Monitoring of Spatio-Temporal Variations of Oil Slicks via the Collocation of Multi-Source Satellite Images

Abstract

Monitoring oil drift by integrating multi-source satellite imagery has been a relatively underexplored practice due to the limited time-sampling of datasets. However, this limitation has been mitigated by the emergence of new satellite constellations equipped with both Synthetic Aperture Radar (SAR) and optical sensors. In this manuscript, we take advantage of multi-temporal and multi-source satellite imagery, incorporating SAR (Sentinel-1 and ICEYE-X) and optical data (Sentinel-2/3 and Landsat-8/9), to provide insights into the spatio-temporal variations of oil spills. We also analyze the impact of met–ocean conditions on oil drift, focusing on two specific scenarios: marine floating oil slicks off the coast of Qatar and oil spills resulting from a shipwreck off the coast of Mauritius. By overlaying oils detected from various sources, we observe their short-term and long-term evolution. Our analysis highlights the finding that changes in oil structure and size are influenced by strong surface winds, while surface currents predominantly affect the spread of oil spills. Moreover, to detect oil slicks across different datasets, we propose an innovative unsupervised algorithm that combines a Bayesian approach used to detect oil and look-alike objects with an oil contours approach distinguishing oil from look-alikes. This algorithm can be applied to both SAR and optical data, and the results demonstrate its ability to accurately identify oil slicks, even in the presence of oil look-alikes and under varying met–ocean conditions.

1. Introduction

The recent expansion of the global economy has led to a significant increase in maritime traffic and offshore oil production, causing multiple negative consequences for marine life and impacting coastal activities and the people who rely on sea resources. According to [1], oil spills, in particular, have a profound effect on marine biodiversity, harming animals and plants in two main ways: through the toxic effects of the oil itself and the response or cleanup operations. Spilled oil is harmful because its chemical constituents are poisonous, affecting organisms internally through ingestion or inhalation and externally through skin and eye irritation. Additionally, oil can smother small species of fish or invertebrates and coat birds’ feathers and mammals’ fur, reducing their ability to maintain their body temperatures. Chemical toxins from oil spills can also devastate large areas of seagrass and coral reefs, which are crucial for marine biodiversity. Cucco et al. [2] conducted numerical and experimental studies to quantify the risk of oil spill impacts on biodiversity in the Tuscan Archipelago, a high oil-spill-density area in the Western Mediterranean. Their research highlighted a direct relationship between the temporary reduction in maritime traffic due to pandemic restrictions and the likelihood of oil spill damage in the archipelago. Consequently, detecting oil spills and monitoring oil drifts in near real-time contexts have become critical for many coastal nations.

Oil spill detection has significantly improved in recent years, thanks to the increased use of remote sensing devices that can provide high-precision images. Among these technologies, Earth Observation (EO) satellites equipped with Synthetic Aperture Radar (SAR) and optical sensors are among the most widely used, due to their high spatial resolution and wide swath. In general, SAR can image sea surface roughness at small scale through the Bragg scattering resonance mechanism. As oil slicks appear on the sea surface, their viscosity dampens small-scale surface waves, and therefore the backscattering intensity captured by radars on the oil-contaminated sea surface is low (they are seen as dark objects) [3,4,5]. The primary challenge in SAR-based oil detection lies in distinguishing between actual oil and look-alikes, as both often exhibit similar radar backscatter or Normalized Radar Cross Section (NRCS), appearing as dark objects in different SAR images, including C-band Sentinel-1 [6], X-band Cosmo-SkyMed [7], and small SAR ICEYE-X [8], as well as polarimetric SAR [9]. The detection of oil spills in SAR images generally comprises segmentation, feature extraction, and classification procedures [10], in which the segmentation is the first step, employed to detect oil and look-alikes, while the discrimination between oil and look-alikes involves leveraging different features calculated from NRCS [11] and employing classification methods such as adaptive thresholding [12], machine learning (ML) [13], and deep learning (DL) [11,14]. In order to cope with the speckle, the segmentation and detection of the dark areas are usually performed patch-wise, either using traditional methods like Markov-random field [15] or semi-empirical threshold modelling [16], or using DL approaches, e.g., YOLOv8 [17] (Wu et al., 2024) and YOLOv4 [18] object detectors. A significant challenge with supervised approaches lies in the requirement to encompass a wide array of SAR oil spill scenes, including those obtained under challenging conditions in which look-alike examples are also present, in order to effectively train the model. To solve this issue, recent efforts have focused on augmenting the training data. One such approach proposed by [19] involves the development of a self-evolving deep learning model. Additionally, they utilized an adaptive thresholding method to generate supplementary high-quality training data.

In addition to SAR, oil spills can be detected from optical satellite imagery, e.g., Sentinel-2 [20], Landsat-8 [21], and MODIS [22]. Depending on the oil thickness, oil slicks may appear as dark or bright objects in optical images [23]. The oil index computed from reflectance values of spectral bands has been widely used for oil spill detection. It can be computed from the ratio of blue and SWIR bands [20], or from the combination of red–green–blue bands [24,25]. Just as with SAR data, the detection of oil slicks is based on a three-stage approach: (1) identification of darker or brighter regions with respect to a background; (2) feature extraction from identified dark or bright regions; and (3) identification of oil slicks and look-alikes [26]. This is implemented using classification methods ranging from more conventional approaches such as adaptive thresholding [20] to those based on deep learning [27,28]. Classification approaches dealing with optical data share many of the challenges that SAR-based methods face.

Apart from detecting oil slicks, there has been a recent increase in interest in monitoring oil drift. This involves leveraging various types of satellite imagery, which is especially crucial for rescue operations. However, only a limited number of studies [29,30,31] have delved into this area. Study [31] proposed the combination of multi-band SAR and optical images to monitor floating oil drift off the coast of Norway. While yielding new insights, the proposed approach faces challenges when applied at a large scale due to limited access to critical devices, such as UAV-equipped SAR and optical sensors. Likewise, the low quality of some airborne images makes it difficult to observe variations in oil shapes. The study focused on short-term observations (6–11 h), lacking overnight measurements. Study [32] used Sentinel-1 images acquired on three different days (10, 16, and 22 August 2020) to observe oil spills caused by a shipwreck accident near Mauritius Island. However, a significant time lag between these Sentinel-1 images makes the evaluation of the impact of met–ocean variables on oil drift more complicated, as various factors, e.g., oil-slick cleanup and a discontinuity of the oil leak, may emerge between two sequential Sentinel-1 acquisitions and thus impact oil-drift observation.

The usage of multi-sensor data can significantly enhance the timeliness of oil spill detection, thereby improving the monitoring of oil-slick evolution and drift. While supervised machine learning methods can achieve high performance in oil-slick detection when models are properly specified and trained, they are susceptible to domain and distribution shifts. For example, a neural network trained on SAR data may not perform well on optical data. Even within the same data modality, such as training on Sentinel-1 data and testing on ICEYE-X, performance degradation can occur due to differences in wavelength, spatial resolution and radiometric calibration. Collecting training data from multiple sensors is both time-consuming and labor-intensive. Furthermore, acquiring sufficient training data from commercial sensors is challenging and costly. These challenges can impede the adoption of supervised algorithms for oil-slick mapping and monitoring using multiple sensors. In this paper, we introduce an unsupervised and sensor-agnostic method for oil spill detection that can be applied across various sensors (e.g., both SAR and optical). This method can detect oil slicks over large areas in a short time frame. To the best of our knowledge, this is the first study to utilize multiple sensors for oil spill detection and to analyze oil evolution and drift with auxiliary data on wind speed and sea current velocity.

