Detection of Massive Oil Spills in Sun Glint Optical Imagery through Super-Pixel Segmentation

Abstract

Large volumes of crude oil accidentally released into the sea may cause irreversible adverse impacts on marine and coastal environments. Large swath optical imagery, acquired using platforms such as the moderate-resolution imaging spectroradiometer (MODIS), is frequently used for massive oil spill detection, attributing to its large coverage and short global revisit, providing rich data for oil spill monitoring. The aim of this study was to develop a suitable approach for massive oil spill detection in sun glint optical imagery. Specifically, preprocessing procedures were conducted to mitigate the inhomogeneous light field over the spilled area caused by sun glint, enhance the target boundary contrast, and maintain the internal homogeneity within the target. The image was then segmented into super-pixels based on a simple linear clustering method with similar characteristics of color, brightness, and texture. The neighborhood super-pixels were merged into target objects through the region adjacency graph method based on the Euclidean distance of their colors with an adaptive termination threshold. Oil slicks from the generated bright/dark objects were discriminated through a decision tree with parameters based on spectral and spatial characteristics. The proposed approach was applied to oil spill detection in MODIS images acquired during the Montara oil spill in 2009, with an overall extraction precision of 0.8, recall of 0.838, and F1-score of 0.818. Such an approach is expected to provide timely and accurate oil spill detection for disaster emergency response and ecological impact assessment.

1. Introduction

Oil spills from oil production (e.g., from drilling rigs) or during oil transportation (e.g., from oil tankers) represent a large portion of large oil spill accidents. Such large volumes of oil discharge may cause irreversible adverse impacts on marine and coastal environments. For example, research related to the 2010 Deepwater Horizon oil spill demonstrated that the oil spill caused significant declines in the abundance and diversity of coastal wetlands, deep-sea coral reefs, fish species, sea turtles, and cetaceans [1,2,3,4,5].

Recent advancements in technology and sensor availability have enabled remote sensing to play a vital role in oil spill detection and monitoring, taking advantage of its synoptic observation over a wide area and frequent revisit capabilities compared to traditional observation methods [6,7,8,9,10]. Of the various remote sensing techniques, synthetic-aperture radar (SAR) and optical remote sensing are most frequently used in oil spill detection on the ocean surface. Oil films on the ocean surface would dampen surface capillary and short gravity waves, creating a “smoother” surface compared to the uncontaminated nearby water, reducing the Bragg scattering signal under optimal wind conditions, and appearing darker than the surrounding water in SAR imagery [11]. SAR has day and night capabilities and can penetrate through clouds, enabling highly effective detection in oil spill responses. While SAR is primarily used to detect the presence of oil, recent development has demonstrated the capabilities of discriminating thick oil emulsions using quad-polarized SAR images [11,12,13].

Optical remote sensing is also frequently used in oil spill detection, especially for large oil spill accidents. Wide-swath sensors such as the moderate-resolution imaging spectroradiometer (MODIS) and visible infrared imaging radiometer (VIIRS) provide frequent revisit globally (1–2 days in equatorial orbit), which would largely compensate for the weaknesses of lack of coverage under clouds [14,15]. There are two main principles behind the optical detection of oil spills. First, the oil-dampening effect can modulate reflected sunlight, which enhances the oil–water spatial contrast under sun glint. Distinct from the dark spots observed in SAR imagery, oil-covered regions in optical sun glint imagery display either positive or negative contrast with nearby water depending upon solar/viewing geometry and surface roughness [15,16,17,18]. However, it should be noted that when sun glint is negligible or absent (e.g., L_GN < 10⁻⁶ sr⁻¹ for MODIS), oil films may be undetectable [18]. Second, oil has unique optical properties, including high absorption in the blue spectral band, and exponential decay toward longer wavelengths in crude oil. Mixtures of oil–water emulsions enable high scattering, creating higher reflectance in the red, near-infrared (NIR), and shortwave infrared bands. As the multi-spectral information in optical remote sensing can be used to rule out oil look-alikes [19,20], classify oil types [20,21,22], and quantify oil concentration or volume [23,24,25,26], it shows high promise for future research on oil detection through optical remote sensing.

