Multi-Scale Encoding (MSE) Method with Spectral Shape Information (SSI) for Detecting Marine Oil-Gas Leakages

Abstract

Remote sensing technologies are suitable for detecting marine oil-gas leakages on a large scale. It is important to structure an accurate method for detecting marine oil-gas leakages in varied remote sensing images. However, traditional spectral indexes have limited applicability. Machine learning methods need plenty of training and testing samples to establish the optimized models, which is too rigorous for satellite images. Thus, we proposed a multi-scale encoding (MSE) method with spectral shape information (SSI) to detect the oil-gas leakages in multi-source remote sensing data. First, the spectral amplitude information (SAI) and SSI of the original spectra were encoded into a series of code words according to the scales. Then, the differential code words of the marine oil-gas leakage objects were extracted from the SAI and SSI code words. Finally, the pixels of the encoded hyperspectral image (HSI) and multispectral image (MSI) would be determined by the differential code words. Seven images captured by different platforms/sensors (Landsat 7, Landsat 8, MODIS, Sentinel 2, Zhuhai-1, and AVIRIS) were used to validate the performance of the proposed method. The experimental results indicated that the MSE method with SSI was convergent and could detect the oil-gas leakages accurately in different images using a small set of samples.

1. Introduction

Marine oil-gas leakage incidents usually cause ecological disasters [1,2,3,4,5]. For example, in April 2010, the Deepwater Horizontal (DWH) platform in the Gulf of Mexico (GoM) exploded. It caused extremely serious marine pollution. During the leakage incident, 53,000 to 62,000 barrels of crude oil leaked into the GoM. Over 17,725 km² of the sea surface was covered by the oil slicks [6]. A large number of marine life had died from the polluted seawater. On 25 July 2020, a Japanese cargo named Wakashio ran aground off the southeast coast of Mauritius. A large amount of fuel oil (~1000 tons) leaked into the offshore area in this cargo leakage incident [7]. Twenty-eight kilometers along the coastline was seriously polluted by the oil leakage incident. The marine life would suffer from the leaked venomous oil for a long time [8]. In September 2022, the Nord Stream pipelines exploded, and a large amount of natural gas leaked into the atmosphere. The main component of natural gas was methane, which is a potent greenhouse gas. It meant that this incident would contribute to global warming. In these leakage incidents, the leaked oil and gas seriously harmed the marine and global environment. However, on the other hand, natural oil-gas leakages can indicate the undersea reservoirs [9,10,11,12,13,14]. This makes it conducive to explore marine oil-gas resources rapidly. Thus, monitoring marine oil-gas leakages is important for both pollution control and oil-gas exploration.

Satellite remote sensing technologies can be used to observe the Earth on a large scale very well [15,16]. Aerial remote sensing technologies can be used to observe sudden incidents on the Earth’s surface [17]. It is efficient and feasible to detect marine oil-gas leakages by satellite and aerial remote sensing technologies. Object recognition methods are essential for detecting oil-gas leakages in multispectral images (MSI) and hyperspectral images (HSI). However, traditional spectral indexes just took advantage of a little bit of spectral feature information [18,19]. They ignored a lot of spectral details. Thus, the spectral indexes were difficult for detecting oil-gas leakages in different datasets or study areas [10,11,12,13,14,15,16,17,18,19,20,21,22,23]. Machine learning methods, such as artificial neural networks (ANNs) and support vector machine (SVM), had an excellent ability to detect ground objects [24,25,26]. However, they needed a large number of samples to build the optimal models, which was difficult for detecting marine oil-gas leakages, because the leaked oils and natural gas are usually distributed in limited areas in MSI and his [27,28]. There were not enough samples to build the machine learning methods.

In the face of a large amount of oil-gas observation datasets from different sources, it is significant to construct an accurate and applicable detection method. Spectral encoding methods can describe the spectral details in MSI and HSI [29]. It is beneficial to distinguish objects with similar spectral signals in remote sensing images by spectral encoding methods. In 1993, Jia and Richards proposed a fast spectral matching and classification method based on binary coding (BC) [30]. It indicated that the spectral encoding method could be used to detect ground objects efficiently. However, BC was not suitable for detecting marine objects because it only used the spectral amplitude information (SAI) roughly. In 2006, Chang and Chakravarty proposed a spectral derivative feature coding method to describe the spectral shape information (SSI). The spectral shape patterns proposed in their research benefited in translating the SSI [31]. In 2012, Jiao et al. proposed a spectral DNA encoding method to describe the SAI and SSI at the same time [32]. The experimental results indicated that the encoding results by the spectral DNA encoding method were effective for detecting objects in HSI.

