Automated Quality Control of AERONET-OC LWN Data

Giuseppe Zibordi aJoint Research Centre, European Commission, Ispra, Italy

Search for other papers by Giuseppe Zibordi in
Current site
Google Scholar
PubMed
Close
,
Davide D’Alimonte bAEQUORA, Lisbon, Portugal

Search for other papers by Davide D’Alimonte in
Current site
Google Scholar
PubMed
Close
, and
Tamito Kajiyama bAEQUORA, Lisbon, Portugal

Search for other papers by Tamito Kajiyama in
Current site
Google Scholar
PubMed
Close
Open access

Abstract

Quality control (QC) practices are a fundamental requirement for any measurement program targeting the delivery of high-quality data. In agreement with such a need, the Ocean Color component of the Aerosol Robotic Network (AERONET-OC) includes a number of QC steps ensuring the delivery of normalized water-leaving radiance LWN spectra at incremental accuracy levels identified as level 1.0, level 1.5, and level 2.0. Currently, the final QC step allowing for rising level 1.5 LWN spectra to level 2.0 implies the execution of an expert-based procedure, which is extremely time consuming and naturally undergoes subjective decisions on dubious cases. These limitations solicited the development of an automated procedure, so-called A–QCLWN, mimicking the steps supporting an expert analyst during the final QC of AERONET-OC LWN spectra. A–QCLWNapplies hierarchical tests to check (i) the relative consistency of level 1.5 LWN spectra (called candidates) with respect to LWN reference spectra (called prototypes) constructed using LWN spectra formerly and independently quality controlled; (ii) the absence of any pronounced spectral feature in portions of each LWN candidate spectrum expected to exhibit a regular shape; and additionally, when applicable, (iii) the temporal consistency of the LWN candidate spectrum with respect to close-in-time spectra as a criterion to further strengthen the quality of data. A–QCLWN performance has been verified using LWN spectra from AERONET-OC measurement sites representative of various water types embracing oligotrophic/mesotrophic waters dominated by chlorophyll-a concentration and coastal waters exhibiting increasing levels of optical complexity. A–QCLWN has shown an acceptance rate of AERONET-OC level 1.5 LWN candidate spectra varying between approximately 89% and 93% with agreement in the range of 88%–93% with respect to the LWN spectra independently quality controlled through the expert-based procedure. The additional capability of A–QCLWN to rank the fully quality-controlled LWN spectra combining weights depending on the various tests, anticipates the possibility to best support applications with diverse accuracy needs. Finally, acceptance rates of A–QCLWN for LWN prototype spectra built using level 1.5 data, an alternative to fully quality-controlled level 2.0, have shown values generally increased by less than 1%. This indicates the possibility to lessen the constraint implying the existence of reference level 2.0 LWN data for the relative-consistency test at the expense of a fairly low reduction in accuracy.

© 2022 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Giuseppe Zibordi, giuseppe.zibordi@eoscience.eu

Abstract

Quality control (QC) practices are a fundamental requirement for any measurement program targeting the delivery of high-quality data. In agreement with such a need, the Ocean Color component of the Aerosol Robotic Network (AERONET-OC) includes a number of QC steps ensuring the delivery of normalized water-leaving radiance LWN spectra at incremental accuracy levels identified as level 1.0, level 1.5, and level 2.0. Currently, the final QC step allowing for rising level 1.5 LWN spectra to level 2.0 implies the execution of an expert-based procedure, which is extremely time consuming and naturally undergoes subjective decisions on dubious cases. These limitations solicited the development of an automated procedure, so-called A–QCLWN, mimicking the steps supporting an expert analyst during the final QC of AERONET-OC LWN spectra. A–QCLWNapplies hierarchical tests to check (i) the relative consistency of level 1.5 LWN spectra (called candidates) with respect to LWN reference spectra (called prototypes) constructed using LWN spectra formerly and independently quality controlled; (ii) the absence of any pronounced spectral feature in portions of each LWN candidate spectrum expected to exhibit a regular shape; and additionally, when applicable, (iii) the temporal consistency of the LWN candidate spectrum with respect to close-in-time spectra as a criterion to further strengthen the quality of data. A–QCLWN performance has been verified using LWN spectra from AERONET-OC measurement sites representative of various water types embracing oligotrophic/mesotrophic waters dominated by chlorophyll-a concentration and coastal waters exhibiting increasing levels of optical complexity. A–QCLWN has shown an acceptance rate of AERONET-OC level 1.5 LWN candidate spectra varying between approximately 89% and 93% with agreement in the range of 88%–93% with respect to the LWN spectra independently quality controlled through the expert-based procedure. The additional capability of A–QCLWN to rank the fully quality-controlled LWN spectra combining weights depending on the various tests, anticipates the possibility to best support applications with diverse accuracy needs. Finally, acceptance rates of A–QCLWN for LWN prototype spectra built using level 1.5 data, an alternative to fully quality-controlled level 2.0, have shown values generally increased by less than 1%. This indicates the possibility to lessen the constraint implying the existence of reference level 2.0 LWN data for the relative-consistency test at the expense of a fairly low reduction in accuracy.

© 2022 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Giuseppe Zibordi, giuseppe.zibordi@eoscience.eu

1. Introduction

“Good (practically useful) data do not collect themselves. Neither do they magically appear on one’s desk, ready for analysis and lending insight into how to improve processes” (Vardeman and Jobe 2016): this statement well introduces the efforts behind any type of data collection, quality assurance (QA), and quality control (QC) practices. These entail process-oriented QA actions ensuring the correct execution of measurements, complemented by product-oriented QC steps embracing all postgeneration actions supporting the provision of high-quality data (Bushnell et al. 2019, 2020). By focusing on QC, the application of solutions benefitting of automated procedures is definitively invaluable for datasets resulting from a large number of field measurements such as time series from a variety of sites.

The Ocean Color component of the Aerosol Robotic Network (AERONET-OC) provides a clear example of QC effort on the production of in situ normalized water-leaving radiances LWN from globally distributed measurement sites. An intrinsic difficulty characterizing the QC of AERONET-OC data is the diversity of LWN spectra due to the variety of natural water types: a diversity that keeps increasing over time due to the addition of potentially unique measurement sites. Because of this, the QC of LWN spectra exhibiting exclusive features from novel sites would challenge any procedure relying on generic thresholds or LWN reference spectra. This implies the need for a priori knowledge of the spectral features characterizing LWN data from each measurement region and consequently some dedicated effort to create reference data for each measurement site.

QC is often implemented as a process leading to the inclusion or exclusion of data from successive quality levels. Specifically, AERONET-OC data are accessible at three incremental accuracy levels identified as level 1.0, level 1.5, and level 2.0 fully quality checked data. These incremental levels allow for (i) real-time data applications not yet benefitting of full quality control through level 1.5, and (ii) deferred applications requiring higher-quality measurements through level 2.0. However, aside this separation, level 2.0 fully QC data could also benefit of incremental rankings satisfying different applications demanding higher levels of confidence and naturally trading off between expected accuracy and number of quality-controlled data. In the specific case of ocean color, applications requiring fully quality-controlled input data, but with increasing level of confidence, include (i) the generation of empirical bio-optical algorithms linking radiometric data and the concentration of water optically significant constituents, (ii) the assessment of the accuracy of satellite data products within validation exercises, (iii) the quantification of uncertainties affecting satellite data products, and (iv) the indirect calibration of satellite data through system vicarious calibration.

Recalling that an expert-based procedure is currently applied for the final QC step to rise to level 2.0 the qualified level 1.5 LWN spectra, this work presents and discusses a novel procedure allowing for an entirely automated QC of AERONET-OC LWN data directly supporting the potential for an incremental ranking of the level 2.0 data products.

