Automated abovewater hyperspectral radiometer systems currently or recently used for measurement of water reflectance include: 3-sensor TRIOS/RAMSES with fixed azimuth (Arabi et al., 2018); rotating azimuth TRIOS/RAMSES So-Rad (Simis and Olsson, 2013); rotating azimuth DALEC (Slivkoff, 2014; Brando et al., 2016); 3-sensor Seabird/HyperSAS with rotating azimuth (Carswell et al., 2017); the WispStation with six optical paths at two fixed azimuth directions (Peters et al., 2018); and the OSPREY system (Hooker et al., 2012), which includes both zenith and azimuth pointing and both multispectral and hyperspectral detectors. For the WATERHYPERNET a crucial design choice was to use a pan-tilt unit, allowing both azimuth- and zenith-pointing, in contrast to all prior hyperspectral systems except the OSPREY.

WATERHYPERNET currently accepts two abovewater hardware systems: a) the PANTHYR system based on the mature TRIOS/RAMSES radiometer, and b) the HYPSTAR^® system based on a newly-designed radiometer.

The PANTHYR (PAN and Tilt HYperspectral Radiometer) system, shown in Figure 2A, is described in detail by (Vansteenwegen et al., 2019), and consists of two TRIOS/RAMSES radiometers (one irradiance, one radiance with 7° Field of View; 400–900 nm at 10 nm Full Width Half Maximum, FWHM) with external camera mounted on a FLIR PTU-D48 E pan-tilt unit controlled by a single-board Beaglebone Black Industrial computer and associated custom-built electronics.

The HYPSTAR^® (HYperspectral Pointable System for Terrestrial and Aquatic Radiometry) system, shown in Figure 2B and described in more detail in (Kuusk et al., 2024), consists of a newly-designed hyperspectral radiometer (380–1020 nm at 3 nm FWHM) with integrated radiance and irradiance fore-optics and embedded RGB camera, mounted on a Will-Burt Bowler Rx pan-tilt unit, and controlled by a rugged Cincoze DE-1000 PC. The system has an integrated LED light source for relative calibration monitoring during long deployments–see Figure 2B. This light source is outside the HYPSTAR^® system optical path and so monitors not just changes in spectrometer responsivity, but also any contamination of the fore-optics, e.g., from dust or animals (spider webs, bird faeces, etc.).

Both systems include auxiliary sensors for ambient light and rain detection and the HYPSTAR^® has an ambient light sensor measuring continuously during radiometry. Power supplies (grid, solar + battery) and data transmission (cabled internet, wifi, 2G/3G/4G) are site dependent.

The systems are programmed to acquire data typically every 20 min during daylight, although site-specific adjustments between 15 and 30 min repeat period and/or limiting to a few hours around local solar noon are possible if justified by scientific needs or power limitations. Both systems follow an abovewater radiometry acquisition protocol based on (Mobley, 1999), termed hereafter M1999. Measurements are acquired at 90° and/or 135° relative azimuth to Sun and potentially both left and right of Sun, when permitted by the local mounting structure and its shadows/reflections. Data are transmitted to land in near real time for automated, centralised processing and quality control. Extension of the processing to generate uncertainty estimates for data value is in progress following the work of the FRM4SOC project (Banks et al., 2020; Ruddick et al., 2019). A data portal is under development to distribute data publicly to users such as satellite mission validation entities and developers of atmospheric correction algorithms. Pending implementation of these developments in an operational processing environment, some prototype datasets have been distributed via ZENODO, and are listed at https://waterhypernet.org/data/.

An important feature of both these systems, also present on the precursor multispectral CIMEL CE318-T-OC (Seaprism) and the hyperspectral OSPREY system but not on other hyperspectral systems, is the use of a pointing system with both pan and tilt possibilities. While many other hyperspectral systems recognise the importance of panning to achieve the desired relative azimuth to Sun, tilting has three important advantages compared to typical fixed zenith/nadir angle systems: 1. Both sky and water radiance measurements can be made with the same radiometer, thus saving on acquisition cost and ensuring identical wavelength scale and radiometer sensitivity and characterisation for both sky and water radiance; 2. When not measuring, the radiometers can be “parked” pointed downwards to reduce fore-optics contamination from atmospheric deposition; 3. It is possible to adopt new pointing scenarios with different zenith/nadir angles from the standard M1999 protocol, e.g., scanning the principal plane into sunglint to better estimate the effective Fresnel reflectance coefficient (Goyens and Ruddick, 2023) or scanning the skydome to check for clouds and/or obstructions or to estimate aerosol properties or Sun/moon pointing for calibration monitoring.