However, traditional spectral encoding methods just roughly expressed the SAI and SSI. It was necessary to propose a spectral encoding method which could fully mine the spectral differential details to detect marine objects. Inspired by the former studies, a multi-scale encoding (MSE) method with SSI was proposed in this research to detect oil-gas leakages in HSI and MSI. First, the SAI and SSI of the images and samples were encoded into code words according to the scales. Then, the SAI and SSI code words were joined together. The joined code word sequences were used to describe the spectral details of marine objects. Next, the differential code words between the seawater and oil-gas leakage objects were sought out from the samples. In this step, the overall and partial code word similarities would be calculated to distinguish the oil-gas leakage objects from seawater. Finally, the pixels were determined according to the calculation results. Since each encoding scale would produce a detection result, the best result produced by the optimal encoding scale would be output as the final detection result. The innovation of the research was that it proposed a novel spectral translation method by spectral encoding. The proposed method shielded the spectral heterogeneity of oil and gas leakage objects in different MSIs and HSIs. It meant that the proposed method could directly describe the spectral differences in different MSIs and HSIs. Moreover, the marine oil-gas leakages could be detected by simply matching the overall/partial code words.

Seven marine oil-gas leakage (including exploitation leakage, ship fuel oil leakage, natural gas pipeline leakage, and natural crude oil seepage) remote sensing images captured by diverse platforms (Landsat 7, Landsat 8, MODIS, Sentinel 2A, Zhuhai-1, and AVIRIS) were used to validate the accuracy and applicability of the MSE method with SSI. The experimental results indicated that the proposed method could detect marine oil-gas leakages accurately using a small set of samples.

2. Materials and Methods

2.1. Marine Oil-Gas Leakages

Figure 1 shows the schematic diagram of the marine oil-gas leakages. A large amount of oil and gas resources exist at the bottom of the oceans [33]. Oil exploitation platforms are built up to obtain crude oils from the oil reservoirs. If incidents occurred in the oil exploitation processes, crude oil would continuously leak into the marine environment. The crude oil in the reservoirs tends to leak into the marine environment gradually. In the process of natural seepage, the crude oil arrives at the sea surface by natural oil-gas bubbles [34]. In addition, oil and gas resources are inevitably transported through the marine environment [35]. Oil-gas pipelines and cargo ships are common transportation technologies. If incidents occurred during the transportation process, the oil (crude oil, refined oil, fuel oil, etc.) and gas would leak into the marine environment immediately [36,37,38].

According to the Bonn Agreements, the thickness of oil slicks is classified from the thickest to the thinnest into five types: continuous true color oil slicks (code 5), discontinuous true color oil slicks (code 4), metallic oil slicks (code 3), rainbow oil slicks (code 2), and silver oil slicks (code 1). Emulsions (emulsified oil slicks) are not coded by the Bonn Agreement, because their thicknesses are not stable [39]. Because of the density and viscosity of the crude oil components (alkane, aromatic hydrocarbon, olefin, etc.), a large scope of the sea surface would be covered by the leaked oil [40]. The heavy components of the leaked oil will collect and sink to the bottom of the ocean under the influence of gravity. This damages the marine ecosystem for a long time [41]. The oil slicks formed by natural leakages have limited scopes [42]. Affected by the currents and wind, the natural oil slicks resemble feathers [43]. Gas leakages form a large number of bubbles by the huge pressure in the pipes. The seawater splashes when the natural gas bubbles arrive to the sea surface. Since the bubbles near the leakage point tended to coalesce and the circumjacent bubbles tend to drift randomly, the splashing area on the sea surface usually consists of a core area and a surrounding area. The physical and chemical reactions of the oil and gas leakages make the spectral signals of oil-gas leakages complex and changeable [44,45,46].

