The HYPSTAR sensor has been developed in the frame of the H2020 HYPERNETS project and is part of the HYPERNETS network (Goyens et al., 2022). The model designed for water applications is the HYPSTAR-SR radiometer which takes radiance (field of view (FOV) = 2°) and irradiance (FOV = 180°) measurements from 380 to 1020 nm with a spectral sampling of 0.5 nm (i.e., Full Width Half Maximum is 3 nm).

As part of the HYPERNETS project, a system has also been developed to operate the HYPSTAR in autonomous mode and outdoor conditions. This system is mainly composed of a pan-tilt unit, to accurately point the radiometer towards specific directions, and a water-proof main box. This box contains an electronic card to switch ON/OFF the system, a rugged PC, a GPS antenna and a 3G card for data transmission. Ancillary sensors are used to measure humidity, temperature and pressure inside the main box. Deported sensors are used to detect rainfall and measure ambient light conditions. Two webcams, one pointing towards the system, the other towards the target/surrounding waters, are used for remote control and detection of potential presence of floating debris, animals or boats which could contaminate the radiometric measurements, respectively. The whole system is powered by a 12V battery connected to a solar panel for long-term autonomous operation. The hypernets_tools software ³ is installed on the rugged PC to execute predefined measurement sequences when environmental conditions (e.g., enough light and no rain) are favorable and store recorded data. After the execution of each measurement sequence, the recorded data is also automatically transferred (3G) to a remote server for quality-controls and further processing (i.e., computation of the water-leaving reflectance signal) (see Goyens et al. (2022) for details). The 3G communication also allows to modify remotely the data acquisition procedure. At night the system is OFF to minimize power consumption with the HYPSTAR pointing to nadir. The system wakes up every morning to execute measurement sequences at predefined times, typically every 15 or 30 min.

When executing a “water standard” measurement sequence, the system first detects the Sun position to adopt a +/- 90° or +/- 135° azimuth angle. The sensor is then successively pointed towards the zenith to measure the downwelling irradiance, (Ed), then at zenith angles of 140° and 40° to measure respectively the sky (Ls) and upwelling (Lu) radiances. These viewing angles are the ones recommended by Mobley (1999) based on computations and by Ruddick et al. (2006) based on field measurements. The duration of this sequence varies from 2 to 3 min, depending on light conditions and contains 6 measurements of each parameter recorded in this order: 3 Ed, 3 Ls, 6 Lu, 3 Ls and 3 Ed spectra.

The data transferred to servers are processed using the HYPERNETS Processor ⁴ . The data quality controls and computation of the remote sensing reflectance signal (Rrs in sr⁻¹) which is multiplied by π sr to obtain the water-leaving reflectance (Rhow, dimensionless) spectra are detailed in Goyens et al. (2022). Spectra are distributed to users as Rhow and Rhow_nosc values, i.e., with or without applying the NIR similarity correction (Ruddick et al., 2005), which was proved to be no longer valid in highly turbid waters (Doron et al., 2011). However, note that the quality checks are slightly different from one HYPERNETS site to another (and in some conditions from one application to another). Here, we have used quality checks that do not rely on the NIR similarity correction.

Satellite data considered in the present study were selected based on their specifications in terms of spatial, temporal and spectral resolutions which should be adapted to monitor water quality parameters in estuaries and coastal lagoons. As already highlighted (Doxaran et al., 2009; Ody et al., 2016; 2022; Novoa et al., 2017; Renosh et al., 2020), the combined S2A&B-MSI and S3A&B-OLCI satellite sensors from the European Space Agency (ESA) have great capabilities in this scope. S2-MSI has 12 wavebands from the visible to the shortwave-infrared (SWIR) with different spatial resolutions ranging from 10 to 60 m. The S2A&B-MSI sensors in activity since 2015 and 2017, respectively, provide a revisit time period of about 4 days at mid-latitudes. Comparatively, the S3A&B-OLCI sensors in activity since 2016 and 2018, respectively, provide daily observations at 300 m spatial resolution with 21 wavebands in the visible and NIR spectral regions. Level-1 data and Level-2 products were downloaded from CREODIAS ⁶ .

The L8 and L9-OLI sensors provide since 2013 and 2021, respectively, observations at high spatial resolution (30 m), with 9 wavebands from the visible to the SWIR and a revisit time period of about 9 days at mid-latitudes. These sensors are operated by NASA, the National Aeronautics and Space Administration. Data and products were downloaded from USGS ⁷ .