While many semi-/automatic oil spill extraction algorithms have been developed based on SAR [27,28,29,30,31,32], there are few semi-/automatic oil spill detection algorithms working for optical imagery. Currently, oil spill extent detection in optical remote sensing images largely relies on visual interpretation and manual delineation [33,34,35], which require extensive experience and expertise; and the derived oil spill map is also subject to interpretation. Other than visual interpretation, methods including pixel-based indices and deep learning have been developed to identify and classify oil slicks based on optical imagery [36,37,38,39,40,41,42,43,44]. Such methods have achieved good results in small- to mid-sized oil spills (~5–20 km in size). While sun glint reflectance varies depending on the surface roughness and solar/viewing geometry, the above methods may not apply to massive/large oil spills (hundreds of km, or covering a large part of the image swath), because of differences in the ambient light field caused by sun glint over the covered area. A general robust satellite technique based on the historical long-term time series of satellite data and local anomaly detection in the NIR band has been applied to massive oil spill detection in sun glint MODIS imagery during the Deepwater Horizon oil spill [43,44]. However, such an approach has difficulties in discriminating the negative-contrast oil slicks because of the low range of variability in reflectance.

Therefore, the objective of this study is to develop an oil spill detection approach for large oil spill accidents to overcome the large reflectance variability over the spilled area and to accomplish the identification of both the negative and positive contrast oil slicks in sun glint imagery. In this paper, an object-based method consisting of super-pixel segmentation and neighborhood merging is proposed to identify oil slicks through object feature analysis. The approach was tested and applied to the extraction of oil slicks with different spatial sizes and under different sun glint intensities during the Montara oil spill, and the extraction performance was evaluated objectively by comparing it with manually delineated oil slicks.

2. Data

2.1. Study Area

The Montara oil field is located in the Australian sector of the Timor Sea, ~260 km off the northern shoreline of Western Australia. A blowout of oil and gas happened on 21 August 2009, at a water depth of 77 m (Figure 1), due to accidental loss of control of the H1 well of the Montara wellhead. Subsequent oil slicks were found in the Timor Sea. The uncontrolled well continued to discharge oil into the Timor Sea until it was finally capped on 3 November 2009. An estimated 29,600 barrels of crude oil were discharged into the Timor Sea, making the Montara oil spill one of the worst oil disasters in Australian history [45,46]. The study area of this paper includes the main polluted area of the oil spill (14–11° S, 123.5–126.5° E, Figure 1).

2.2. Data and Preprocessing

MODIS L1A data were acquired from the NASA ocean color data archive. The data were processed to Rayleigh-corrected reflectance (R_rc, dimensionless) through the SeaWiFS data analysis system (SeaDAS, version 7.5.3, NASA Goddard Space Flight Center, Greenbelt, MD, USA), with bands centered at 469, 555, 645, 859, 1240, and 1640 nm. These bands were used for spectral analysis, where bands centered at 859, 645, and 555 nm were used for segmentation, and the 859-nm band was used for additional object analysis in this study. Here, the land bands rather than the ocean bands were selected because the ocean bands tend to saturate under strong sun glint. Sea surface temperature (SST, in units of °C) data were also acquired through the SeaDAS processing. In this study, sun glint strength was defined as the normalized sun glint radiance (L_GN, in units of sr⁻¹). L_GN was also acquired through the SeaDAS processing, and calculated via the Cox and Munk model based on the input from the wind speed (to calculate surface roughness) and the solar/viewing geometry [47,48,49]. The datasets were then mapped to a rectangular projection and all the bands were resampled to 250-m resolution. Appropriate image preprocessing procedures were used to mitigate the inhomogeneous light field over the image, enhance the contrast at the target boundary, and maintain homogeneity inside the target. The series of procedures conducted for preprocessing after the generation of R_rc data are summarized in Figure 2. Firstly, a selected cloud-free image was used to mask the land pixels, which was achieved with the normalized difference water index (NDWI) calculated using the MODIS green and NIR bands centered at 555 and 859 nm, respectively, with a threshold of <−0.05 for land and >0.5 for submerged islands [50]. Land mask, SST, and L_GN datasets were then used as auxiliary data to discriminate oil slicks.

