Oil Spill Detection and Mapping Using Sentinel 2 Imagery - MDPI

Journal of

Marine Science and Engineering Article

Oil Spill Detection and Mapping Using Sentinel 2 Imagery Polychronis Kolokoussis *

ID

and Vassilia Karathanassi

Remote Sensing Laboratory, National Technical University of Athens, 15780 Athens, Greece; [email protected] * Correspondence: [email protected]; Tel.: +30-210-772-2599 Received: 4 December 2017; Accepted: 3 January 2018; Published: 6 January 2018

Abstract: Two object-based image analysis methods are developed for detecting oil spills from known natural outflows as well as light oil spill events using Sentinel 2 imagery. The methods are applied to Sentinel 2 images of a known area of natural oil outflow as well as on a Sentinel 2 image of a recent oil spill event along the south coast of Athens, Greece. The preliminary results are considered very successful and consistent, with a high degree of applicability to other Sentinel 2 satellite images. Further testing and fine tuning of the proposed object-based methodology should be carried out using atmospheric correction and ground truth. Keywords: oil spill; Sentinel-2; OBIA

1. Introduction Several times multispectral imagery has been used for the detection of oil spills. Although clouds affect the temporal resolution of such images, the potential of optical sensors for detecting oil pollution phenomena is considerable. An oil spill’s spectral signature depends on favorable lighting and sea conditions, oil optical properties, film thickness, and the optical properties of the water column. Based on spectro-radiometric measurements, Brown et al. [1] observed that oil has no specific characteristics that distinguish it from the background. Klemas [2], in his study, states that oil sheen shows up as silvery and reflects light over a wide spectral region. Heavy oil appears brown, peaking in the 600 to 700 nm region, while mousse looks red-brown and peaks closer to 700 nm. A study of oil spectra in shallow seawater showed that for crude oils and light petroleum product, a thin oil spill has similar spectral characteristics to background spectra [3]. In the case of open ocean images [4,5], the organic compounds in the oil and oil/water mixtures have absorption features that are distinct from those from water and clouds. Simulations that have been performed by varying the water optical properties and considering different crude and refined oils provided some general rules [6]: very thin oil films (sheens) are more easily detected at viewing directions near the sun-glint zone, while very thick films are more likely to be detected at viewing angles away from the sun. For films of intermediate thickness, the detectability depends mainly on the oil’s specific optical properties. Using Landsat 7 Enhanced Thematic Mapper (ETM) imagery, Zhao et al. [7] observed features of weathered oil suspended in the water column with bright contrast, and oil film floating on the sea surface with dark contrast. Lee et al. [8] developed an algorithm for discriminating film-like oil from thick oil using DubaiSat-2 and Landsat Operational Land Imager (OLI) data. Taravat et al. [9] performed ratio operations on Landsat 7 ETM images to enhance oil spill features. They concluded that the bands’ difference between 660 and 560 nm, division at 660 and 560, and division at 825 and 560 nm, normalized by 480 nm, provide the best result. Then these ratio images fed the input nodes of a multilayer perceptron neural network in order to perform a pixel-based supervised classification. Bradford et al. [10] performed selective color band combinations, contrast enhancement, and histogram

J. Mar. Sci. Eng. 2018, 6, 4; doi:10.3390/jmse6010004

www.mdpi.com/journal/jmse

J. Mar. Sci. Eng. 2018, 6, 4

2 of 12

warping before applying an image segmentation process that separates out contiguous regions of oil and provides a raster mask to an analyst. Corucci et al. [11] exploited many different machine learning techniques such as simple statistical classifiers, radial basis function and multilayer perceptron neural networks, adaptive neuro-fuzzy inference system (ANFIS), and support vector machine (SVM) in order to investigate the potential of oil spill classification from optical satellite images. Object-based image analysis (OBIA) has often been applied, mainly to synthetic aperture radar (SAR) imagery, for oil spill detection. The object-based process for typical SAR-based oil spill detection consists of three steps: image segmentation for dark-spot identification, feature extraction, and oil spill and look-alike discrimination. Segmentation techniques are used to create objects that exhibit high local contrast and thus may be considered dark spots [12]. Then fine designed and appropriately combined gray, geometric, and textural features, as well as fuzzy logic, are employed to classify dark spots as oil spill or seawater. Although there are many recent studies [13–15] that profit from object-based SAR image analysis to perform oil spill detection, studies employing object-based analysis on multispectral image are rarely found in the literature. There are, however, certain limitations to detecting oil spills in SAR images as oil spill detection relies on the fact that the oil film decreases the backscattering of the sea surface, resulting in a dark formation that contrasts with the brightness of the surrounding spill-free sea. For C band SAR images, like Sentinel 1a and 1B, a minimum wind field of 2–3 m/s creates sufficient brightness in the image and makes the oil film visible. On the other hand, when the wind speed is too high, it causes the spill to disappear [16,17]. Thus, an alternative method for detecting oil spills with other types of data is considered very useful. Maianti et al. [18] applied OBIA on Moderate Resolution Imaging Spectroradiometer (MODIS) bands and MODIS-based indexes for monitoring oil spill dynamics. Kolokoussis et al. [19] developed an object-based method for oil spill detection using very high multispectral images such as Ikonos, QuickBird, RapidEye, and WorldView2, as well as high-resolution satellite images such as Landsat TM. Reiche et al. [20] created an object-based mapping and classification system for terrestrial oil spill pollutions in West Siberia using QuickBird data. Sentinel-2A and Sentinel-2B high-resolution multispectral data are available since June 2015 and March 2017, respectively. Nevertheless, their efficiency in oil spill detection has not yet been sufficiently investigated. In this study, motivated by the high-quality results obtained by the OBIA method presented in [19], we adapted the developed method for exploiting Sentinel-2 data characteristics. Oil spills resulting from the underwater natural oil outflow close to Zakynthos Island, Greece, and the oil spill caused by the “Agia Zoni II” tanker sunk in the Saronic Gulf, Greece, have been used for evaluation of the results. 2. Materials and Methods 2.1. Study Areas Two study areas (Figure 1) have been used in this work; the first (area A) is an area with a known natural oil outflow, 6 km south of Zakynthos Island in Greece. This area has been pointed out by [19] and is known to provide 15 km long natural oil spills during the summer period every year. The second area (area B) is the Saronic Gulf, east of Salamina Island and south of Athens. In this area an oil spill has been caused by the “Agia Zoni II” tanker sunk on 10 September 2017. As shown in Figure 1, the sinking of the “Agia Zoni II” tanker had a largeenvironmental impact, affecting most of the southern coastal area of Athens from Piraeus to Agios Kosmas (shown with red arrows in Figure 1). The 45-year-old vessel has been carrying 2500 tons of crude oil.

J. Mar. Sci. Eng. 2018, 6, 4 J. Mar. Sci. Eng. 2018, 6, 4

3 of 12 3 of 11

J. Mar. Sci. Eng. 2018, 6, 4

3 of 11

Figure 1. 1. Study area A, A, the the purple purple arrow arrow shows shows the the location location of of the the natural natural Figure Study areas areas A A and and B. B. Within Within area outflow. Within area B, the purple arrow shows the position where the “Agia Zoni II” tanker outflow. Within area B, the purple arrow shows the position where the “Agia Zoni II” tanker sunk, sunk, while the the red redarrows arrowsshow showthe thecoastal coastalareas areas affected spill (coordinate reference system of while affected byby thethe oiloil spill (coordinate reference system of the Figure 1. Study areas zone A and34B.North). Within area A, the purple arrow shows the location of the natural the map is WGS84/UTM map is WGS84/UTM zone 34 North).

2.2. 2.2.

outflow. Within area B, the purple arrow shows the position where the “Agia Zoni II” tanker sunk, while the red arrows show the coastal areas affected by the oil spill (coordinate reference system of Datasets Datasets the map is WGS84/UTM zone 34 North).