2.2. MSE Method

The MSE method focuses on translating the SAI of the original spectral signals into code word sequences. It transforms the implied spectral information into code words that can be contrasted directly. When encoding at a low scale, only a small number of code elements are used to describe the original SAI. The encoded results tend to be rough. When encoding at a high scale, a large number of code elements are used to describe the original SAI. The encoded results tend to be detailed. The calculations of the MSE method are exhibited below (Formulas (1)–(3)).

$$T^M=\alpha \times \frac{{\displaystyle \sum R_i}}{N}$$

(1)

$$T_{i,j}^{LH}=R_{Min}+j\times \frac{T^M-R_{Min}}{0.5\times E-1}$$

(2)

$$T_{i,k}^{UH}=T^M+k\times \frac{R_{Max}-T^M}{0.5\times E-1}$$

(3)

In the formulas, T^M is the middle threshold of the spectral signals. It is used to divide the SAI into two different parts. N is the spectral band number of the samples and images. It is used to record the number of spectral signals. R_i is the i-th spectral reflectance of the samples and images. Alpha is the control parameter of the SAI encoding. With a smaller alpha (<1), the encoding results tend to describe the lower reflectance spectral signals in detail; with a bigger alpha (>1), the encoding results tend to describe the higher reflectance spectral signals in detail. T^LH are the thresholds of the lower spectral signals, and j is the index of the thresholds at the i-th band. R_Min is the minimum reflectance of the spectral signals. E is the number of code elements, which is determined by the encoding scales. A bigger E means more code elements will be used to describe the spectral signals. T^UHs are the thresholds of the upper spectral signals, and k is the index of the thresholds at the i-th band. R_Max is the maximum reflectance of the spectral signals.

The MSE method can translate the SAI of remote sensing images excellently. Figure 2 shows the spectral encoding results of an oil slick sample in an AVIRIS image on different encoding scales. Figure 2a shows the floating oil slick and its original spectrum. Figure 2b–e are the encoding results (alpha = 1) at the scales of 2, 4, 6, and 36, respectively. It can be observed intuitively from the MSE results that, with the encoding scales increasing, more and more code elements are used to describe the original spectra of the marine oil slicks. Then, the spectral details are revealed gradually. When the encoding scale is 36, the encoded code word sequences are very close to the original spectra. The spectral absorption features can be observed clearly in the encoding results. It indicates that the MSE method can describe the characteristics of the original spectra signals. This is beneficial for detecting ground objects based on the MSE and pattern feature matching.

2.3. MSE Method with SSI

It is difficult to effectively and stably distinguish the objects with high spectral similarity only by the SAI information [47]. SSI can enhance the ability of the image information expression. Encoding the SAI and SSI enables the original spectral details to be fully expressed. This is beneficial to distinguish the objects with high spectral similarities in the MSI and HSI. The spectral derivative coding method was adopted in this research to translate the SSI [31]. It defined the spectral shape patterns by three adjacent bands. The spectral shape patterns are exhibited in

Table 1. The spectral shape patterns and their code words.

Spectral Shape Pattern	Code Word	Spectral Shape Pattern	Code Word
	0		5
	1		6
	2		7
	3		8
	4

. Formula (4) is used to calculate the spectral gradient threshold. Formulas (5) and (6) are used to calculate the similarity between the encoded samples and pixels.

$$T^S=\beta \times \frac{{\displaystyle \sum _{i=1}^{N-1}(R_{i+1}-R_i)}}{N-1}$$

(4)

$$f(C_i^{sample},C_i^{pixel})=\left\{\begin{array}{l}0 , if C_i^{sample}\ne C_i^{pixel}\\ 1 , if C_i^{sample}=C_i^{pixel}\end{array}\right.$$

(5)

$$S=\raisebox{1ex}{${\displaystyle \sum f(C_i^{sample},C_i^{pixel})}$}\!\left/ \!\raisebox{-1ex}{$(2N-2)$}\right.$$

(6)