To detect oil slicks from SAR and optical satellite datasets, we developed a two-step algorithm that integrates an object detection algorithm, known as the Hierarchical Split-Based Approach (HSBA) [33], with oil contour determination. Originally designed to categorize image pixel values into “target” and “background” classes (suitable for delineating oil slicks), HSBA has been extensively utilized across various applications [34,35,36]. HSBA can be applied to different satellite imagery types (SAR and optical) and used to identify oil objects of different sizes within an entire image, achieving this in an unsupervised manner. Indeed, the areas affected by oil spills typically occupy only a fraction of an image, and therefore, leveraging adaptive object detection techniques rooted in local information becomes crucial. To differentiate oil-slick objects from similar-looking ones, we propose the use of oil contours as an additional step. This refinement is particularly crucial in complex scenarios where both oil objects and look-alikes (such as low-intensity wind areas, algae/seaweed, coastal freshwater, convective cells, clouds, and high-intensity reflectance areas) exhibit comparable input values, potentially leading to misclassifications by the HSBA algorithm.

Leveraging the availability of different satellite sensors, this paper focuses on the development of a multi-source systematic and automatic monitoring system from various image types. Most datasets used in this paper (i.e., Sentinel-1/2/3 and Landsat-8) can be downloaded at no cost via public data sources. Moreover, they have a high spatial resolution, e.g., a 10 m pixel-size for Sentinel-1/2, and a wide swath, e.g., 250 km (Sentinel-1) and 290 km (Sentinel-2). Over several regions (e.g., the Persian Gulf) where the cloud impact is less significant, we can obtain a high number of optical images enabling oil spill detections at a high frequency. In particular, due to the crossing of Sentinel-1/2 and Landsat-8/9, we can combine the diverse images to monitor oil drift both in the short term (spanning several hours) and the long term (over 24 h). The met–ocean data used for evaluating these influences on the spatio-temporal variations of the oil spills can be obtained through the CMEMS program [37,38]. Although their spatial resolution may be relatively large (25 km for surface wind [37] and 9 km for current [38]), they can offer data with a high temporal resolution (one hour for wind and 0.5 h for current) for most regions of the world. Other high-resolution met–ocean data sources can be considered for some specific regions, if necessary. Specifically, in this paper, we combine Sentinel-1 (SAR), Sentinel-2, Landsat-8/9 (optical), and Sentinel-3 (optical) images for the detection of floating oil slicks and the observation of their short-term (4–8 h) and long-term (24–32 h) evolution in the Persian Gulf (off the coast of Qatar). Likewise, for emergency cases like the shipwreck off the coastal region of Mauritius, we collocate SAR data, including Sentinel-1 IW (Interferometric Wide), EW (Extra Wide), and ICEYE-X, to observe the oil spills emanating from a stationary source (shipwreck) at a given time and every 13 to 21 h thereafter. To estimate the short- and long-term directions, distances, and velocities of oil drifts, we superpose the detected oil slicks from different satellite data. This approach also allows us to observe changes in the shape and size of the oil entities.

Section 2 presents the collocation and pre-processing of various satellite images (Sentinel-1/2/3 and Landsat-8) for oil-slick detection off the shore of Qatar and Sentinel-1 IW/EW and ICEYE-X for oil spill observation off the shore of Mauritius. Section 3 presents the methodology for detecting dark areas and oil contours. Section 4 and Section 5 illustrate the oil-slick evolution observation in areas offshore of Qatar and Mauritius. Section 6 validates the proposed algorithm. Section 7 and Section 8 comprise the Discussion and Conclusion.

2. Data Preparation

2.1. Data Collocation

This study utilized different satellite images for oil-slick detection and oil-drift observation, including SAR (C-band Sentinel-1 IW/EW and X-band ICEYE-X) and optical (Sentinel-2, Landsat-8, and Sentinel-3) technologies. The SAR images are acquired with vertical co-polarization (VV-pol). The spatial resolution (pixel sizes) and swath widths of these data are described in

Table 1. Description of multi-source satellite imagery used for oil-slick detection.

Sensor	Type	Mode/Band	Pixel Size	Swath	Repeat Cycle
Sentinel-1	SAR (C-band)	IW (VV-pol)	10 m	250 km	6 days
		EW (VV-pol)	30 m	400 km
ICEYE-X	SAR (X-band)	Strip mode (VV-pol)	3 m	80 km	1–22 days
Sentinel-2	Optical	Red-Green-Blue (RGB)	10 m	290 km	5 days
Landsat-8	Optical	Red-Green-Blue (RGB)	30 m	185 km	16 days
Sentinel-3	Optical	OLCI	300 m	1270 km	27 days

. The ICEYE-X images have the highest resolution (3 m) but smallest swath (80 km), while Sentinel-3 offered the lowest resolution (300 m) but the largest swath (1270 km). Sentinel-1 SAR (IW) and Sentinel-2 (RGB) optical images have a similar resolution (10 m) and close swaths (250 km and 290 km).

The long repeat cycle of most satellites shows that the collocation of different sensors proposed in this study is an effective solution for detecting oil slicks and observing oil drifts for short- and long-term durations.

Table 2. Datasets of all satellite images for oil-slick observation in areas offshore of Qatar and Mauritius.

Floating Oil Slick		Offshore Qatar
Case study	Date	Sensors and observation time (UTC)			Time lag (hour)
#Q1	27–28 March 2021	Sentinel-2 06:56:21 27/03	Sentinel-3 06:34:43 28/03	Sentinel-114:33:0628/03	S-2/S-3: 24 h
					S-3/S-1: 8 h
					S-2/S-1: 32 h
#Q2	5–6 July 2021	Sentinel-1 02:23:25 05/07	Sentinel-2 06:56:21 05/07	Landsat-8 06:58:26 06/07	S-1/S-2: 4 h
					S-2/L-8: 24 h
					S-1/L-8: 28 h
#Q3	3 September 2021	Sentinel-1 02:23:54	Sentinel-2 06:56:21		S-1/S-2: 4 h
Fixed-source oil slick		Offshore Mauritius
10–11 August 2020		Sentinel-1 IW 01:37:55 10/08	Sentinel-1 EW 14:36:16 10/08	ICEYE-X 11:12:41 11/08	IW/EW: 13 h
					EW/ICE: 21 h
					IW/ICE: /34 h

presents the collocation of various satellite data for observing the evolution of oil slicks in time and space. For offshore Qatar (Figure 1a), we focused on three cases (#Q1–3) based on the different combinations of Sentinel-1/2/3 and Lansat-8 images. Case #Q1 combines Sentinel-2/3/1 images with time lags of about 24, 8, and 32 h, respectively. Case #Q2 collocates Sentinel-1/2 and Landsat-8 images with differences of approximately 4, 24, and 28 h, respectively. Finally, case #Q3 combines Sentinel-1/-2 images with a time lag of about 4 h. To observe the oil spills offshore Mauritius (Figure 1b), we combined Sentinel-1 IW/EW and ICEYE-X images, with time lags of approximately 13, 21, and 34 h, respectively.