2. AERONET-OC data

AERONET-OC, a subnetwork of the Aerosol Robotic Network (AERONET; Holben et al. 2001; Giles et al. 2019) was designed to support the validation of satellite ocean color radiometric products with globally distributed and highly accurate in situ reference data of normalized water-leaving radiance LWN and aerosol optical depth τa determined with consolidated and standardized instrumentation, measurement, and processing methods. Details on AERONET-OC are provided in Zibordi et al. (2009, 2022). Key elements of AERONET-OC measurements and processing are summarized in the following subsections to ensure self-consistency to this work.

a. Measurements and processing

AERONET-OC, equivalent to AERONET, relies on CE-318 and CE-318T radiometers designed to provide spectral measurements in the ultraviolet, visible, and near-infrared regions of the solar spectrum. Specifically, these radiometers perform measurements with a full-angle field of view of 1.2° at spectral bands relevant for ocean color applications by autonomously acquiring on a band-by-band basis the following quantities (see also Zibordi et al. 2009, 2022):

  1. the direct solar irradiance E(θ0, ϕ0, λ), required to determine the spectral aerosol optical depth τa(λ), function of the center wavelength λ of each spectral band and of the illumination geometry determined by the solar zenith θ0 and azimuth ϕ0 angles;

  2. NT sea-viewing measurements (with NT = 11) to determine the sea radiance LT(θ, φ, λ) including both the water and surface reflectance contributions with θ indicating the sea-viewing angle and φ the relative azimuth between solar and sensor planes; and

  3. Ni sky-viewing measurements (with Ni = 3) to determine the sky radiance Li(θ′, φ, λ) with θ′ = 180° − θ indicating the sky-viewing angle.

For each measurement sequence, Li(θ′, φ, λ) is determined by averaging the Ni sky-viewing radiance data, while LT(θ, φ, λ) is determined from the average of a fixed percent of the NT sea-viewing data exhibiting the lowest radiance levels (2 out of 11).

The water-leaving radiance LW(θ, φ, λ), i.e., the radiance emerging from the sea surface, is quantified from LT(θ, φ, λ) and Li(θ′, φ, λ), according to
LW(θ,φ,λ)=LT(θ,φ,λ)ρ(θ,φ,θ0,W)Li(θ,φ,λ),
where ρ is the sea surface reflectance function of the measurement and illumination geometries, and of the sea state expressed through the wind speed W (Mobley 1999). In agreement with current measurement protocols the LT(θ, φ, λ) and Li(θ′, φ, λ) values are determined at θ = 40° and φ = 90° (IOCCG 2019). The use of the percent of the NT sea-viewing measurements exhibiting the lowest radiance levels is effective in partially compensating a systematic underestimate of the ρ values applied (D’Alimonte et al. 2021).
The normalized water-leaving radiance LWN(λ), the radiance that would be measured with nadir view, no atmosphere, the sun at the zenith and at the mean Earth distance, is computed as
LWN(λ)=LW(θ,φ,λ)Cfr(θ,φ,θ0,λ,τa,IOP,W)[D2td(λ)cosθ0]1,
where D2td(λ)cosθ0 is an estimate of the downward Ed(λ) to mean extra-atmospheric E0(λ) irradiance ratio (Zibordi et al. 2004), with D accounting for the sun–Earth distance as a function of the day of the year and td(λ) the atmospheric diffuse transmittance (Tanré et al. 1979). The correction factor Cfr(θ, φ, θ0, λ, τa, IOP, W) minimizes the impact of bidirectional effects due to the off-nadir view of the radiometer and of the illumination geometry, with τa and IOP indicating the atmospheric aerosol optical depth and the inherent optical properties of water, respectively. It is recalled that the AERONET-OC corrections include (i) those determined assuming that the IOPs are solely defined by chlorophyll-a concentration, which are specific for the so-called Case 1 waters (Morel et al. 2002); and alternatively (ii) those relying on the determination of the IOPs from the water-leaving radiance spectra LW, likely applicable to any water type (Lee et al. 2011; Talone et al. 2018).

b. QC of AERONET-OC data

The product-oriented QC process of AERONET-OC data aims at removing LWN spectra affected by (i) large environmental perturbations due to clouds or heavy sea state, (ii) elements of the deployment structures seen by the sensor field of view, (iii) light field perturbations due to the vicinity of the sensor footprint to the deployment structure, and also (iv) large changes in sensor responsivity during the deployment period. The measurements sequentially performed on a band-by-band basis are certainly a drawback of the CE-318 and CE-318T systems. In fact, they naturally introduce spectral inconsistencies in LT measurements due to random wave perturbations in the different spectral bands. Still, their impact is minimized by identifying and removing those LWN spectra affected by large intraband variability that may lead to spectral inconsistencies.

AERONET-OC data, equivalent to the AERONET atmospheric products (Holben et al. 2001), are accessible at three incremental quality levels. Level 1.0 includes LWN(λ) data benefitting of a basic QC: (i) τa(λ) has been determined; (ii) the NT sea-viewing radiance and Ni sky-viewing radiance measurement sequences do not exhibit missing data; (iii) the dark values are below a minimum threshold; (iv) the value of ϕ0 is within site-dependent limits determined to ensure minimum superstructure perturbations in LT(θ, φ, λ); and (v) W is below the maximum threshold of 15 m s−1.

Level 1.5, mostly proposed for real-time applications, includes LWN(λ) derived from level 1.0 data passing additional automated QC tests: (i) cloud screened AERONET τa(λ) exists at level 1.5 in the AERONET database (Smirnov et al. 2000; Giles et al. 2019), (ii) a series of empirical thresholds are satisfied [e.g., LWN(λ) > −0.01 mW cm−2 μm−1 sr−1 indicating absence of exceedingly negative values at any λ], (iii) LWN(412) < LWN(443) at coastal sites, (iv) LWN(1020) < 0.1 mW cm−2 μm−1 sr−1 in regions not affected by very turbid waters to exclude measurements perturbed by the presence of obstacles in the sight of the sea-viewing sensor; and (v) the NT sea-viewing radiance measurements and Ni sky-viewing radiance measurements exhibit low variance indicating low wave effects and negligible cloud contamination, respectively.

Finally, level 2.0 supporting deferred applications, includes LWN(λ) determined from level 1.5 products for which (i) the level 2 AERONET τa(λ) exists; (ii) the NT sea-viewing radiance measurements and Ni sky-viewing radiance measurements satisfy lower variance thresholds with respect to those applied for level 1.5; (iii) the differences between pre and postdeployment calibration coefficients for the AERONET-OC radiometer exhibit values smaller than 5% (see Zibordi et al. 2022); and finally, (iv) LWN(λ) do not show questionable values during the final spectrum-by-spectrum assessment performed by an experienced analyst. This final step has been often supported by an early automated QC process (D’Alimonte and Zibordi 2006) aiming at rejecting LWN(λ) spectra exhibiting (i) low statistical representativeness within the dataset itself and (ii) anomalous features with respect to a generic dataset of reference LWN(λ) spectra. This automated procedure has however shown limits in determining the quality of spectra not statistically represented in either the reference dataset or the AERONET-OC data going to be screened. Because of this, the final QC step put in place until now to raise LWN level 1.5 data to level 2.0 relies on the final assessment of each LWN(λ) spectrum by an experienced analyst (Zibordi et al. 2022). This expert-based QC procedure (hereafter identified as E–QCLWN) would ideally require knowledge of a number of details such as measurement protocol, information on deployment structures, instrument functioning, in addition to regional atmospheric and marine optical properties. E–QCLWN, aside exposing results to subjective decisions on dubious cases, is extremely time consuming. This solicited the development and implementation of a fully automated procedure replacing E–QCLWN for the final quality control of AERONET-OC LWN level 1.5 data candidate to level 2.0.