Secondly, the false color composite (FRGB, centered at 859, 645, and 555 nm) was generated after Gaussian stretching [51,52]. Here, instead of R_rc data, the FRGB was used in order to increase the computation efficiency. As can be seen from the transect in Figure 3a, the dark oil area generally shows lower brightness in the NIR band than in the surrounding water. However, the ambient light field is not homogeneous across the image: the NIR band brightness of the water decreases gradually from the left to the right side of the image because of the decreased sun glint intensities (L_GN in Figure 3d). Moreover, such a gradual change in magnitude due to sun glint overwhelms the local contrasts between the oil and water, making it difficult to extract oil slicks based on the fixed threshold method.

Thirdly, the gradual color intensity variation across the image caused by sun glint needed to be minimized. This was achieved by subtracting a gradient background, which was modeled using the SFIT function in Interactive Data Language (IDL, version 8.5, Exelis Visual Information Solutions, Inc., a subsidiary of Harris Corporation, Boulder, CO, USA) software to determine the polynomial fit to the three channels of FRGB [53]. Taking the first channel (converted from R_rc_859) as an example, the brightness of the water became more homogeneous after subtraction of the gradient background caused by sun glint (Figure 3b,e vs. Figure 3a,d). Finally, the target boundary contrast was enhanced, and the target internal homogeneity was achieved through additional subtraction of median filtering with the kernel size determined through trial and error [54,55]. While the water brightness over the transect became more homogeneous, the brightness contrast between the oil and water was enhanced, as displayed in Figure 3c,f. The parameter settings in the preprocessing procedures are listed in

Table 1. Parameter settings for data preprocessing and segmentation.

Data preprocessing	SFIT function	Maximum degree of fit	4
	Median filter	Kernel size	233
Segmentation	SLIC	Segmentation number K	200,000
		Maximum color distance Nc	10
		Maximum iteration	10

.

3. Methods

3.1. Work Flow

The approach proposed in this study employs the commonly used chain structure flow in oil spill detection (Figure 4), which generally has three steps [28].

The first step, which consisted of two stages, was to generate the various objects: first, the image was segmented into super-pixels with similar color, brightness, and texture through the simple linear clustering (SLIC) method [56]; then, the super-pixels were merged into objects with different scales based on the region adjacency graph (RAG) method [57,58]. This is further explained in Section 3.2.

The next step was to separate the color anomaly objects from the background by analyzing the relationship between the sun glint reflectance and the lightness (in the CIELAB color space) of the background. Then, it needed to be decided whether the object belonged to the bright, or dark, object category. This is further explained in Section 3.3.

Finally, oil look-alikes needed to be excluded from the derived objects. These false positives included cloud shadow and sea surface roughness anomalies in dark objects, and seawater under sun glint and clouds in bright objects. These oil-look-alike objects were excluded through a decision tree whose nodes use various parameters including object-based spectral and spatial characteristics. This is further explained in Section 3.4.

3.2. Object Generation Based on Segmentation

SLIC segmentation was conducted after the image preprocessing to generate super-pixels from the FRGB image. A super-pixel is a small region consisting of multiple pixels that are located next to each other and have similar characteristics including color, brightness, and texture. These small regions (super-pixels) retain effective information and generally do not destroy the boundary information of objects in the image [59,60]. The SLIC method consists of three parts: color space transformation, distance metric construction, and iterative clustering. Firstly, the pixels of the FRGB image are transformed into five-dimensional feature vectors in the CIELAB color space and XY coordinates. The CIELAB color model can be considered a re-projection of the RGB color model, whose colors are composed of lightness (L), the green-to-red component (A), and the yellow-to-blue component (B). CIELAB is designed to approximate human vision, with the L component closely matching the human perception of lightness, which defines black at 0 and white at 100. Secondly, the similarity of two pixels can be measured by the distance between their corresponding five-dimensional vectors. The larger the distance, the smaller their similarity. The algorithm first generates K seed points, then searches the space surrounding each point, and finally classifies similar pixels into the same class with the seed until all pixels are classified. Thirdly, the average vector value in these K super-pixels is calculated to derive a recalculated center of the K clusters. The new K centers are then used to search for the most similar pixels around them. The above procedures are conducted iteratively until convergence, where the K centers do not shift [56,61,62]. Figure 5b shows the segmentation results of the SLIC method. Here, the color of the super-pixel represents the average brightness of the corresponding pixels in the RGB color space. The water areas were divided into various rectangular super-pixels with similar size and color. However, water super-pixels adjacent to other targets (such as oil slicks or clouds) were irregularly shaped, with their boundaries fitting the shape of the other targets. The parameter settings of the SLIC segmentation are listed in Table 1.

