3.1.

Data Coming from Laboratory Measurements

Transmission measurements were performed with a PerkinElmer spectrophotometer, taking into account the reflection at each interface.¹² Each product was poured into a quartz cell with a thickness that was adapted to the product. These measurements enabled us to derive the attenuation coefficient and then from the attenuation coefficient, a theoretical minimal thickness that induces a spectral attenuation effect on reflectance between oil and water and a theoretical maximal thickness beyond which spectral reflectance no longer evolves with thickness.

Concerning reflectance, measurements were performed with the ONERA spectro-goniometer dedicated to large samples (“Banc de BRDF Grands Echantillons”) as shown Fig. 1. It covers the 0.35- to $2.5\text{-}\mu \mathrm{m}$ spectral range with a 3-nm spectral resolution in the VNIR (0.35 to $1\mu \mathrm{m}$) and 10 nm in the SWIR (1 to $2.5\mu \mathrm{m}$) spectral ranges from an ASD Fieldspec spectroradiometer. The zenith angles of the source (SZA) and of the spectrometer collection optics (VZA) can vary between 0 deg and 60 deg with an accuracy of 1 deg. The azimuthal angle of the spectrometer collection optics (VAA) can vary between 0 deg and 180 deg with an accuracy of 1 deg. Thus, the setup enables the recording of a spectral radiance for a sequence of angular configurations over nearly a full quarter of a sphere. However, due to the occultation of the source by the collection optics, measurements are not possible in the backscattering direction $\pm 10\mathrm{deg}$. Our measurements are limited to nadir ($\mathrm{VZA}=0\mathrm{deg}$). For most of them, SZA varies from 10 deg to 50 deg with a step of 10 deg for an arm angle equal to 180 deg corresponding to $\mathrm{VAA}=180\mathrm{deg}$. For a small set of products, in italics in Table 1, measurements were performed for only one angular configuration: $\mathrm{SZA}=40\mathrm{deg}$, $\mathrm{VZA}=0\mathrm{deg}$, and $\mathrm{VAA}=180\mathrm{deg}$ for the arm with the collection optics. Measurements were performed for various thicknesses ranging from 4 to 10 mm: five thicknesses (4, 5.5, 7, 8.5, and 10 mm) for the small set of products in italics in Table 1 and three thicknesses (5, 7.5, and 10 mm) for the others.

Fig. 1

ONERA spectro-goniometer dedicated to large samples (Banc de BRDF Grands Echantillons).

For non-opaque and non-scattering products, a model of the reflectance measurement was established according to Fig. 2 and led to Eq. (1). This model, more complete than in Ref. 12, enables artificial modification of the background and thickness and therefore generation of adequate signatures.

Fig. 2

Ray path used for reflectance measurement modeling for non-opaque and non-scattering products.

Fig. 3

Comparison of computed reflectance for sample thickness from 4 to 12 mm (various colors) and measured (bold black) laboratory reflectance for diesel: approximative expected diesel thickness (a) 5 mm and (b) 10 mm.

A validation of the model and of the measurements was done for some products. The attenuation coefficient value deduced from transmission measurements was used in Eq. (1). Various constant values for the product optical index were used to compute the $T$ terms. As the actual thickness was only roughly known (measurement with a simple ruler), the computation was performed for several thickness values. The computed reflectances (colored lines) were compared with the measured one (bold black line) as shown in Fig. 3 for diesel. The expected diesel thickness (according to the ruler) is 5 mm in Fig. 3(a) and 10 mm in Fig. 3(b). This is in agreement with the model for a 4-mm and 8-mm thickness, respectively (in the SWIR region more sensitive to the thickness). This shows a fairly good agreement between model (including inputs) and measurements.

3.2.