3. The fully automated QC process

The fully automated process matter of a recent development and verification, incorporates in a single process the basic QC steps applied to LWN level 1.5 data alongside with a generalization of the QC tests supporting the expert-based E–QCLWN procedure. This automated QC procedure replacing E–QCLWN and hereafter identified as A–QCLWN, implies the existence of archived LWN(λ) data at level 2.0 for each specific site or, alternatively, for sites exhibiting analogous measurement conditions entailing equivalent radiometric products in terms of LWN spectral shape and amplitude. This suggests that A–QCLWN can be reliably applied to data from AERONET-OC sites for which historical fully quality-controlled data already exist or alternatively exist for an equivalent measurement site. As already anticipated, this ideal constraint is advised by the variety of measurement conditions characterizing the AERONET-OC sites, which include LWN spectra representative of a large diversity of water types in addition to peculiar illumination geometries spanning from subpolar to equatorial regions.

A–QCLWN comprises various steps hereafter summarized referring to the schematic shown in Fig. 1.

Fig. 1.
Fig. 1.

Schematic of the A–QCLWN procedure.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

a. Data ingestion and preprocessing

Data ingestion allows uploading level 1.5 and level 2.0 data from the AERONET-OC version 3 database and additionally a number of parameters essential for data preprocessing. For instance, recalling that the basic quality checks used for the removal of Li(λ) and LT(λ) data exhibiting large variability of intraband measurements entail stricter thresholds for level 2.0 with respect to level 1.5 (Zibordi et al. 2022), only the level 1.5 data satisfying the same thresholds applied to level 2.0 data are retained for successive analysis. Consequently, these thresholds indicating maximum standard deviations allowed for intraband data in each measurement sequence are included among the required A–QCLWN input processing parameters. Additional generic inputs are the dates identifying the periods for which level 1.5 and level 2.0 data are ingested. When level 1.5 and level 2.0 data refer to the same period and consequently level 2.0 data are a subset of level 1.5 data, the A–QCLWN procedure allows comparing its performance with respect to the already quality-controlled level 2.0 data. Conversely, if level 1.5 and level 2.0 refer to diverse periods, A–QCLWN allows identifying those level 1.5 LWN spectra not benefitting of existing QC results. Specific input parameters for A–QCLWN are detailed in the following subsections for each QC step.

b. Automated QC criteria for level 1.5 data

Actual QC of each level 1.5 LWN spectrum is performed applying diverse automated tests aiming at verifying (i) the relative consistency of each spectrum to be quality controlled with respect to a reference, (ii) the intraband spectral consistency; and finally, when applicable, (iii) the temporal consistency of each spectrum with respect to those close in time.

1) Relative consistency

The relative consistency aims at verifying the existence of level 2.0 LWN spectra exhibiting statistical similarity with the level 1.5 LWN spectrum under test. Specifically, it is ideally assumed that any LWN spectrum from a specific site must have features already represented in a fully quality-controlled dataset (typically, but not exclusively, from the same site) already existing at level 2.0 as a result of the expert-based E–QCLWN process.

The relative-consistency test implemented in A–QCLWN implies the use of a reference LWN spectrum hereafter called prototype, for each spectrum to be quality controlled called candidate. Prototype LWN spectra are produced using the similarity matrix relating level 1.5 LWN spectra to level 2.0 LWN spectra. This similarity matrix S is defined by Sij entries representing the Euclidean distance between the ith row of each level 1.5 LWN spectrum and the jth row of a level 2.0 LWN spectrum, and consequently allows associating each level 1.5 LWN spectrum to a number of level 2.0 LWN spectra exhibiting close spectral features.

The relative-consistency test implies the following:

  1. The identification through the similarity matrix S of a fixed number of level 2.0 LWN reference spectra best approximating the level 1.5 candidate and consequently warranting statistical representativity to the prototype. Five level 2.0 reference spectra have been used to QC candidate LWN spectra from current AERONET-OC sites.

  2. The construction of a relative-consistency prototype spectrum obtained by averaging the level 2.0 data records best approximating the level 1.5 candidate, with the standard deviation σRC(λ) at each center wavelength λ determining the spectral statistical representativity of the prototype.

  3. The analysis of the LWN spectral differences between candidate and prototype. The candidate LWN spectrum passes the relative-consistency test when the spectral differences between candidate and prototype at each center wavelength λ are explained by the spectral statistical representativity σRC(λ) determined for the prototype and the spectral standard uncertainty uC(λ) assigned to the candidate. This latter uncertainty is spectrally determined in agreement with the scheme proposed in Zibordi et al. (2022) as a sole function of the radiance values, regardless of center wavelength and measurement site with uC(λ) = 0.0091 + 0.0405LWN(λ).

Assuming a coverage factor k = 2 and negligible the correlation between errors contributing to uncertainties, the candidate spectrum passes the relative-consistency test when the difference ΔRC[LWN(λ)] between relative-consistency prototype and candidate LWN(λ) at each λ satisfies the test (Immler et al. 2010):
|ΔRC[LWN(λ)]|<2σRC2(λ)+uC2(λ),
where σRC(λ) approximates the standard uncertainty affecting the relative-consistency prototype.

Still, regardless of the former test, candidate LWN(λ) spectra are rejected when σRC(λ)>3uC(λ) at any center wavelength λ. This test only aims at excluding from level 2.0 those candidates poorly represented in the reference dataset as determined by a very high value of σRC(λ).

2) Spectral consistency

The spectrally asynchronous measurements performed by the AERONET-OC radiometers may introduce spectral inconsistencies. Some of these inconsistencies may not be captured by the relative-consistency test. Because of this, an additional test is applied to the LWN candidate spectra to identify the presence of excessive unexplained maxima or minima in a given spectral interval. This exclusive test verifies the presence of positive convexities with features exceeding a minimum user definable threshold. The spectral-consistency test is commonly restricted to the 442–560 nm interval generally exhibiting an LWN(λ) spectral shape not characterized by pronounced features. Implications associated with the selection of the spectral interval for this test are addressed in the discussion section.

The candidate spectrum passes the spectral-consistency test when the absolute change rate of LWN(λ) per unit wavelength |ΔLWNλ| determined nearby the center wavelength corresponding to a local minimum of LWN(λ), does not exceed a threshold. A value of 0.0001 mW cm−2 μm−2 sr−1 has been applied for the sites analyzed in this study.

3) Temporal consistency

The temporal-consistency test aims at identifying the level 1.5 LWN spectra exhibiting features significantly deviating from those of the previous or subsequent spectra in a given time interval. The rationale for such a test is the assumption that any relatively short temporal change affecting LWN spectra should not exhibit sharp variations in trends in any spectral band.

The temporal-consistency test implies

  1. The identification of level 1.5 LWN spectra close in time to the candidate based on specific constraints such as the time interval including the candidate and the minimum number of LWN close-in-time spectra. The time interval and the related number of samples required to ensure reliability to the test may differ across measurement sites due to local environmental variabilities. Still, a 2-h interval and a minimum of nine level 1.5 LWN spectra appeared to ensure robustness to the test regardless of the AERONET-OC site when applied to data collected with CE-318T radiometers programmed to perform six sea measurement sequences per hour.

  2. The application of a smoothing function to the selected level 1.5 LWN spectra to construct a temporal-consistency LWN prototype. The spectral statistical representativity of this LWN prototype is determined by the standard deviation σTC(λ) of the LWN(λ) data contributing to the computation of its value. A smoothing function with a five-sample smoothing box has been applied to data from current AERONET-OC sites.

Similar to the relative consistency, the candidate LWN spectrum passes the temporal-consistency test when the spectral differences between candidate and prototype are explained by the spectral statistical representativity assigned to the prototype σTC(λ) and the spectral standard uncertainty uC(λ) determined for the candidate. Again, assuming a coverage factor k = 2, the candidate spectrum passes the temporal-consistency criterion when the spectral differences ΔTC[LWN(λ)] between temporal-consistency prototype and candidate LWN(λ) at each λ satisfies the test:
|ΔTC[LWN(λ)]|<2σTC2(λ)+uC2(λ),
where σTC(λ) approximates the standard uncertainty affecting the temporal-consistency prototype.