RAG is a typical topological data structure to characterize the adjacency of super-pixels [57,58]. Based on graph theory, RAG mergence was then conducted to merge super-pixels of a similar color into one object based on the Euclidean distance of three-dimensional feature vectors in the RGB color space. The over-segmented super-pixels with a similar color were gradually merged until no similar pairs of regions were found, which was operated by judging if the color difference was larger than a termination threshold. In this study, an adaptive termination threshold of neighborhood mergence was used, which was determined as follows:

$$T_{NM}=0.35\times \sqrt{N_{\mathrm{m}}^2+R_{\mathrm{m}}^2+G_{\mathrm{m}}^2}$$

(1)

where T_NM represents the adaptive termination threshold for neighborhood merging, and N_m, R_m, and G_m represent the median values of the three channels of the FRGB image.

3.3. Bright/Dark Object Recognition

As mentioned, oil slicks show either a positive contrast or a negative contrast with the water under the influence of sun glint, depending on the surface roughness and solar/viewing geometries. Therefore, the derived objects were first separated from the background into bright objects and dark objects. As most pixels within the study area were seawater, the median value of L in the CIELAB of the image was used to define the category of background seawater. When L_GN is weak, the contrast between the dark object lightness (L_o) and the seawater background lightness (L_bg) is small. The contrast becomes larger with increased L_GN values. The same applies to the contrast of bright objects with background seawater. The process of bright/dark object recognition can be divided into two steps. Firstly, objects that meet the conditions set by the different L_GN are classified as bright targets if their lightness is greater than L_bg. Secondly, objects that satisfy the conditions set by the different L_GN are classified as dark targets if their lightness value is smaller than L_bg. The rules for separating bright and dark objects with different sun glint conditions are shown in Figure 6.

As shown in Figure 5d, most of the oil slick objects were successfully segmented into objects other than the background. However, oil slicks were divided into several different objects due to the non-homogeneity within the slick object. After the bright/dark object recognition, these connected parts of the oil slick were merged into one through an eight-neighborhood connectivity analysis.

3.4. Oil Spill Identification

The characteristics of the derived objects were used to discriminate oil slicks from the derived dark and bright objects, which included four parameters and one decision condition: (1) the mean value of auxiliary data (L_GN, SST) of the object, (2) the mean difference between the outer boundary pixels and the inner boundary pixels of the objects in the NIR reflectance (R_rc at 859 nm), (3) the mean reflectance difference between the object and a calculated shift object in the NIR band, (4) the size of the object, and (5) whether or not the object was located near land.

The discrimination of oil slicks was separated for bright and dark objects. For the discrimination of dark targets, false positive targets included surface roughness anomalies and cloud shadows. Sea surface roughness anomalies were ruled out through the object contrast with the background water. While sea surface roughness anomaly regions in low-wind conditions usually display an extensive dark area with scattered dark spots in the boundary area in SAR imagery, oil-slick-covered areas tend to show a sharp contrast at the slick boundary [63]. A similar phenomenon can be observed in optical imagery under sun glint. Therefore, the object boundary tends to display a larger contrast at the oil-slick-covered area than at the surface roughness anomaly area. In this study, the NIR mean reflectance difference between the inner boundary pixels (brown pixel with a red frame in Figure 7b) and the outer boundary pixels (blue pixel with a black frame in Figure 7b) of the bright/dark objects, i.e., MDN (mean difference to neighborhood pixels), was used to discriminate oil slicks from the surface roughness anomaly area. Under the same sun glint condition, the contrast between cloud and cloud shadow was larger than that between oil slicks and the background water. Due to the different local overpass times of MODIS Terra and Aqua, cloud shadow in the study area is usually observed on the west side of the cloud with Terra and the east side of the cloud with Aqua. Such cloud shadow was recognized through the reflectance difference in the NIR band between the bright and dark objects. A potential cloud region was searched by shifting the dark objects. Taking the Aqua image as an example, a potential cloud area (purple pixels with a black frame in Figure 7a) was searched by shifting N pixels to the west of the dark object (gray pixels in Figure 7a). N is the average value of the width of the dark object. The maximum value of N was set to 5, as a statistical analysis showed that the mean width of cloud shadows in the study area rarely exceeded 5 pixels. The mean difference in the NIR reflectance between dark objects and the potential cloud area, i.e., MDC (mean difference to cloud object), was calculated. Through a statistical analysis, we found that the larger the L_GN, the larger the threshold values of MDN and MDC. The MDN and MDC parameter threshold settings for dark oil detection under different L_GN values are shown in Figure 8. The dark objects from shear zones and the leeward side of the island were discriminated by checking whether the dark object was the neighborhood of the land mask. If any pixel within the dark object was less than a two-pixel distance from the land mask, the dark object was identified as a non-oil object.