In this this work In work three three Sentinel Sentinel 2A 2A and and one one Sentinel Sentinel 2B 2B images images have have been been used used (Table (Table 1). 1). Furthermore, Furthermore, 2.2. Datasets two Sentinel 1A/1B images (Table 1) have initially been used for identifying the oil spill caused by by the the two Sentinel 1A/1B images (Table 1) have initially been used for identifying the oil spill caused “Agia Zoni II” tanker. Unfortunately, due to very low wind speeds in the days following the “Agia In this work three Sentinel 2A and oneto Sentinel 2B images been (Tablefollowing 1). Furthermore, “Agia Zoni II” tanker. Unfortunately, due very low wind have speeds inused the days the “Agia two Sentinel 1A/1B images (Table 1) have initially been used for identifying the oil spill caused by the 2). Zoni II” sinking, the oil spill could not be safely detected on the Sentinel 1A and 1B images (Figure Zoni II” sinking, the oil spill could not be safely detected on the Sentinel 1A and 1B images (Figure 2). “Agia Zoni II” tanker. Unfortunately, due to very low wind speeds in the days following the “Agia Zoni II” sinking, the oil spill could not be safely detected on theimages Sentinel 1A and 1B images (Figure 2). Table 1.List Table 1. List of of the the Sentinel Sentinel satellite satellite images used. used.

Geographic Area

Date Satellite Table 1.List ofSentinel the Sentinel satellite images used. Sentinel-2 Image Geographic Area Date Sentinel Satellite Sentinel-2 Image Zakynthos 7 August 2017 2A L1C_T34SDG_A011101_20170807T092725 Geographic Area Date Sentinel Satellite Sentinel-2 Image Zakynthos 7 August 2017 2A L1C_T34SDG_A011101_20170807T092725 Zakynthos 27 August 2017 2A L1C_T34SDG_A011387_20170827T092818 Zakynthos 7 August 2017 2A L1C_T34SDG_A011101_20170807T092725 Zakynthos August 2017 2A L1C_T34SDG_A011387_20170827T092818 Salamina 1327September 2017 2A L1C_T34SGH_A011630_20170913T091832 Zakynthos 27 August2017 2017 2A L1C_T34SDG_A011387_20170827T092818 Salamina 13 September 2A L1C_T34SGH_A011630_20170913T091832 Salamina 8 September 2017 2B L1C_T34SGH_A002650_20170908T091906 Salamina 13 September 2017 2A L1C_T34SGH_A011630_20170913T091832 Salamina 8 September 2017 2B L1C_T34SGH_A002650_20170908T091906 S1A_IW_GRDH_1WDV_20170915T042323 Salamina 8 September2017 2017 2B L1C_T34SGH_A002650_20170908T091906 Salamina 15 September 1A S1A_IW_GRDH_1WDV_20170915T042323 _20170915T042348_018381_01EF05_7D8E Salamina 15 September 2017 1A S1A_IW_GRDH_1WDV_20170915T042323 _20170915T042348_018381_01EF05_7D8E Salamina 15 September 2017 1A _20170915T042348_018381_01EF05_7D8E

(a)(a)

(b)(b)

Figure 2. Sentinel and imagesofofthe the Saronic Saronic Gulf: Sentinel 1B1B image of 14 September 2017; Figure 2. Sentinel Sentinel 1A 1A and 1B1B images Gulf:(a) (a) Figure 2. 1A and 1B images of the Saronic Gulf: (a) Sentinel Sentinel 1B image image of of 14 14 September September 2017; 2017; (b) Sentinel 1A image of 15 September 2017. Both images present large dark areas, caused by low (b) Sentinel Sentinel1A 1Aimage imageofof1515 September 2017. Both images present caused bywind low (b) September 2017. Both images present largelarge darkdark areas,areas, caused by low wind speeds, which cannot be associated with oil spills. wind speeds, be associated with oil spills. speeds, whichwhich cannotcannot be associated with oil spills.

For the object-based image analysis of the Sentinel 2 images, bands 1, 2, 3, 4, 8, 9, and 11 were For the object-based image analysis of theofSentinel 2 images, 1, 2,bands 3, 4, 8, 9,shown and 11 were used. The spatial and spectral characteristics the Sentinel 2A andbands 2Bimage in were For the object-based image analysis of the Sentinel 2 images, bands 1, 2, 3, 4, are 8, 9, and 11 used.Figure The spatial and spectral characteristics of the Sentinel 2A and 2Bimage bands are shown in 3. The images were preprocessed (import, resampling same resolution, Sentinel used. The spatial and spectral characteristics of the Sentinelto2A and 2Bimagesubset) bandsinare shown in Figure 3. The images were preprocessed (import, resampling to same resolution, subset)module in Sentinel Application Platform (SNAP) 5.0, but were not atmospherically corrected since the relevant

Application Platform (SNAP) 5.0, but were not atmospherically corrected since the relevant module

J. Mar. Sci. Eng. 2018, 6, 4

4 of 12

Figure 3. The images were preprocessed (import, resampling to same resolution, subset) in Sentinel J. Mar. Sci. Eng. 2018, 6, 4 4 of 11 Application Platform (SNAP) 5.0, but were not atmospherically corrected since the relevant module J. Mar. Sci. Eng. 2018, 6, 4 4 of 11 sen2cor sen2cor is not appropriate for water surfaces. stillexperimental experimental tools like ACOLITE and C2RCC is not appropriate for water surfaces.Other, Other, still tools like ACOLITE and C2RCC sen2cor is not appropriate for prerelease) water surfaces. Other, still experimental tools likeofACOLITE andsubsets C2RCC (the latter available in SNAP 6.0 will bebetried in thefuture. future. OBIA of appropriate subsets of (the latter available in SNAP 6.0 prerelease) will tried in the OBIA appropriate (the available in 2B SNAP 6.0 prerelease) willbeen be tried inout the future. OBIA of appropriate subsets of thelatter Sentinel 2A images (Figure 4) has carried out with with eCognition Developer 8.7. the Sentinel 2A and 2B and images (Figure 4) has been carried eCognition Developer 8.7. of the Sentinel 2A and 2B images (Figure 4) has been carried out with eCognition Developer 8.7.

Figure Thebands bands of 2A2A andand 2B. 2B. Figure 3. 3.The ofSentinel Sentinel Figure 3. The bands of Sentinel 2A and 2B.

(a) (a)

(b) (b)

(c)

(d)

(c) (d)2017; (b) Zakynthos 27th Figure 4. The Sentinel 2A subsets used for OBIA: (a) Zakynthos, 7th August Figure 4. The(c)Sentinel subsets used9th forSeptember OBIA: (a) 2017, Zakynthos, August 2017;and (b) (d) Zakynthos 27th August 2017; Saronic2A Gulf (Salamina) before 7th the oil spill event Saronic Gulf August (c)September Saronic (Salamina) September thecomposite oil spill and (d) Saronic Figure (Salamina) 4. The2017; Sentinel 2AGulf subsets (a) before Zakynthos, 7 event August 2017; (b)Gulf Zakynthos 13th 2017, afterused the9th oilfor spillOBIA: event.2017, Natural color 432(RGB). (Salamina) 13th September after the 9oilSeptember spill event. Natural color composite 432(RGB). 27 August 2017; (c) Saronic Gulf2017, (Salamina) 2017, before the oil spill event and (d) Saronic Gulf 2.3. The Developed Methodology (Salamina) 13 September 2017, after the oil spill event. Natural color composite 432(RGB).

