that light from objects surrounding the water body can modify the radiance that reaches the sensor, but also that large portions of the sky may be blocked by land surface (e.g., vegetation). Although Aguieira is characterized by flat areas, steep slopes appear in zones of conversion with other water courses, and various types of vegetation are present throughout the margins of the reservoir [
38,
40,
41].
A final issue concerns the temporal dimension, which should always be considered, particularly when discussing the dates when the samplings are performed compared to the dates when the satellite images are captured. Images for Aguieira in spring of 2020 were taken 9 days after in-situ sampling was carried out. Given this temporal distance, the dates of the observations presented in
Figure 5 were assessed. The observations for this period coincide with values further from the dashed x = y line and, therefore, it is important to assess what effect this has on the results. Hence, NRMSE and NMBE results for the Aguieira reservoir, without considering observations from spring of 2020, are presented in
Table 4. With these changes, relative error was similar in terms of Chl
a but decreased from 80.3% to 53.7% in terms of TSS, using the 20 m product. With the same product, systematic error went from underestimating in-situ values to overestimating, increasing from −16.4% to 18.9%, for Chl
a, and from −15.8% to 31.6%, for TSS. Using the 60 m product, relative and systematic errors were consistently lower than before, where both statistics for both variables decreased in value. Ideally, both NRMSE and NMBE results should be the lowest possible, indicating low relative and systematic errors, respectively. This would indicate that the satellite variables are precise estimations of their in-situ counterparts, therefore validating the methods used.
In addition, it is important to consider the services provided by the studied reservoirs. Among reservoirs, those built for generating hydroelectricity usually have the most pronounced fluctuations in water level. These fluctuations result from variations in the electricity demand [
64]. Also, reservoirs built for water storage aim to sustain the flow in the river downstream and level out ordinary fluctuations in discharge [
64]. The Aguieira and Alqueva reservoirs were built for these functions. In
Figure A1 (see
Appendix C), an evident temporal variation of the shape and size of both reservoirs is recorded. Therefore, given the regular changes that occur in a reservoir, it is ideal to collect samples on the same day when satellite images will capture the reservoir. Kyryliuk and Kratzer (2019) were able to plan this aspect in their work. They used the weather app “Weather Pro” to screen, with 7-day forecasting, for cloud-free dates closer to the “overpass” time of the satellite over their study area, successfully avoiding cloud interference that would result in low-quality match-ups, or no match-ups at all [
57]. ESA also provides a tool that allows to predict the “overpass” time of a satellite over an AOI [
65].
Particularly in cloudier seasons such as autumn and winter, there is less availability of suitable satellite images, i.e., images with no cloud, haze or cirrus interference and that capture the reservoir in its entirety. If the field campaign to retrieve water samples is not planned considering the “overpass” of the satellite over the AOI, there may not be suitable images to match with in-situ data. In turn, this will affect the accuracy of the results, or even impede the study altogether. While in Alqueva it was possible to use satellite imagery of the exact date or the day after samplings were performed, in Aguieira that was only possible for autumn of 2018. For the remaining seasons, the dates differed from 6 days to 9 days, and in the period of autumn of 2019 there was no imagery with enough quality to be used, near the in-situ data collection. This may be the reason behind the results of Aguieira, and the results presented in
Table 4 prove that relative and systematic errors are generally lower when using images from dates not too far apart or in the same day as in-situ samplings were carried out. In conclusion, ideally the dates should be the same for both retrievals because the availability of suitable satellite imagery can be a limiting factor when not considered beforehand.
The inland-water community is smaller in number, more fragmented and less well-funded than the ocean-colour community, particularly when one considers the number and complexity of the challenges currently faced. In general, the wider scientific community has been slow to fully recognise the importance of freshwater ecosystems to global-scale processes and the provisioning of ecosystem services upon which human society relies [
55]. Although inland waters comprise a small fraction of the Earth’s surface water, it is becoming increasingly clear that they are of disproportionate importance to the global biosphere [
66]. Despite a large amount of valuable inland water remote sensing research having been overlooked because it was either published in the predigital era or in the grey literature, i.e., conference proceedings, PhD theses, etc., the current advancements in this field of study have marked improvements in the accuracy, applicability, and robustness of RS products for inland waters [
55]. By studying the validity of applying C2RCC to these two reservoirs we hope not only to contribute to these improvements, but also bring forth new knowledge concerning inland waters, and particularly reservoirs.
In the last few years, several large projects on RS of inland waters have been funded (particularly within the EU), including: the ESA Diversity II project [
67] and the EC FP7 eartH
2Observe project [
68]. This funding is fundamental for the collective growth and improvement of the limnology and RS communities, as satellite remote sensing has been proven to be a low-cost and rapid alternative for monitoring water quality. In a direct follow-up to this study, it would be compelling to explore some aspects. Firstly, it would be interesting to use a different neural net—the standard or the new C2X-Complex-Nets—in search of better results. Secondly, it could be of interest to study the Aguieira reservoir with more detail in order to assess the issues previously mentioned. Ideally, new water samples would be collected on the same day as satellite images are captured, and regionally specific IOPs would be collected and used for a better parameterization of the C2RCC AC processor.