False-positive targets in bright objects include clouds and water under strong sun glint. Large-size clouds were identified through the auxiliary SST data, as the difference in the SST between cloud and seawater was significantly larger than that between oil slicks and seawater [64]. If the mean SST of the bright object (SST_o) differs from the median value of the auxiliary SST of the image by a threshold SST_d, the object would be identified as a non-oil object. However, such cloud discrimination cannot be applied to clouds of small size because SST data with a resolution of 1 km are not sufficient to detect small-sized clouds. Instead, small-sized clouds were discriminated using procedures similar to the cloud shadow discrimination. Taking the Aqua image as an example, a potential cloud shadow area (gray pixels with a black frame in Figure 7c) was searched by shifting the bright object (purple pixels in Figure 7c) N pixels eastward without intersecting with the bright object. The mean difference between the bright object and the potential cloud shadow in the NIR band reflectance, i.e., MDCS (mean difference to cloud shadow object), was calculated. The MDCS in bright objects was lower for the oil slick area than for the cloud area. False positives from water under high sun glint were discriminated from the bright oil through the MDN, with the same principle as that of the discrimination of the roughness anomaly seawater. Through a statistical analysis, we found that the larger the L_GN, the larger the threshold values for the MDN and MDCS parameters. The positive-contrast oil slick identification rules and the threshold settings of the SST_d, MDN, and MDCS parameters under different L_GN values are shown in Figure 9.

Finally, small-sized clouds or cloud shadow objects that were not identified through the above processes were excluded through a sliding window method. The sliding window length/width and step were set to 1/6 of the image length/width, and the threshold was set to 30 pixels. The average size of the objects in the sliding window was calculated. If the average size was less than the threshold, the objects in the window were discriminated as non-oil objects. By this means, the dispersed cloud and associated cloud shadows were removed.

3.5. Performance Evaluation

When evaluating the performance of the approach, both the positive contrast and negative contrast oil objects were regarded as oil targets, while other targets were regarded as background targets. To objectively evaluate the performance of the approach, the oil slicks in the images were visually interpreted and manually delineated in ArcMap software (version 10.8, Environment System Research Institute, Redlands, CA, USA) [33]. These interpreted oil slicks were used to assess the performance using the approach proposed in this study. Three evaluation metrics widely used for target recognition were chosen in this study for evaluation [65,66,67]: precision, recall, and F1-score, which are defined as

$$Precision=\frac{TP}{TP+F{P}^{\prime }}$$

(2)

$$Recall=\frac{TP}{TP+F{N}^{\prime }}$$

(3)

$$F_1=\frac{2\times TP}{2\times TP+FN+F{P}^{\prime }}$$

(4)

where TP (true positives) denotes the number of target objects (pixels) classified into the correct category, TN (true negatives) is the number of non-target objects classified into the correct category, FN (false negatives) is the number of target objects classified into the wrong category, and FP (false positives or false alarms) is the number of objects from a non-target category incorrectly classified as a target. High precision indicates high detection accuracy, and high recall suggests a high number of positive objects detected. F₁ is an indicator used in statistics to measure the performance of a binary model, which considers both the precision and recall of the classification results, and it can be regarded as a weighted average of model precision and recall with a maximum value of 1 and a minimum of 0.