2.2. Data Preprocessing

All Sentinel-1/2/3, Landsat-8, and ICEYE-X images collected were preprocessed to obtain the required inputs, namely, the Sentinel-1/ICEYE-X NRCS, Sentinel-2/Landsat-8 oil index, and Sentinel-3 T865 variable [39].

To obtain the NRCS used for oil-slick detection, the Level-1 Sentinel-1 and ICEYE-X SAR images were preprocessed, applying the following steps:

• Border and thermal noise removal [40];
• Radiometric calibration [41];
• Speckle noise filtering [42];
• Land masking;
• Terrain correction.

The images were filtered using a Boxcar filter with a window size of 5 × 5 pixels, which represents a good compromise between preservation of spatial detail and signal-to-noise ratio [42]. Figure 2 illustrates the NRCS elements (in dB) of Sentinel-1 and ICEYE-X images after pre-processing. Oil slicks on the two images are observed as dark objects with low NRCS.

Optical oil index = (Red + Green + Blue)/3

(1)

Regarding optical imagery, Equation (1) delineates the calculation of the optical oil index selected, i.e., the average values of Red (R), Green (G), and Blue (B) bands extracted from Sentinel-2 or Landsat-8 Level-2 images.

We tested different RGB combinations, including those proposed in the referenced studies [20,24,25]. We selected the averages of the RGB bands in order to preserve the oil features observed on RGB images, i.e., dark and bright pixels, depending on the oil thickness (Bonn Agreement Oil Appearance Code), and reduce the number of bands to be processed. Figure 3 presents an oil slick, as captured in a Sentinel-2 RGB image (extracted scene) and via the oil index (in dB) calculated from the averages of RGB bands. In the two images, the oil slick consists of dark (low dB) and bright pixels (high dB). The difference in oil pixel values is likely attributable to variations in oil thickness. This topic is not explored here in detail because it falls beyond the paper’s scope. Instead, we focus on detecting dark and bright-oil pixels using the average RGB oil index.

In addition to evaluating Sentinel-2 and Landsat-8, we explored various combinations of Sentinel-3 bands and identified the T865 variable (extracted from Sentinel-3 Level-2 data) as being particularly relevant for this investigation. Figure 4a presents an oil slick observed on a Sentinel-3 OLCI tristimulus image with RGB channels obtained from the combinations of different bands, i.e., bands 1–10 for red, 3–10 for green, and 1–5 for blue. Figure 4b illustrates the T865 variable [39] (in dB), with the oil slick appearing as a dark object (low dB).

3. Methodology

The oil-spill classification is split into three subsequent steps: (1) dark, in SAR and optical data, and bright, in optical data, object detection; (2) feature extraction; and (3) oil-spill and look-alike separation. Below, in Figure 5, we detail how our method approaches these steps in order to detect oil slicks across three distinct datasets, including SAR and optical imagery).

(1) The first task can be framed as a classification problem, distinguishing between the target class (oil-slick and look-alike) and the background (oil-free sea). The oil films on the sea surface are observed as dark objects in Sentinel-1/ICEYE-X images (T1), dark and bright pixels (T2 and T3) on Sentinel-2/Landsat-8 images, and dark objects on Sentinel-3 images (T865 variable) (T4).

This involves binary classification, with the goal of distinguishing the target (T1–T4) from the background (BG). One often-used method for this task is histogram thresholding, which involves selecting an appropriate threshold. This step has a direct impact on the classification outcomes [43]. Assuming both T and BG follow a Gaussian distribution, parametric methods involve estimating parameters of the following Gaussian mixture density [44], or Probability Density Function (PDF):

$$\begin{array}{ll}PDF\left(x\right)& ={\omega }_T{PDF}_T\left(x\right)+{\omega }_{BG}{PDF}_{BG}\left(x\right)\\ & ={\omega }_T{\displaystyle \frac{1}{\sqrt{2\pi {\delta }_T^2}}}{e}^{-{\displaystyle \frac{{\left(x-{\mu }_T\right)}^2}{2{\delta }_T^2}}}+{\omega }_{BG}{\displaystyle \frac{1}{\sqrt{2\pi {\delta }_{BG}^2}}}{e}^{-{\displaystyle \frac{{\left(x-{\mu }_{BG}\right)}^2}{2{\delta }_{BG}^2}}}\end{array}$$

(2)

${\omega }_T$ and ${\omega }_{BG}$ are weights of the two distributions, ${\mu }_T$ and ${\mu }_{BG}$ are means, and ${\delta }_T^2$ and ${\delta }_{BG}^2$ are standard deviations of two distributions, respectively.

The accuracy of classification is significantly impacted by the proportions of classes present in the scene and the degree of overlap between PDF_T and PDF_BG. When there is a significant imbalance between T and BG, it becomes challenging to accurately parameterize their PDFs. The degree of overlap between the PDFs directly impacts the accuracy of under- and over-detection. To address the two mentioned limitations, we employed an adaptive threshold method that was previously established for mapping floodwater [33]. The technique consists of two key steps. The algorithm begins by parameterizing the PDF_T and PDF_BG. It then iteratively uses thresholding and region-growing techniques to determine the optimal threshold for seeds (THS) and the threshold for stopping the region growth process (THSRG). The parameterization of the PDFs is carried out using HSBA [33], which finds specific regions, or tiles, of the image where the PDF_T and PDF_BG can be fitted with more reliability, meaning that the histogram of a specific tile is clearly bimodal. The size of the tiles is determined by the ability to parameterize the PDFs associated with two distinct classes. HSBA begins by analyzing the entire image and subsequently decreases the size of the tiles by a quad-tree decomposition of the image (Figure 6).

In the second stage, spatial information is incorporated into the selection of the optimal threshold to mitigate the impact of class overlap on the final classification. The latter is achieved by a region-growing method, assuming that the pixels forming T are clustered rather than distributed randomly over the entire image. Thus, we initially categorize as T those pixels that exhibit a high probability of belonging to T. Subsequently, we include in T those pixels that have a lower degree of belonging to T but are spatially connected to the initial predictions. In order to accomplish this, we employ the region growing method, in which the PDF_T and PDF_BG will determine the selection of THS and THSRG. We can assign the threshold (THS) to the average value of the PDF_T, which represents pixels that have a high likelihood of belonging to T. To choose THSRG, many thresholds are evaluated. The selection is made by minimizing the root-mean-squared error between the PDF_T and the histogram produced by the region growth process. The two thresholds are automatically determined within the regions specified by HSBA, and subsequently applied to the full image to obtain the ultimate classification.

The algorithm is applied separately to detect the four different targets, T1–T4, which reflect the oil-slick/look-alike class in the different input images.