In Formula (4), T^S represents the threshold of the spectral shape gradient. It is used to describe the waveforms of the spectral signals. The spectral shape patterns should be defined by the T^S of adjacent signals. R_i is the spectral reflectance of the i-th band. N is the band number of the original spectral signals. The parameter of beta is used to control the adaptability of the spectral shape gradient. Large beta (> 1) helps to filter the spectral shape changes caused by noises. A small beta (<1) helps to expose the small spectral shape details. Since different remote sensing platforms collect different spectral signals for ground objects, adjusting the value of the beta according to the actual remote sensing dataset is beneficial to detect marine objects accurately. In Formulas (5) and (6), C is the SAI and SSI code word sequences. f is used to judge the consistency between the encoded samples and pixels at the i-th code word. If the i-th code words of them were the same, the result of f would be 1, otherwise, the result of f would be 0. S is the similarity of the encoded samples and pixels. The more code words the encoded samples and pixels have in common, the greater the value of S is.

Figure 3a shows the original spectra of the oil slicks formed by the leaked crude oils in the GoM oil spill incident. Figure 3b shows the encoding results (scale = 12) of the oil slicks produced by the MSE method with SSI. It can be clearly found that the spectral code words of the oil slicks are obviously different. A large amount of different code words can be observed in their SAI and SSI encoding results. This makes it easy to distinguish them by matching the spectral encoding results. It should be noted that the proposed method in this research is a kind of relative encoding method instead of an absolute encoding method. In addition, the maximum and minimum value of the samples and pixels are independent. Thus, in the encoding results, the ground objects with large spectral reflectance differences may share some same code words in the SAI encoding results (the sheens and oil slicks of code 4 in Figure 3). This ensures that the ground objects with low spectral reflectance can fully display their spectral details. It is very suitable for detecting marine oil-gas leakages.

Figure 4 shows the processes of detecting oil-gas leakages in HSI and MSI by the MSE method with SSI. Step 1: select samples from the remote sensing images. Calculate the thresholds according to the encoding scales. Then, encode the SAI of the samples and remote sensing images by the MSE method at different scales. Step 2: calculate the spectral shape gradient thresholds. Then, encode the SSI of the samples and pixels according to the spectral shape patterns. Step 3: produce the complete spectral encoding results by combining the SAI and SSI code word sequences. Step 4: mine the differential code words of the oil-gas leakages and match the encoded samples and pixels by the differential code words. Calculate their similarity. Step 5: detect the leakages according to the similarity calculating results. Step 6: pick out the optimal result from the detecting results produced by different scales as the final result of the whole process.

It should be noticed that the experimental satellite data were level 1 images. They contained more distinctive ground reflectance information. It contributed to differentiating ground objects. The airborne and spaceborne reflectance images were corrected by the fast line-of-sight atmospheric analysis of the spectral hypercube (FLAASH) model.

3. Results

In this research, seven different remote sensing datasets were used to validate the accuracy and applicability of the proposed method. The information of the experimental datasets is listed in

Table 2. The information of the experimental datasets.

NO.	Source	Incident	Oil Spill Type
1	Landsat 7	Deepwater Horizon drilling platform explosion	Exploitation leakage
2	Landsat 8	Japanese oil tanker stranding around Mauritius	Ship leakage
3	Sentinel-2	Drilling platform oil spill	Exploitation leakage
4	MODIS	Deepwater Horizon drilling platform explosion	Exploitation leakage
5	Zhuhai-1	Natural oil spill in the South China Sea	Natural leakage
6	AVIRIS	Deepwater Horizon drilling platform explosion	Exploitation leakage
7	Sentinel-2	Nord Stream 2 pipeline explosion	Pipeline leakage

. The experimental datasets contain natural oil seepage, exploitation oil spills, ship oil spill, and natural gas pipeline leakage (caused by the Nord Stream 2 pipeline explosion). They can effectively evaluate the performance of the MSE method with SSI for detecting oil-gas leakages.

3.1. Oil Spill Detection in Landast 7 Data

The experimental image was captured by the Landsat 7 satellite (six bands, spatial resolution 30 m) in the GoM on 1 May 2010. The oil slicks were leaked by the DWH oil exploitation platform (Figure 5) [48]. Since the DWH oil spill incident occurred on 20 April 2010, the leaked oil covered a large scope of the sea surface. In the experimental image, the majority of the oil slicks were weathered.