Following the successful AERONET-OC approach, and in the interests of reducing costs and achieving a high degree of standardisation, the WATERHYPERNET data acquisition, transmission, processing and distribution is fully automated in Near Real-Time (NRT, <24 h between data acquisition and data availability), at least for the data using pre-deployment radiometer calibration and default wind speed. Data distribution is foreseen from the www.waterhypernet.org data portal, but, during the current prototype phase, where the automated quality control does not yet meet the desired long-term standard, public release of the NRT data is not yet implemented. Despite this prototype status, while the data currently acquired and processed lack some of the features that will be implemented (full measurement uncertainty estimation, better characterisation of optical perturbations from the platforms, links to spectral convolution and BRDF correction tools), these data are already considered to be very valuable for assessing the quality of satellite data for wavelengths or geographical regions not covered by AERONET-OC.

The WATERHYPERNET v0 data are, for example, probably superior in quality to many shipborne reflectance measurements, because they are less subject to tilt, and certainly provide many more hyperspectral matchups per year than shipborne validation data sources. The WATERHYPERNET data should ideally have arrived at maturity at least 7 years ago, e.g., to support the validation of Sentinel-2 and Landsat 8, and are critically needed now for the validation of hyperspectral missions such as PRISMA, EnMAP, EMIT and PACE. Faced with this dilemma of having v0 prototype data that are considered useful but not of the final quality that is expected for routine operations in 2 years time, the current approach is to publicly release limited datasets via www.zenodo.org with appropriate disclaimers on quality and without the “WATERHYPERNET” branding.

In the present paper examples are given of prototype datasets and their application, for both PANTHYR and HYPSTAR^® systems.

For both the PANTHYR and the HYPSTAR^® system the data acquisition protocol is based closely on the M1999 measurement method. For each azimuth angle, ∆ ϕ , measurements are made of downwelling irradiance, E d ∆ ϕ , λ , upwelling water radiance, L u ∆ ϕ , θ v , λ and downwelling sky radiance, L d ∆ ϕ , 180 ° − θ v , λ for nadir-viewing angle θ v = 40 ° and wavelength λ . The sequence of measurements adopted here is: 3* E d , 3* L d , X* L u , 3* L d , 3* E d , where X = 6 for HYPSTAR^®, X = 11 for PANTHYR. With the single radiance sensor design concept these measurements cannot be made simultaneously, in contrast to typical shipborne supervised measurements (Ruddick et al., 2006). This non-simultaneous acquisition has the slight disadvantage that measurements will be contaminated if there is significant time variation of illumination conditions, beyond the correctible Sun zenith angle variation, during the full sequence, which lasts typically <2 min for PANTHYR and <7 min for HYPSTAR^®. Such contamination occurs primarily if there are scattered clouds near the Sun disk, and can generally be detected when the first and last 3* E d (and, for HYPSTAR^® also first and last 3* L d ) are different by more than a pre-defined threshold and/or when the E d is much lower than a clear sky model E d as described in the next two subsections.

During the above-mentioned sequence a simple ambient light sensor measures continuously and can be used for identifying any variation in illumination during the sequential radiometric measurements.

The primary radiometric measurand produced by WATERHYPERNET is the (directional) water-leaving radiance reflectance as represented by ρ w hereafter, and defined by: ρ w ∆ ϕ , θ v , λ = π L w ∆ ϕ , θ v , λ E d λ (1) where L w is the water-leaving radiance (with air-water interface reflection removed), and E d is the planar downwelling irradiance just above the water surface.

L w is estimated by correction of L u for light reflected at the air-water interface, assumed to be a multiple of L d , using the approach of (Mobley, 1999) where the wind-roughened air-water interface is modelled via an “effective Fresnel coefficient”, ρ F : L w = L u − ρ F L d   (2) ρ F is modelled using the look-up table of (Mobley, 1999) with inputs for the solar zenith angle, viewing zenith angle, relative azimuth angle between Sun and sensor, and the wind speed. See section 4 of (Ruddick et al., 2019) and references therein for a detailed discussion of this approach.

The wind speed is retrieved from the 0.25° gridded, 6-hourly nowcast wind speed provided by the National Centers for Environmental Prediction (NCEP) GDAS (https://rda.ucar.edu/datasets/ds083.3/citation/).

For clear and moderately turbid, but not extremely turbid, water sites, a wavelength-independent “NIR Similarity Spectrum (SimSpec)” correction is applied, following (Ruddick et al., 2005), using the expected constant ρ w ratio in the NIR (i.e., 780 and 870 nm) to correct for residual Sun glint.

As a follow-up to the study of (Vanhellemont, 2020), Sentinel-2 (A&B) (S2) data are compared to PANTHYR data for the Acqua Alta and Oostende sites.