Equivalent to relative consistency, candidate LWN(λ) spectra are rejected when σTC(λ)>3uC(λ) at any center wavelength λ. This criterion only aims at excluding from level 2.0 those candidates characterized by a low temporal stability, as determined by a very high value of σTC(λ).

Considering the irregular temporal distribution of LWN(λ) due to clouds and sea state perturbations, the temporal-consistency criterion is not a critical element impacting the raise of candidate level 1.5 data to level 2.0. Nevertheless, it allows enforcing further confidence in the accuracy of data.

c. Assessment and ranking

Besides the automated process, the uniqueness of the proposed A–QCLWN procedure is the potential for ranking the quality of data using independent indices whose weighted contributions can be combined into a single value defining the quality of each fully QC LWN spectrum.

The independent tests applied to assess each level 1.5 candidate LWN spectrum lead to the generation of 0 and 1 flags indicating failure or success conditions, respectively.

In summary,

  • the relative-consistency test checks if the difference between the level 1.5 candidate and the level 2.0 prototype LWN(λ) data does not exceed spectral uncertainty thresholds determined for the candidate and the prototype at each center wavelength λ in the 400–1020 nm interval;

  • the spectral-consistency test checks if the candidate level 1.5 LWN(λ) in a given spectral interval, does not exhibit a pattern affected by positive convexities suggesting unexplained spectral features;

  • the temporal-consistency test, when applicable, verifies if the difference between the candidate and the time-matching level 1.5 LWN(λ) prototype obtained smoothing the level 1.5 values preceding and following the candidate, do not exceed spectral uncertainty thresholds determined for the candidate and the prototype at each center wavelength λ in the 400–1020 nm interval.

The overall rank of the level 1.5 candidate LWN spectrum is determined by
R=(RCWR+TCWT)SC,
where RC, TC, and SC indicate the relative-, temporal-, and spectral-consistency flags with values of 0 or 1, while WR and WT are weighting coefficients set to 0.6 and 0.4, respectively. Consequently, ℜ exhibits values comprised between 0 and 1, with 0.6 assumed to be the minimum value satisfying level 2.0 quality requirements.

The values of the weighting coefficients WR and WT have been simply suggested by the need to consider as quality controlled any LWN(λ) spectrum satisfying minimum QC criteria. So, considering an arbitrary ranking scale from 0 to 1 (with 1 indicating the highest data quality), and taking acceptable (i.e., quality controlled) any spectrum exhibiting ranking exceeding 0.5, the value of 0.6 is assigned to data satisfying the relative-consistency test assumed as the minimum, but also sufficient, criterion when the spectral-consistency test is also passed. In other words, Eq. (5) shows that SC is exclusive, and consequently SC = 0 forces ℜ = 0.0, which excludes the candidate LWN spectrum from level 2.0. Conversely, RC and TC contribute with different weights to ℜ. Specifically, in the current implementation of A–QCLWN, when SC = 1 and only the relative-consistency test is satisfied (i.e., RC = 1 and TC = 0), the candidate LWN spectrum is considered qualified for level 2.0 with ℜ = 0.6. When the temporal-consistency test is also satisfied (i.e., RC = 1 and TC = 1), the candidate is qualified for level 2.0 with rank ℜ = 1.0.

Outputs of the LWN assessment and ranking steps are (i) a file listing the level 1.5 LWN candidates not qualified for level 2.0 and (ii) a log file including statistics on the overall QC process with details on results from each test applied to the level 1.5 candidate LWN(λ).

4. Results from the application of the automated A–QCLWN procedure

The performance of the automated A–QCLWN procedure varies from site to site as a result of features intrinsic to the level 1.5 LWN spectra or of the quality of the level 2.0 LWN spectra supporting the relative-consistency test. The A–QCLWN performance has been evaluated using data solely collected with the recent CE-318T instruments, which benefit of an extended number of spectral bands and can ensure an increased number of measurement sequences per unit time. Specifically, the A–QCLWN performance has been evaluated exploiting LWN(λ) data from AERONET-OC sites representative of different water types. These include (see Fig. 2) (i) the Casablanca Platform (CPL) in the western Mediterranean Sea often exhibiting Case-1 oligotrophic–mesotrophic waters; (ii) the Acqua Alta Oceanographic Tower (AAOT) in the northern Adriatic Sea and the Galata Platform (GLT) in the Black Sea characterized by optically complex waters exhibiting various concentrations of suspended sediments and chromophoric dissolved organic matter (CDOM); and finally, (iii) the Irbe Lighthouse Tower (ILT) in the Baltic Sea characterized by extremely high CDOM concentrations. Figure 3 shows LWN representative spectra for the considered sites.

Fig. 2.
Fig. 2.

AERONET-OC sites considered in the study: Casablanca Platform (CPL) in the western Mediterranean Sea, Acqua Alta Oceanographic Tower (AAOT) in the northern Adriatic Sea, Galata Platform (GLT) in the Black Sea, and Irbe Lighthouse Tower (ILT) in the Baltic Sea.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

Fig. 3.
Fig. 3.

Mean LWN spectra for the considered AERONET-OC sites (CPL, AAOT, GLT, and ILT) determined from level 2.0 LWN spectra produced between 2017 and 2019. The error bars indicate ±1 standard deviation of the averaged LWN spectral values. Some center wavelengths have been shifted by 2 nm to increase readability of the spectral values.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

The A–QCLWN performance analysis relies on AERONET-OC level 2.0 data quality controlled with the E–QCLWN procedure summarized in section 2 and comprehensively detailed in Zibordi et al. (2022). The number of AERONET-OC level 2.0 LWN spectra used in the following analyses, all corrected for bidirectional effects applying the Morel et al. (2002) scheme, largely varies across the diverse deployment sites. Nevertheless, the number of spectra only reflects the data produced with CE-318T radiometers programmed to perform at least six sea-viewing measurement sequences per hour in the 400–1020 nm spectral range, whose quality was formerly controlled with the E–QCLWN procedure.

a. Summary results from the application of A–QCLWN

Results from the application of A–QCLWN are summarized in Table 1 with rank ℜ determined for the LWN spectra from each AERONET-OC site. Recalling that level 1.5 LWN spectra are assumed qualified for level 2.0 when ℜ ≥ 0.6, Table 1 shows that A–QCLWN generally qualifies for level 2.0 more than 90% of the level 1.5 candidate LWN spectra with values varying from approximately 89% for GLT up to 93% for ILT. Consistently, the cumulative number of level 1.5 candidate LWN spectra excluded from level 2.0 exhibiting values of ℜ ≤ 0.4 varies from approximately 7% for ILT up to 11% for GLT.

Table 1

Ranking of the candidate level 1.5 LWN spectra from CPL, AAOT, GLT, and ILT sites. N indicates the number of spectra evaluated. The integers in each column indicate the number of level 1.5 LWN candidate spectra satisfying requirements leading to ℜ = 1.0, 0.6, 0.4, and 0.0. The decimal numbers in parentheses indicate the percentage with respect to the total. The “total accepted” values indicate the sum of the percentage of level 1.5 LWN candidate spectra exhibiting ℜ = 1.0 or 0.6.

Table 1

Sample outputs from each QC step implemented in A–QCLWN are illustrated in Figs. 47. Specifically, Fig. 4 shows the candidate LWN spectrum passing the relative-, spectral-, and temporal-consistency tests. Conversely, Fig. 5 provides the example of a candidate LWN spectrum passing the relative- and spectral-consistency test but not the temporal-consistency one at the single 400 nm center wavelength. Failure of the temporal-consistency test illustrated in Fig. 5 is also documented through the LWN(λ) time series displayed in Fig. 6. The candidate spectrum identified by the vertical bar, clearly shows that the value at 400 nm deviates from the temporal trend of the close-by data. This spectral deviation is again explained by the asynchronous spectral acquisition of LT(λ) and to a lesser extent of Li(λ) data. The same explanation applies to the spectral inconsistencies affecting the candidate LWN spectrum displayed in Fig. 7 passing the relative- and temporal-consistency tests, but not the spectral-consistency one. This is notably shown by the positive convexity produced by the relative minimum occurring at the 510 nm center wavelength.