The oil pollution range is very large. Since the Landsat 7 images have invalid stripes [49], the interpolated experimental image is somewhat blurred and uneven (Figure 5a). After encoding the samples by the MSE method with SSI (scale = 36, alpha = 1.5, beta = 0.2), it was found that the second and third code words of the oil slicks were different from that of seawater in the experimental Landsat 7 image. The partial spectral code words were used to detect the leaked crude oils. In the detection results (Figure 5b), the leaked oil slicks were detected and marked in red. Since the interpolation method could not restore the distorted strips in Landsat 7 images, the detection results in the interpolated parts were unsatisfactory. However, most of the oil leakages were detected correctly. It indicated that the partial code words produced by the MSE method with SSI were suitable to detect oil leakages in Landsat 7 images.

3.2. Oil Spill Detection in Landsat 8 Data

The experimental image was captured by the Landsat 8 satellite (seven bands, spatial resolution 30 m) in the southeast coast of Mauritius on 7 September 2020 (Figure 6). The oil slicks in the experimental image were leaked by a Japanese oil tanker WAKASHIO [50]. The Japanese oil tanker ran aground on 25 July and started leaking fuel oil on 6 August. The impact of this oil spill incident within Mauritius’s environment and economy is likely to last for decades.

Since the quality of the Landsat 8 images was very high, the stranded oil tanker and the surrounding oil slicks could be observed clearly (Figure 6a). After encoding the samples by the MSE method with SSI (scale = 36, alpha = 0.5, beta = 0.5), it was found that the first, second, ninth, and tenth spectral code words of the leaked oil slicks were different from that of the seawater in the experimental Landsat 8 image. The partial spectral code words were used to detect the leaked fuel oils. In the detection result (Figure 6b), the leaked oil slicks were detected and marked in red. The oil slicks which surrounded the stranded ship were detected accurately. From the experimental result, it was concluded that the partial spectral code words produced by the MSE method with SSI were suitable to detect oil leakages in Landsat 8 images.

3.3. Oil Spill Detection in Sentinel-2 Data

The experimental image was captured by the Sentinel-2 satellite (nine bands, spatial resolution 20 m) in the GoM on 23 April 2016 (Figure 7). The GoM contained vast reservoirs of oil and natural gas. Many oil exploitation platforms distributed around the study area. In addition, the oil slicks in the experimental image were at a large scope, and some of them were collected and emulsified (Figure 7a). Thus, the oil slicks should be leaked by the oil exploitation platform.

After encoding the samples by the MSE method with SSI (scale = 20, alpha = 0.5, beta = 1.2), it was found that the emulsions had differential spectral information at the first, ninth, and sixteenth code words. The other oil slicks had differential spectral information at the first, second, third, eighth, and ninth code words. The oil leakage in the Sentinel-2 image was detected by the partial spectral code words. In the detection result (Figure 7b), the detected emulsions were marked in red and the other oil slicks were marked in yellow. Although some seawater was misidentified as oil slicks because of the sea waves and the belt splicing errors in the Sentinel-2 images, the majority of the oil slicks were detected correctly. This indicated that the partial code words produced by the MSE method with SSI were suitable to detect oil leakages in Sentinel-2 images.

3.4. Oil Spill Detection in MODIS Data

This experimental image was captured by the Terra satellite (MODIS imager, 21 bands, spatial resolution 1 km) in the GoM on 29 April 2010 (Figure 8). The oil slicks were leaked by the DWH oil exploitation platform [51]. Although the image was captured at the early stage of the DWH oil spill incident, a large amount of oil slicks existed on the sea surface. The shape of the leaked oil slicks resembled an “S” on the sea surface (Figure 8a).

After encoding the samples by the MSE method with SSI (scale = 8, alpha = 1, beta = 1), it was found that the leaked crude oil slicks had differential spectral information at the fourth, eleventh, thirteenth, eighteenth, twenty-second to twenty-fourth, and thirty-fourth code words (a total of eight code words). The leaked crude oil slicks were detected and marked in red in the results (Figure 8b). Since the spatial scope of the study area is very large, sun glint exists in the experimental image. It is difficult to distinguish the leaked crude oil slicks from the sun-glint-polluted seawater [52]. Although only the main body of the leaked oil slicks was detected, the detection result indicated that the partial code words produced by the MSE method with SSI were suitable to detect oil leakages in MODIS images.