Satellite images must be preprocessed before being used as input for the proposed algorithm, including the determinations of NRCS (from Sentinel-1/ICEYE), average RGB oil index (from Sentinel-2/Landsat-8), and T865 variable (from Sentinel-3). These inputs are normalized and converted into dB with the mean value of each scene in order to remove anomalies and possible changes in the acquisition conditions, i.e., incidence angle and wind speed/direction for SAR data, as well as sun glint for optical images. For instance, the NRCS decreases with increasing incidence angles and reduced wind speed. Likewise, strong surface winds may significantly affect the homogeneity of oil films, which can produce a large difference in NRCS intensity of oil pixels, as shown in Figure 2a.

(2) In complex scenarios where oil objects and look-alikes have similar input values, HSBA may struggle to distinguish between them. To differentiate between oil slicks and their look-alikes, we have considered the variation in backscattering coefficients or surface reflectance near the boundaries of the oil as a crucial parameter for distinguishing them. Therefore, to enhance the HSBA-based oil-slick detection, we applied the determination of oil contours [11,45,46] as an additional technique to distinguish oil slicks from look-alikes. To this end, first, we utilized the two non-linear filters, i.e., mean filter [47] and standard deviation filter [48], to input data (NRCS, oil index, and T865 variable). The mean filter is defined as follows:

$$G\left(x,y\right)=\left\{{\left[\prod _{i,j\in S}S\left(i,j\right)\right]}^{{\displaystyle \frac{1}{mn}}}\right\}$$

(3)

$G\left(x,y\right)$ is the output image, $S\left(x,y\right)$ is the original image, and the filter mask is m by n pixels. This technique [47] helps to reduce the noise, which enhances the differences between dark/bright areas and the sea background. The mean filter is essential for object contour determination, and is based on the standard deviation filter [11,45,46] generally defined as follows:

$$K\left(x,y\right)=\left\{\sqrt{{\displaystyle \frac{1}{N}}\sum _{i,j\in S}^N{\left(G\left(i,j\right)-\overline{G}\right)}^2},\mathrm{W}\left(p,q\right)\right\}$$

(4)

$G\left(x,y\right)$ and $K\left(x,y\right)$ are the input and output images of the standard deviation filter in Equation (4), respectively. $N$ is the pixel number of the image/sub-image, and $\mathrm{W}\left(p,q\right)$ is the window size of the filter.

The selection of the parameters of the mean and standard deviation filters (i.e., window size and moving step) [47,48] is crucial to obtain satisfactory results. Although increasing the window size and moving step gives us images with reduced noise, there is a risk of removing some of the oil pixels. Therefore, the selection of the filter parameters should depend on the extents of the oil slicks.

(3) We use the adaptive thresholding approach [49] to identify oil contours in the T1-T4 classes identified by HSBA on the filtered images. In general, oil slicks exhibit higher contrast values compared to look-alikes, making adaptive thresholding an effective approach for classifying the two objects based on differences in contour values. In this final step, oil objects identified by HSBA are overlapped with oil contours determined using the two non-linear filters and adaptive thresholding. Segments identified by HSBA that lack any intersecting contours delineated in the second-step procedure are discarded as look-alike objects.

The introduced algorithm combines HSBA with oil contour determination, effectively addressing challenges in oil spill detection from satellite imagery. It leverages the strengths of both approaches, which is especially valuable in discerning between oil slicks and look-alikes and in scenarios where oil-slick homogeneity is affected by strong surface winds. Moreover, HSBA excels in automatically detecting oil-spill or look-alike objects in an unsupervised manner, regardless of the satellite image extension and swath width. It is important to note that these last two aspects are significant. In many patch-wise detection algorithms, determining the patch size or identifying the appropriate area for applying the segmentation algorithm are non-trivial tasks.

To study oil drifts, we collocate and superimpose oil slicks identified by the proposed method among different satellite datasets. Despite differences in spatial resolution, e.g., Sentinel-1/2 (10 m pixel-size) vs. Landsat-8/9 (30 m pixel-size), and oil features, e.g., Sentinel-1 SAR vs. Sentinel-2/Landsat-8/9 optics, observations of oil drifts in terms of changes in oil shape and location, as well as oil-drift direction and velocity, remain robust over short-term and long-term periods, allowing for evaluation of the effects of met–ocean conditions on oil drifts, using the CMEMS data for surface wind fields and current vectors.

4. Observation of Floating-Oil-Slick Evolution Off the Shore of Qatar

4.1. Case #Q1: 27–28 March 2021

Figure 7 presents the oil slicks detected for Case #Q1. On the Sentinel-2 image (Figure 7a-left), we applied HSBA and object contour determination twice, since the oil slick includes dark and bright pixels. In Figure 7a-right, the dark-oil pixels detected consisted of a white oil object and green oil contour, and the bright-oil pixels consisted of only a red oil contour. The HSBA could not classify the bright-oil object since its oil index values were very close to the background. The bright and background-value distribution is not bimodal [33]. The dark-oil part detected lacked some object pixels due to the inhomogeneity of the oil slick. One look-alike (on the left) with low oil index values was also detected by HSBA (in white); however, we can classify it as a look-alike due to the lack of oil contour. Figure 7b,c illustrate the oil slicks detected, including oil object and contour, from Sentinel-3/1 images, respectively. We realized that, as was the case with the oil slicks detected by Sentinel-2, the contours of the oil objects identified by Sentinel-1/3 are fragmentary due to inhomogeneity.

Figure 8 collocates Sentinel-2/3/1 images to observe the changes (in shape and position) of the oil slicks detected over about one day (Figure 8a), approximately 8 h (Figure 8b), and about 32 h (Figure 8c). To facilitate the oil-drift observation, we merged the oil objects and contours, as well as the dark and bright-oil pixels detected by Sentinel-2. In Figure 8a, the collocation of Sentinel-2/3 images indicates that the oil slicks detected moved about 18.5 km southeastward. Compared to the initial oil shape on the Sentinel-2 image, that identified on the Sentinel-3 image evolved significantly after about one day. It split into two parts, and many oil pixels disappeared, presumably due to strong winds and/or evaporation. In Figure 8a, the oil had moved slightly westward, by about 5.2 km, after 8 h. It is hard to accurately estimate the oil movement distance in this case because many oil parts did not move in exactly the same direction and with a similar velocity. The lack of many oil pixels also changed the oil shape significantly, probably due to strong winds and/or evaporation. Figure 8c indicates that the oil detected had moved southward after 32 h; however, it is hard to estimate the distance accurately due to the significant evolution in the oil shape.