Fig. 4.
Fig. 4.

Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests indicated by Rc and Tc, respectively) with ℜ = 1.0 determined by RC = 1 and TC = 1 (sample from the AAOT at 1259:27 UTC 20 Jul 2019). The error bars indicate 2 times the standard uncertainty determined for the candidate [i.e., 2uC(λ)], 2 times the standard deviation of the reference level 2.0 spectra contributing to the relative-consistency analysis for the Rc prototype [i.e., 2σRC(λ) ], and 2 times the standard deviation of the level 1.5 spectra contributing to the temporal-consistency analysis for the Tc prototype [i.e., 2σTC(λ) ].

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

Fig. 5.
Fig. 5.

Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests indicated by Rc and Tc, respectively) exhibiting ℜ = 0.6 determined by RC = 1 and TC = 0 (sample from AAOT at 1000:47 UTC 21 Aug 2018). The error bars have been determined as in Fig. 4.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

Fig. 6.
Fig. 6.

Comparison of LWN(λ) candidate and temporal-consistency time series (sample from AAOT at 1000:47 UTC 21 Aug 2018). The filled squares indicate actual LWN(λ) data while filled circles indicate smoothed LWN(λ) values. Colors from cyan to black indicate center wavelengths from 400 to 670 nm. The vertical bar highlights the LWN candidate spectrum exhibiting temporal inconsistency at 400 nm (the enlarged cyan filled square) leading to TC = 0.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

Fig. 7.
Fig. 7.

Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) with RC = 1 and TC = 1, but implying exclusion from level 2.0 because of SC = 0 determined by the local minimum of the candidate LWN(λ) at 510 nm (sample from AAOT at 1020:17 UTC 28 Nov 2018). The error bars have been determined as in Fig. 4.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

b. Evaluation of the A–QCLWN performance

An evaluation of the performance of the A–QCLWN procedure is possible through confusion matrices determined using LWN spectra already quality controlled with the expert-based E–QCLWN procedure. Results from this comparison are summarized in Tables 25 for data from the CPL, AAOT, GLT, and ILT sites, respectively. Results show agreement A (i.e., the sum of the candidate spectra rejected or accepted by both procedures) varying between approximately 88% at CPL and up to 93% at AAOT.

Table 2

Confusion matrix for CPL data determined with N = 464 level 1.5 LWN candidate spectra with agreement A = 87.7% between results from the automated A–QCLWN and the expert-based E–QCLWN procedures. Levels 1.5 and 2.0 data are from 2 Apr to 25 Aug 2019.

Table 2
Table 3

As in Table 2, but for AAOT levels 1.5 and 2.0 data from 4 Oct 2017 to 20 Jul 2019 (N = 4919 and A = 92.8%).

Table 3
Table 4

As in Table 2, but for GLT levels 1.5 and 2.0 data from 5 Nov 2018 to 6 Aug 2019 (N = 1375 and A = 88.9%).

Table 4
Table 5

As in Table 2, but for ILT levels 1.5 and 2.0 data from 4 Jul 2018 to 17 Sep 2019 (N = 1342 and A = 91.8%).

Table 5

Still, the agreement A cannot be considered an index for the absolute assessment of A–QCLWN. In fact, while E–QCLWN largely depends on qualitative comparisons between spectra either indicating relative consistency or temporal consistency, A–QCLWN performs the same tests relying on thresholds specifically determined for candidate and prototype spectra. This means that while E–QCLWN may be affected by subjective decisions, A–QCLWN strongly depends on the thresholds applied. Nevertheless, when looking at the full independence of the two quality control procedures and at the results obtained from their application to AERONET-OC data from sites characterized by very different water types and measurement conditions, values of A in the range of 88%–93% appear extremely satisfactory. Definitively, in virtue of its scheme relying on sample-specific thresholds, A–QCLWN logically outclasses E–QCLWN in terms of precision as shown in the previous section.

c. Examples of A–QCLWN and E–QCLWN conflicting results

The confusion matrices provided in Tables 25 indicate disagreements varying between approximately 7% and 12% across the various sample cases. The source of this disagreement between the A–QCLWN and E–QCLWN results is largely explained by the lower precision and “subjective” decisions of the expert performing the quality control process. Two examples are provided below.

The first example is illustrated in Fig. 8 and shows well matching candidate and prototype spectra. However, the relative-consistency test is not satisfied for the value of LWN at the 665 nm center wavelength, thus preventing to raise the candidate spectrum to level 2.0. The analysis performed by the expert through the E–QCLWN procedure did not identify any spectral issue, and consequently allowed for qualifying the candidate spectrum for level 2.0.

Fig. 8.
Fig. 8.

Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) exhibiting ℜ = 0.4 determined by RC = 0 due to the value of the LWN(λ) candidate at the 665 nm center wavelength, but still exhibiting TC = 1 as a result of passing the temporal-consistency test. The same candidate spectrum (sample from AAOT at 1053:14 UTC 11 Jan 2019) was formerly passing the E–QCLWN scrutiny. The error bars have been determined as in Fig. 4.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

A second example of disagreement between A–QCLWN and E–QCLWN is illustrated in Fig. 9. The candidate spectrum passes all the A–QCLWN tests and consequently has been raised to level 2. Still, it was not considered qualified for level 2.0 when quality checked through the E–QCLWN procedure because of the assumed large temporal change affecting the candidate LWN(λ) (see Fig. 10). It is recalled that E–QCLWN raises LWN candidate spectra to level 2.0 by combining “subjective” decisions on both statistical representativity and, when data are available, temporal stability (see Zibordi et al. 2022).

Fig. 9.
Fig. 9.

Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) exhibiting ℜ = 1.0 determined by RC = 1 and TC = 1 (sample from AAOT at 0942:07 UTC 2 Jun 2019). The same candidate spectrum did not formerly pass the E–QCLWN scrutiny due to a temporal change considered too large by the expert (see also Fig. 10). The error bars have been determined as in Fig. 4.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

Fig. 10.
Fig. 10.

Comparison of LWN(λ) candidate and temporal-consistency time series (sample from AAOT at 0942:07 UTC 2 Jun 2019). The filled squares indicate actual LWN(λ) data while filled circles indicate smoothed LWN(λ) values. Colors from cyan to black indicate center wavelengths from 400 to 670 nm. The vertical bar highlights the LWN candidate spectrum passing the A–QCLWN temporal-consistency test and consequently leading to TC = 1.

Citation: Journal of Atmospheric and Oceanic Technology 39, 12; 10.1175/JTECH-D-22-0029.1

The above examples show the higher precision of the A–QCLWN with respect to the E–QCLWN procedure. In particular, A–QCLWN prevents subjective decisions on the quality of the data. Further, the robustness introduced by the application of A–QCLWN should prevent a “deterioration” of the QC process over time. In fact, it appears to warrant the exclusion of poor data in the QC process, which could successively affect the process itself.

5. Discussion

The availability of QC level 2.0 LWN spectra for each site has been anticipated to be a requirement for the A–QCLWN procedure. This choice ensures access to reference spectra best strengthening the relative-consistency test by avoiding the use of LWN(λ) data that may exhibit measurement artifacts. The requirement, however, becomes definitively a limitation each time there is a novel site because it entails the application of an independent and likely nonautomated quality control process (e.g., the E–QCLWN one). It may become an even more relevant constraint when candidate LWN spectra are not statistically represented in level 2.0, which is a condition that cannot be excluded for multiyear data from sites characterized by environmental changes.