2.3. The Developed Methodology Initial step for any OBIA methodology is the segmentation of the remote sensing images in order Initial step for any OBIA methodology is themerging segmentation of thein remote sensing images order to create image objects. Segmentation involves the pixels the image into imageinobject 2.3. The Developed Methodology to create image objects. Segmentation involves pixels in theand image into image object primitives called objects or segments withmerging certain the homogeneity geometric criteria. primitives objectsmethodology oris asegments with certain homogeneity and criteria.in order Initial step forcalled any OBIA thecreates segmentation of the remote sensing Multiresolution segmentation processisthat several levels with imagegeometric objects atimages various Multiresolution segmentation is a process that creates several levels with image objects at scales. Lower-level objects (sub-objects) smaller in whilein higher-level objects to create image objects. image Segmentation involvesare merging thesize pixels the imageimage intovarious image object scales. Lower-level image objects (sub-objects) are smaller in size while higher-level image objects (super-objects) are or larger. There is a topological restriction in and multiresolution segmentation that primitives called objects segments with certain homogeneity geometric criteria. Multiresolution (super-objects) are larger. There is a topological restriction in multiresolution segmentation that

segmentation is a process that creates several levels with image objects at various scales. Lower-level image objects (sub-objects) are smaller in size while higher-level image objects (super-objects) are larger. There is a topological restriction in multiresolution segmentation that smaller objects are contained but never intersect with larger objects. In eCognition the criteria used during multiresolution segmentation

J. Mar. Sci. Eng. 2018, 6, 4

5 of 12

are: (a) the scale, which defines the maximum size of the objects; (b) the color and shape weights (with sum to 1 constraint), which give priority to spectral homogeneity or to shape characteristics; and (c) the smoothness and compactness weights (with sum to 1 constraint), which have to do with the shape characteristics and force the segmentation to create regular shaped or compact objects. Additionally, the weights of the image layers (bands), which are taken into consideration during segmentation, can be adjusted appropriately and it is a common practice that, unless needed, the lower-resolution image bands are not used during segmentation in order to achieve better image object shapes. The basic spectral signature of an image object can be derived from the mean value of its pixels for every image band. Nevertheless, image objects have a large variety of useful features besides their spectral information; a few of them are listed later on since they have been used in this study (Tables 2 and 3). Standard deviation (StdDev) can be calculated within an image object for every band. Brightness difference of an object to its neighbor, sub- or super-objects, or even to the whole scene can be calculated for every image band as well. Any custom ratio (e.g., B2/B11) or index (e.g., NDVI, NDWI, etc.) can be calculated for an object based on its mean values at the relevant image bands. Image objects also have geometric features like area, border length, length, width, length/width, etc. Additionally, in OBIA class-related features (e.g., distance to a certain class) are also available and can be applied after an initial classification. Thus, the final classification of an image object can depend on the classes of its neighbor, sub- or super-objects. In order to take these class-related features into consideration, OBIA classification procedures are usually iterative processes. After several trials for the best segmentation of Sentinel 2 images for oil spill detection, it has been decided that the most appropriate procedure is the multiresolution segmentation of the images in two levels. The scope of the first level was to create small objects based mostly on spectral behavior, while the objects of the second level were much larger but as compact as possible in order to create uniform objects. The second level with the large objects was necessary as it has been shown [19] that one of the oil spill detection features is the brightness difference of the smaller with the larger objects. The segmentation parameters that have been used can be summarized as follows:

•

Level-1 segmentation (small objects) Weight for high resolution (10 m) bands 2, 3, 4, and 8:1 Weight for lower resolution (60 and 20 m) bands 1, 9, and 11:0 Scale 10, color 0.8—shape 0.2, compactness 0.5—smoothness 0.5

•

Level-2 segmentation (large objects) Weight for high resolution (10 m) bands 2, 3, 4, and 8:1 Weight for lower resolution (60 and 20 m) bands 1, 9, and 11:0 Scale 80, color 0.1—shape 0.9, compactness 1.0—smoothness 0.0

Using the above scale parameters and segmentation criteria, the size of the image objects of the lower level (Level 1) ranged from 0.001 to 0.15 km2 , while the image objects of the upper level (Level 2) ranged from 0.3 to 1.5 km2 . After segmentation, interpretation of the images and examination of various features of the image objects in various band ratios has been performed. After several trials it has been decided that the following ratios had to be used:

• • •

the normalized difference water index (NDWI)Sentinel 2 equivalent: (B2 − B11)/(B2 + B11), for the discrimination of sea from land (Figures 5a, 6a, and 7a), the ratio B2/B11 (Figures 5c, 6c, and 7c), and the custom feature: StdDev(B2)*B2/B11, which multiplies the band 2 standard deviation of an image object with the previous ratio value, which has been calculated for the same object (Figures 5d, 6d, and 7d).

J. Mar. Sci. Eng. 2018, 6, 4

6 of 12

J. Mar. Sci. Eng. 2018, 6, 4

6 of 11

As already pointed out in [19], the oil spill objects present higher standard deviation and higher J. Mar. Sci. Eng. 2018, 6, 4 6 of 11 brightness compared their out much larger super-objects (Figures 5b, standard 6b, and deviation 7b). In addition, the ratio As already to pointed in [19], the oil spill objects present higher and higher brightness compared to their much larger super-objects (Figures 5b, 6b, and 7b). In addition, the ratio B2/B11 proved to perform well the detection oil spill areas when away deviation from sun-glint, while the As already pointed out in in [19], the oil spillof objects present higher standard and higher B2/B11 to perform well in proved the detection of a oilmuch spill areas when away from sun-glint, while theeven in custom brightness featureproved StdDev(B2)*B2/B11 to be better indicator of oil spill objects compared to their much larger super-objects (Figures 5b, 6b, and 7b). In addition, the ratio custom feature StdDev(B2)*B2/B11 proved to be a much better indicator of oil spill objects even in sun-glint areas. Theirtocombined use the results. B2/B11 proved perform well in provided the detection of best oil spill areas when away from sun-glint, while the sun-glint areas. Their combined use provided the best results. custom feature StdDev(B2)*B2/B11 proved to be a much better indicator of oil spill objects even in sun-glint areas. Their combined use provided the best results.

(a)

(b)

(a)

(b)

(c)

(d)

Figure 5.Image object features (c) calculated for the Sentinel 2A subset of the (d)Zakynthos image acquired

Figure 5. Image object features calculated for the Sentinel 2A subset of the Zakynthos image on 7 August 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) object features calculated(b) for Band the Sentinel 2A subset of the Zakynthos acquired acquiredFigure on 75.Image August 2017: (a) NDWI; 2 mean difference to super image objects; (c) B2/B11; StdDev(B2)*B2/B1. on 7 August 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) and (d) StdDev(B2)*B2/B1. StdDev(B2)*B2/B1.

(a)

(b)

(a)

(b)

(c)

(d)

Figure 6.Image object features Zakynthos image acquired (c) calculated for the Sentinel 2A subset of the (d) on 27 August 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) Figure 6.Image object features calculated for the Sentinel 2A subset of the Zakynthos image acquired StdDev(B2)*B2/B11. Figure 6. Image object features calculated for the Sentinel 2A subset of the Zakynthos image on 27 August 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) acquiredStdDev(B2)*B2/B11. on 27 August 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11;

and (d) StdDev(B2)*B2/B11.

J. Mar. Sci. Eng. 2018, 6, 4

7 of 12

J. Mar. Sci. Eng. 2018, 6, 4

7 of 11

J. Mar. Sci. Eng. 2018, 6, 4

7 of 11

(a)

(b)

(a)

(b)

(c)

(d)

(c) (d) Gulf image acquired Figure 7.Image object features calculated for the Sentinel 2A subset of the Saronic Figure 7. Image object features calculated for the Sentinel 2A subset of the Saronic Gulf image acquired on 13 September 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) 7.Image2017: object features calculated for the2 Sentinel 2A subset ofto thesuper Saronic Gulf image acquired and on 13Figure September (a) NDWI; (b) Band mean difference objects; (c) B2/B11; StdDev(B2)*B2/B11. on 13 September 2017: (a) NDWI; (b) Band 2 mean difference to super objects; (c) B2/B11; and (d) (d) StdDev(B2)*B2/B11. StdDev(B2)*B2/B11.