Surface wind and current fields corresponding to the Sentinel-2/3/1 scenes (Figure 7a–c) are shown in Figure 9. Wind and current directions are displayed with the up-north reference (0°), meaning that wind or current blows or moves from south to north. From 06:00 to 21:00 on 27 March, wind intensity was relatively low (1–5 m/s), and the wind direction changed significantly between the northwest and northeast. Then, from 21:00 to 06:00 the following day, the wind speed quickly increased from 1 m/s to 10.5 m/s, blowing mainly southeastward. Meanwhile, the current also moved in different directions, from the southeast at 06:00 on 27 March, then southwest and northwest, at 18:00. From 21:00 on 27 March, to 06:00 on 28 March, it moved between the northeast and north. The wind and current observations indicate that the oil detected by Sentinel-2 moved in the same direction as the current (between the southeast and southwest) in the first phase (06:00–18:00, 27 March) since the wind was blowing between the northwest and northeast at that time with a low wind speed (about 2 m/s). Then, in the second phase, from 21:00 on 27 March, to 06:00 on 28 March, the oil moved in the same direction as the wind (southeast) since the current was moving between the northeast and north. In particular, the significant increase in wind speed (from 2 m/s to 10.5 m/s) in the second phase is assumed to be the main contribution to the oil drift toward the southeast. Likewise, strong wind is assumed to have split the oil detected in the Sentinel-2 into two parts, which were observed on the Sentinel-3 image.

From 06:00 to 15:00 on Mar. 28, corresponding to the oil slicks detected by Sentinel-3/1, the wind continued blowing southeastward, with an increase in intensity from 10.5 m/s to 12.5 m/s, while the current moved toward the northwest. Note that the oil shape detected on the Sentinel-3/1 images (quite) significantly evolved due to the strong wind. However, the oil-drift distance and direction between Sentinel-3/1 images could not be accurately determined since the two superimposed oil slicks were very close. We argue that the difference between wind and current directions, i.e., southeast vs. northwest, respectively, is the main reason for the oil-drift stagnation. While the oil tended to move in the same direction as the current (northwest), it was also impacted by the wind blowing southeast at a velocity of 12 m/s. As a result, one oil part moved in a northwesterly direction, while the other one moved in a direction between southwestward and westward.

Case #Q1 showed that it was not easy to accurately conclusively determine the wind and current impact on the oil drift for both short-term (i.e., several hours) and long-term (about one day) observations if the wind and current fields significantly changed, and/or the two did not move in the same direction. In addition, the large scale (and perhaps the oil thickness) of the oil detected was probably another factor affecting its movement. The results also showed that the oil-drift velocity and direction were mostly influenced by the current circulation, as the corresponding wind speed was low. As the wind blew with high intensity, e.g., increasing from 2 m/s to 10.5 m/s (Figure 9), it contributed more significantly to the movement of the oil slick (Figure 7a). Likewise, strong winds could impact the evolution of the oil shapes, i.e., splitting oil slicks into many parts, a consideration which also affected the detection of the oil objects and contours (lack of some oil pixels), as shown in Figure 7a–c. Additionally, as wind and current vectors were the main contributors to the oil drift, the difference between the wind and current directions might decelerate the oil movement.

4.2. Case #Q2: 5–6 July 2021

Figure 10 illustrates the oil slicks detected for Case #Q2. In Figure 10a, the proposed method can identify and delineate all oil objects with oil contours from the Sentinel-1 image. The oil slicks detected by Sentinel-2 include dark- and bright-oil pixels (Figure 10b). As for the dark-oil parts, they are classified as white (dark-oil object) or pink (dark-oil contours). The HSBA cannot identify some small oil parts and misclassifies some look-alikes. However, we can detect these look-alikes thanks to the lack of oil contours. As in the previous case, HSBA failed to identify bright-oil objects, although we can determine bright-oil contours (classified in red). Some cloud pixels are also misclassified in red since they have index values similar to the bright-oil ones. Figure 10c presents the oil slicks detected by Landsat-8, which only consisted of bright-oil pixels. We make some assumptions to explain the lack of dark-oil parts, including the coarse spatial resolution of Landsat-8 images (30 m vs. 10 m of Sentinel-2), sun glint conditions, and the evaporation of the dark-oil parts. The oil objects detected by Landsat-8 are classified as white (for oil objects) and green (for oil contours).

Figure 11 presents oil-drift observation by collocating Sentinel-1/2 and Landsat-8 images. The time lags between Sentinel-1/2, Sentinel-2/Landsat-8, and Sentinel-1/Landsat-8 are about 4, 24, and 28 h, respectively. To facilitate the oil movement observations, we delineated the small oil objects that were not detected by the proposed method. In Figure 11a, compared to the oil observed by Sentinel-1, the slick identified on the Sentinel-2 image (about 4 h later) had moved slightly, in a direction between westward and northwestward, by about 1.5–4.5 km (depending on different oil objects). Some small-scale oil objects disappeared, probably due to surface wind and evaporation, and their shapes evolved more significantly than the moderate-scale ones. Concretely, the long and narrow tail of the oil detected by Sentinel-1 fragmented into many small parts, as shown on the Sentinel-2 image, while the size of its head section was reduced. Figure 11b illustrates the oil movement after one day, from 5 July, 06:56:21 to 6 July, 06:58:26. The oil continued moving northwestward for a distance of about 12.3–12.9 km. Some small-scale oil objects seemed to have completely evaporated, and the moderate-scale ones reduced their size. Figure 11c indicates that the oil tended to move northwestward, with significant changes observed for both small- and moderate-scale oil objects after 28 h.

Figure 12 shows wind and current fields corresponding to the extracted Sentinel-1/2 and Landsat-8 scenes (Figure 10a–c) and their mean values. From 02:00–07:00 on 5 July, corresponding to the oil detected by Sentinel-1/2 (Figure 11a), winds blew between westward and northwestward, with speeds of 5.5–7.5 m/s. These wind values are quasi-ideal for oil detection from SAR and optical data. The current moves first in a direction between northwestward and northward, at 02:00, and then eastward, at 07:00, with velocities of 0.26–0.31 m/s. The more detailed observations in Figure 12iii show that the oil generally moved in the same direction as the wind and current, i.e., between west and northwest. From 07:00 on 5 July to the following day, the wind continued blowing between northwest and north, with little change in direction, while wind speed decreased from 8 m/s to 3 m/s and then increased again to 8 m/s. At the same time, while the current velocity increased a little, from 0.31 m/s to 4 m/s, its direction changed more significantly, from the east (07:00, 5 July), to the southeast and southwest, and then to the northwest and north (06:00, 6 July). According to this observation, the oil detected by Sentinel-2- and Landsat-8 oil tended to move in the same direction as the wind and current (06:00, 6 July).

Case #Q2 indicated that a wind speed of 5–8 m/s was an optimal condition for detecting oil slicks, especially for Sentinel-1 SAR (Figure 10a). Regarding their evolution, the small-scale oil objects quite significantly changed after several hours with the normal met–ocean conditions, i.e., wind speeds of 5–8 m/s and current velocities of 0.26–0.31 m/s. Evaporation was assumed to be the principal factor explaining the disappearance of the small-scale oil and the reduction in the size of the moderate-scale oil (after one day of observation). Compared to Case #Q1, the impacts of the wind and current on the (assumed) movement of the oil detected in Case #Q2 could be more easily observed, since both wind and current speeds were relatively stable. Furthermore, the wind and current directions had only a few changes after 28 h (except for the current from 06:00 on 5 July, to 00:00 on 6 July). Additionally, the small and moderate scales of the oil detected might be another factor used to more easily determine the oil-drift direction.