3.5. Oil Spill Detection in Zhuhai-1 Data

This experimental image was captured by the Zhuhai-1 satellite (32 bands, spatial resolution 10 m) in the South China Sea, near the Ryukyu Island, on 30 March 2021 (Figure 9). The study area was located in the sea area of the Southwest Taiwan Basin, which contained abundant submarine oil and gas resources. The oil slicks in this experimental image were limited to a small area and resembled a feather on the sea surface (Figure 9a). According to the characteristics of natural oil slicks [53], the oil slicks in the experimental image should came from the marine oil reservoir.

Figure 9 shows experimental image 5 and its natural oil leakage detection results. After encoding the samples by the MSE method with SSI (scale = 24, alpha = 1.2, beta = 1.2), it was found that the natural oil slicks had differential spectral information at the twenty-fourth, forty-ninth, fiftieth, fifty-fifth, and fifty-sixth code words. The natural oil slicks were accurately detected by the partial spectral code words (Figure 9b). However, since the Zhuhai-1 satellite could capture abundant, high-quality spectral signals, the oil slicks were also accurately detected by matching the integral spectral code words (Figure 9c). From the detection results, it was found that the partial or integral spectral code words produced by the MSE method with SSI were both suitable to detect the natural seepages in Zhuhai-1 satellite data.

3.6. Oil Spill Detection in AVIRIS Data

The image was captured by the AVIRIS (192 bands) in the GoM on 17 May 2010 (Figure 10). This dataset was provided by the jet propulsion laboratory (JPL). The AVIRIS image has excellent spatial and spectral resolutions (spatial resolution 7.6 m, spectral resolution 10 nm) [54]. Different thicknesses of oil slicks can be detected from the AVIRIS image.

Since the original AVIRIS image was too large to handle, four subareas were selected to validate the proposed method. In the first subarea (Figure 10a), oil slicks of code 5 were black, strides of emulsions looked orange, oil slicks of code 4 had a patchy distribution, and the sheens surrounded the thicker oil slicks [55,56]. A clear convergence of emulsions existed in the second subarea (Figure 10c), with flaky oil slicks of code 4 and sheens distributed around the emulsion convergence. In the third subarea (Figure 10e), small patches of oil slicks of codes 4 and 5 existed. In addition, a stripe of emulsions distributed at the bottom right corner of the image. The fourth subarea (Figure 10g) was the boundary of the oil pollution. The sheens in the fourth subarea slightly altered the color of the unpolluted seawater. Since the spectral differences of the oil slicks in the AVIRIS images can be fully expressed by the MSE method with SSI (Figure 3), the oil slicks in the experimental AVIRIS image were detected accurately (Figure 10b,d,f,h) by matching the integral spectral code words produced by the MSE method with SSI (scale = 4, alpha = 1, beta = 1). All the results were obtained by the same sample set. Even the sheens at the pollution boundary could be distinguished from the seawater by the code words. Thus, it was concluded from the results that the integral spectral code words produced by the MSE method with SSI were suitable to detect the oil slicks of different thicknesses in AVIRIS images.

3.7. Natural Gas Leakage Detection in Sentinel-2 Data

This image was captured by the Sentinel-2 satellite (nine bands, spatial resolution 20 m) in the Baltic Sea on 30 September 2022 (Figure 11). The gas leakage in the experimental image was caused by the explosion of the Nord Stream 2 pipeline. It was located about 24.3 km southeast of the Bornholm Island.

Huge pressure in the pipe made the natural gas spill out of the seawater. The bubbles created by the natural gas flow formed a leakage area on the sea surface. The leakage area consisted of a core area and a surrounding area. The core area was a circle with a radius of around 100 m, and the shape of the surrounding area resembled a water drop (Figure 11a). The spectra of the core area were very different from the spectra of the seawater, and the spectra of the surrounding area were similar to the spectra of the seawater. After encoding the samples by the MSE method with SSI (scale = 24, alpha = 1, beta = 0.5), it was found that the natural gas leakage had many differential spectral code words compared to seawater. Thus, the natural gas leakage areas were detected by the integral spectral code words (Figure 11b). In the detection result, the core area and the surrounding area were detected accurately. The gas leakage could be totally distinguished from the seawater. It was concluded that the integral spectral code words produced by the MSE method with SSI were suitable to detect natural gas leakages caused by pipeline incidents in Sentinel-2 satellite data.