Figure 8 summarizes the OBIA method developed within this study for the detection of oil spills

using Sentinel 2 imagery,the for processing. Thewithin criterion objects” Figure 8 summarizes OBIA method developed within this study for the land detection Figure 8 summarizes thesingle OBIA date method developed this“Distance study for to theneighbor detection of oil spillsof oil has been set in order to avoid misclassification of shallow waters as oil spills due to bottom spillsusing using Sentinel 2 imagery, for single date processing. The “Distance criterion to “Distance neighbor Sentinel 2 imagery, for single date processing. The criterion neighbor to land objects” land contributions in the reflectance values. has beenbeen set in to avoid misclassification of shallow waters as oilasspills due to bottom objects” has setorder in order to avoid misclassification of shallow waters oil spills due to bottom contributions in the reflectance values. contributions in the reflectance values.

Figure 8. Single-date OBIA method. Figure8.8.Single-date Single-date OBIA Figure OBIAmethod. method.

Table 2 summarizes the fuzzy rule set values that have been used for the image object features used.

J. Mar. Sci. Eng. 2018, 6, 4

8 of 11

J. Mar. Sci. Eng. 2018, 6, 4

8 of 12

J. Sci. 6, J. Mar. Mar.Table Sci. Eng. Eng. 2018, 6, 4 4 2 2018, summarizes

8 8 of of 11 11 the fuzzy rule set values that have been used for the image object features

J. Mar. Sci. Eng. 2018, 6, 4 used.

8 of 11

J. Mar. Sci. Eng. 2018, 6, 4

8 of 11

Table 2 summarizes the fuzzy rule set values that been the object 2 2018, summarizes the fuzzy values thatofhave have been used used for forOBIA the image image object features features J. Mar.Table Sci. 6,values 4 8 of 11 Table 2.Eng. Feature used forrule the set fuzzy rules the single-date method. used. Table 2 summarizes the fuzzy rule set values that have been used for the image used. Table 2. Feature values used for the fuzzy rules of the single-date OBIA method.object features used.Table 2 summarizes the fuzzy rule set values that have been used for the image object features J. Mar. Sci.J.Eng. 6, 42018, Class Feature ValueValue Fuzzy Limits 8 of8 11 Feature Decision Fuzzy Limits Mar. 2018, Sci. Eng. 6,Class 42. Feature values used of 11 Table for the fuzzy rules ofDecision the single-date OBIA method. Table 2. Feature values used for the fuzzy rules of the single-date OBIA method. used.Table 2 summarizes the fuzzy rule set values that have been used for the image object features Seawater Table NDWI >0.4 Seawater NDWI >0.4Value 0.35 0.45 2. Feature values used for the fuzzy rules of the single-date OBIA method. 0.35 0.45 Class Feature Decision Fuzzy Limits Limits Class Feature Decision Value Fuzzy used. Table 2 Table summarizes the fuzzy rule set values that have beenbeen used for the image object features 2 summarizes the fuzzy rule set values that have used for the image object features

Table the fuzzy rulesobjects of the single-date OBIA method. B2 used mean for difference to super >11Value 10 Limits 12 Feature Decision Fuzzy J. Mar. Sci. Eng. 2018, 6,Class 42. Feature values 8 of 11 >11 >0.4 Seawater NDWI 0.35 0.45 used. 12 10 Seawater B2 mean difference to super NDWIobjects >0.4 0.35 0.45 StdDev(B2)*B2/B11 >135 OBIA method. 130 140 Table 2. Feature values used for the fuzzy rules of the single-date Class Feature Decision Value Fuzzy Limits StdDev(B2)*B2/B11 >135 Seawater NDWI >0.4 0.35 0.45 B2 mean mean difference difference to super super objects objects >11 10 12 130 140 Oil spill the fuzzy B2/B11 >13 12 14 B2 to >11 10 12 Table 2 summarizes rule for set the values that havethe been used for the method. image object features 2. Feature used fuzzy rules single-date OBIA Table Feature values values usedB2for theStdDev(B2)*B2/B11 fuzzy of theofsingle-date OBIA method. StdDev(B2)*B2/B11 >135 130 140 Class Feature Fuzzy Oil spill 2.Table Area inrules pixels >20 19 Limits 21 B2/B11 >13 >135 Seawater NDWI >0.4 0.45 130 140 mean difference to super objects Decision >11Value 0.35 10 12 12 14 used. Oil spill B2/B11 >13 12 14 Distance to Land objects (meters) >350 300 400 Oil spill B2/B11 >13 Value 12 Limits 14 StdDev(B2)*B2/B11 >135 130 140 Class Feature Decision Fuzzy Area in pixels >20 Value NDWI >0.4 0.35 0.45 Class Seawater Feature Decision Fuzzy Limits B2 mean difference to super objects >11 10 12 21 19 Area in >20 19 21 Area in pixels pixels >20 19 21 Oil spill B2/B11 >13 12 14 Distance to oil spill (meters) 135 130 140 Seawater NDWI >0.4 0.35 0.45 Distance toDistance Land objects (meters) >350 >350 Table 2. Feature values used fuzzy rules of the single-date Seawater NDWI >0.4 0.35 0.45 Distance to the Land objects (meters) >350 300 400 B2 mean for difference to super objects >11 OBIA method. 10 12 300 400 Possibly dissolved oil spill to Land (meters) 300 400 Area in objects pixels >20 19 21 Oil spill B2/B11 >13 12 12 14 B2 mean difference to super objects >11 >135 10 130 StdDev(B2)*B2/B11 140 >125 120 130 Distance to Land objects (meters) >350 300 400 B2 mean difference to super objects >11 10 12 Distance to oil spill (meters) 13513 12 140 Possibly dissolved spill Possibly dissolved oil using spill single-date TherePossibly is aoil drawback images: of>135 oil spills in 130 areas of14 high bottom StdDev(B2)*B2/B11 140 Distance to Land objects (meters) >350 300 400 Distance toNDWI oil (meters) 250 300 B2/B11 >13 125 120 130 StdDev(B2)*B2/B11 >125 >0.4 0.45 Area in spill pixels >20 19 14 21 StdDev(B2)*B2/B11 >125 120 130 130 120 Possibly dissolved oil spill Oil spill happens close to theB2/B11 >13 12 14 reflectance. This mostly seashore, but could also happen in areas of low depth Area in >20 350 There is a to using single-date false detection oil in areas high bottom Distance toimages: oil spill (meters) 300 StdDev(B2)*B2/B11 120 130 TherePossibly is a drawback drawback to using single-date images: false detection of of>125 oil spills spills and in250 areas of21 high Area pixels >20 19 B2 mean difference touse super objects >11 10 12 dissolved oil avoid spill elsewhere. The best way to these errors isinto multitemporal imagery require thatbottom the oil Distance to Land objects (meters) >350 300 400 reflectance. This mostly happens happens close to the seashore, but could also happen in areas of low depth Distance to Land objects (meters) >350 300 400 StdDev(B2)*B2/B11 >135 130 140 There is a drawback to using single-date images: false detection of oil spills in areas of high bottom Distance to oil spill (meters) 13 12 14 elsewhere. The best way to single-date avoid these errors is to use multitemporal imagery and require that the oil Possibly dissolved spillhappens reflectance. This close to the seashore, but could also happen in areas of low depth The best way to avoid these errors is to use multitemporal imagery and require that the oil There is a drawback using single-date images: false detection of oil spills in areas of high bottom Distance to oil spill (meters) methodology 125 is not 120 130130 StdDev(B2)*B2/B11 >125 120 case application of the multitemporal mode of the proposed possible, it is better to Possiblyonly dissolved oil image. spill Area in pixels >20 19 21 Figure 9. In spill is detected in one The combined use of multitemporal images is shown in elsewhere. The best way to avoid these errors is to use multitemporal imagery and require that the oil spill is detected only in one image. The combined use of multitemporal images is shown in Figure 9. In bottom reflectance. This mostly happens close to the seashore, but could also happen in areas reflectance. This mostly happens close to the seashore, but could also happen in areas of low depth is detection a drawback using single-date images: false detection oil spills in“distance areas of130 high bottom keep oil spill at to a sufficiently large from the seashore using from land”of low There isThere a drawback to using single-date images: false detection of oilofspills inthe areas of high bottom StdDev(B2)*B2/B11 >125 120 Distance to distance Land objects (meters) >350 300 400

used.