Actually, the requirement on the availability of quality-controlled level 2.0 LWN data could be omitted if major measurement artifacts can be excluded from level 1.5 LWN data, which is a reasonable assumption considering that poor LWN spectra are largely removed by the basic QC tests applied when rising level 1.0 LWN data to level 1.5. This less stringent approach has been verified using level 1.5 in place of the level 2.0 data as input for the tests implemented in A–QCLWN. Summary results from such an exercise are provided in Table 6. They only indicate a slight increase of accepted level 1.5 LWN candidate spectra (not exceeding 1.2% for the considered sites and periods) with respect to the analyses relying on level 2.0 reference data. This result is explained by the larger variability statistically affecting the prototype spectra due to the use of LWN(λ) data not previously quality controlled and likely leading to larger values of the standard deviation σRC(λ), which naturally increases the probability of raising to level 2.0 candidate level 1.5 LWN spectra satisfying Eq. (3). The increase of the accepted LWN candidate spectra shown by the use of level 1.5 data alternative to level 2.0 for the relative-consistency test may then slightly decrease the accuracy of the QC process. Nevertheless, the above results appear to support the statement that “often, wisely and purposely collected data carry such clear message that they essentially ‘analyze themselves’” (Vardeman and Jobe 2016).

Table 6

Ranking of the level 1.5 LWN candidate spectra from the CPL, AAOT, GLT, and ILT candidate samples, determined exclusively using level 1.5 in place of level 2.0 LWN spectra. N indicates the number of spectra evaluated. The integers in each column are the number of level 1.5 candidate LWN spectra satisfying requirements leading to ℜ = 1.0, 0.6, 0.4, and 0.0. The decimal numbers in parentheses are the percentage with respect to the total. The “total accepted” values indicate the sum of the percentage of level 1.5 candidate LWN spectra exhibiting ℜ = 1.0 or 0.6. The percentage values in parentheses associated with the “total accepted” ones are the increments resulting from the use of level 1.5 in place of level 2.0 LWN.

Table 6

An additional element requiring discussion is the spectral range considered for the spectral-consistency test. The basic assumption supporting the validity of such a test is the absence of any positive convexity in the shape of the LWN candidate spectrum within the selected spectral interval. This implies, for instance, the inapplicability of this test to LWN spectral regions in the presence of peculiar biological events such as blooms producing positive convexities in the blue part of the spectrum due to an increase of LWN from 443 toward 400 nm. This limitation, more likely affecting sites such as the Baltic Sea ones characterized by summer cyanobacteria blooms, could be mitigated with the implementation of additional specific tests accounting for the amplitude of the spectrum or the frequency and uniformity of relative minima at specific center wavelengths. Still, any site specific investigation is out of the scope of this work.

It is evident that LWN candidates not affected by spectral inconsistencies and only passing the temporal-consistency test may suggest a poor representativity of the candidate spectrum among the input reference ones (e.g., the level 2.0 LWN). Because of this, the optional capability to check critical cases such as those resulting from the rejection of candidates with quality only supported by TC = 1, but exhibiting RC = 0, has been included in the current implementation of A–QCLWN. This discretionary expert-based step may allow modifying the value of ℜ, and consequently force the acceptance of potentially novel level 1.5 LWN spectra not yet represented in the level 2.0 dataset.

Finally, it is emphasized that the spectral standard uncertainty uC(λ) assigned to LWN(λ) is determined from a linear relationship statistically linking the two quantities (Zibordi et al. 2022). This practical solution has been suggested by results from a best effort to quantify uncertainties affecting AERONET-OC LWN(λ) from sites located in the Adriatic Sea, Baltic Sea, and Black Sea (Gergely and Zibordi 2014). It is then expected that new and more comprehensive investigations on AERONET-OC LWN(λ) uncertainties may provide new findings likely allowing to more accurately quantify uC(λ) and consequently improve QC at sites exhibiting different water types.

6. Summary and conclusions

The application of QA and QC best practices is essential to ensure a confident use of measurements. Focusing on QC, which aims at supporting the provision of high-quality data through postmeasurement actions, this work has shown the potential for a fully automated procedure to quality control AERONET-OC LWN data. Specifically, attempting to reproduce the steps allowing an expert analyst to raise level 1.5 LWN(λ) data to level 2.0, an automated procedure called A–QCLWN has been implemented and tested.

The proposed procedure, which complements a number of basic tests supporting QC of level 1.0 and level 1.5 LWN spectra, verifies (i) the relative consistency of level 1.5 LWN spectra (called candidate) with respect to LWN reference spectra (called prototype) constructed using LWN data from level 2.0 independently quality controlled; (ii) the absence of any pronounced spectral feature in portions of LWN spectra expected to show a regular shape; and finally, when applicable, (iii) the temporal consistency of each LWN candidate spectrum with respect to close-in-time spectra relying on the assumption that during a relatively short time interval LWN spectra should not exhibit any sudden change in the trend of spectral values. This last step is considered valuable to increase confidence on the quality of LWN candidate spectra already passing the former two basic steps. Consequently, the temporal-stability criterion is only considered relevant for future AERONET-OC developments ideally allowing to better address the exploitation of level 2.0 data products with an accuracy better targeting application requirements.

The performance of A–QCLWN has then been tested using LWN spectra from sites representative of very different water types comprising oligotrophic/mesotrophic waters dominated by chlorophyll-a concentration and a variety of optically complex coastal waters. Such an evaluation, restricted to data collected with CE-318T radiometers programmed to perform six sea-viewing measurement sequences per hour in the 400–1020 nm spectral interval, indicates acceptance rates of LWN candidate spectra varying between approximately 89% and 93% across the various AERONET-OC sites. These results exhibit agreement in the range of 88%–93% with LWN spectra independently quality controlled by an expert analyst. When considering the more qualitative assessment provided by an expert analyst compared to the precision distinguishing the A–QCLWN relying on uncertainties assigned to LWN candidates and prototypes, the performance of A–QCLWN certainly outclasses that relying on the expert analyst.

It is remarked the potential for ranking the fully quality-controlled AERONET-OC LWN spectra in view of supporting their targeted exploitation as a function of different application requirements. Taking into account that the spectral-consistency test leads to the exclusion of any LWN candidate exhibiting spectral inconsistencies, in the current implementation of A–QCLWN the LWN candidate spectra passing both the relative- and temporal-consistency tests obtain the highest rank, followed by those candidates only exhibiting relative consistency. Conversely, LWN candidates only showing temporal consistency are not considered qualified for level 2.0.

An apparent constraint of the current A–QCLWN implementation is the expected availability of already QC level 2.0 LWN spectra for the relative-consistency test. This limitation has however shown a fairly low impact on the QC process. In particular, the use of LWN prototype spectra built from level 1.5 data alternative to the level 2.0 has led to an increase of the acceptance rate varying from 0% to 1.2% across the sites considered. This result is explained by the basic QC tests embedded in the AERONET-OC data processing already removing LWN spectra affected by major artifacts from level 1.5. The slight increase in the acceptance rate is explained by the presence of non–fully quality-controlled LWN spectra among those used to construct relative-consistency prototypes. This increases the spectral variability assigned to the LWN prototypes, and consequently the probability of matching the spectral features of the LWN candidates.

The above results suggest confidence in the application of A–QCLWN to AERONET-OC LWN(λ) data with the major benefit of removing subjective decisions in the process naturally affecting any expert-based screening, still recognizing that the reliability of any statistical QC approach depends on the accuracy of the input data and parameters.

Acknowledgments.

This work benefitted of support from the Joint Research Centre through the COLORS project and from the European Commission Directorate-General for Defence Industry and Space through the Copernicus Programme. The authors thank the AERONET Team for processing and distributing the AERONET-OC data.