4.3. Case #Q3: 3 September 2021

Figure 13 presents the oil slicks detected for Case #Q3. The first scene (left) illustrates an oil spill leaked from a rig off the shore of Qatar, and the other two (middle and right) show floating oil slicks close to the first one. The method proposed can detect all oil slicks from the three Sentinel-1 scenes quite accurately, including oil objects and contours, which are classified in light and dark blue, respectively. In the second and third cases, HSBA misclassified the look-alikes (low wind-intensity areas) as oil objects; however, we could identify the oil slicks based on their oil contours. Just as in Cases #Q2 and #Q3, we applied the HSBA and oil contour approach to the Sentinel-2 dark- and bright-oil pixels twice (Figure 13c,d). For the dark-oil parts, HSBA could identify oil objects (in white) but misclassified some look-alikes (Figure 13d, scenes #1 and #3). However, the oil could be distinguished from look-alikes thanks to the oil contours classified in green. For scene #2 (Figure 13d, middle), as the oil is similar to the background, HSBA failed to identify it. Only the oil contours were identified and classified as green. One can observe the same result for the bright-oil pixel observation. As HSBA could not detect oil objects, only oil contours were detected and classified in red. The bright-oil part in scene #2 (Figure 13c, middle) only consists of oil contours.

Figure 14 presents the collocation of the Sentinel-1/2 scenes (Figure 13) for observing the oil drift after about 4 h. In the first scene (Figure 14a), it is hard to accurately determine the oil-drift direction based only on the changes in location between Sentinel-1/2 images since this oil is leaked from a rig and is not the same type as found in floating oil slicks. The size of the oil mass detected is generally reduced after about 4 h. Some light (or thin) oil pixels disappeared, probably due to evaporation. In the second scene (Figure 14b), the oil moved southward about 2.95 km. Many oil pixels detected by Sentinel-1 disappeared, probably due to evaporation, after 4 h. Practically, only the oil contours can be observed on the Sentinel-2 image. This observation strengthened the determination of oil contours used to distinguish oil from look-alikes. In the last scene (Figure 14c), the oil slicks, including small- and moderate-scale slicks, are assumed to have moved northeastward by about 3.75–4.8 km. Additionally, compared to scene #2 (Figure 14b) and Case #Q2 (Figure 11a), the shapes and sizes of the small- and moderate-scale oil slicks evolved less significantly in this case. They tended to extend vertically and narrow horizontally in the same direction as their movements. Some small oil parts were still present after 4 h.

Figure 15a–c presents wind and current fields and their mean values, corresponding to the extracted Sentinel-1/2 scenes (Figure 13). In the first scene (Figure 15–left), the wind blew between the north and northwest with a wind velocity of 2–4 m/s, and the current moved in the same direction. As noted in the comment above, as it is not easy to estimate the oil-drift direction in this case; we cannot accurately evaluate the impact of wind and current on the oil movement. In the second case (Figure 15–middle), the wind blew between the northwest and north with a velocity of 4–6.5 m/s while the current moved in a direction between the southeast and south, and the oil (Figure 14b) moved southward. The latter tended to move in the same southern direction as the current but stagnated due to the wind blowing in the northwest direction. In the last scene (Figure 15–right), the surface wind, current, and oil detected (Figure 14c) moved in the same northeastern direction. The wind velocity was relatively low (2–4 m/s), and this may explain the small changes in the oil shape, while the oil-drift distance is similar to the other scenes because the current speed is almost the same.

Case #Q3 showed that a wind speed of 3–6 m/s was optimal for oil-slick detection, especially from Sentinel-1 SAR images (Figure 13a,b). It also indicated that low wind speed (below 3 m/s) had little effect on the changes in oil shape and size (except for tiny objects) and contributed minimally to the oil drift, which was more influenced by the current. The impact of the moderate winds (4–6 m/s) on the oil drift (Figure 14b) was not well observed; however, it may affect the change in oil shape. Like Case #Q1, the oil spread more quickly if the wind and current moved in the same direction slowly, and stagnated when the directions of the wind and current were different.

5. Observation of Oil Spill Drift Off the Shore of Mauritius

Figure 16 presents the detection of oil spills associated with the MV Wakashio shipwreck off the shore of Mauritius in August 2020. This oil spill scenario differs from those observed off the shore of Qatar, which consisted of floating oil slicks. However, we applied a similar methodology (Figure 5) to this case. The oil detected included oil objects and contours as found for the floating oil #Q1–3. Figure 16a–c illustrate the scenes extracted (NRCS) from Sentinel-1 IW, EW, and ICEYE-X images for oil spill detection. The time lags between these images are about 13, 21, and 34 h. The ICEYE-X image has the highest spatial resolution (about 3 m), the Sentinel-1 EW image has the lowest (40 m), and the Sentinel-1 IW spatial resolution is 10 m. Figure 16d–f present the oil detected in the extracted scenes (Figure 16a–c). The oil detection based on the Sentinel-1-EW image is less accurate than the others, i.e., more false positives from the classified oil pixels, arguably due to its lower spatial resolution (40 m vs. 3–10 m), which makes the distinguishing between oil pixels and sea background more complicated and less accurate. Between Sentinel-1 IW and EW images (13 h time lag), the detected oil has undergone several changes; however, the oil surface, when comparing Sentinel-1 EW and ICEYE-X images (21 h time lag), is more extended along the coast.

Figure 17 presents wind and current fields corresponding to the three extracted scenes (Figure 16a–c) and their mean values. Due to the coarse spatial resolution of CMEMS data, we do not have accurate current fields near the coast. From 01:00 to 15:00 on 10 August, the wind was blowing in a northwestern direction with wind velocity relatively stable at about 6–7 m/s, while the current moved southwestward with current speeds of 0.2–0.35 m/s. The wind direction matched the oil-drift direction. Likewise, due to the stability of wind velocity, the oil detected on the Sentinel-1-IW image changed little after 13 h. From 15:00 on 10 August, to 11:00 on 11 August, the wind continued blowing in a northwestern direction but at a lower velocity (2–5 m/s). The oil surface in this period (Figure 16e,f, between Sentinel-1 EW and ICEYE-X images) was more significant than in the previous period (Figure 16d,e, between Sentinel-1 IW/EW images). This result leads to an assumption that the low wind speed may have slowed down the movement and dispersion of the oil spill toward the coast. Additionally, due to low wind velocity, current (unfortunately not available at a high spatial resolution for this tiny area) may be one of the factors causing the oil spill to spread over a large surface. This assumption is aligned with the observations in Section 4 (for the movement of the floating oil). It is more general and adequate than the conclusion in [32], in which only the surface wind speed is mentioned as a primary source of this oil drift.