4. Discussion

4.1. Encoding Scales and the Parameters

In the process of MSE, the encoding scale largely controls the ability to express the original spectral information. With the encoding scale increasing, the SAI encoding results will be closer and closer to the original spectra (Figure 3). Before the encoding scale reaches the limit, there must be a best encoding scale which can detect marine oil-gas leakages accurately. This process is called encoding convergence. Figure 12a–d showed the oil slick detection results in the experimental AVIRIS image by the MSE method on different scales. It could be found that when the encoding scale was small (4 and 8), only small bits of the code elements were used to describe the original SAI. The spectral details would be hidden by the rough spectral encoding process. The oil slicks with similar spectral signals (sheens and seawater) would be misidentified. When the encoding scale became larger (12 and 24), more code elements were used to describe the original SAI. The hidden spectral details would be revealed by the delicate encoding process, and the sheens and seawater were distinguished correctly.

The MSE method with SSI produces not only the SAI code words, but also the SSI code words. Figure 12e–h showed the detection results produced by the MSE method with SSI. It was found that the detection result produced by the MSE method on scale 4 contained some pixels of misidentified sheens (Figure 12a, blue box). However, the MSE method with the SSI differentiated the sheens from seawater accurately on scale 4 (Figure 12e). This indicated that the SSI code words enhanced the expression ability for the original spectra. In addition, it was found that the MSE method converged on scale 12, and the MSE method with SSI converged on scale 4. It meant that the SSI code words made the speed of encoding convergence faster. However, if the encoding scale exceeded the best scale, the spectra would be overfitted and the detection results tended to become worse. Thus, it could be observed that the emulsions were accurately detected by the MSE method with SSI on scale 4 (Figure 12e), but the emulsion detection performances on scales 8, 12, and 24 (Figure 12f–h) were poor.

Figure 13 shows the best encoding scales of different datasets in this research. It can be observed that the best encoding scales for different datasets were different. This meant that the speed of encoding convergence was affected by the datasets. The aerial remote sensing data with a large amount of spectral bands (AVIRIS) tended to converge very fast (scale = 4). The satellite remote sensing data with few spectral bands (Landsat 7 and Landsat 8) converged slowly. However, they would converge until the scale of 36. In addition, the convergence speeds of the satellite remote sensing data with more spectral bands (Sentinel-2, MODIS, and Zhuhai-1) were diverse. Since the remote sensing images captured by the Sentinel-2 and Zhuhai-1 satellites had similar spatial resolutions (20 m and 10 m), their convergence speeds were similar too (scale = 20 and 24). The spectral reflectance signals of the MODIS data were more homogeneous than that of Sentinel-2 and Zhuhai-1 datasets. Thus, the convergence speed of MODIS data was faster than that of Sentinel-2 and Zhuhai-1 datasets. However, the pixels of the MODIS image were very likely to be composed of mixed marine objects because of the low spatial resolution (1 km). Thus, only the main body of the oil leakage in the MODIS images could be detected (Figure 8).

Except for the encoding scale, the encoding parameters (alpha in Formula (1) and beta in Formula (6)) play an important role in the process of describing the original spectra.

Table 3. The parameters of the MSE method with SSI for the experimental datasets.

Parameter	Experimental Dataset No.
	1	2	3	4	5	6	7
Source	Landsat 7	Landsat 8	Sentinel-2	MODIS	Zhuhai-1	AVIRIS	Sentinel-2
Bands	6	7	9	21	32	192	9
Alpha	1.5	0.5	0.5	1	1.2	1	1
Beta	0.2	0.5	1.2	1	1.2	1	0.5

listed the parameters of the MSE method with SSI for the experimental datasets. It was intuitively observed from the table that the SAI encoding parameter (alpha) and the SSI encoding parameter (beta) were varied for different datasets. The SAI encoding parameter ranged from 0.5 to 1.5, and the SSI encoding parameter ranged from 0.2 to 1.2. In addition, the parameters of the remote sensing datasets with abundant spectral bands tended to approach 1. This might be probably because the datasets with rich spectral information did not require additional needs of encoding tendency.