case application of the multitemporal mode of the proposed methodology is not possible, it is better to

spill is detected only into one image. Themode combined of multitemporal images is shown inand 9.oil In case application of the multitemporal of is the proposed methodology is not possible, itFigure is better to that elsewhere. The best way to these errors touse use multitemporal imagery and require that the reflectance. This mostly happens close to the seashore, but could happen in areas low depth depth elsewhere. The best way avoid errors is to use multitemporal require rule This (Figure 8). reflectance. happens closethese to the seashore, but could alsoalso happen inimagery areas ofofof low depth Theremostly is a drawback to avoid using single-date images: false detection of125 120 130 case application of the multitemporal mode of the proposed methodology is not possible, it is better to spill is detected only in one The combined ofmultitemporal multitemporal images is shown in Figure 9. In elsewhere. The to image. avoid these errors is to use imagery and require that 9. the detected only inbest one image. The combined use ofuse multitemporal images is shown in Figure Inoil Figure 9.spill In iscase of way the multitemporal mode offrom thedetection proposed methodology notbottom possible, it ruleapplication (Figure There is8). a drawback to using single-date images: false of oil spills inpossible, areas ofis high keep oil spill detection at a sufficiently large distance the seashore using the “distance from land” case application of the multitemporal mode of the proposed methodology is not it is better to spill is detected only in one image. Theof combined use of methodology multitemporalis images is shownitin case application of the multitemporal mode the proposed not possible, isFigure better 9. toIn reflectance. This mostly happens close to the seashore, but could also happen in areas of low depth is better to keep oil spill detection at a sufficiently large distance from the seashore using the “distance rule (Figure keep oil spill8). detection at a sufficientlymode large of distance from the seashore using thepossible, “distanceitfrom land” ofathe multitemporal the proposed methodology is not is better to keep oilcase spillapplication detection at sufficiently large distance from the seashore using the “distance from land” elsewhere. The best way to avoid these errors is to use multitemporal imagery and require that the oil rule (Figure 8). from land” rule (Figure 8). keep oil spill detection at a sufficiently large distance from the seashore using the “distance from land” rule (Figure 8). spill is detected only in one image. The combined use of multitemporal images is shown in Figure 9. In rule (Figure 8). case application of the multitemporal mode of the proposed methodology is not possible, it is better to keep oil spill detection at a sufficiently large distance from the seashore using the “distance from land” rule (Figure 8).

Figure 9. Multitemporal OBIA method. 9. OBIA Figure 9. Multitemporal Multitemporal OBIA method. method. OBIA method. These values have Table 3 summarizes the fuzzyFigure rule set values for the multitemporal Figure 9. Multitemporal OBIA method. been applied to all the Sentinel images used in this work, providing successful oil spill detection results. Table rule values OBIA method. These values have Table 3 3 summarizes summarizes the the fuzzy fuzzyFigure rule set set values for for the the multitemporal multitemporal 9. Multitemporal OBIA method. OBIA method. These values have been applied to all the Sentinel images used in this work, providing successful spill detection results. 3 summarizes the values fuzzy rule set values for rules theproviding multitemporal OBIAoil method. These values been Table applied to all the Sentinel images used in work, successful oil spillmethod. detection results.have Table 3. Feature used for thethis fuzzy of themethod. multitemporal OBIA Figure 9. Multitemporal OBIA been Table applied to all the Sentinel images used in this work, providing successful oil spill detection results.have 3 summarizes the fuzzy rule set values for the multitemporal OBIA method. These values Class Decision OBIA Value method. Fuzzy Limits 9. Multitemporal Table 3. Feature valuesFigure used for the Feature fuzzy rulesOBIA of themethod. multitemporal

Table 3. Feature values used for in thethis fuzzy rules of the multitemporal OBIA been Table applied toSeawater all the Sentinel images used work, successful oil spillmethod. detection results.have 9. Multitemporal method. 3 summarizes theFigure fuzzy rule set values forOBIA theproviding multitemporal OBIA method. These values NDWI >0.4 0.35 0.45

Figure 9. Multitemporal OBIA Table 3. Feature values used for the fuzzy rules of method. the multitemporal OBIA method.

Mar. Sci. Eng.Class 2018, 6, 4 8 of 11 Class Feature Decision Value Fuzzy Limits B2rule mean diff. sup. obj. (oil spill date) successful >11 10 Limits 12 Feature Decision Value Fuzzy beenJ.Table applied to all the Sentinel images used intothis work, providing oil spill detection results.have 3 summarizes the fuzzy set values for the multitemporal OBIA method. These values

Seawater NDWI >0.4 0.35 0.45 Table 3. Feature used for the fuzzy rules ofspill themethod. multitemporal method. B2set mean diff. to sup. obj. (no oil date) 0.4 0.35 0.45 have Figure 9. Multitemporal OBIA Class Feature Decision Fuzzy Limits Table 3 summarizes the fuzzy values rule values the multitemporal OBIA method. These been applied toSeawater all the Sentinel images used infor this work, providing successful oilValue spill detection results.features Oil spill B2 mean diff. to sup. obj. (oil spill date) >11 10 12 Table 2 summarizes the fuzzy rule set values that have been used for the image object StdDev(B2)*B2/B11 (oil spill date) >135 130 140 B2 mean diff. to sup. obj. (oil spill date) >11 10 12 Seawater NDWI >0.4 0.35 0.45 Table 3. Feature values used for the fuzzy rules of the multitemporal OBIA method. been applied to all the Sentinel images used in this work, providing successful oil spill detection results. Class Feature Decision Value Fuzzy Limits Table 3 summarizes the fuzzy ruleB2set values forobj. the OBIA method. mean diff. to sup. (nomultitemporal oil spill date) 135Value Fuzzy 130 Limits 140 Class Feature Decision StdDev(B2)*B2/B11 >135 130 140 B2 images mean diff. toused sup. obj.(oil (no oil spill date) 11 10 12 been Table applied allthe the values Sentinel images used intothis work, providing successful oil spill detection results. oil spill Oil spill StdDev(B2)*B2/B11 (no spill date) 0.4 0.35 0.45 StdDev(B2)*B2/B11 (no oilspill spill date) 135 130 140 Feature Decision Value Fuzzy Limits Table values used for the fuzzy rules of the single-date OBIA method. J. Mar. Sci. Eng.Class 2018, 6, 42. Feature 8 of 11 B2 mean diff. to sup. obj. (no oil spillhave date) been used 11 10 12 spill StdDev(B2)*B2/B11 0.4 0.35 0.45

Class

Feature StdDev(B2)*B2/B11 (oil spill date)

Decision Value >135

Fuzzy 130Limits140

3. Feature useddiff. for the fuzzy rules thedate) multitemporal B2 mean to sup. obj. (no oilofspill 11 10 0.45 12 spill (no 0.4 0.35 130 Table Oil 2 summarizes theB2StdDev(B2)*B2/B11 fuzzy ruletoset that have been used the image object features StdDev(B2)*B2/B11 (oil spill date) >135 for 140 Sci. Eng.Class 2018, 6, 4 used for 8 of 11 TableJ.3.Mar. Feature values the fuzzy rules of the multitemporal OBIA method. B2 mean diff. to sup. obj. (no oil spill date) Decision 0.4 0.35 0.45 B2 meanStdDev(B2)*B2/B11 diff. to sup. obj. (oil spill date)date) >11 135 0.4