The S2 satellites have onboard a 13 band MultiSpectral Instrument (MSI) spanning the VSWIR with four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution. MSI has two bands that are not processed to surface reflectance, one at 945 nm for estimation of water vapour, and one at 1.3 µm for the detection of cirrus clouds.

S2 satellite data for both sites were collected as top-of-atmosphere reflectance ( ρ t , “L1C”, orthorectified and tiled) and Sen2Cor (Main-Knorn et al., 2017) surface reflectance (“L2SR”) for a 3 × 3 km region of interest (ROI), as defined by a bounding box in latitude and longitude, from the Google Earth Engine (GEE) archive (Gorelick et al., 2017) on download dates between 2023-08-10 and 2023-08-13, including acquisitions from 2019-10-01 for Acqua Alta and 2022-02-27 for Oostende and up to 2023-08-11 for both sites. The GEE download uses the latest available processing baseline (from N0208 in late 2019 through N0509 early 2023, see Supplementary Data Sheet S1 for processor versions).

Imagery was processed to water-leaving radiance reflectance ( ρ w ) using ACOLITE_DSF using an ACOLITE GitHub clone dated 2023-08-10, with “3bfe8d8” commit (https://github.com/acolite/acolite). Inside the ROI the aerosol optical thickness (AOT) estimation is assumed uniform in this implementation (pixel-by-pixel processing is also possible in ACOLITE but not used here). Output resolution was 10 m, replicating pixels from the 20 and 60 m bands to fill the 10 m grid. S2 imagery was processed using the following two processor options:

1. ACOLITE/DSF with AOT estimated from VNIR bands (no SWIR) (termed “ACOLITE_DSF”)

Other processor options (including ancillary datasets for pressure, ozone and water vapour) are used according to the defaults documented in (Vanhellemont, 2020).

Matchups with PANTHYR (using only 225° relative azimuth) were made using a 60 min full width window (i.e., +/-30 min around overpass time), linearly interpolated to the overpass time if possible, i.e., when two bounding measurements are available. PANTHYR data are spectrally convoluted to the Sentinel-2 (A&B) bands in reflectance space with spectral response functions S2-SRF_COPE-GSEG-EOPG-TN-15-0007_3.0 (dated 2018-01).

Satellite data are mean averaged over 11 × 11 10 m pixels centred on a reference location. The reference location used for extraction is located as specified in (Vanhellemont, 2020) for Oostende, i.e., 90 m East of the platform, and for Acqua Alta by (Vanhellemont, 2019a, Supplementary Material S2) avoiding platform and near platform pixels. Quality control was made using a threshold on the 95th percentile (P95) of the 11 × 11 pixel box, rejecting matchups where in this box P95 ρ t 1610  nm ≥ 0.05 or ρ t 1375  nm ≥ 0.005 , respectively filtering non-water and cirrus near the matchup location. The 3 × 3 km subscene data were also used, filtering out subscene P95 ρ t 443  nm ≥ 0.3 , thus removing scenes with nearby clouds or other very bright objects.

In situ and satellite matchup data are provided in Supplementary Data Sheet S2.

Reduced Major Axis (RMA) regression lines and squared correlation coefficients, R 2 are provided for the comparison between in situ and satellite ρ w λ measurements, denoted x i and y i respectively, where i = 1 … n . Error statistics were computed for the Root Mean Squared Difference (RMSD), the Mean Difference (MD), and Mean Absolute Percentage Difference (MAPD) between the in situ and satellite measurements as follows: R M S D = 1 n ∑ i = 1 n y i − x i 2 (3) M D = 1 n ∑ i = 1 n y i − x i (4) M A P D = 100 % n ∑ i = 1 n y i − x i 0.5 ∗ y i + x i (5)

Sentinel-3_OLCI/A&B data between February 2021 and March 2023 were downloaded from the EUMETSAT Data Store as Level 2 Water Full Resolution (WFR) products processed using the OLCI L2 processor IPF-OL-2 version 07 (EUMETSAT, 2021; Zibordi et al., 2022) for the HYPSTAR^® deployments at six sites: Acqua Alta Oceanographic Tower (VEIT), Lake Garda (GAIT), Etang de Berre (BEFR), Gironde (MAFR), La Plata (LPAR) and Zeebrugge (M1BE).

An adaptation of the Matchup Data Base (MDB) approach proposed by EUMETSAT (EUMETSAT, 2022) was used to organise the satellite and in situ data, and perform the validation analysis (Gonzalez Vilas et al., 2024). Validation protocols were based on the recommendation available in the literature for medium-resolution satellites (Concha et al., 2021), applying the same protocols for all the sites except for the adaptations noted below (1 valid pixel for GAIT and MAFR, negative satellite reflectance masking for BEFR, MAFR and LPAR).

The 15 OLCI bands between 400 and 865 nm are included in the analysis. The 1020 nm band is excluded because the low signal causes unreliable measurements.