6. Evaluation and Validation

To evaluate the accuracy of oil spill detection from multi-source satellite data based on the proposed algorithm, we utilize the quantitative metrics, including Precision, Recall, and F1-score, defined as follows:

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{TP}{TP+FP}}$$

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{TP}{TP+FN}}$$

$$\mathrm{F}1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

TP, FP, and FN stand for True Positive, False Positive, and False Negative, respectively. In the context of oil spill detection, Precision assesses the algorithm’s ability to differentiate between actual oil slicks and look-alikes. Recall, on the other hand, measures the algorithm’s effectiveness in identifying all oil pixels, ensuring none are missed. F1-score is the harmonic mean of the Precision and Recall. A perfect model has an F1-score of 1. The ground truth has been determined through visual assessment of the available dataset.

Table 3. Quantitative evaluation of the proposed algorithm’s performance (HSBA, combined with oil contour determination) in the Persian Gulf vs. manually segmented ground truth data.

Case	Sensor	Precision	Recall	F1-Score
#Q1 27–28 March 2021	Sentinel-2	0.978	0.578	0.726
	Sentinel-3	0.917	0.538	0.735
	Sentinel-1	0.980	0.476	0.641
#Q2 5–6 July 2021	Sentinel-1	0.92	0.87	0.894
	Sentinel-2	0.658	0.588	0.621
	Landsat-8	0.95	0.378	0.529
#Q3 3 September 2021 (three scenes)	Sentinel-1	0.98	0.90	0.938
	Sentinel-2	0.878	0.743	0.805

shows the Precision, Recall, and F1-score of the proposed method for oil spill detection from different satellite images, for the three cases in the Persian Gulf (Q#1-2-3). The Precision rates are quite high for most cases, meaning that our algorithm can accurately distinguish oil objects from look-alikes, while the Recall shows lower values (Case Q#1 and Q#2), especially for oil spill detection from optical images (Sentinel-2/3 and Landsat-8), meaning that the algorithm misses detecting some small oil objects, especially from optical images. However, the F1-scores of the three cases are satisfactory, especially for Case Q#3.

To highlight the advantages of the proposed two-step algorithm, Figure 18i,ii compares the oil slicks identified from the Sentinel-2 (scene S#3–Figure 13c) and Sentinel-1 (scene S#2–Figure 13a) images, respectively, for Case Q#3, 3 September 2021, by only HSBA (Figure 18i–a, Figure 18ii–d), those identified from HSBA plus oil contour (Figure 18i–b, Figure 18ii–e), and those manually segmented (Figure 18i–c, Figure 18ii–f). HSBA exhibited misclassifications between oil objects and look-alikes in both instances when compared to manually identified oil slicks. Specifically, the precision of the HSBA-based approach is only 6.7% for the Sentinel-2 image (Figure 18i–a), primarily due to a significant number of false positives, and approximately 55.15% for the Sentinel-1 image (Figure 18ii–d), also owing to a notable presence of false positives. However, in the two-step approach, most look-alikes are eliminated, resulting in detected oil slicks (Figure 18i–b, Figure 18ii–e) closely resembling those manually segmented (Figure 18i–c, Figure 18ii–f) for both Sentinel-1 and Sentinel-2 images. As a result, the precision scores are significantly improved, reaching 96.15% for Sentinel-2 and 93.55% for Sentinel-1.

The Recall score values, as shown in Table 3, are lower due to the missing of some oil pixels detected in comparison to the manually segmented ones. Figure 18 and Figure 19 compare the oil slicks detected by the algorithm against those manually segmented from Sentinel-2 imagery on 5 July 2021, at 06:56:21 (Figure 19), and Landsat-8 imagery on 6 July 2021, at 06:58:26 (Figure 20). The proposed algorithm successfully identifies all oil objects on both Sentinel-2 and Landsat-8 images. However, it fails to detect some oil pixels when the two figures are superimposed (Figure 19c and Figure 20c). Specifically, the Recall scores of the proposed algorithm for oil-slick detection are 58.87% for the Sentinel-2 image (Figure 19) and 37.88% for the Landsat-8 image (Figure 20). The primary reasons for the undetected oil pixels include the inhomogeneity of oil objects and the large variations in oil pixel values observed by SAR (due to sea surface roughness) and optical sensors (due to sea surface reflectance). This issue is particularly pronounced under strong surface winds, as demonstrated in Case #Q1, which evinced an increase in sea surface roughness.

Table 4. Quantitative evaluation of the proposed algorithm’s performance (HSBA combined with oil contour determination) for oil spill detection off the shore of Mauritius vs. manually segmented ground truth data.

Case	Sensor	Precision	Recall	F1-Score
Mauritius 10–11 August 2020	Sentinel-1 IW	0.90	0.862	0.88
	Sentinel-1 EW	0.438	0.98	0.609
	ICEYE-X	0.851	0.814	0.832

presents the Precision, Recall, and F1-score values of the proposed algorithm, compared to ground truth data, for oil spill detection off the coast of Mauritius. With the exception of Sentinel-1 EW, the Precision scores for Sentinel-1 IW and ICEYE-X are quite high and closely align with those in Table 3. The Recall values are also good in this case, leading to satisfactory F1-scores. The lower Precision rate for Sentinel-1 EW is likely due to the algorithm’s overestimation, which may result from the coarse spatial resolution of this image mode (30 m pixel-size) compared to the 10 m pixel-size of Sentinel-1 IW and the 3 m pixel-size of ICEYE-X.

7. Discussion

The oil spill detection algorithm presented in this paper is a two-step procedure. The strength of this approach is that it can be applied to different parameters extracted from different satellite images, i.e., SAR NRCS, optical oil index, and T865 variable. Its flexibility is particularly useful when processing new datasets since it is fully automatic and adaptive. In the first step, HSBA was applied to the input variables used to identify oil slicks, observed as dark objects on Sentinel-1/3 images and dark and bright entities on Sentinel-2 and Landsat-8 images). As shown in Figure 18, HSBA is generally effective in accurately delineating oil objects when they exhibit significant differences from the sea surface background. However, it may encounter challenges in correctly classifying some look-alikes that have pixel values similar to those of oil objects. To address this, in the second step, we employed a method for determining oil contours by applying non-linear filters (mean and standard deviation) along with an adaptive thresholding approach. This technique is grounded in the hypothesis that oil objects possess a more uniform structure than look-alikes due to oil viscosity, making them better separated from sea clusters. This distinction is particularly evident in the Sentinel-2 image (Figure 13c, middle scene), in which the oil contour enables the differentiation between look-alikes and actual oil objects.

The results of oil-slick detection in Figure 7 highlighted the advantages of the two-step procedure technique proposed in this paper, especially in cases with surface wind speeds exceeding 7 m/s, which are not optimal for oil spill detection. In instances of strong surface winds (Figure 9), the damping of the oil slicks observed in Figure 7 was less pronounced, resulting in oil objects with inhomogeneous characteristics and a notable difference in NRCS. If only the determination of oil-slick contours were to be applied, this issue would significantly impact the identification of the shapes of oil objects. However, with the application of HSBA, which selectively chooses oil objects that are clearly distinguishable as fitting the PDF of the target class used for classification over the entire image, the oil objects were accurately identified, even in cases where the objects did not have a homogeneous structure.

Table 5. Overview of the floating oil drifts off the shore of Qatar, and their relationship to surface wind and current.

