Quality control of automated hyperspectral remote sensing measurements from a seaborne platform

S. P. Garaba, M. R. Wernand, and O. Zielinski University of Bremen, Department of Geosciences, P.O. Box 330440, 28334 Bremen, Germany Institute of Marine Resources, Department of Marine Physics and Sensors, Bussestraße 27, 27570 Bremerhaven, Germany Royal Netherlands Institute for Sea Research, Physical Oceanography, Marine Optics & Remote Sensing, P.O. Box 59, 1790AB Den Burg, Texel, The Netherlands

sea surface and sky images.The first three flags are based on meteorological conditions, to select erroneous incoming solar irradiance (E S ) taken during dusk, dawn, before significant incoming solar radiation could be detected or under rainfall.Furthermore, the relative azimuthal angle of the optical sensors to the sun is used to identify possible sunglint free sea surface zones.A total of 629 spectra remained after applying the meteorological masks (first three flags).Based on this dataset, a fourth flag for sunglint was generated by analysing and evaluating water leaving radiance (L W ) and remote sensing reflectance (R RS ) spectral behaviour in the presence and absence of sunglint salient in the simultaneously available sea surface images.Spectra conditions satisfying "mean L W (700-950 nm) < 2 mW m −2 nm −1 Sr −1 " or alternatively "minimum R RS (700-950 nm) < 0.010 Sr −1 ", mask the most measurements affected by sunglint, providing efficient flagging of sunglint in automated quality control.It is confirmed that valid optical measurements can be performed 0 • ≤ Φ ≤ 360 • although 90 • ≤ Φ ≤ 135 • is recommended.

Introduction
Remote sensing has become a key tool in the investigation of marine biochemical and geophysical characteristics on a regional or global scale (IOCCG, 2000;Schofield et al., 2004;Dierssen, 2010).During the last few decades, technical progress in remote sensing has made it possible to carry out optical measurements from in-water, abovewater, airborne and satellite platforms (Moore et al., 2009).However, automated and unmanned optical measurements from any of these platforms are likely to be erroneous Figures due to meteorological conditions (e.g.rainfall, cloud cover, humidity, and dusk/dawn conditions), sunglint, and sensor setup (Gordon and Jacobs, 1977;Wernand, 2002;Zhang and Wang, 2010).Therefore quality control to eliminate these disturbing factors is a crucial procedure in determining colour of seawater.Sunglint is a phenomenon resulting from a direct beam of sunlight reflected from a seawater surface directly into the down looking optical sensor (Morel and Gordon, 1980;Ottaviani et al., 2008).Studies suggest that sunglint is caused by Fresnel reflection from a number of "dancing facets" on the wind affected seawater surface and is controlled by the position of the sun, optical sensor viewing angle, water refractive index, wind direction and speed (Kay et al., 2009;Zhang and Wang, 2010).In a recent audit of sunglint correction models for optical measurements in marine environments it is conceded that despite their benefits most of these models rely partly on the black pixel assumption and therefore tend to moderately correct glint pixels (Kay et al., 2009).This black pixel assumption, water leaving radiance is insignificant in the near infra-red spectrum, has been reported to be inconsistent to some degree in coastal and turbid waters, which contributes also to the fact that correction models come with a probability of over-or underestimation of apparent and inherent optical properties (Siegel et al., 2000;Shi and Wang, 2009).This framework provides the motivation to develop a sunglint flag with the objective of masking affected data, hence minimising the probability of errors likely to occur when using correction models.Wernand (2002) described a meteorological quality flagging method to optimise automated hyperspectral measurements for coastal and shelf seas.To minimise sunglint he suggests the use of two optical sensors looking in different azimuthal directions to measure water surface leaving radiance and utilises the lowest water surface leaving radiance for each measurement, so as to select the spectrum with the least sunglint.
While this setup is useful and minimises sunglint effect on measurements, it requires additional sensors to be installed with the possible risk of sunglint still affecting both sensors.Introduction

Conclusions References
Tables Figures

Back Close
Full This study presents a set of four quality control flags for automated and unmanned above-water hyperspectral measurements, a flag to mask sunglint is generated here to compliment meteorological flags reported by Wernand (2002).These hyperspectral measurements are supplemented by simultaneous sky and sea surface images from a camera system.Parallel to the image and spectrum analysis, optimum azimuthal zones for sensor measurement were assessed with the goal to identify sunglint affected and non affected zones.The sea surface images act as "sea-truth" in the validation step of spectra and the Solar Position Algorithm, SPA (Reda and Andreas, 2004) generated estimate azimuthal angles of the sun during the field campaign.
In the next section, the measurement methodology will be explained followed by a presentation of the results and a discussion of the study findings.Finally, the new flagging method will be presented along with possible future developments.