Satellite measurements for each band were computed as the average, excluding outliers, for a measurement window of three by three pixels around the site location, using a strict criterium of nine valid pixels for all the sites except for GAIT and MAFR, where only one valid pixel was required because of the proximity of the coastline. Outliers are identified (EUMETSAT, 2022) when the pixel value is lower (or greater) than mean minus (or plus) 1.5 standard deviations, where mean and standard deviation are computed using valid (non-flagged) pixels in the 3 × 3 extraction window. For masking, the default flag list proposed by EUMETSAT (EUMETSAT, 2021) is used and is based on the Water Quality and Science Flags (WQSF) dataset for all the sites. As WQSF RNEG flags allow for low negative values up to a defined threshold, pixels with negative satellite reflectance between 400 nm and 442 nm are also masked for BEFR, MAFR and LPAR.

In situ measurements for each satellite band were extracted as the HYPSTAR^® L2 reflectances convoluted using the mean spectral response function for each Sentinel-3 mission (https://sentinels.copernicus.eu/web/sentinel/technical-guides/sentinel-3-olci/olci-instrument/spectral-characterisation-data, accessed on 8 April 2024) without applying the NIR similarity spectrum correction for the validation of sites with highly turbid waters (i.e., MAFR, LPAR and M1BE).

In situ and satellite matchup data are provided in Supplementary Data Sheet S8.

Reduced Major Axis (RMA) regression lines and squared correlation coefficients, R 2 are provided for the comparison between in situ and satellite ρ w λ measurements. Error statistics were computed for the Root Mean Squared Difference (RMSD), the Mean Difference (MD), and Mean Absolute Percentage Difference (MAPD) by Eqs 3–5.

Reflectance data from the autonomous systems were used to derive phytoplankton parameters in Belgian North Sea waters over the spring-summer season in 2020 (PANTHYR at Oostende, O1BE, Supplementary Data Sheet S9) and 2023 (HYPSTAR^® at Zeebrugge_MOW1, M1BE, Supplementary Data Sheet S10). In this region, Phaeocystis globosa is considered as a non-toxic but undesirable phytoplankton species because of the unsightly and sometimes dangerous (Philippart et al., 2020) generation of foam, and because the gelatinous mucus has potential impact on species composition at higher trophic levels (Rousseau et al., 2000).

Chlorophyll a concentration (Chl-a hereafter) was calculated using the CRAT method (Ruddick et al., 2001) from reflectance spectra (without SimSpec correction) that have passed quality control. This algorithm, designed for hyperspectral data, is based on the red-NIR reflectance spectrum but contrary to typical semi-analytical red-edge algorithms it avoids calculation of NIR backscattering for moderate-high Chl-a (>13.45 μg/L). This algorithm is less suitable for low Chl-a, but is relevant for the bloom events considered here.

Two indices for P. globosa were calculated from reflectance spectra based on existing algorithms which are described in detail in the cited papers, and briefly summarised here.

The Lubac Index (LI), defined in (Lubac et al., 2008), is a binary algorithm (yes/no) indicating if the phytoplankton assemblage is dominated by P. globosa, and is based on the shape of the second derivative reflectance spectrum between 420 and 560 nm. The second derivative was calculated following the formulation described by (Lubac et al., 2008; Lavigne et al., 2022) for intervals of 2.5 nm for the PANTHYR data and of 1 nm for the HYPSTAR^® data. Before making the second derivative calculation, ρ _w spectra were smoothed to avoid strong outliers in the second derivative. A five-point window (12.5 nm) running average was applied to all PANTHYR spectra, and a nine-point window (8 nm) running average was applied 3 times to all HYPSTAR^® data.

The Modified Astoreca Line Height index (MALH), defined in (Lavigne et al., 2022), and based on prior work by (Astoreca et al., 2009) is a line height difference algorithm (Eq. 6) measuring the absorption anomaly at λ 2 = 482.5  nm with respect to a non-linear baseline between λ 1 = 470  nm and λ 3 = 490  nm : M A L H = 1 ρ w λ 2 − 1 ρ w λ 1 1 − w × 1 ρ w λ 3 w × a w   N I R × ρ w N I R (6) where w = λ 2 − λ 1 λ 3 − λ 1 . a w   N I R and ρ w   N I R are respectively the pure water absorption and the water-leaving radiance reflectance at a near infrared band, here chosen as λ N I R = 700  nm , giving a w   N I R = 0.57   m − 1 according to (Kou et al., 1993).