Finally, the Aqua-MODIS (Moderate-resolution Imaging Spectroradiometer) sensor (NASA) was also considered as it provides daily observations since 2002 at spatial resolutions of 250, 500 and 1000 m (19 wavebands from the visible to the SWIR). Level-1 data were downloaded from the NASA ocean color data portal ⁸ to generate level-2 products at 250 m spatial resolution using the SeaWiFS Data Analysis System (SeaDAS) l2gen function. This processing, however, is known to be limited in confined areas such as estuaries, bays and lagoons as only two wavebands are available at 250 m and since the wavebands at 1000 m spatial resolution are used to apply atmospheric corrections (i.e., water pixels partly on land will be masked).

The list of high-quality matchups, identified in 2021 and 2022 between satellite data and HYPERNETS measurements, is presented in Table 2, for each satellite sensor and HYPERNETS station. These quality matchups were identified based on the availability of HYPERNETS field data and cloud-free satellite data.

Before comparing satellite-derived and field-measured Rhow values, the HYPERNETS Rhow spectra were convolved with the relative spectral responses of each satellite sensor to obtain band-equivalent Rhow values (e.g., see section 2.3.1 in Renosh et al. (2020) for details). The quantitative comparisons were then based on five different statistical indicators: the slope, intercept and determination coefficient (R ²) of the best-fitted linear regression, the root mean square error (RMSE) (Eq. 1) and mean absolute percentage error (MAPE) (Eq. 2): R M S E = ∑ i = 1 n R h o w _ s a t λ − R h o w _ H Y P λ   2 n (1) M A P E % = 1 n ∑ i = 1 n R h o w _ s a t λ − R h o w _ H Y P ( λ R h o w _ H Y P λ × 100 (2) Where Rhow_sat(λ) is the water-leaving reflectance derived from the satellite at the waveband λ, and Rhow_HYP(λ) is the water-leaving reflectance derived from the HYPERNETS field measurements within the same waveband λ.

Around the HYPERNETS station operated in Berre, the spatial dynamics of the water reflectance was characterized using 20 S2-MSI and 20 S3-OLCI images here corrected for atmospheric effects using ACOLITE-GLINT (Figure 3). Independently of the spatial resolution (20 m S2-MSI, and, 300 m S3 OLCI, respectively), mean values of Rhow (555) in the 3 × 3 and 5 × 5 pixels boxes, respectively, show a very good agreement with the Rhow (555) values of the central pixel (where is located the station). This homogeneity is even confirmed when mapping the variations (in %) of Rhow (555) compared to the Rhow (555) value in the central pixel over a larger area (Figure 3). These variations are typically lower than 15% on S2-MSI images (despite some apparent residual glint effects) and lower than 7% on S3-OLCI images (Figure 3). Now looking at the temporal variations of field-measured Rhow (555) values (HYPERNETS data), between 10 h and 14 h (local time) over a 10 months period, these variations within +/-30 and +/-60 min time windows do not usually exceed 6%. Rare maximum variations of 10% are also detected. This could be explained by the presence of algal blooms causing higher diurnal variability in the optical properties of the water (i.e., August). This period of time also corresponds to higher variability in the HYPERNETS in-situ reflectance spectra (see Figure 2). Therefore, as a local adaptation of matchup protocols, extracting the mean satellite-derived value over the 3 × 3 pixels box and field measurements averaged within a +/-30 min time windows around the satellite data acquisition time seems appropriate in the central part of the Berre lagoon.

The spatiotemporal dynamics of the water reflectance at the mouth of the Gironde Estuary significantly differ from that observed in Berre (Figure 4). Around the HYPERNETS station located at the extremity of the pontoon (left shore), the high spatial resolution of S2-MSI pixels (7 images) show limited variations of the water reflectance signal (Rhow (665) within boxes of 1–5 × 5 pixels boxes (once the pontoon is accurately masked). This is no longer the case on S3-OLCI pixels which coarser spatial resolution results in a more significant scatter, as Rhow (665) values within boxes of 3 × 3 to 5 × 5 pixels are partly contaminated by the shore and the pontoon itself (Figure 4). As for the temporal variations of the water reflectance measured in the field, time windows of maximum +/-30 min must be considered to avoid variations larger than 20% (as the turbidity of the water, and consequently the water reflectance, strongly varies with tidal currents [e.g., Doxaran et al. (2009)]. This is also confirmed by Figure 2 with the very high variability in water reflectance expected at the Gironde especially in May and, slightly less, in November. Therefore, at the mouth of the Gironde Estuary, the site-specific matchup protocol consists in extracting the satellite-derived value of the closest pixel to the HYPERNETS station and average the field measurements within a +/-15 min time window.