130 140 100.35 12 0.45

B2 rule mean difference to that super objectsbeen used >11 10 12 (no 11 for the 130 10 12 Class Feature Decision Value Fuzzy Limits StdDev(B2)*B2/B11 (oil spill date) >135 130 method. Class Feature Decision Fuzzy Limits >135Value 130 140 140 Table 2. Feature values usedStdDev(B2)*B2/B11 for the fuzzy rules of the single-date OBIA Oil spill B2 mean diff. to sup. obj. (no oil spill date) 0.4 NDWI >0.4 0.35 0.45 StdDev(B2)*B2/B11 (oil spill date) >135 130 140 0.45 Feature Decision Fuzzy Limits B2/B11 Area >13 >20 Value12 0.35 in pixels 1914 21 (nofuzzy oil spill date) 11 >11 1021 400 12 Area in pixels >20 >350 19 0.35 Distance to Land objects (meters) 300 B2 mean diff. to sup. obj. (oil spill date) Seawater NDWI >0.4 0.45 10 12 StdDev(B2)*B2/B11 >135 130 140 Feature Decision Value Fuzzy Limits B2Class mean diff. to sup. obj. (no oil spill date) 13 12 14 Possibly dissolved StdDev(B2)*B2/B11 >135 130 140 StdDev(B2)*B2/B11 (oil spill date) >135 >125 Seawater NDWI >0.4 0.35 0.45 Area in pixels >20 19 21 StdDev(B2)*B2/B11 120 130 130 140 Oil spill B2/B11 >13 12 14 Distance to Land objects (meters) >350 300 400 There is a drawback to using single-date images: false detection of oil spills in areas of B2 mean difference to super objects >11 10 12high bottom Area in pixels >20 19 21 StdDev(B2)*B2/B11 >135 130 140 Distance to oil spill (meters) 350 300 400 Possibly dissolved oil spill Oil way spill to avoid these errorsB2/B11 >13 12require 14 that the oil elsewhere. The best is to use multitemporal imagery and StdDev(B2)*B2/B11 >125 120 130 Distance to oil (meters) 20 19 21 Possiblyonly dissolved oil image. spill spill is detected in one The combined use false of multitemporal images in Figure 9. In There is a drawback to using single-date images: detection of>350 oil spillsis inshown areas of high bottom Distance to Land objects (meters) 300 400 StdDev(B2)*B2/B11 >125 120 130

Seawater

case application the multitemporal mode of seashore, the proposed is not in possible, it is better to reflectance. This of mostly happens close to the but methodology could also happen areas of low depth Distance toimages: oil spill (meters) 250 There is detection a drawback to using single-date false detection of125 images is shown in Figure 9. In

Table 2 summarizes the fuzzy rule set values that have been used for the image object features used. J. Mar. Sci. Eng. 2018, 6, 4

8 of 11

Table 2. Feature values used for the fuzzy rules of the single-date OBIA method.

Table 2 summarizes the fuzzy rule set values that have been used for the image object features J. Mar. Sci. Eng. 2018, 6,Class 4 8 of 11 Feature Decision Value Fuzzy Limits J. Mar. Sci. Eng. 2018,J.used. 6,Mar. 4 Sci. Eng. 2018, 6, 4 9 of 12 8 of 11 Seawater NDWI >0.4 0.35 0.45 Table 2 summarizes the fuzzy rule set values that have been used for the image object features J. Mar. Sci. Eng. 2018, 6, 4 8 of 11 Table 2. Feature values the fuzzy rulesobjects of the single-date OBIA method. B2 used mean for difference super >11for 10 12 used.Table 2 summarizes the fuzzy rule set valuesto that have been used the image object features StdDev(B2)*B2/B11 >135 130 140 Table Class Feature that have been Decision Value Fuzzy Limits used.Table 2 summarizes the fuzzy rule 3. setCont. values used for the image object features

3.

Oil spill B2/B11 >13 12 14 Table J. Mar. Sci. Eng. 2018, 6, 42. Feature values used for the fuzzy rules of the single-date OBIA method. 8 of 11 used. Seawater NDWI >0.4 0.35 0.45 Area in pixels >20 19 21 Table 2. Feature values used for the fuzzy rules of the single-date OBIA method. Class Feature Decision Value Fuzzy Limits Distance to Land objects (meters) >350Value 300 Limits 400 Class Feature Decision Fuzzy B2 rule mean set difference to superhave objectsbeen used >11for the image 10 12 Table 2 summarizes the values fuzzy values object features Table 2. Feature used forspill the fuzzythat rules of the single-date OBIA method. Class Decision Value Fuzzy Limits >135 130 140 StdDev(B2)*B2/B11 (no StdDev(B2)*B2/B11 oil 13 Value 12 Limits 14 Class Feature Decision Fuzzy B2/B11 Seawater NDWI >0.4 0.35 0.45 Oil spill 120 130 12 14 B2 meanStdDev(B2)*B2/B11 difference to super objects >13 >125 >11 10 12 Area in pixels >20 19 21 StdDev(B2)*B2/B11 >135 130 140 Area in pixels >20 2. Feature values used for the fuzzy rules of the single-date OBIA method. Seawater NDWI >0.4 0.35 0.45 There is aTable drawback to using single-date images: false detection of oil spills in areas of high bottom B2 mean difference to super objects >11 10 12 21 19 Distance to Land objects (meters) >350 300 400 Oil spill happens close to B2/B11 >13 12 14 StdDev(B2)*B2/B11 >135 130 140 reflectance. This mostly the seashore, but could also happen in areas of low depth B2 mean difference to super objects 11Value 10 Limits 12 ClassDistance to oil Decision Fuzzy J. Mar. Sci. Eng. 2018, 6, 4 spill (meters) Distance toFeature oil spill (meters) 20 19 21 Oil way spill to B2/B11 >13 12 14 that the oil Possibly dissolved oil avoid spill these errors elsewhere. The best isinto use multitemporal >135 imagery and StdDev(B2)*B2/B11 130require 140 Possibly dissolved oil spill Distance to Land >350 300 400 StdDev(B2)*B2/B11 (oil spill date) Seawater NDWI >0.4 0.35 0.45 Area in objects pixels (meters) >125 >125 >20 19 21 StdDev(B2)*B2/B11 120 130 120 130 Oil spill B2/B11 >13 12 14 spill is detected only in oneDistance image. to The use of multitemporal images is250 shown in Figure 9. In oilcombined spill (meters) 350 300 Distance tospill Land objects (meters) 300 400 StdDev(B2)*B2/B11 (no oil date) 125>11 is not120 130 Distance to Land objects (meters) >350 300 400 StdDev(B2)*B2/B11 >135 130 140 Distance to oil spill (meters) 13 12require 14 that the oil elsewhere. The best is use multitemporal imagery Distance toimages: oilto spill (meters) 300 rule (Figure 8). StdDev(B2)*B2/B11 >125 120 TherePossibly is a drawback false detection of20 19 21 spill is detected only in one image. The combined use of multitemporal images is shown in Figure 9. In reflectance. This mostly happens close to the seashore, but could also happen in areas of lowbottom depth There is a drawback to using single-date images: false detection of oil spills in areas of high StdDev(B2)*B2/B11 >125 120 130 Distance to Land objects (meters) >350 300 400 3. Results