Data and methods
Underway optical measurements in the North Sea, Scotland Sea, Irish Sea and Celtic Sea were performed aboard R/V Heincke cruise HE302 between 21 April and 14 May 2009.Ocean Data View (Schlitzer, 2010) was used to generate the study area map, Fig. 1, which shows the ship track for which above-water hyperspectral measurements were obtained.

Instrumentation
A RAMSES-ACC hyperspectral cosine irradiance meter (TriOS, Germany) was used to measure incoming solar radiation, E S (λ) and two RAMSES-ARC hyperspectral radiance meters with a field-of-view of 7 • in air, were used to detect respectively the sea surface radiance L sfc (θ sfc , Φ, λ) and sky radiance L sky (θ sky , Φ, λ).A frame (see Fig. 2) designed to hold the irradiance sensor facing upwards, the sky and sea surface radiance sensors at zenith angels θ sfc = 45 • and θ sky = 135 • , was fixed to the mast of Introduction

Conclusions References
Tables Figures

Back Close
Full the ship facing the starboard, 12 m above sea surface.These spectral measurements were automatically collected at 15 min intervals over a spectral range λ = 320-950 nm in steps of 5 nm.Spectrum classification, visible (VIS), λ = 320-700 nm, and near infrared (NIR), λ = 700-950 nm, was implemented.A DualDome D12 (Mobotix, Germany) camera system with customised lense objective L43 with a field-of-view of 45 • , was used to capture simultaneous images of the sky and sea surface during hyperspectral measurements as illustrated in Fig. 2. The positioning height of the camera system and optical sensors proved to be unaffected by sea spray.The ship's position and heading were recorded using a Differential Global Position System (DGPS) and sampling times were logged in UTC.

Conclusions References
Tables Figures

Back Close
Full optical measurements used and computed in Eq. ( 1) are available on PANGAEA (Garaba et al., 2010a, b, c, d).

Sunglint flag
The derived L W and R RS from Eq. ( 1) were utilised in the testing and evaluation of the new sunglint flag intended to mask measurements affected by sunglint.Parallel to this investigation, the SPA and DGPS data were utilised along with Eq. ( 2) as complimentary tools to derive the relative angle of the optical sensors to the sun's azimuth position, Φ derived ; where Φ sun is the sun's azimuthal position, Φ ship ship's azimuthal heading, and the sensor azimuthal heading Φ sensor = 85 • .In order to maintain the circular nature of angle 0 • ≤ Φ derived ≤ 360 • , the modulo arithmethic was applied.A test was then performed on Φ derived , using corresponding sea surface images as "sea truth", to identify the number of optical measurements consistent with 90 • ≤ Φ ≤ 135 • (Fougnie et al., 1999;Mueller et al., 2003;Deschamps et al., 2004).This test was also aimed to answer: is it possible to obtain useful optical measurement from an automated and unmanned seaborne platform for 0 • ≤ Φ derived < 90 • and 135 The sunglint flag was developed on the premise that water absorbs light in the NIR, but reflects light due to a wind roughened sea surface and/or scattering influenced by optically active water constituents.In Fig. 3  mean spectra for the two sets over the whole spectrum range (λ = 320-950 nm), to obtain a general overview on typical spectra for set Nns and Ns, 3. The goal of this step was to eliminate spectra with images in Ns while conversely retaining as much as possible spectra in Nns.Hence utilising the findings from step 2, i.e. spectrum shape, behaviour and magnitude variations of both L W and R RS in the VIS and NIR, a combination of inequality equations and band ratios were implemented to assess and generate the sunglint flag; -first evaluation test used the NIR mean values for set Nns and Ns to obtain threshold values specific to each set, this test was then repeated using the minimum values in the NIR, -next test involved using the classical band ratioing based on dominant spectral bands both in the VIS and NIR i.e. λ = 400 nm, 460 nm, 760 nm, 940 nm (Halthore et al., 1997;Wernand, 2002;Kay et al., 2009).
4. In the last step the sunglint flag was produced after quality and quantity control check from step 3.