Only few (5) quality matchups between L8&9-OLI satellite products and HYPERNETS field data are available in both the Berre lagoon and Gironde Estuary to assess the performance of existing processing algorithms and estimate the uncertainties associated to the water reflectance products. These results, however, tend to confirm those obtained with S2-MSI, with quite satisfactory results obtained with C2RCC in Berre (Figure 7; Table 5) and with ACOLITE-GLINT in the Gironde (Figure 8; Figure 5B).

Comparatively to S2-MSI and L8&9-OLI, numerous matchups were available (20 and 15, respectively, in the Berre lagoon and Gironde Estuary) to assess the validity of S3-OLCI satellite products generated using different processing algorithms.

Doxaran et al. (2009) used the “surface reflectance” MODIS land product at full spatial resolution (250 m) to retrieve and map concentrations of suspended particulate matter (SPM) in the whole Gironde Estuary. Here, the more conventional NIR-SWIR atmospheric correction method was applied to estimate water reflectance values at the mouth of the estuary characterized by moderately to highly turbid waters. The SeaDAS l2gen function allowed retrieving Rhow values at 250 m spatial resolution at all MODIS visible to NIR wavebands, by interpolation, for comparisons with HYPERNETS Rhow field measurements.

Unfortunately, the use of the l2gen function is problematic in inland and nearshore turbid waters such as estuaries and bays as atmospheric corrections and application of land, cloud and other masks rely on wavebands at 1000 m resolution. Therefore, as expected, few MODIS products satisfy matchup criteria and only 11 images provide matchups with MAFR HYPERNETS data. Including all MODIS wavebands, the matchup results show a significant but not satisfactory linear correlation (R ² = 0.62) between satellite-derived and field-measured Rhow values (Figure 11). The NIR-SWIR processing clearly tends to underestimate actual Rhow values, by about 30%. Results are inverse and satisfactory (slope of 1.2, negligible intercept and R ² = 0.92) only at 645 nm, a band with a native spatial resolution of 250 nm and wavelengths sensitive to moderately to highly turbid waters (Table 12). Therefore, despite the potential of specific MODIS wavebands, accurately retrieving Rhow values in confined and turbid estuarine waters remains challenging using the ocean color l2gen function.

As part of the HYPERNETS project, a new network of autonomous field optical stations has been developed and is already in operation around the world for the validation of multi-sensor satellite products. It notably includes coastal and inland water sites such as estuaries, lagoons and reservoirs, i.e., optically complex water bodies where only few measurements are available in this scope (e.g., Pahlevan et al., 2021). These stations provide daily hyperspectral above-water radiance and irradiance measurements, recorded with a precise viewing geometry and used to accurately compute the water-leaving reflectance signal (Goyens et al., 2022). The autonomous system operating the HYPSTAR radiometer was already proved to be robust as successfully operated in wet and windy environments for more than 2 years over a wide range of air temperatures. This system guaranties accurate pointing geometries and provides useful information on environmental conditions (e.g., illumination, rainfall, humidity, air pressure, floating debris) to help operators optimizing their sampling strategy and anticipate technical problems.

The HYPERNETS network will certainly help characterizing the complex optical properties of nearshore coastal waters and confined inland waters (coastal lagoon, estuaries, rivers). These new field datasets will be of great benefit to complement existing ones such as AERONET-OC datasets (Zibordi et al., 2018; 2020), the SeaWiFS Bio-optical Archive and Storage System (SeaBASS), GLORIA (a globally representative hyperspectral in situ dataset for optical sensing of water quality) (Lehmann et al., 2023) or Lake Bio-optical Measurements and Matchup Data for Remote Sensing data (LIMNADES) (Pahlevan et al., 2021). The field datasets in generation will certainly enlarge our knowledge in terms of optical (and biogeochemical) properties of natural waters, extending the ranges of water dynamics and turbidity currently covered for the training of reflectance models used to develop inversion algorithms.