4.2. The Superiority of the MSE Method with SSI

Figure 14 shows the spectra of oil-gas leakage objects and their spectral encoding results by the MSE method with SSI. Although the spectra of oil-gas leakages are disparate, differential spectral code words are mined from the heterogeneous spectral datasets. The oil-gas leakages can be accurately detected using these differential spectral code words. However, it was found that the detection strategies were different for the heterogeneous datasets. For the remote sensing images containing a little bit of spectral information (Landsat 7, Landsat 8, and Sentinel-2), the oil leakage objects should be detected by matching the partial spectral code words. For the remote sensing images containing a large amount of spectral information (AVIRIS), the oil leakage objects could be detected by matching the integral spectral code words. As for the Zhuhai-1 satellite images (32 bands), both the partial spectral code words and integral spectral code words could be used to produce correct detection results of oil leakages. However, the oil leakages in the MODIS image (21 bands) could not be detected by matching the integral spectral code words. This might be because the spatial resolution of MODIS was too low to determine a single kind of marine object. The partial spectral code words were suitable for detecting oil leakages in MODIS images. Since the natural gas leakages caused by pipeline incidents had particular spectral signals, the integral spectral code words produced by the MSE method with SSI were suitable to detect them correctly. Thus, it could be concluded that the MSE method with SSI was suitable for detecting oil-gas leakages in different remote sensing images.

The proposed method adopted the pattern-matching strategy to detect marine oil-gas leakages. Only a small set of samples were needed to distinguish the oil-gas leakage objects from seawater.

Table 4. Samples of the experimental datasets.

Sample Number	Experimental Dataset No.
	1	2	3	4	5	6	7
Seawater	140	20	10	33	22	27	21
Oil leakage	187	20	41	58	48	-	-
Gas leakage/Core area	-	-	-	-	-	-	10
Gas leakage/Surrounding area	-	-	-	-	-	-	10
Emulsions	-	-	-	-	-	46	-
Code 5	-	-	-	-	-	63	-
Code 4	-	-	-	-	-	27	-
Sheens (Code 1–3)	-	-	-	-	-	66	-

lists the samples of the seven experiments in this research. Except for Landsat 7 data, which used more than 100 samples since the study area was too wide, all the other satellite and airborne data used less than 70 samples. In particular, Landsat 8 satellite data only needed 20 samples to distinguish oil leakages from seawater. Even though only a small set of samples were used to construct the MSE method with SSI, the differential spectral details could be extracted to detect the oil-gas leakages. Thus, it could be concluded that the MSE method with SSI was robust. It could detect marine oil-gas leakages in a small number of samples.

5. Conclusions

Marine oil-gas leakages are harmful to the marine ecological environment. Remote sensing technologies can monitor marine oil-gas leakages on a large scale. It is important to build a flexible method for detecting marine oil-gas leakages accurately. Traditional methods cannot mine abundant spectral differences to distinguish the oil-gas leakages from seawater stably; thus, they cannot accurately detect the oil-gas leakages in diverse remote sensing images. In this research, an MSE method with SSI was proposed to detect marine oil-gas leakages. First, the SAI and SSI of the images and samples were encoded into code words according to the scales. Then, the SAI and SSI code words were joined together to describe the spectral details of the marine objects. Next, the differential code words of the oil-gas leakages were mined. Finally, the pixels were determined by calculating the code word similarity on different scales. Seven datasets captured by different remote sensing platforms were used to validate the accuracy and applicability of the proposed method. The experimental results indicated that the MSE method with SSI could fully explore the spectral differences between oil-gas leakages and seawater. Oil and gas leakages could be correctly detected by the proposed method using a small set of samples. The experimental results indicated that the convergence speeds of the proposed method were varied across different datasets and the dataset with rich spectral information had a high convergence speed. In addition, the proposed method tended to converge before the scale of 36 for all the experimental datasets.