Finally, because MALH depends on the concentration of P. globosa, the ratio MALH/Chl-a is calculated as a proxy for the P. globosa fractional contribution to phytoplankton absorption. This ratio is calculated from the slope of linear regressions between MALH and Chl-a data available 48 h before and after the measurement date. Hence, a positive value suggests a high fraction of P. globosa in the phytoplankton community, and a negative value a low fraction of P. globosa. When the regression slope was not significant (p-value ≤5%) results are shown in grey.

The scatterplots comparing Sentinel-2 (A&B) water-leaving radiance reflectance with matchup PANTHYR in situ measurements from the Acqua Alta (VEIT) and Oostende (O1BE) validation sites are shown in Figure 5 for selected spectral bands. Scatterplots for all spectral bands can be found in Supplementary Data Sheet S3. Results for Sentinel-2A and Sentinel-2B were very similar, and are presented together in this scatterplot. The results shown in Figure 5 are quite different for the two validation sites, suggesting different algorithm performance issues for the different turbidity ranges (Acqua Alta moderately turbid, Oostende highly turbid).

For the Sen2Cor processor, the vertical dispersion of points in these scatterplots, and strong positive bias (MD) at all wavelengths with a striking overestimation of ρ _w at 865 nm suggests a general underestimation of aerosol reflectance and/or uncorrected sunglint–this is not surprising for an algorithm which is designed for atmospheric correction over land, taking aerosol optical thickness from dense dark vegetation (if present in scene) or from the fall-back external meteorological data (CAMS) and with no sunglint, or even skyglint, correction.

For the ACOLITE_DSF processor, without sunglint correction, results are improved somewhat compared to Sen2Cor with reduction of the positive bias (MD) at all bands, suggesting better estimation of aerosol reflectance, but with many positively biased outliers, most obvious at 865 nm, where the expected and in situ measured ρ _w is systematically low, especially at Acqua Alta.

For the ACOLITE_DSF_GC processor, with a sunglint correction, many of those positively biased outliers are now well-corrected at 865 nm, but little difference is found at 443 nm.

The spectral RMSD between Sentinel-2 (A&B) and PANTHYR measurements at Acqua Alta and Oostende for these matchups is shown in Figure 6. The general decrease of RMSD with increasing wavelength for the Sen2Cor and ACOLITE_DSF_GC processors is quite different from the “water-like” RMSD spectrum found in a validation study in the La Plata estuary - see Figure 4A of (Dogliotti et al., 2023). The RMSD spectrum of Figure 6 is typical of situations where the dominant error source is imperfect correction of atmospheric path reflectance (aerosols and/or Rayleigh)–see Figure 9 of (Vanhellemont and Ruddick, 2021) and associated discussion in that paper for similar experience with atmospheric correction of Sentinel-3/OLCI. This error is greater for Sen2Cor as discussed above. Interestingly, comparison of ACOLITE_DSF with ACOLITE_DSF_GC in Figure 6 shows that the sunglint correction has successfully reduced RMSD for 665–865 nm but not for 443–560 nm. This suggests that the remaining dominant error for ACOLITE_DSF_GC is related to atmospheric path reflectance but not sunglint. From the ACOLITE_DSF_GC 443 nm scatterplot in Figure 5 it seems that the atmospheric path reflectance error is positively biased for the Acqua Alta site but negatively biased for the Oostende site. The leading hypotheses for these biases are:

• ACOLITE_DSF_GC gives positively biased ρ _w 443 − 560  nm at the Acqua Alta site because of spatial variability of the atmosphere and/or air-water interface over the ROI. The fundamental DSF assumption is that the atmospheric composition (esp aerosols) and air-water interface are spatially invariant over the ROI, and can be estimated from the darkest pixel, where the water (or land surface) reflectance is negligible at least for a single wavelength. If the darkest pixel is “too dark”, e.g., because of sensor noise, a darker than average wave facet, cloud shadowing of the atmosphere, etc., then the AOT will be underestimated, and ρ _w overestimated. If the validation pixel has a brighter atmosphere than the darkest pixel, e.g., because of patchy haze or undetected thin clouds, then a similar overestimation of ρ _w will occur. This is an inherent positive bias to the DSF assumption of spatial uniformity of atmosphere and air-water interface over the ROI. This bias could perhaps be reduced by reducing the ROI (while preserving the need for a sufficiently large ROI to find an appropriate dark pixel) or otherwise detecting situations where the assumption of spatial uniformity is violated.

Full testing of these hypotheses is beyond the scope of the present paper, which serves to demonstrate the usefulness of WATERHYPERNET in situ measurements in general and specifically to provide clues to how Sentinel-2 data quality might be improved when using ACOLITE_DSF.

In addition to the abovementioned biases, the scatterplots of Figure 5 suggest an important number of outlier cases. Validation outliers are not always studied in detail, particularly when there are a large number of matchups. However, these outliers contain very important information on the real quality of satellite data that reach end-users as well as vital clues on how to improve processing. While some of the outliers for Sen2Cor and ACOLITE_DSF can be attributed to the lack of a sunglint correction, many outliers remain in the ACOLITE_DSF_GC processing.