Results and discussion
In the previous section the methods used in spectra measurement during the field campaign were explained and in this section the new proposed quality control will be presented along with other findings.A total of 629 spectra measurements satisfied the test conditions of the meteorological flagging.Sea surface image assessment of these corresponding to the unmasked 629 spectra, revealed 501 images free of sunglint, Nns and 128 images affected by sunglint, Ns.Introduction

Conclusions References
Tables Figures

Back Close
Full

Sunglint image analysis
Automated and unmanned optical measurements from a seaborne platform are challenging.It is difficult to adhere to recommended sensor setups for θ sfc , θ sky , Φ; thus inevitable to collect erroneous measurements affected by sunglint, whitecap and foam (Gordon, 1985;Moore et al., 1998;Fougnie et al., 1999;Hooker and Morel, 2003;Kay et al., 2009).In this study, underway measurements were made aboard R/V Heincke on the track illustrated in Fig. 1. Figure 4 demonstrates typical sources of error, revealed by the image inspection in step 2 of Fig. 3.The image inspection was performed by two investigators with an additional referee.Main sources of contamination were identified as sunglint, whitecaps and foam, the latter two resulting in similar sunlight influenced spectral patterns.

Sunglint flag
The assessment of the spectra from the sample sets Nns and Ns indicated that R RS (NIR) and L W (NIR) were significantly higher in the presence of sunglint.L W values over the whole measured spectrum, λ = 320-950 nm, were higher for the set Ns relative to set Nns in both VIS and NIR ranges.R RS values were enhanced for the set Ns compared to set Nns.In Fig. 5 normalised sample mean spectra shapes for Nns (501 spectra) and Ns (128 spectra) are presented to illustrate the aforesaid findings.Data normalisation was applied to maintain the spectral shape while simplifying comparison of spectra, dividing each L W and R RS measurement by the maximum value for each measurement.However, for determining the new flag, the actual computed spectral measurements L W and R RS were used.
The mean Ns spectrum in Fig. 5 shows normL W and normR RS values elevated by sunglint compared to mean Nns spectrum.To further investigate and understand sunglint spectral characteristics, 13 of the 128 sea surface images in Ns were identified to be completely affected by sunglint and in Fig. 6 their spectral shapes reveal the same trend shown in Fig. 5.The findings confirm that in the presence of sunglint,

Conclusions References
Tables Figures

Back Close
Full sea water will reflect light due to a wind roughened sea surface, evident in the sea surface images, and/or multiple scattering by optically active water constituents.This was a unique characteristic in both L W and R RS for optical measurements from this study which aided in developing the new sunglint flag.Spectral band ratioing, a classical method in quantitative colour of seawater interpretation that is known to reveal diversity in R RS shapes of seawater surface features (Lee and Carder, 2000), was implemented to help reveal the differences in Nns and Ns.
To summarise and evaluate the tests implemented in obtaining the best sunglint flag/mask Table 1 was generated.A ranking was introduced so as to order the methods according to their performance when eliminating or selecting the most sunglint affected measurements.
The methods in Table 1 provide an easy and simplified evaluation procedure because they include basic mathematical operations; division and inequalities.The threshold criteria was, chosen because it is a widely used technique in developing flagging and validation algorithms e.g.(Lavender et al., 2005), performed by iterative testing, i.e. first using mean values as threshold for sets Nns and Ns and then adjusting to obtain better performance.Adjusting these threshold values was aimed at; (i) masking/eliminating as many measurements in the sunglint affected set Ns, and (ii) unmasking/keeping as many measurements in the sunglint free set Nns.The performance test summarized in Table 1, revealed the best sunglint flag conditions; "mean (L W ) NIR < 2 mW m −2 nm −1 Sr −1 ", at least 90% effective in sets Nns (97.21%) and Ns (91.41%) but falls short as 14 valid spectra (2.79% of Nns) were Introduction

Conclusions References
Tables Figures

Back Close
Full masked and 11 erroneous spectra (8.59% of Ns) kept; or alternatively "minimum (R RS ) NIR < 0.010 Sr −1 ", which despite the relatively low performance in set retaining measurements in Nns (94.01%) it however reduces the number of erroneous spectra kept after validation to 7 spectra (5.47% of Ns).