Data Coming from Tank Filled with Sea Water

For the measurements, a small part of a large tank ($1900{\mathrm{m}}^2$) located in CEDRE (Center for Research and Experimentation for Accidental Water Pollution) at Brest, France, fairly deep (3 m) and filled with sea water, was used to generate successive small slicks. For each product, a known volume was poured into a frame floating on the tank as in Fig. 4(a). Two hyperspectral cameras fixed about 15 m above the frame imaged each slick as shown in Fig. 4(b). The features of the cameras are given in Table 2. The water surface in the frame was cleaned between each product. The cameras also imaged the Spectralon^© reference panels set on the quay. Computing the radiance ratio of pixels from the slick to pixels from the reference panel and multiplying this ratio by the reflectance of the reference panel provides spectral signatures. The thickness assessment is described in Refs. 15 and 16. It relies on slick detection using indices. A mean thickness is computed by dividing the volume poured in the frame by the surface of the slick. In the case of NOFO2015 mixed slick, with a thin and a thick part, the assessment is refined. From the thickness assessment of several products, a common mean thickness of $25\mu \mathrm{m}$ has been found for light crude oil (HC1541), gasoil, and unleaded gas 98. The VNIR image [Fig. 5(a)] enables the detection of the whole slick, and the SWIR image [Fig. 5(b)] enables the detection of the thicker part only. The difference gives the surface covered by the thin part. Assuming a thickness of $25\mu \mathrm{m}$ for the thinner part, the volume of this thinner part is then computed. This thinner part volume is subtracted from the total volume poured into the frame to assess the volume corresponding to the thicker part of the slick. The mean thickness of the thicker part is then assessed by dividing the volume of this part with the corresponding surface. It leads to 1.1 mm for the NOFO2015 emulsion thick part.

Fig. 4

Tank experiment: (a) pouring a known volume of product (rapeseed oil) and (b) imaging the slick with two hyperspectral cameras at the top of a cherry picker.

Table 2

Features of the two hyperspectral pushbroom cameras.

Fig. 5

NOFO2015 emulsion in the tank: (a) VNIR image and (b) SWIR image.

3.3.

Data Coming from Sea

Hyperspectral imaging of marine slicks was performed in 2015 during a campaign organized by NOFO, a Norwegian entity (Norwegian Clean Seas Association for Operating Companies). The campaign objective was to test devices to clean a slick. The ONERA imaging system named SETHI onboard a FALCON20 aircraft was allowed to image the slicks during the recovery exercises.¹⁷ The optical images were acquired with the same cameras as used for the tank experiment. Their features are given in Table 2.

During the experiment corresponding to results presented in Sec. 5, the sea state was 3 on the Douglas scale.

3.4.

Extended Database

All of the data presented previously constitutes the extended database. It contains not only spectral signatures (laboratory and/or tank and/or at sea spectral reflectance) but also the attenuation coefficient, corresponding theoretical minimal and maximal thicknesses from an optical point of view, and thickness derived from the tank experiment when available. The current content regarding thicknesses is illustrated by Fig. 6.

Fig. 6

Illustration of the extended database content for thicknesses.

In this figure, the att_MIN and att_MAX thicknesses correspond to theoretical optical boundaries computed by the attenuation coefficient. For the att_MIN, the value is taken in the SWIR domain, around $1.7\mu \mathrm{m}$. For the att_MAX, it is taken either in SWIR (for UN1267, TBTS, NOFO2015 emulsion, HC1541, DALIA, USAN which includes all crude oils) or in VNIR (for the others) depending on the spectral shape of the attenuation coefficient and its availability (some of our measurements, for very opaque oils, were not reliable for wavelengths $<0.7\mu \mathrm{m}$). Considering only the VNIR domain would probably lead to much smaller att_MAX values for all crude oils in agreement with values from Ref. 10. The att_MAX thickness corresponds to the theoretical skin depth and represents the largest thickness that enables ray crossing of the full layer above the water, forward and backward. For thicknesses larger than att_MAX, the deepest part of the layer does not contribute to the measured reflectance; therefore reflectance cannot be inverted with an optical model to restitute the thickness. The att_MIN value represents the theoretical smallest thickness that induces a detectable spectral attenuation for ray crossing the full layer forward and backward. It is noteworthy that slicks with smaller thicknesses than att_MIN may be detectable due to other phenomena (index difference between product and water and possible interferences).