We define here “outlier” as a matchup where the difference between satellite and in situ exceeds the RMSD difference over all matchups for that site, i.e., Δ ρ w 443  nm > ρ w R M S D − s i t e 443  nm , where Δρ _w (443 nm) is the difference between in situ and satellite measurements of ρ _w (443 nm), ρ w R M S D − V E I T 443  nm = 0.01894 and ρ w R M S D − O 1 B E 443  nm = 0.01380 . All outliers (40/124 for VEIT and 11/31 for O1BE) were analysed subjectively by two experts on the basis of “Validation Diagnostic sheets” (Supplementary Data Sheet S4–S7) showing the spectral plot for satellite ρ w (all three processors) compared to the in situ measurement, both hyperspectral and Sentinel-2 bands together with imagettes of the ROI for:

• Surface reflectance, ρ s , RGB composite (665 nm, 560 nm, 443 nm)

All outliers were positively biased at ρ w 443   n m for VEIT and negatively biased at ρ w 443  nm for O1BE.

Figure 7 shows an example Validation Diagnostic sheet for the matchup with Sentinel-2/A acquisition over Acqua Alta (VEIT) on 2022-11-21T10:03:44. Inspection of the ρ s 865  nm and ρ t 2202  nm imagettes suggests that there are thin (undetected) patchy clouds in the pixels near the reference location, marked as a red cross in Figure 7. Furthermore cloud shadow at the dark pixel may add further low bias to the AOT estimate.

Figure 8 shows an example Validation Diagnostic sheet for the matchup with Sentinel-2/A acquisition over Oostende (O1BE) on 2022-06-12T10:59:33. Inspection of the ρ s R G B and ρ t 2202  nm imagettes suggests that there is a cloud/contrail shadow over the reference location. Inspection of the full image (not shown) showed indeed a suitably shaped and positioned contrail just South of this ROI.

On the basis of this subjective expert analysis, the hypotheses for all outliers are:

• For VEIT, 37/40 outlier cases it is thought that the AOT used at the validation pixel is too low, although the underlying reason can be diverse, and is sometimes unclear. The dark pixel may be “too dark”, e.g., because of sensor noise, a darker than average wave facet, cloud shadowing of the atmosphere, etc., or the validation pixel may have a brighter atmosphere/interface than the dark pixel, e.g., because of patchy haze or undetected thin clouds. In 3/40 outlier cases, the cause of the outlier is difficult to discern. In 38/40 outlier cases (and most non-outlier cases) the DSF approach uses the 865 nm band (8A) for the dark pixel.

Figure 9 shows the scatterplot of matchups between Sentinel-3_OLCI/A&B Level 2 Water Full Resolution (WFR) and the HYPSTAR^® deployments at six sites (VEIT, GAIT, BEFR, MAFR, LPAR and M1BE, see Section 4.2) for each individual band and grouped by satellite sensor.

Overall, better fits were achieved between 490 nm and 885 nm ( R 2 > 0.88 ), whereas a higher dispersion and worse correlation were observed in the blue part of the spectrum (400 nm, 412.5 nm, 442.5 nm). A negative bias (MD) is observed in all the bands. While the regression slope is close to 1 (between 0.9 and 1.0) for bands 709–885 nm, the lower regression slope (between 0.79 and 0.89) for bands 442–681 nm suggests a systematic difference between satellite and in situ measurements that warrants further attention.

The spectral shape of the bias (MD) and RMSD (Figures 10B, C) for 442–885 nm is similar to a turbid water reflectance spectrum. While some atmospheric correction algorithms have uncertainty which can be proportional to water reflectance, this is not expected for the WFR algorithm, suggesting that this difference may be dominated in the range 442–885 nm either by a systematic in situ measurement error or by space (or, less likely, time or angular) differences between the satellite and the in situ measurement. Both satellite sensors show similar performance.

The global scatter plot of matchups of Sentinel-3_OLCI/A&B Level 2 and HYPSTAR^® including all the bands is shown in Figure 10A. There is an overlap of data points from both sensors, with a good fit and high correlation ( R 2 = 0.93 ).

The number of valid matchups (whole spectra) from both sensors was very similar (S3A: 295, S3B: 300), as well as the global metrics including all the bands (Table 3). Spectral variation metrics (Figures 10B–D) follow similar patterns for both S3A and S3B. In fact, RMSD and bias (MD) differences between both satellites are always lower than 0.0006 and 0.0002, respectively. The single satellite metric values are also close to those obtained combining both sensors (see R ² and MD in Table 3).

