Automatic Identification of Sunken Oil in Homogeneous Information Perturbed Environment through Fusion Image Enhancement with Convolutional Neural Network

Abstract

With the development of the marine oil industry, leakage accidents are one of the most serious problems threatening maritime and national security. The spilt crude oil can float and sink in the water column, posing a serious long-term threat to the marine environment. High-frequency sonar detection is currently the most efficient method for identifying sunken oil. However, due to the complicated environment of the deep seabed and the interference of the sunken oil signals with homogeneous information, sonar detection data are usually difficult to interpret, resulting in low efficiency and a high failure rate. Previous works have focused on features designed by experts according to the detection environments and the identification of sunken oil targets regardless of the feature extraction step. To automatically identify sunken oil targets without a prior knowledge of the complex seabed conditions during the image acquisition process for sonar detection, a systematic framework is contrived for identifying sunken oil targets that combines image enhancement with a convolutional neural network (CNN) classifier for the final decision on sunken oil targets examined in this work. Case studies are conducted using datasets obtained from a sunken oil release experiment in an outdoor water basin. The experimental results show that (i) the method can effectively distinguish between the sunken oil target, the background, and the interference target; (ii) it achieved an identification accuracy of 83.33%, outperforming feature-based recognition systems, including SVM; and (iii) it provides important information about sunken oil such as the location of the leak, which is useful for decision-making in emergency response to oil spills at sea. This line of research offers a more robust and, above all, more objective option for the difficult task of automatically identifying sunken oils under complex seabed conditions.

1. Introduction

With the continuous expansion of offshore oil production and utilization, the number of oil spills around the world has gradually increased. For example, between 1973 and 2020, there were 156 oil spills involving vessels with a capacity of more than 10 tons in coastal areas of China [1]. In 2010, an oil spill occurred in the US Gulf of Mexico, which caused great concern in the international community [2]. Marine oil spill accidents not only lead to the waste of marine oil resources but also cause environmental pollution of varying degrees, which poses an immediate threat to marine safety [3]. In the event of a large oil spill, the spilled crude oil and heavy oil may float in the water column and sink due to factors such as oil density, water turbulence caused by waves or currents, and sediment floating in the water [4]. Due to changes in temperature, hydraulics, and the marine environment, the sunken oil may be dispersed by ocean wave currents and float to the sea surface or to land seasonally, resulting in secondary pollution. As a result, sunken oil can impact the marine environment for longer periods of time and over greater distances than floating oil. Therefore, it is critical to detect and identify sunken oil at the first attempt after an oil spill and take corrective emergency measures to completely remove it.

Recent research has explored various techniques for detecting sunken oil. Visible light and SAR satellite remote sensing effectively detect floating oil, but their effectiveness decreases with depth due to wavelength attenuation. This limitation has spurred research into alternative detection techniques. Diver observations, for example, offer direct and accurate results but are restricted by depth, visibility, and human endurance [5]. Laser fluorosensors can detect oil on or near the water surface, but their effectiveness decreases with depth, limiting their use for sunken oil detection [6]. In contrast, bottom trawls are effective for detecting oil on the seafloor but potentially damaging to benthic ecosystems [7]. Comparative studies have evaluated these methods based on efficiency, time consumption, cost, and operational limitations. Experimental research has demonstrated that acoustic technology, particularly high-frequency sonar, outperforms other methods for sunken oil detection [8]. In 2005, an oil spill occurred in the US Gulf of Mexico, and as the oil spill area gradually expanded, side-scan sonar technology became the only way to detect sunken oil [9]. In 2015, an accident occurred in the Mississippi River where two oil tankers collided, causing tons of crude oil to spill into the river downstream. The side-scan sonar successfully located two sunken oil targets [10]. Given its demonstrated effectiveness and ongoing technological improvements, acoustic technology, specifically high-frequency sonar, positions itself as the most promising solution for addressing the challenges of sunken oil detection.

Despite these advances, challenges remain in distinguishing sunken oil from other seafloor features and accurately quantifying oil volumes. Current research focuses on integrating multiple sensing technologies and developing advanced data processing algorithms to address these issues. Recent advancements in sonar technology have further improved detection accuracy. For instance, Snellen et al. [11] developed a multi-beam sonar system that achieved seabed sediment classification in controlled experiments. Similarly, Tang et al. [12] combined high-frequency sonar with neural networks, enhancing classification accuracy to 88% in field trials. The above methods are effective when it comes to detecting sunken oil in shallow waters. As the depth of the water increases, the signal for detecting sunken oil is not only significantly weakened, but it also becomes mixed up with signals from other objects floating in the water and on the seabed. Therefore, interpreting and identifying the sunken oil heavily relies on the expert’s prior knowledge. In this context, developing an automated approach to identify the sunken oil under complex deep seabed conditions is feasible and critical to supporting emergency response to offshore oil spills.

In recent years, great progress has been made in automatically identifying underwater targets based on sonar detection images, thanks to advances in the machine learning and deep learning communities. For example, Bian et al. used the pixel value importance functions to initially filter the targets; they then used a CNN as a secondary classification method to achieve underwater target identification [13]. Wang et al. proposed a novel deep learning model using an adaptive weighting convolutional neural network (AW-CNN) to classify underwater sonar detection images [14]. Williams et al. used convolutional neural networks to perform underwater target classification in Synthetic Aperture Sonar (SAS) images [15]. However, the following challenges exist when using computer vision methods to detect and identify sunken oil: (i) the image conveys more complicated information, such as background with sand and other sediments, and (ii) the sunken oil targets vary in size and shape. Therefore, more powerful approaches need to be developed to accurately identify the information about sunken oil targets.