4. Results

4.1. Optical Characteristics of Oil Spill

With the same glint strength at the oil–water boundary, the oil and water contrast can be clearly observed due to the oil’s modulation of roughness (Figure 10a–c). The images in Figure 10a,b were taken at different times on the same day. The sun glint strength decreases from left to right in Figure 10a (0.052~0.002 sr⁻¹), while the opposite is found in Figure 10b (0.0002~0.025 sr⁻¹). Different oil–water contrasts can be found at the same oil slick, such as dark oil (blue circle point in Figure 10a) and bright oil (green circle point in Figure 10b). Here, the oil slick or area of the slick in the negative contrast with the water is referred to as dark oil, while the positive contrast slick or area of the slick is referred to as bright oil in this paper.

Oil look-alikes can also display similar positive/negative contrast characteristics, such as clouds, water under high sun glint, cloud shadows, and roughness anomaly seawater. Oil slicks and oil look-alikes under different sun glint conditions are presented in Figure 10a–c, and the corresponding reflectance spectra of the selected targets are displayed in Figure 10d. When sun glint strength differs, reflectance spectra of different targets may resemble each other (Figure 10d). For example, the sea surface roughness anomaly (Figure 10c, pink square point) and cloud shadow (Figure 10a, pink triangle point) resembled oil (Figure 10a, blue circle point) spectra in shape, and their reflectance magnitude was also close to that of oil slicks. This dark roughness anomaly region was caused by the local low wind (2.2 m/s in the pink square point vs. 4.5 m/s in the green square point; reflectance spectra displayed in Figure 10d). Seawater under high sun glint (Figure 10b, red square point) and clouds (Figure 10c, red triangle point) resembled oil (Figure 10c, red circle point) spectra in shape, with the reflectance magnitude also being close to that of oil slicks (Figure 10d). The target reflectance magnitude and shape also changed with the sun glint strength. As can be seen from the selected water locations in Figure 10, the water reflectance spectra decrease in all spectral bands with the decreased sun glint strength (L_GN = 0.04 sr⁻¹ at the red square point vs. 0.03 sr⁻¹ at the green square point vs. 0.02 sr⁻¹ at the blue square point).

4.2. Extracted Oil Slicks

Figure 11 shows the results of the extracted oil slicks using the approach proposed in this study. Most of the detected oil slicks were within 100 km of the platform location for the duration of the spill, with almost no slicks observed 210 km away from the platform. The detected area of oil slicks ranged from 1131 to 5310 km² in the MODIS images, with an average area of 3314 km². It can be seen that most of the oil slicks were sufficiently extracted, with false positive targets in the figure (clouds, water under strong sun glint, cloud shadows, and sea surface roughness anomaly area) being sufficiently removed. The oil slicks in Figure 11a,b,e,g are characterized by many dispersed slick patches in the image. They are composed of either one major slick with small scattered oil patches (Figure 11e,g) or several large slicks with small scattered patches that are distributed some distance from each other (Figure 11a,b,e). This approach achieved a good performance in extracting these types of oil slicks, with some minor omissions, demonstrating the approach’s capabilities in dealing with oil slicks of multiple scales. Figure 11c,d,f,i present the negative contrast of oil slicks with seawater. Oil slicks in these images are composed of one or several large patches close to each other. As can be seen, the extraction results show a clean slick delineation and are very close to the manual interpretation results. An oil slick with positive contrasting features can be observed in Figure 11h, and both positive and negative contrasting features can be observed in Figure 11e,g. The overall extraction results are strong, with all major slicks being identified, but with some false positive detections from clouds and cloud shadows (Figure 11e,g,h).

4.3. Performance Evaluation

Table 2. Evaluation of the extracted oil spills. Data and time are in coordinated universal time (UTC).