This analysis is reassuring as to the interoperability of the A and B units of the Sentinel-3/OLCI constellation, achieved by good harmonisation of space hardware and ground processing elements, and for the satellite ρ _w in the NIR bands 709–885 nm. For the spectral range 442–681 nm a better understanding is needed of in situ measurement uncertainty before conclusions can be drawn on the atmospheric correction algorithm performance.

Although designed for the launching application of satellite validation, the WATERHYPERNET data can be used without satellite data for single point monitoring of water quality parameters, including phytoplankton.

Figure 11 shows a time series of phytoplankton parameters derived from the PANTHYR data at Oostende (O1BE) in 2020, extended from the previous analysis of Figure 5 of (Lavigne et al., 2022) to include MALH per unit chlorophyll a as a biomass-independent indicator of species fraction of P. globosa. The striking difference in MALH between the end-April/beginning-May bloom (positive ⇒ dominated by P. globosa), and the end-June bloom (negative ⇒ not dominated by P. globosa) is matched by the LI flag, and can be traced back to subtle differences in curvature (second derivative) of the water-leaving radiance reflectance - see (Lavigne et al., 2022) for full details.

Figure 12 shows a similar time series of phytoplankton parameters, but now derived from the HYPSTAR^® data at Zeebrugge (M1BE) in 2023. In this dataset, the phytoplankton blooms are less strong, and the MALH shows fewer positive values, although MALH, LI and MALH/Chl-a all indicate a phytoplankton bloom dominated by P. globosa in early/mid-Apr.

While the MALH algorithm was originally designed (Astoreca et al., 2009) and refined (Lavigne et al., 2022) for application to hyperspectral satellite data, and the WATERHYPERNET network was designed for validation of satellite data, these time series of in situ measurements both reinforce the expected potential of future hyperspectral satellite data, and raise the possibility of using standalone WATERHYPERNET data for single-point time series of water quality parameters. A similar study showed the potential for using HYPSTAR^® data for detecting the presence of cyanobacteria in a drinking water reservoir (Goyens et al., 2022).

The results of Section 5.1 comparing PANTHYR data from Oostende and Acqua Alta with Sentinel-2 data show that:

• The ACOLITE_DSF atmospheric correction algorithm performs better than the Sen2Cor atmospheric correction (designed for land, but still the only standard Sentinel-2 product for coastal waters) for all spectral bands, for both clear and turbid waters (Figure 5; Figure 6).

The results of Section 5.2 comparing HYPSTAR® data from six validation sites with Sentinel-3 data show that:

• Reasonable results are achieved at all spectral bands (RMSD<0.014) with very similar performance for the A and B units of the Sentinel-3/OLCI, confirming that good interoperability of the constellation has been achieved by harmonisation of space hardware and ground processing elements.

The results of Section 5.3 using the PANTHYR and HYPSTAR® data in Belgian waters (without satellite data), as a follow-up of work by (Lavigne et al., 2022), show that:

• Phytoplankton biomass and some information on dominant species (here Phaeocystis globosa) can be monitored at high frequency from these hyperspectral data.

This finding is of interest both in its own right for pointwise monitoring of water quality, and as a precursor for information that might be retrieved from the new generation of hyperspectral satellite missions (Dierssen et al., 2020), provided that the treatment of sub-resolution scale spectral features such as absorbing atmospheric gases renders second derivative spectra of adequate quality (Ruddick et al., 2023).

This paper has described the WATERHYPERNET, a federated network of automated in situ measurements of hyperspectral water reflectance designed for satellite validation. The medium-term ambition is to provide a sufficient quantity of high quality water reflectance data over sufficiently diverse water, atmosphere and Sun conditions to satisfy the needs for radiometric validation of all VIS/NIR (380–900 nm) spectral bands of all current and future optical satellite missions used for aquatic applications. The list of satellite missions expected to use WATERHYPERNET data is long, and includes: dedicated “water colour” missions such as Sentinel-3A&B (&C&D), MODIS, VIIRS; “land” missions repurposed for coastal and inland water applications such as Sentinel-2A&B (&C&D) and Landsat 8&9; recent and future hyperspectral missions such as PRISMA, ENMAP PACE … CHIME, SBG and the geostationary GLIMR; and the emerging “Newspace” cubesat constellations pioneered by the PlanetScope Doves and Superdoves.

The network is currently at the stage of a proof-of-concept prototype, with two functioning hardware systems, PANTHYR and HYPSTAR^®, automated data acquisition, transmission and processing of data and demonstration datasets. Although significant improvements are expected in the next 2 years with evolution of the in situ data processing, particularly regarding quality control and estimation of measurement uncertainties, the prototype datasets are already considered by satellite validation/performance team users to be relevant for showing some performance issues and by water quality managers for describing phytoplankton dynamics, including some indication of species composition.