Remote sensing reflectance
In the previous section, a sunglint flag was generated which can be implemented using L W or R RS .In Table 2, the meteorological flagging (Wernand, 2002) conditions and sunglint flag conditions are conjointly presented.The first three flags rely on the incoming radiation, E s , thus masking measurements taken during dusk or before significant incoming solar radiation can be detected (Flag 1), dawn (Flag 2), or under rainfall (Flag 3).Equation ( 1) is then implemented to derive water leaving radiance, L W and remote sensing reflectance, R RS , followed by the sunglint flag validation.In this study the sunglint flag was implemented using L W (Flag 4a) because it is the first product of Eq. ( 1) but can be replaced by R RS (Flag 4b) as summarised in Table 2. Proposing Flag 4a or Flag 4b was aimed at not limiting flagging to one remote sensing product e.g.L W or R RS only.This proposed quality control method (meteorological and sunglint flagging) aims to eliminate erroneous spectral measurements collected from an unmanned and automated platform.When all the flags were applied to the available measurements, Flag 1-Flag 4a, a subset was selected to illustrate typical spectra for case 2 waters along the cruise track (Fig. 7).

Relative azimuth angle between the sun and the sensor
In the previous section, sunglint flags were developed using L W and R RS .Using SPA and sea surfaces images as complimentary tools, an additional evaluation based on spectra kept after running the flagging proposed in Table 2 and the matching Φ derived was performed.The main goal was to answer these two questions; how many of Introduction

Conclusions References
Tables Figures

Back Close
Full the optical measurements of this study conform to the recommended 90 • ≤ Φ ≤ 135 • ?At which Φ can valid and useful measurements be collected from an unmanned and automated platform?According to the available measurements, valid spectra can be collected at 0 • ≤ Φ ≤ 360 • .However, the validity of the spectra is dependent on the accuracy of air-sea interface reflection coefficient estimate.Table 3 summarises these findings, it is shown that ∼85% of the optical measurements at the recommended 90 In the other regions 0 • ≤ Φ derived < 90 • and 135 • < Φ derived ≤ 360 • valid measurements were also collected.This can be attributed to the R/V "pitch, roll and yaw", also valid for ferries equipped with radiometers, motions causing fluctuations in optical sensor setup; azimuth and zenith angle changes (Mishra and Nath, 1999;Aas, 2010).The dynamic changes in the optical sensor setup on an unmanned and automated seaborne platform influence the collection of valid and invalid measurements as shown in Table 3.

Conclusions
Hyperspectral optical measurements from an unmanned and automated seaborne platform are often affected by meteorological conditions and sunglint.To overcome these disturbances, the new proposed sunglint flag, masks spectra affected by sunglint based on either L W or R RS , and the meteorological flags (Wernand, 2002), mask spectra affected by rainfall, dusk and dawn based on E S to mitigate collectable erroneous measurements.A camera system, to obtain sea surface images (evidence for sunglint, whitecaps, foam) and sky images (evidence for cloud cover or overhead sun), and the SPA to estimate the azimuthal angle of the sun relative to optical sensors, can both be used as complimentary tools.Optical measurements collected as proposed in this study, e.g. the planned permanent installation of this system on R/V Heincke, will provide test datasets for further validation of the meteorological and sunglint flagging for different water bodies.