Data availability statement.

AERONET-OC data are accessible at https://aeronet.gsfc.nasa.gov/cgi-bin/draw_map_display_seaprism_v3.

REFERENCES

  • Bushnell, M., C. Waldmann, S. Seitz, E. Buckley, M. Tamburri, J. Hermes, E. Henslop, and A. Lara-Lopez, 2019: Quality assurance of oceanographic observations: Standards and guidance adopted by an international partnership. Front. Mar. Sci., 6, 706, https://doi.org/10.3389/fmars.2019.00706.

    • Search Google Scholar
    • Export Citation
  • Bushnell, M., and Coauthors, 2020: QARTOD—Prospects for real-time quality control manuals, how to create them, and a vision for advanced implementation. U.S. Integrated Ocean Observing System Rep., 22 pp., https://doi.org/10.25923/ysj8-5n28.

  • D’Alimonte, D., and G. Zibordi, 2006: Statistical assessment of radiometric measurements from autonomous systems. IEEE Trans. Geosci. Remote Sens., 44, 719728, https://doi.org/10.1109/TGRS.2005.862505.

    • Search Google Scholar
    • Export Citation
  • D’Alimonte, D., T. Kajiyama, G. Zibordi, and B. Bulgarelli, 2021: Sea-surface reflectance factor: Replicability of computed values. Opt. Express, 29, 25 21725 241, https://doi.org/10.1364/OE.424768.

    • Search Google Scholar
    • Export Citation
  • Gergely, M., and G. Zibordi, 2014: Assessment of AERONET-OC LWN uncertainties. Metrologia, 51, 4047, https://doi.org/10.1088/0026-1394/51/1/40.

    • Search Google Scholar
    • Export Citation
  • Giles, D. M., and Coauthors, 2019: Advancements in the Aerosol Robotic Network (AERONET) version 3 database—Automated near-real-time quality control algorithm with improved cloud screening for sun photometer aerosol optical depth (AOD) measurements. Atmos. Meas. Tech., 12, 169209, https://doi.org/10.5194/amt-12-169-2019.

    • Search Google Scholar
    • Export Citation
  • Holben, B. N., and Coauthors, 2001: An emerging ground‐based aerosol climatology: Aerosol optical depth from AERONET. J. Geophys. Res., 106, 12 06712 097, https://doi.org/10.1029/2001JD900014.

    • Search Google Scholar
    • Export Citation
  • Immler, F. J., J. Dykema, T. Gardiner, D. N. Whiteman, P. W. Thorne, and H. Vömel, 2010: Reference quality upper-air measurements: Guidance for developing GRUAN data products. Atmos. Meas. Tech., 3, 12171231, https://doi.org/10.5194/amt-3-1217-2010.

    • Search Google Scholar
    • Export Citation
  • Lee, Z. P., K. Du, K. J. Voss, G. Zibordi, B. Lubac, R. Arnone, and A. Weidemann, 2011: An inherent-optical-property-centered approach to correct the angular effects in water-leaving radiance. Appl. Opt., 50, 31553167, https://doi.org/10.1364/AO.50.003155.

    • Search Google Scholar
    • Export Citation
  • Mobley, C. D., 1999: Estimation of the remote-sensing reflectance from above-surface measurements. Appl. Opt., 38, 74427455, https://doi.org/10.1364/AO.38.007442.

    • Search Google Scholar
    • Export Citation
  • Morel, A., D. Antoine, and B. Gentili, 2002: Bidirectional reflectance of oceanic waters: Accounting for Raman emission and varying particle scattering phase function. Appl. Opt., 41, 62896306, https://doi.org/10.1364/AO.41.006289.

    • Search Google Scholar
    • Export Citation
  • Smirnov, A., B. N. Holben, T. F. Eck, O. Dubovik, and I. Slutsker, 2000: Cloud-screening and quality control algorithms for the AERONET database. Remote Sens. Environ., 73, 337349, https://doi.org/10.1016/S0034-4257(00)00109-7.

    • Search Google Scholar
    • Export Citation
  • Talone, M., G. Zibordi, and Z. Lee, 2018: Correction for the non-nadir viewing geometry of AERONET-OC above water radiometry data: An estimate of uncertainties. Opt. Express, 26, A541A561, https://doi.org/10.1364/OE.26.00A541.

    • Search Google Scholar
    • Export Citation
  • Tanré, D., M. Herman, P. Y. Deschamps, and A. De Leffe, 1979: Atmospheric modeling for space measurements of ground reflectances, including bidirectional properties. Appl. Opt., 18, 35873594, https://doi.org/10.1364/AO.18.003587.

    • Search Google Scholar
    • Export Citation
  • Vardeman, S. B., and J. M. Jobe, 2016: Statistical Methods for Quality Assurance. Springer-Verlag, 437 pp.

  • Zibordi, G., F. Mélin, S. B. Hooker, D. D’Alimonte, and B. Holben, 2004: An autonomous above-water system for the validation of ocean color radiance data. IEEE Trans. Geosci. Remote Sens., 42, 401415, https://doi.org/10.1109/TGRS.2003.821064.

    • Search Google Scholar
    • Export Citation
  • Zibordi, G., and Coauthors, 2009: AERONET-OC: A network for the validation of ocean color primary products. J. Atmos. Oceanic Technol., 26, 16341651, https://doi.org/10.1175/2009JTECHO654.1.

    • Search Google Scholar
    • Export Citation
  • Zibordi, G., K. J. Voss, B. C. Johnson, and J. L. Mueller, 2019: Protocols for satellite ocean colour data validation: In situ optical radiometry. IOCCG Protocol Series Vol. 3, 72 pp., https://doi.org/10.25607/OBP-691.

  • Zibordi, G., M. Talone, and F. Mélin, 2022: Uncertainty estimate of satellite-derived normalized water-leaving radiance. IEEE Geosci. Remote Sens. Lett., 19, 1502905, https://doi.org/10.1109/LGRS.2021.3134876.

    • Search Google Scholar
    • Export Citation
Save
  • Bushnell, M., C. Waldmann, S. Seitz, E. Buckley, M. Tamburri, J. Hermes, E. Henslop, and A. Lara-Lopez, 2019: Quality assurance of oceanographic observations: Standards and guidance adopted by an international partnership. Front. Mar. Sci., 6, 706, https://doi.org/10.3389/fmars.2019.00706.

    • Search Google Scholar
    • Export Citation
  • Bushnell, M., and Coauthors, 2020: QARTOD—Prospects for real-time quality control manuals, how to create them, and a vision for advanced implementation. U.S. Integrated Ocean Observing System Rep., 22 pp., https://doi.org/10.25923/ysj8-5n28.

  • D’Alimonte, D., and G. Zibordi, 2006: Statistical assessment of radiometric measurements from autonomous systems. IEEE Trans. Geosci. Remote Sens., 44, 719728, https://doi.org/10.1109/TGRS.2005.862505.

    • Search Google Scholar
    • Export Citation
  • D’Alimonte, D., T. Kajiyama, G. Zibordi, and B. Bulgarelli, 2021: Sea-surface reflectance factor: Replicability of computed values. Opt. Express, 29, 25 21725 241, https://doi.org/10.1364/OE.424768.

    • Search Google Scholar
    • Export Citation
  • Gergely, M., and G. Zibordi, 2014: Assessment of AERONET-OC LWN uncertainties. Metrologia, 51, 4047, https://doi.org/10.1088/0026-1394/51/1/40.

    • Search Google Scholar
    • Export Citation
  • Giles, D. M., and Coauthors, 2019: Advancements in the Aerosol Robotic Network (AERONET) version 3 database—Automated near-real-time quality control algorithm with improved cloud screening for sun photometer aerosol optical depth (AOD) measurements. Atmos. Meas. Tech., 12, 169209, https://doi.org/10.5194/amt-12-169-2019.