Date	Oil Scale	Time Lag between Two Sensors (Hour)	Oil-Slick Shift (km)	Wind Speed (m/s)	Current Speed (m/s)	Direction (Up-North Reference)
						Oil Drift	Wind	Current
#Q1 27–28 March 2021	Large	24 (S-2/3)	18.5	1.0–10.7	0.23–0.32	SE	NW/NE/SE	SE/SW/NW/NE/N
		8 (S-3/1)	5.2	10.7–12.7	0.27–0.33	NW/SW	SE	NW
#Q2 5–6 July 2021	Medium–small	4 (S-1/2)	4.5	5.5–7.8	0.25–0.30	NW	W/NW	NW/N/E
		24 (S-2/L-8)	12.3–12.9	3.0–6.0	0.30–0.39	NW	NW/N	SE/SW/NW/N
#Q3 3 September 2021	Large–Medium–small	4 (S-1/2)	N/A	2.6–4.4	0.29	N/A	NW	NW
			2.95	3.5–6.3	0.26–0.27	S	NW/NE	SE/S
			3.75–4.8	1.8–4.2	0.26–0.27	NE	NE	NE/E

offers an overview of the oil detected from Sentinel-1/2/3 and Landsat-8 images off the shore of Qatar, and its movement for 4, 8, and 24 h of observations. The detected oil encompasses various scales, ranging from large to medium and small oil objects. As discussed earlier, the factors influencing oil drift include surface wind and current, as well as characteristics of the oil itself, including shape, size, and features (thin, thick, homogeneous, etc.). Table 5 shows that the oil drift may move in the same direction as wind and/or current, depending on wind/current velocity, as well as the difference or similarity of the wind and current directions.

In Case #Q1, accurately estimating the movement direction of the large-scale oil is challenging due to significant changes in the velocities and directions of wind and current after 24 h (Figure 9). Likewise, wind and current do not always move in the same direction. These factors may explain the stagnation of the oil drift after 32 h and the differences in the direction of movement of some oil parts. However, we can observe the significant impact of strong wind (above 10 m/s) on the change in oil shape. Indeed, at high wind speeds (10.7–12.7 m/s), the oil is split into one big part and many smaller ones, and some of them disappeared, probably due to evaporation. The impact of wind and current on the oil-drift direction is more evident in Case #Q2. The oil moved in the same direction as wind and current after 4 and 24 h. The wind velocity of about 8 m/s (07:00, Figure 12) may have significantly impacted the change in oil shape (medium scale). It is split into many small parts, especially as to the oil tail. The oil continued changing at wind speeds of 3–6 m/s but was still present after 24 h. For Case #Q3, we observe that at a low wind velocity (below 4 m/s), the shape of the medium-scale oil only changed a little, while the small one may evolve more significantly, but it is still present on the sea surface. This result differs from Cases #Q1 and #Q2, for which the small-scale oil disappears at higher wind speeds (above 6 m/s). The change in oil shape is also observed for Case #Q3 (middle scene, Figure 14b) at wind speeds of 4–6 m/s. The impact of wind and current on the oil drift for Case #Q3 is aligned with the results observed for Cases #Q1 and #Q2. Indeed, it is not easy to determine the drift direction and distance for the large-scale oil (first scene, Figure 14a). The oil drift in the second scene (Figure 14b) stagnated due to the difference between wind and current directions. The oil tended to move in the same direction as the current, since wind velocity (4–6 m/s) was not high enough to affect the oil movement. For the third scene (Figure 14c), the oil moved in the same direction as the wind and current, and the oil drift tended to accelerate due to the similarity of the wind and current directions.

Regarding the oil spill off the coast of Mauritius, the spreading of the oil surface changed very little after 13 h with a stable wind speed of 6–7 m/s; however, the oil surface expanded significantly after 21 h with a lower wind velocity of 2–4 m/s. The dominant factor contributing to this change is presumed to be the current.

The coarse spatial resolution of ERA-5 wind (~25 × 25 km) and CMEM current data (~9 × 9 km) may have complicated the assessment of the impacts of surface wind and current on the oil drift, especially for the oil spill off the shore of Mauritius, and when wind and current fields changed significantly, as observed for Cases #Q1 and #Q2. Therefore, other wind and current data sources with a high spatial resolution derived from numerical models will be considered for future studies.

Apart from met–ocean conditions and oil size, factors such as oil type (e.g., soybean or mineral oil) and thickness could potentially influence variations in oil-drift velocity and direction, and their dissipation over hours or days, as shown in [31]. However, accurate information on the detected oil types in our cases is unavailable, preventing a detailed analysis in this context. Likewise, the differences in the spatial resolution of satellite sensors may impact the evaluation of changes in the shape of oil slicks, especially between Sentinel-1/2 and Sentinel-3 and Landsat-8. Additionally, the difference in the observation of oil spills between SAR and optical sensors may contribute to the oil size and shape variations. Since images acquired by both SAR and optical sensors with comparable spatial resolution and simultaneous acquisition time and area are unavailable, definitive conclusions on this matter cannot be drawn in this paper. This aspect should be thoroughly investigated in future studies.

The oil-drift variables—such as direction, velocity, and changes in oil size and shape—obtained by combining multi-source satellite data—are highly practical for rescue operations aimed at protecting marine biodiversity. This is particularly crucial off the coast of Mauritius, where the ecosystem is vulnerable, and recovery is prolonged. This approach provides accurate and timely information, which is essential for short-term emergency responses like cleanup and contaminated area zoning, as well as for long-term solutions such as depollution and recovery.

8. Conclusions

This paper introduces an innovative method that can be applied to datasets with diverse characteristics (active and passive, and varying swath width, spatial resolution, frequency, and so on) for the detection of oil spills in various scenarios, such as floating oil off the coast of Qatar and oil leakage from a shipwreck off the coast of Mauritius. It demonstrates effective discrimination between oil and look-alikes, particularly in challenging conditions with strong surface wind. Furthermore, due to its ability to handle data from many satellite missions, it considerably reduces the latency in accessing information, which is critical during emergencies. Through collocating detected oil slicks from different satellite images acquired at different times and combining them with sea surface wind and current data, we were able to observe that oil-drift velocity and direction are notably influenced by wind and current, irrespective of oil slick sizes. Wind contributes significantly to oil shape and size evolution, while current has a more pronounced impact on oil movement, especially at low wind speeds (below 4–5 m/s). The oil drift tends to accelerate when wind and current move in the same direction and decelerates noticeably when they move in different directions.

The integration of multi-source satellite imagery for oil spill detection and monitoring, along with the assessment of how met–ocean variables influence oil drift, as proposed in this paper, is crucial for minimizing the impact of oil spill incidents on biodiversity and ecosystems, particularly in coastal and vulnerable regions. This approach offers an effective and cost-efficient solution for observing oil spills and their movement in both short-term (several hours) and long-term (more than 24 h) scenarios.

In future studies, the combination of Sentinel-1/2/3 and Landsat-8 images with different time lags for oil-drift observations will be integrated with numerical simulations from models such as GNOME and OpenOil. Additionally, the remote sensing-derived data from this study may serve as valuable training/validation datasets for the development of Machine Learning/Deep Learning models.