The thicknesses observed in our tank experiment correspond to the blue triangles. The experimental thickness is in the att_MIN-att_MAX range for only two cases (Colza and 15W40). This means that, for these two cases, the spectral reflectance and optical model looking at attenuation could theoretically enable one to assess thickness. For TBTS and emulsionNOFO2015 cases, the actual thickness is above the att_MAX; therefore thickness cannot be derived from an optical model. For the other cases (HC1541, Gasoil, SP98, Tournesol, and Plantoil), the actual thickness is smaller than the att_MIN, and thickness cannot be derived from attenuation terms of an optical model. In several cases, for highly transparent oils, such as unleaded gas, the att_MAX values are very high and, of course, the actual thickness is expected to be much lower than this theoretical boundary. The whole content is summarized in Table 3.

Table 3

Overview of the extended database content. L corresponds to laboratory measurement, T to tank experiment measurements, and S to measurements at sea. As in Table 1, the italic L corresponds to the case with only one angular configuration. Y stands for yes and indicates that information is present. N stands for no and indicates that information is not present.

4.1.

Detection

Detection relies on spectral indices that correspond to spectral band ratio or linear combination. The definitions of the indices used in this paper are provided in Table 4. An index image is computed for each case. The corresponding histogram is used to separate water and oil.

Table 4

Definition of spectral indices. ρ stands for reflectance, and numbers correspond to wavelength expressed in micrometer.

4.2.

Identification

Product identification is performed from spectral libraries. Spectral reflectances collected in the laboratory constitute the lab spectral library, and the ones collected during the tank experiment constitute the tank spectral library.

For the non-opaque and non-scattering products, the laboratory model was used to replace the background used for the laboratory measurements with the water of the tank.

Three spectral distances were selected to compare the image with the libraries: Spectral Information Divergence (SID),²³ Spectral Angle Mapper (SAM),²³ and CHI2. A threshold can be given by the user to limit the classification to really similar spectra.

4.3.

Quantification

For identified products, volume assessment can be done due to thickness being deduced from signature adjustment (by adjusting the exponential terms as explained in Sec. 3.1) or using the thickness stored in the database. Knowing the pixel size for the image, the surface covered by the product is simply computed as the product of the number of pixels found by the detection tool with the pixel size. This surface is multiplied by the thickness found by signature adjustment or associated with the product in the database:

5.1.

Detection

5.1.1.

Tank experiment case

Figure 7 shows the results for the various detection indices and automatic thresholding in the case of NOFO2015 emulsion in the tank. The three VNIR indices enable the detection of the whole slick, whereas the SWIR indices enable the detection of its thick part. For VNIR indices, the automatic thresholding can also be applied to the slick itself to distinguish its thin and thick parts.

Fig. 7

Comparison of detection results for various indices for NOFO2015 emulsion in the tank.

5.1.2.

Sea case

The detection result for an airborne image is presented in Fig. 8 for VNIR indices and in Fig. 9 for SWIR ones. In these figures, the upper part corresponds to the indices images, and just below are the corresponding detection masks generated with automatic thresholding on the indices images. The lower part corresponds to the histograms of indices images with the threshold found. Only FI enables the detection of the whole slick for the marine tank. RAI and nFI seem to detect only the thicker emulsified part. For the SWIR indices, they seem to detect the very thick part of the slick and the automatic thresholding reinforces the underdetection. The automatic thresholding on the FI index has been applied to the slick itself to distinguish its thin and thick parts.

Fig. 8

Comparison of detection results for various VNIR indices for NOFO2015 emulsion at sea.

Fig. 9

Comparison of detection results for various SWIR indices for NOFO2015 emulsion at sea.

5.2.

Identification

5.2.1.

Tank experiment case

Figure 10 shows the results for the identification step in the case of NOFO2015 emulsion in the tank using, respectively, the tank library for the upper part and the lab library for the lower part of the figure. In the lab library case, SAM, CHI2, and SID enable the identification of the thicker part of the slick. In the tank library case, a signature for the thinner part of the slick is included, which enables the detection of this part, but for the thicker part, the identification seems to be less efficient than for the lab library. For some pixels, there is confusion with sunflower oil. Figure 11 illustrates the content of each library.

Fig. 10

Comparison of identification results for various spectral distances for NOFO2015 emulsion in the tank using the tank library (upper part) and the lab library (lower part).