    • Search Google Scholar
    • Export Citation
  • Holben, B. N., and Coauthors, 2001: An emerging ground‐based aerosol climatology: Aerosol optical depth from AERONET. J. Geophys. Res., 106, 12 06712 097, https://doi.org/10.1029/2001JD900014.

    • Search Google Scholar
    • Export Citation
  • Immler, F. J., J. Dykema, T. Gardiner, D. N. Whiteman, P. W. Thorne, and H. Vömel, 2010: Reference quality upper-air measurements: Guidance for developing GRUAN data products. Atmos. Meas. Tech., 3, 12171231, https://doi.org/10.5194/amt-3-1217-2010.

    • Search Google Scholar
    • Export Citation
  • Lee, Z. P., K. Du, K. J. Voss, G. Zibordi, B. Lubac, R. Arnone, and A. Weidemann, 2011: An inherent-optical-property-centered approach to correct the angular effects in water-leaving radiance. Appl. Opt., 50, 31553167, https://doi.org/10.1364/AO.50.003155.

    • Search Google Scholar
    • Export Citation
  • Mobley, C. D., 1999: Estimation of the remote-sensing reflectance from above-surface measurements. Appl. Opt., 38, 74427455, https://doi.org/10.1364/AO.38.007442.

    • Search Google Scholar
    • Export Citation
  • Morel, A., D. Antoine, and B. Gentili, 2002: Bidirectional reflectance of oceanic waters: Accounting for Raman emission and varying particle scattering phase function. Appl. Opt., 41, 62896306, https://doi.org/10.1364/AO.41.006289.

    • Search Google Scholar
    • Export Citation
  • Smirnov, A., B. N. Holben, T. F. Eck, O. Dubovik, and I. Slutsker, 2000: Cloud-screening and quality control algorithms for the AERONET database. Remote Sens. Environ., 73, 337349, https://doi.org/10.1016/S0034-4257(00)00109-7.

    • Search Google Scholar
    • Export Citation
  • Talone, M., G. Zibordi, and Z. Lee, 2018: Correction for the non-nadir viewing geometry of AERONET-OC above water radiometry data: An estimate of uncertainties. Opt. Express, 26, A541A561, https://doi.org/10.1364/OE.26.00A541.

    • Search Google Scholar
    • Export Citation
  • Tanré, D., M. Herman, P. Y. Deschamps, and A. De Leffe, 1979: Atmospheric modeling for space measurements of ground reflectances, including bidirectional properties. Appl. Opt., 18, 35873594, https://doi.org/10.1364/AO.18.003587.

    • Search Google Scholar
    • Export Citation
  • Vardeman, S. B., and J. M. Jobe, 2016: Statistical Methods for Quality Assurance. Springer-Verlag, 437 pp.

  • Zibordi, G., F. Mélin, S. B. Hooker, D. D’Alimonte, and B. Holben, 2004: An autonomous above-water system for the validation of ocean color radiance data. IEEE Trans. Geosci. Remote Sens., 42, 401415, https://doi.org/10.1109/TGRS.2003.821064.

    • Search Google Scholar
    • Export Citation
  • Zibordi, G., and Coauthors, 2009: AERONET-OC: A network for the validation of ocean color primary products. J. Atmos. Oceanic Technol., 26, 16341651, https://doi.org/10.1175/2009JTECHO654.1.

    • Search Google Scholar
    • Export Citation
  • Zibordi, G., K. J. Voss, B. C. Johnson, and J. L. Mueller, 2019: Protocols for satellite ocean colour data validation: In situ optical radiometry. IOCCG Protocol Series Vol. 3, 72 pp., https://doi.org/10.25607/OBP-691.

  • Zibordi, G., M. Talone, and F. Mélin, 2022: Uncertainty estimate of satellite-derived normalized water-leaving radiance. IEEE Geosci. Remote Sens. Lett., 19, 1502905, https://doi.org/10.1109/LGRS.2021.3134876.

    • Search Google Scholar
    • Export Citation
  • Fig. 1.

    Schematic of the A–QCLWN procedure.

  • Fig. 2.

    AERONET-OC sites considered in the study: Casablanca Platform (CPL) in the western Mediterranean Sea, Acqua Alta Oceanographic Tower (AAOT) in the northern Adriatic Sea, Galata Platform (GLT) in the Black Sea, and Irbe Lighthouse Tower (ILT) in the Baltic Sea.

  • Fig. 3.

    Mean LWN spectra for the considered AERONET-OC sites (CPL, AAOT, GLT, and ILT) determined from level 2.0 LWN spectra produced between 2017 and 2019. The error bars indicate ±1 standard deviation of the averaged LWN spectral values. Some center wavelengths have been shifted by 2 nm to increase readability of the spectral values.

  • Fig. 4.

    Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests indicated by Rc and Tc, respectively) with ℜ = 1.0 determined by RC = 1 and TC = 1 (sample from the AAOT at 1259:27 UTC 20 Jul 2019). The error bars indicate 2 times the standard uncertainty determined for the candidate [i.e., 2uC(λ)], 2 times the standard deviation of the reference level 2.0 spectra contributing to the relative-consistency analysis for the Rc prototype [i.e., 2σRC(λ) ], and 2 times the standard deviation of the level 1.5 spectra contributing to the temporal-consistency analysis for the Tc prototype [i.e., 2σTC(λ) ].

  • Fig. 5.

    Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests indicated by Rc and Tc, respectively) exhibiting ℜ = 0.6 determined by RC = 1 and TC = 0 (sample from AAOT at 1000:47 UTC 21 Aug 2018). The error bars have been determined as in Fig. 4.

  • Fig. 6.

    Comparison of LWN(λ) candidate and temporal-consistency time series (sample from AAOT at 1000:47 UTC 21 Aug 2018). The filled squares indicate actual LWN(λ) data while filled circles indicate smoothed LWN(λ) values. Colors from cyan to black indicate center wavelengths from 400 to 670 nm. The vertical bar highlights the LWN candidate spectrum exhibiting temporal inconsistency at 400 nm (the enlarged cyan filled square) leading to TC = 0.

  • Fig. 7.

    Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) with RC = 1 and TC = 1, but implying exclusion from level 2.0 because of SC = 0 determined by the local minimum of the candidate LWN(λ) at 510 nm (sample from AAOT at 1020:17 UTC 28 Nov 2018). The error bars have been determined as in Fig. 4.

  • Fig. 8.

    Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) exhibiting ℜ = 0.4 determined by RC = 0 due to the value of the LWN(λ) candidate at the 665 nm center wavelength, but still exhibiting TC = 1 as a result of passing the temporal-consistency test. The same candidate spectrum (sample from AAOT at 1053:14 UTC 11 Jan 2019) was formerly passing the E–QCLWN scrutiny. The error bars have been determined as in Fig. 4.

  • Fig. 9.

    Comparison of LWN candidate and prototype spectra (for the relative- and temporal-consistency tests, indicated by Rc and Tc, respectively) exhibiting ℜ = 1.0 determined by RC = 1 and TC = 1 (sample from AAOT at 0942:07 UTC 2 Jun 2019). The same candidate spectrum did not formerly pass the E–QCLWN scrutiny due to a temporal change considered too large by the expert (see also Fig. 10). The error bars have been determined as in Fig. 4.

  • Fig. 10.

    Comparison of LWN(λ) candidate and temporal-consistency time series (sample from AAOT at 0942:07 UTC 2 Jun 2019). The filled squares indicate actual LWN(λ) data while filled circles indicate smoothed LWN(λ) values. Colors from cyan to black indicate center wavelengths from 400 to 670 nm. The vertical bar highlights the LWN candidate spectrum passing the A–QCLWN temporal-consistency test and consequently leading to TC = 1.

All Time Past Year Past 30 Days
Abstract Views 179 0 0
Full Text Views 501 241 12
PDF Downloads 368 192 10