The validation sites described here constitute a network with reasonable coverage of water and atmosphere conditions where satellite data need to be validated (i.e., everywhere where end-users use final products), including clear and turbid waters, coastal and inland waters, various phytoplankton species, low and moderate Coloured Dissolved Organic Matter (CDOM) absorption, various Sun zenith, cloud and wind conditions, etc. A very complete network could be achieved with 20 appropriately chosen validation sites, although the choice of site is clearly limited by availability of funding, dedicated local scientist(s) and stable mounting platforms. Moreover, the current sites can be affected by occasional and/or long-term downtime associated from diverse causes (hardware failures, recalibration of radiometers, platform and/or ship crew availability, funding, etc.).

If the network manager could choose where to locate additional validation sites it would be relevant to add: better coverage of the Southern hemisphere (for satellite commissioning phases occurring in the Northern hemisphere winter), one or two sites at very high latitude (e.g., >70° to validate for the difficult high zenith angles and high air masses but with increased acquisitions/week), one or two sites at low latitudes (e.g., <30° to validate for the difficult high sunglint conditions), one or more very nearshore/inland/high altitude sites for testing adjacency effect removal and atmospheric pressure (and hence Rayleigh scattering) differences, one or two sites with very high CDOM and better longitudinal coverage (e.g., over the West and East coasts of North or South America and over Western and Eastern Asia to provide validation data for geostationary satellites). However, it is likely that the siting of additional validation sites will be more opportunistic (related to national funding and individual motivated scientists) than strategic.

The main challenges in establishing and consolidating a network such as WATERHYPERNET are organizational (funding, governance/coordination) and hardware-related. The equipment used here consists of one system (PANTHYR) based on a mature/aging COTS radiometer and COTS pan-tilt with non-commercial assemblage of driving electronics and mechanics and one system (HYPSTAR^®) with a newly-designed prototype radiometer and pre-commercial assembly of pan-tilt, host system PC and mechanics with some custom-designed elements (relay board, junction box/cabling). These systems can function autonomously for many months/years in the best cases, but are far from plug-n-play, requiring considerable expertise for preparation and installation and troubleshooting/repairs. The large number of components and the hostile environmental conditions combine to generate a wide range of low probability but high impact failure modes with difficult logistics (safety training and equipment, availability of transfer boats and crew and seaworthy specialized technical/scientific staff) for maintenance visits to most offshore sites.

Abovewater radiometry, led by the AERONET-OC network (Zibordi et al., 2009; Zibordi et al., 2021), is now established as the main source of in situ data for radiometric validation of water colour missions, and there is a growing expertise in laboratory calibration and characterization of hyperspectral radiometers (Talone et al., 2016; Zibordi et al., 2017; Talone and Zibordi, 2018; Vabson et al., 2019; Kuusk et al., 2024), and how to propagate the related uncertainties (Białek et al., 2020; De Vis et al., 2024). However, there are still two elements of the measurement method where the quantification of measurement uncertainties for each individual measurement result can be improved:

• Removal of light reflected at the air-water interface is currently implemented in WATERHYPERNET using the wind speed formulation for effective Fresnel coefficient of (Mobley, 1999), which is a common approach and corresponds to the recommendations of the IOCCG protocols (Zibordi et al., 2019). However, there are many known problems with this model - see Section 4 and Section 6.2 of (Ruddick et al., 2019) and references therein - and no clear consensus on the bias/uncertainty associated with this correction at moderate and high wind speed (>5 m/s), especially for the shorter wavelengths.

In addition to these aspects of the measurement method, the uncertainties relating to space, time and viewing angle differences between the in situ measurement and the satellite measurement need to be included for a full validation uncertainty analysis, although these are beyond the scope of WATERHYPERNET, which aims to estimate first the uncertainty of the WATERHYPERNET measurement for its own space, time and viewing angle coverage. For comparison between satellite and in situ data:

• Spatial differences between satellite and in situ measurements can be quantified using higher resolution satellite data, potentially multispectral or even very broadband, which could provide a site-dependent estimate of spatial differences as a function of satellite pixel size. Some validation sites should only be used for metre or decametre scale satellite validation.

As regards data processing, archiving and distribution and user support:

• The increase in number of sites and data volume will necessitate efficient storage, increased computing power, especially for the time-consuming uncertainty calculations, and careful automation, including exception handling.

In this paper we describe the prototype of the WATERHYPERNET network, and demonstrate its utility for massively multi-satellite radiometric validation. Future work (next 2 years) will focus on:

• Keeping the existing validation sites running as far as possible, and adding a few new sites within the current hardware and human resource limits (especially radiometer manufacturing, calibration and characterization).