The first aim of HYPERNETS stations operated in water sites is to validate atmospheric, sun glint and adjacency corrections applied to multi-sensor high and medium spatial resolution satellite data. Matchups between field and satellite data in dynamic and often turbid nearshore waters is challenging and matchup protocols previously defined for open ocean then coastal waters must be revised (Concha et al., 2021). In the present study, the analysis of field radiometric measurements and ocean color satellite data has highlighted the need to adapt to each site matchup protocols in contrasted environments such as estuaries and coastal lagoons. Strict quality controls are also required and applied to field radiometric measurements to remove any suspicious or contaminated data (Goyens et al., 2022).

Operating two HYPERNETS stations in France for less than 2 years has already provided numerous matchups with high (L8&9-OLI, S2-MSI) and medium (S3-OLCI, Aqua-MODIS) spatial resolution satellite data. These new matchup datasets will grow and rapidly complement recent attempts to validate the processing of radiometric satellite data in coastal and inland waters and assess the uncertainties associated to satellite-derived optical and biogeochemical properties (e.g., Novoa et al., 2017; Warren et al., 2019; Renosh et al., 2020; Pahlevan et al., 2021).

The first results obtained using the two HYPERNETS stations operated in contrasted French coastal waters confirm that the water-leaving reflectance retrieved using most atmospheric correction processors is associated to significant (>20%) to critical (>60%) errors in nearshore waters (here a coastal lagoon and the mouth of an estuary), as already pointed out by previous studies (e.g., De Keukelaere et al., 2018; Ilori et al., 2019; Warren et al., 2019). These errors typically increase from green and/or red wavebands where the Rhow signal is usually maximum towards the blue and NIR spectral regions; several atmospheric correction algorithms, in their current version, simply generate unreliable results (Warren et al., 2019). There is no processor outperforming others, i.e., providing satisfactory results over the wide diversity of coastal and nearshore waters (Pereira-Sandoval et al., 2019), so that a prior optical water type classification is recommended for selecting the most appropriate atmospheric correction processor (e.g., Soomets et al., 2020; Pahlevan et al., 2021). In moderately turbid coastal lagoon waters, our results show that C2RCC and Polymer provide satisfactory results (errors respectively lower than 30% in visible wavebands and lower than 40% from 443 to 800 nm), as already reported by Pereira-Sandoval et al. (2019) for inland waters where glint and adjacency corrections of satellite data may be as important as atmospheric corrections. Now our study actually shows that the current versions of these two processors reach their limits in turbid to highly turbid waters (with errors higher than 40% at all wavebands). The failure most probably comes from the water reflectance models used in these processors, i.e., models which do not yet account for the large variations of Rhow in the NIR part of the spectrum (Knaeps et al., 2018). The HYPERNETS network will provide new validation datasets to complement the lack of matchups between field and satellite data in highly turbid waters (Knaeps et al., 21018; Goyens et al., 2022), but also contribute to tune then adopt adapted water reflectance models (e.g., Lee et al., 2016; Luo et al., 2018) in atmospheric correction processors. In such waters, our results clearly highlight the very satisfactory performances of ACOLITE (with errors lower than 20% over the visible to NIR parts of the spectrum, when the glint correction is applied) but also Sen2Cor (errors lower than 30% in visible wavebands), i.e., processors which do not make (potentially erroneous) assumptions on the water-leaving reflectance signal to be retrieved from the signal recorded at the top of the atmosphere. Similar results were obtained by Vanhellemont and Ruddick (2021) in moderately turbid coastal waters.

Our results therefore complement previous studies quantifying atmospheric correction errors and understanding the implication on downstream data, such as the concentrations of algal and non-algal particles, water turbidity (and transparency) and light absorption by colored dissolved organic matter (CDOM) (e.g., Ansper and Alikas, 2019; Pahlevan et al., 2021). In turbid estuarine waters, several processors certainly provide satisfactory results in order to accurately estimate the water turbidity and SPM concentration, but also Chla concentrations using a NIR to red edge algorithm (Gernez et al., 2017). In moderately turbid lagoon waters, a processor such as C2RCC (neural networks) is expected to provide satisfactory results to estimate SPM concentrations (using Rhow in the green or red waveband) but also Chla concentrations using a blue-to-green ratio (Erena et al., 2019; Jiménez-Quiroz et al., 2021) or preferably using a NIR-red ratio algorithm where waters are more turbid and/or more productive (Tavares et al., 2021; Zhan et al., 2022). The retrieval of CDOM optical properties is certainly more challenging, as all atmospheric correction processors are associated to significant errors at short visible wavelengths (<450 nm).