application of the multitemporal mode of is the methodology is not in possible, it Zakynthos, is better to elsewhere. The best way to these errors toproposed use multitemporal imagery require that the oil reflectance. This mostly happens close to the seashore, but could also happen areas low depth Figure 10a–ccase presents classification results for Zakynthos, August 2017; There is a the drawback to avoid using single-date images: false detection of7125 120 the 130 August 2017; and the Saronic respectively. The Sentinel images of Zakynthos and case application ofGulf, the of the proposed methodology is notis possible, itSaronic is better to spill is detected inmultitemporal one Themode combined multitemporal images shown in Figure 9.oil In elsewhere. The best way to image. avoid these errors is touse useof2multitemporal imagery and require that the Gulf have been processed with bothmultitemporal the single-date and multitemporal OBIA methods, with the latter There is detection awith drawback to using single-date images: false detection of oil spills in“distance areas high bottom keep spill athe sufficiently large distance from the seashore using the from land” case application of the of the proposed methodology is not it islatter better to Gulf have been processed both single-date and multitemporal OBIA methods, withofin the spill isoildetected only inat one image. Themode combined use of multitemporal images ispossible, shown Figure 9. In reflectance. mostlyattruth happens closelarge to the seashore, but could also happen in“distance areas of from low depth providing better results. NoThis ground was for any ofthese these areas, but: (a) natural the rule (Figure 8). keep oil spill detection atruth sufficiently distance from seashore using the land” case application of the multitemporal mode of the proposed methodology is not possible, it isnatural better to oil providing better results. No ground was available available for any ofthe areas, but: (a) the oil elsewhere. The best way to avoid these errors is to use multitemporal imagery and require that the oil rule (Figure 8). spill spill of Zakynthos is aoil area whose has been studied by Kolokoussis et al. [19], keep detection at a sufficiently large distance from the seashore the “distance from land” of Zakynthos iswell-known aspill well-known area whose behavior behavior has been studied byusing Kolokoussis et al. [19], spill is detected only in one image. The combined use of multitemporal images is shown in Figure 9. In rule (Figure 8).be drawn and accurate results can be drawn by photointerpretation; and (b) the oil spill detections on the Saronic and accurate results can by photointerpretation; and (b) the oil spill detections on the case application of the multitemporal mode of the proposed methodology is not possible, it is better to Gulf Saronic image are and in accordance with the seashore pollution. to land” careful Gulfrealistic image and accordance with the from reported seashore pollution. According keep oilare spillrealistic detection at ain sufficiently largereported distance the seashore using theAccording “distance from to careful photointerpretation of the images, the Zakynthos classification results, at least for the rule (Figure 8). photointerpretation of the images, the Zakynthos classification results, at least for the known natural natural flow, are successful. considered very successful. Moreover, smalleroutflows natural outflows flow,known are considered very Moreover, evidence of evidence smaller of natural close to the close to the shore (Figure 10a, b) is also provided. The Saronic Gulf image oil spill classification shore (Figure 10a,b) is also provided. The Saronic Gulf image oil spill classification seems toseems present a to present a significant commission error close to the refinery area at the north of Salamina Island, significant commission error close to the refinery area at the north of Salamina Island, which cannot which cannot be linked to the sinking of the Agia Zoni II. If indeed false, this detection has to be be linked to the sinking of the Agia Zoni II. If indeed false, this detection has to be studied further. studied further. Since this is a more complex and unstudied area, ground truth would be necessary Sinceinthis is atomore complex and unstudied area, ground truth would be necessary in order estimate order estimate the classification accuracy and, thus, no further processing was carried out to at this the classification accuracy and, thus, no further processing was carried out at this point. point.

Figure 9. Multitemporal OBIA method.

Table 3 summarizes the fuzzy rule set values for the multitemporal OBIA method. These values have been applied to all the Sentinel images used in this work, providing successful oil spill detection results. Figure 9. Multitemporal OBIA method. Table 3. Feature values used for the fuzzy rules of the multitemporal OBIA method.

Table 3 summarizes the fuzzy rule set values for the multitemporal OBIA method. These values have Figure 9. Multitemporal OBIA method.Decision Value Fuzzy Limits Class Feature been applied toSeawater all the Sentinel images used in this work, providing successful oil spill detection results. NDWI >0.4 0.35 0.45 Figure 9. Multitemporal OBIA method.

mean diff. to sup. obj. (oilmultitemporal spill date) >11method. These 10 12 Table 3 summarizes the fuzzyB2 rule set values for the OBIA values have Figure 9. Multitemporal OBIA Table 3. Feature values useddiff. for the fuzzy themethod. multitemporal B2 mean to sup. obj.rules (no oilofspill date) 135 130 140 Class Feature Decision Fuzzy Limits been Table applied to all the Sentinel images in this work, providing successful oilValue spill detection results.have 3 summarizes the fuzzy ruleused set values for OBIA values StdDev(B2)*B2/B11 (nothe oil multitemporal spill date) 0.4OBIA method. 0.35 0.45 3. Feature values used for thethis fuzzy of themethod. multitemporal Figure 9. Multitemporal been appliedTable toSeawater all the Sentinel images used in work, providing successful oil spill detection results. B2used meanfor diff. to fuzzy sup. obj. (oil spill date) >11OBIA method. 10 12 Table 3. Feature values the rules of the multitemporal (a) (b) Class Feature Decision Value Fuzzy Limits

B2 mean diff. to sup. obj. (no oil spill date) 0.4method. 0.35 0.45 Class Decision Value method. Fuzzy StdDev(B2)*B2/B11 (oil spill >135 130 Limits 140 been applied toSeawater all the Sentinel images used in this work, providing successful oil spill detection results. B2 mean diff. to NDWI sup. (no obj. oil (oilspill spilldate) date) >11 10 12 >0.4 0.35 0.45 StdDev(B2)*B2/B11 0.4

Fuzzy Limits 10 12 10 12 0.35 0.45

Oil spill Class Oil spill Seawater

StdDev(B2)*B2/B11 spilldate) date) StdDev(B2)*B2/B11(no (oiloilspill B2 mean diff. to sup. obj. (no oil spill date) Feature StdDev(B2)*B2/B11 (no oil spill date) StdDev(B2)*B2/B11 NDWI(oil spill date) StdDev(B2)*B2/B11 B2 mean diff. to sup. (no obj. oil (oilspill spilldate) date)

135 0.4 11 135 135 130 140 Table 3. Feature values used for the fuzzy rules thedate) multitemporal B2 mean diff. to sup. obj.(oil (nospill oilofspill 11 10 12

Oil spill

B2 mean diff. to sup. obj. (no oil spill date) StdDev(B2)*B2/B11 (oil spill date) StdDev(B2)*B2/B11 (no oil spill date)

10 130 130

12 140 140

(c) Figure 10. OBIA classification results for: (a) Zakynthos, 7 August 2017; (b) Zakynthos, August 27

Figure 10. OBIA classification results for: (a) Zakynthos, 7 August 2017; (b) Zakynthos, 27 August 2017; 2017; and (c) Saronic Gulf, 13 September 2017. and (c) Saronic Gulf, 13 September 2017.

4. Discussion Sentinel 2 imagery offers fine spatial and spectral resolution for oil spill detection, and the proposed methodology seems to provide successful results in the case of natural oil outflows as well as light oil spill events. The proposed oil spill detection methodology has been based on OBIA, which is