OSD Introduction Conclusions References
Tables Figures

Back Close
Full In this study it is assumed sunglint was the main cause of error in the collected measurements after applying the meteorological flagging (Wernand, 2002).However, the influence of whitecaps and foam has been reported to cause both; (i) a decrease in reflectance in the NIR due to radiation absorption by large air bubbles (Whitlock et al., 1982;Frouin et al., 1996), or physical coolness of residual foam (Marmorino and Smith, 2005), (ii) enhanced reflectance occurring as soon as waves break generating thick strong reflecting foam (Moore et al., 1998).Image analysis revealed that in sea surface images influenced by whitecaps/foam, sunglint was also present and it was not possible to distinguish how each contribute to the spectra.It was therefore assumed that for the available measurements sunglint and whitecaps/foam led to erroneous measurements.
It is therefore important that in future studies the contribution of whitecaps and foam be specifically investigated with respect to sunglint flagging.Automated and unmanned above-water optical measurements based on recommend optical sensor setup for example the reported 90 • ≤ Φ ≤ 135 • (Mueller et al., 2003) cannot be completely achieved and does not guarantee valid measurements, also reported by Aas (2010).In this study it has been shown that valid measurements can be obtained for 0 • ≤ Φ ≤ 360 • from an unmanned seaborne platform.It is recommended that prior to the automated and unmanned optical measurements the SPA be utilised to predict or identify optimal Φ despite the limiting factors of R/V "pitch, roll and yaw" motions.Here the idea is not to necessarily change the cruise track but to perfom underway measurements avoiding sunglint.Alternatives approaches to minimise sunglint, but increasing technical requirements, would be (a) to turn the sensors by some automatic device or (b) to have more sensors looking at different azimuthal directions.However for fixed platforms, such as piles, these altenative approaches can be avoided if SPA is utilised beforehand to limit sunglint influenced measurements.Furthermore, investigations in coastal and turbid waters using hyperspectral radiometers are still needed to fully understand the sunglint spectral signatures both in the VIS and NIR ranges.

OSD Introduction
Full    The "minimal flag" sets the lower limit for which significant incoming solar radiation can be measured.
Flag 2 The "shape flag" will mask optical measurements influenced by dusk "red colouring of the sky" or dawn radiation.
E S (470 nm)/E s (680 nm) > 1 Flag 3 The "rainfall flag" will mask optical measurements influenced by precipitation or high humidity.

Flag 4a
The "sunglint flag" will mask optical measurements influence by sunglint based on L W .
Mean L W (700-950 nm) < 2 mW m −2 nm −1 Sr −1  These observations were obtained after filtering with the meteorological masks (Wernand, 2002).Full Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | a simplified activity diagram illustrates the steps that were implemented in this sunglint flag investigation and evaluation; 1.Sea surface images were retrieved, matching the unmasked spectra validated with the meteorological flagging (Wernand, 2002), visually inspected and classified into; Nns -image set without sunglint or Ns -image set with sunglint, 2. Analysis of the two sets Nns and Ns was now focussed on identifying unique characteristics using L W and R RS .It involved looking at individual spectra and Introduction Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Flag
Discussion Paper | Discussion Paper | Discussion Paper | Fig. 1.R/V Heincke HE302 cruise track were above-water hyperspectral optical measurements were performed between 21 April and 14 May 2009.The blue line represents the track were optical measurements were collected, and the red line is for the return cruise to Bremerhaven without measurements.The annotated sites are; the German Bight (GB), the Central North Sea (CNS), Atlantic inflow into North Sea (ANS), Irish Sea (IRS), and the Celtic Sea-St.George's Channel (CS-SGC).

Table 1 .
A summarized evaluation of the best sunglint flag conditions ranked according to their performance in masking sunglint spectra.To check for effectiveness, the percentage E (Ns) % and E (Nns) % was derived by dividing the number of spectra masked or unmasked by each condition with the actually number of spectra in sets, sunglint set Ns -128 spectra and non sunglint set Nns -501 spectra.NIR refers to λ = 700-950 nm.

Table 2 .
(Wernand, 2002)orological and sunglint flag conditions.The meteorological(Wernand, 2002)flags are represented by Flag 1-3.The sunglint flag can be implemented either as Flag 4a or Flag 4b depending on availability of measured water leaving radiance, L W or remote sensing reflectance R RS .For the flagging purposes of this study Flag 4a was implemented.

Table 3 .
The percentage of sunglint free observations in relation to the azimuthal angle Φ derived .
derived Number of observations Number of sunglint free % collected at Φ derived observations at Φ derived 0 • < Φ derived ≤ 45 •