Fig. 11

Tank and lab libraries used for results presented in Fig. 10.

5.3.

Quantification

Once the product is identified, the volume can be computed. In the NOFO2015 emulsion case, as explained in Ref. 15, light cannot reach the bottom of the emulsion layer and go back to the sensor. Therefore, the thickness cannot be determined only by the analysis of the in situ spectral reflectance with an optical model (cf Sec. 3.1). The mean thickness observed in the tank is taken as the best estimate of the mean thickness of the slick at sea. According to the tank experiment, the mean thickness for the thinner part of the slick is assumed to be equal to $25\mu \mathrm{m}$ and it leads to 1.1 mm for its thicker part. The corresponding result for the airborne image of the NOFO2015 campaign is presented in Table 5.

Table 5

Slick surface and volume for the NOFO2015 airborne image.

6.1.

Database Content Representativeness

6.2.

Comparison with Other Volume Assessment

A method for assessing volume of oil on water relying on its visual appearance is well known in the offshore community. It corresponds to a code, named Bonn Agreement Oil Appearance Code (BAOAC),²⁷ linking the visual appearance of the oil slick to a thickness range. Although this code is not adequate for emulsion, a volume assessment has been made relying on image photointerpretation.¹⁶ The corresponding results are reported in Table 7. Class 3 corresponds to code 3 when the surface appears “metallic” and thickness is in the range 5 to $50\mu \mathrm{m}$. Class 4 corresponds to code 4 when the surface appears with discontinuous true color and its thickness is between 50 and $200\mu \mathrm{m}$. Class 5 corresponds to emulsion and thus does not correspond to code 5 when the surface appears with continuous true color, but the code 5 lower limit of $200\mu \mathrm{m}$ is taken into account in the volume assessment.

Table 7

Slick surfaces and volumes for the NOFO2015 airborne image.

The slick surface found using photointerpretation is larger than the one found using the index. The sum of class 3 and class 4 surfaces is much larger than the surface corresponding to the thin part of the slick for the index. The surface for class 5 and the thick part of the slick are of the same order of magnitude. As no thickness range is available for class 5, its contribution to the total volume is probably highly underestimated.

There is a need to connect the thin part of the slick to class 3 and class 4, to consolidate the thickness for the thin part(s), and to update the thickness for the thick part if necessary. According to the tank experiment, taking 1.1 mm as the approximate thickness for the thick part of the slick implies the selection of 0.025 mm for its thin part. This is why these values appear with the font (italics) in Table 7 and the corresponding volumes are added to estimate the approximate volume. The thickness upper bound for the thin part of the slick is just taken as the mean thickness of class 4, which also corresponds to a thickness that has been detected using the index.

The translation of the Bonn code thickness ranges into mean thickness that was made by Hu et al.²⁸ cannot be used in or adapted to the NOFO2015 emulsion case. Indeed, the correspondence between reflectance and thickness is, at least, product specific. In the thickness range obtained during our laboratory measurements, the NOFO2015 emulsion reflectance remains the same. Therefore, in contrast to Hu et al.,²⁸ it is not possible to establish a relation between reflectance and thickness by referring to laboratory measurements and transferring it to an image. The cumulative histograms have been computed as by Hu et al.²⁸ to check their shape and the possible identification of classes from inflection points. They are presented in Fig. 12 for the two VNIR wavelengths (470 and 670 nm) used for the FI index and for the two SWIR wavelengths (1240 and 1640 nm) used by Hu et al.²⁸ The histogram obtained for 470 nm would suggest two classes for the emulsion. It is difficult to identify classes for the other wavelengths.

Fig. 12

Cumulative histogram for the VNIR wavelengths included in the FI index and the two SWIR wavelengths used in Ref. 28.

Despite the uncertainty about the thickness of the thinner part of the slick and the effect of the wind and the Langmuir circulation, the laboratory and tank measurements enable one to provide a lower bound and a possible value for the slick volume, whereas the Bonn code only provides a range for the lower bound. They provide a minimum and a possible volume for the thicker part of the slick, which is usually the most important part to collect.