J. Mar. Sci. Eng. 2018, 6, 4

10 of 12

4. Discussion Sentinel 2 imagery offers fine spatial and spectral resolution for oil spill detection, and the proposed methodology seems to provide successful results in the case of natural oil outflows as well as light oil spill events. The proposed oil spill detection methodology has been based on OBIA, which is considered to be of fundamental importance. Some of the main image object features (standard deviation within an object, brightness difference to larger super objects, and size of objects) that were used for the detection cannot be utilized by pixel-based classification methods. Moreover, class-related features are not available in most pixel-based classifiers. Among the two proposed methods, the first method can be applied on single-date images, but its results will suffer from some commission errors in shallow water. The second method is applied on multitemporal data, with one image acquired on a date without oil spills and a second one after an oil spill event. With the latter method, the aforementioned commission errors in shallow water can be successfully avoided. The proposed multitemporal OBIA methodology has provided very accurate results and is in accordance with the previous work of the authors [19], where Ikonos, QuickBird, RapidEye, and Landsat Thematic Mapper (TM) images were used for the detection of natural oil flows. Sentinel 2 gives timely and freely available satellite image data and, due to the fine resolution, provides much better results than Landsat imagery, comparable to those achieved with the high-resolution satellite images in [19]. The proposed methodology has been tested on natural oil outflows as well as light oil spill events but has not been tested on heavy oil spill events like the Deep Horizon oil spill [21], which are known to have different spectral behavior. Atmospheric corrections are a very important preprocessing step in remote sensing, but the existing algorithms for water areas are still experimental and thus were not used at this point, but should be used in future experiments. Ground truth data are of fundamental importance for the development of any remote sensing methodology, but in the case of oil spills it is very difficult to organize and carry out this type of data collection. The main reason for this is that there are absolute restrictions concerning experiments with oil in seawater, and thus someone has to be ready to act only when an oil spill event takes place. Nevertheless, this work has to be validated in the future with ground truth data in order to derive valid conclusions. 5. Conclusions Two OBIA methods have been developed and used on Sentinel 2 imagery in order to detect oil spills, with the latter providing more consistent results. The first method can be applied on a single date Sentinel 2 image, but its results suffer from some commission errors in shallow water areas. For improving the methodology performance for oil-spill events close to shore, a second improved multitemporal method has also been proposed. Using this method, oil spills close to shore, such as those caused by the accident of the Agia Zoni II tanker in the Saronic Gulf, have been successfully detected. The preliminary results are considered successful and consistent, with a high degree of applicability to other Sentinel 2 satellite images. Further testing and fine tuning of the proposed object-based methodology has to be carried out using atmospheric correction and ground truth. Acknowledgments: This study has received funding from the Horizon 2020 research and innovation program under the Marie Sklodowska-Curie grant agreement No. 691071. We would like to thank the European Space Agency (ESA) for supplying the Sentinel-1 and 2 images. Author Contributions: V. Karathanassi had a major contribution to the literature review and scientific background; P. Kolokoussis conceived, designed and performed the experiments; P. Kolokoussis and V. Karathanassi discussed/evaluated the results and wrote the paper. Conflicts of Interest: The authors declare no conflict of interest.

J. Mar. Sci. Eng. 2018, 6, 4

11 of 12

References 1.

2. 3. 4.

5.

6.

7.

8.

9. 10.

11.

12. 13.

14. 15.

16. 17. 18. 19.

Brown, H.M.; Bittner, J.P.; Goodman, R.H. The Limits of Visibility of Spilled Oil Sheens. In Proceedings of the Erim Conferences, Second Thematic International Airborne Remote Sensing Conference and Exhibition, San Francisco, CA, USA, 24–27 June 1996; Volume III, p. 327. Klemas, V. Tracking Oil Slicks and Predicting their Trajectories Using Remote Sensors and Models: Case Studies of the Sea Princess and Deepwater Horizon Oil Spills. J. Coast. Res. 2010, 26, 789–797. [CrossRef] Karathanassi, V. Spectral Unmixing Evaluation for Oil Spill Characterization. Int. J. Remote Sens. Appl. 2014, 4. [CrossRef] Clark, R.N.; Swayze, G.A.; Leifer, I.; Livo, K.E.; Lundeem, S.; Eastwood, M.; Green, R.O.; Kokaly, R.; Hoefen, T.; Sarture, C.; et al. A Method for Qualitative Mapping of Thick Oil Using Imaging Spectroscopy; United States Geological Survey: Reston, GA, USA, 2010. Available online: http://pubs.usgs.gov/of/2010/1101 (accessed on 11 October 2017). Hu, C.; Weisberg, R.H.; Liu, Y.; Zheng, L.; Daly, K.L.; English, D.C.; Zhao, J.; Vargo, G.A. Did the northeastern Gulf of Mexico become greener after the Deepwater Horizon oil spill? Geophys. Res. Lett. 2011, 38, L09601. [CrossRef] Carnesecchi, F.; Byfield, V.; Cipollini, P.; Corsini, G.; Diani, M. An optical model for the interpretation of remotely sensed multispectral images of oil spill. In Proceedings of the SPIE 2008, Remote Sensing of the Ocean, Sea Ice, and Large Water Regions, Cardiff, UK, 15–18 September 2008; Volume 7105. [CrossRef] Zhao, J.; Temimi, M.; Ghedira, H.; Hu, C. Exploring the potential of optical remote sensing for oil spill detection in shallow coastal waters—A case study in the Arabian Gulf. Opt. Express 2014, 22, 13755–13772. [CrossRef] [PubMed] Lee, M.S.; Park, K.A.; Lee, H.R. Detection and dispersion of oil spills from satellite optical images in a coastal bay. In Proceedings of the 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Beijing, China, 10–15 July 2016. Taravat, A.; Del Frate, F. Development of band ratioing algorithms and neural networks to detection of oil spills using Landsat ETM+ data. EURASIP J. Adv. Signal Process. 2012, 2012, 107. [CrossRef] Bradford, N.B.; Sanchez-Reyes, P.J. Automated oil spill detection with multispectral imagery. In Proceedings of the SPIE 8030, Ocean Sensing and Monitoring III, Orlando, FL, USA, 25–29 April 2011; Volume 8030. [CrossRef] Corucci, L.; Nardellib, F.; Cococcionia, M. Oil Spill Classification from Multi-Spectral Satellite Images: Exploring Different Machine Learning Techniques. In Proceedings of the Remote Sensing of the Ocean, Sea Ice, and Large Water Regions 2010, Toulouse, France, 20–23 September 201. Karathanassi, V.; Topouzelis, K.; Pavlakis, P.; Rokos, D. An object-oriented methodology to detect oil spills. Int. J. Remote Sens. 2006, 27, 5235–5251. [CrossRef] Lang, H.; Zhang, X.; Xi, Y.; Zhang, X.; Li, W. Dark-spot segmentation for oil spill detection based on multi-feature fusion classification in single-pol synthetic aperture radar imagery. J. Appl. Remote Sens. 2017, 11, 015006. [CrossRef] Su, T.F.; Li, H.Y.; Liu, T.X. Sea Oil Spill Detection Method Using SAR Imagery Combined with Object-Based Image Analysis and Fuzzy Logic. Adv. Mater. Res. 2015, 1065–1069, 3192–3200. [CrossRef] Chen, Z.; Wang, C.; Teng, X. Oil spill detection based on a superpixel segmentation method for SAR image. In Proceedings of the2014 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Quebec City, QC, Canada, 13–18 July 2014. Topouzelis, K.; Kitsiou, D. Detection and classification of mesoscale atmospheric phenomena above sea in SAR imagery. Remote Sens. Environ. 2015, 160, 263–272. [CrossRef] Topouzelis, K. Oil spill detection by SAR images: Approaches and Algorithms. Invited paper for the special issue: Synthetic Aperture Radar (SAR). Sensors 2008, 8, 6642–6659. [CrossRef] [PubMed] Maianti, P.; Rusmini, M.; Tortini, R.; Gianinetto, M. Monitoring large oil slick dynamics with moderate resolution multispectral satellite data. Nat. Hazards 2014, 73, 473–492. [CrossRef] Kolokoussis, P.; Karathanassi, V. Detection of Oil Spills and Underwater Natural Oil Outflow Using Multispectral Satellite Imagery. Int. J. Remote Sens. Appl. 2013, 3, 145–154.

J. Mar. Sci. Eng. 2018, 6, 4

20. 21.

12 of 12

Reiche, J.; Hese, S.; Schmullius, C. Objektbasierte Klassifikation terrestrischer Ölverschmutzungen mittel shochauflösender Satellitendaten in West-Sibirien. Photogramm. Fernerkund. Geoinf. 2017, 11, 275–288. Liu, Y.; MacFadyen, A.; Ji, Z.G.; Weisberg, R.H. (Eds.) Monitoring and Modeling the Deepwater Horizon Oil Spill: A Record-Breaking Enterprise; Geophysical Monograph Series; AGU/Geopress: Washington, DC, USA, 2011; Volume 195, p. 271. © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).