Serial Number	Date	Time (hh:mm)	Sensor	Area of the Oil Spill by the Approach (km2)	Precision	Recall	F1-Score
a	30 Aug 2009	05:20	Aqua	1468.44	0.969	0.861	0.912
b	17 Sep 2009	02:15	Terra	3338.06	0.981	0.806	0.885
c	24 Sep 2009	05:15	Aqua	5310.13	0.93	0.863	0.895
d	24 Sep 2009	02:15	Terra	4055.06	0.939	0.873	0.905
e	03 Oct 2009	02:15	Terra	3559.56	0.848	0.837	0.843
f	10 Oct 2009	05:15	Aqua	4293.94	0.877	0.826	0.851
g	17 Oct 2009	05:20	Aqua	1131.44	0.639	0.783	0.704
h	19 Oct 2009	02:15	Terra	2521.19	0.448	0.809	0.576
i	21 Oct 2009	02:05	Terra	2208.38	0.992	0.78	0.873
j	30 Oct 2009	01:55	Terra	5255.06	0.423	0.879	0.571
Overall	/	/	/	/	0.8	0.838	0.818

displays the quantitative assessment of the extraction results against the manually delineated slicks through visual interpretation. The statistical results demonstrate an overall precision value of 0.8, which means that 80% of the oil pixels were detected correctly. The overall recall value of 0.838 demonstrates that 83.8% of the oil spill pixels in the image were successfully detected.

For individual images, the recall of the approach ranged from 0.780 to 0.879. The main factors affecting the recall included the existence of clouds in the image and inhomogeneity within the oil slick. Cloud cover on top of an oil slick causes it to be divided into several parts. However, the dark oil objects would be mistakenly identified as cloud shadow if the relationship between the dark oil and cloud objects meets the third rule in Section 3.4, as shown by the omitted oil slicks in the regions marked by red ellipses in Figure 11a,b,i,j. The coexistence of bright and dark oil objects would also lead to similar effects described above (recognized as clouds and cloud shadow), leading to the omission of oil slicks in the extraction results (Figure 11e,f,g).

The precision of the extracted results ranged from 0.423 to 0.969 for individual images. Among them, three of the images had a precision of <0.8. The main factor affecting the precision is the sea surface roughness anomaly adjacent to the clouds, such as the regions marked by a red rectangle in Figure 11. Although there were only a few misidentified objects, the impact on the precision of the approach was great because of the large area of misidentified objects.

5. Discussion and Conclusions

The detection performed excellently for situations where the oil slick distribution in the image is concentrated, and where the oil spill area is not covered by clouds, as shown in Figure 11a–d,f,i. The extraction results of these situations display a good performance, with F1-scores of over 0.85. In circumstances where there is a multiple-target distribution in the image, e.g., the coexistence of surface roughness anomalies and clouds, the extraction result shows a low F1-score due to a large number of detected false positives and the small number of undetected false negatives. As shown in Figure 11h,j, the extraction results of these two situations are the worst among the 10 images, with F1-scores of 0.576 and 0.571.

As presented in this study, precise cloud/shadow detection and masking before the segmentation can significantly improve the performance of the approach proposed in this study. However, the MODIS cloud mask products would mistakenly recognize the bright oil slicks as clouds, prohibiting the application of such a cloud mask product to the approach, which calls for more precise cloud mask algorithms under strong sun glint conditions.

In conclusion, an object-based approach was proposed in this study to extract massive oil slicks in sun glint MODIS imagery of the Montara oil spill that occurred in the Timor Sea in 2009. After appropriate imaging preprocessing procedures including Gaussian stretching, median filtering, and sun glint gradient modeling, the SLIC and RAG algorithms were used to segment the image into super-pixels and to merge the neighborhood super-pixels into target objects with varying sizes. False positive detections from the objects were ruled out through a decision tree using object spectral and spatial characteristics. The approach achieved a good performance in extracting the oil slicks from the Montara oil spill, with a slick extraction precision of 0.8, recall of 0.838, and F1-score of 0.818. The images used in this study cover a broad range of situations including different sun glint strengths and oil slicks of multiple sizes with various distribution patterns, accompanied by various oil slick look-alikes. The performance assessment proved the robustness of the approach, which represents a useful approach for oil slick extraction using sun glint optical images during massive oil spills. Moreover, as the approach proposed in this study is based mainly on reflectance data in the 859, 645, and 555 nm bands, which can be commonly found in other optical sensors, this approach may apply to other optical images as well.