Currently, there are few studies on the application of deep learning methods and CNN algorithms to identify sunken oil. This paper presents a novel sunken oil identification scheme based on CNN. This multidisciplinary study integrates underwater high-frequency sonar detection technology, image processing technology, and CNN algorithms. In addition, the present study improves the technical capabilities of marine oil spill detection and identification methods and provides technical support for marine safety.

The remaining parts of this paper are organized as follows: Section 2 explains the proposed framework for sunken oil identification including the image preprocessing methods for sonar detection and the CNN architecture. Section 3 presents the details of sunken oil release experiments and case studies on experiment data to evaluate the effectiveness of the framework. The conclusions are summarized in Section 4.

2. Proposed Method for Sunken Oil Identification

Due to the background noise and other interfering targets, the sonar detection image has a low signal-to-noise ratio and low resolution. Therefore, this paper proposes a systematic framework for sunken oil identification based on CNN; the proposed method is divided into two stages to accurately identify the sunken oil target. The first step is to preprocess the original sonar detection image. The goal of this phase is to minimize background noise in the image and segment the suspected sunken oil target areas (including actual sunken oil targets and other nuisance targets). In the second stage, a pre-trained CNN classification model is used to classify and identify the suspected sunken oil target areas, exclude background and interference targets, and determine the actual sunken oil target. The detailed algorithm implementation process is shown in Figure 1.

2.1. Stage i: Homogeneous Interference Signal Attenuation and Target Signal Amplification

The preprocessing of the attenuation of the homogeneous interference signal and target signal amplification includes three processes: homogeneous image filtering, target image amplification, and threshold segmentation, which are described in detail below.

Gaussian filtering is used for homogeneous interference signal suppression [16]. Assuming the size of the filter kernel is n × n, then the two-dimensional Gaussian filter $G(i, j)$ is defined by (1).

$$\begin{array}{c}\left\{\begin{array}{c}G\left(i,j\right)={\displaystyle \frac{1}{2\pi {\sigma }^2}}{e}^{-\left(i^2+j^2\right)/\left(2{\sigma }^2\right)}\\ \mathrm{where} i \in [0, n], j \in [0, n]\end{array}\right.\end{array}$$

(1)

where $\sigma$ is the standard deviation of the normal distribution. The output of Gaussian filtering is the Gaussian-weighted average of the gray values of all pixels in the filter kernel. The calculation for Gaussian filtering is shown in (2).

$$\begin{array}{c}{F}_{i,j}={\displaystyle \frac{1}{{\sum }_{i=1}^n{\sum }_{j=1}^{n}G\left(i,j\right)}}{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}G\left(i,j\right)g(i,j)\end{array}$$

(2)

where $G(i, j)$ is the gray value of pixels at the location described as a coordinate (i, j) in the filter kernel, $g(i,j)$ is the gray value of the input image pixel at coordinates (i, j) within the filter window, and $F_{i,j}$ is the gray value of the pixel in the middle of the filter kernel.

In this study, the Gaussian filter with a filter kernel of 35 × 35 is used to filter sonar detection images. Figure 2b shows the image filtering results obtained when the Gaussian filter with a 35 × 35 filter kernel was applied to the original image. It is clear that the background noise of the filtered image has been largely eliminated.

It is clear from Figure 2b that the filtered image reduces the local contrast of the image and, unfortunately, by suppressing the background noise, the signal in the submerged oil target area is also weakened. Therefore, it is necessary to improve the image. This article applies the histogram equalization algorithm. The algorithm converts the grayscale histogram of the input image from a specific grayscale interval in the comparison set to a uniform distribution throughout the grayscale range, thereby improving the image contrast, improving the image dynamic range, and making the image information more prominent [17]. The effect of improving the sonar detection image is shown in Figure 3. Compared to Figure 2b, the sunken oil area in Figure 3 showed greater contrast to the background image.

After the improved processing, some useless information was incorrectly improved. Therefore, it is also necessary to perform threshold segmentation on the image to segment the sunken oil target area from the background. The sunken oil appears primarily as dark spots compared to the background (see Figure 2b). This feature suggests the use of a threshold segmentation method [18]. The current main approach to threshold segmentation can be categorized into fixed threshold segmentation, adaptive threshold segmentation (adaptive mean threshold segmentation and adaptive Gaussian threshold segmentation), and the Otsu method [19].

The grayscale histogram of the sonar detection images showed a Gaussian distribution; therefore, a fixed threshold segmentation approach was used, and the segmentation threshold was set to the tertile of the grayscale distribution. Next, the image was eroded to obtain a more accurate segmentation result. The post-erosion binary segmentation image is shown in Figure 4.

After threshold segmentation, the Canny operator [20] is used to detect the edge of dark areas in Figure 4 (as shown in Figure 5). Then, the outer rectangle is drawn from dark areas. These areas are segmented from the original image and treated as suspected sunken oil targets because they may contain both the sunken oil targets and the background. The segmented areas are then identified using the CNN model (proper scaling is applied to adjust the CNN input dimensions) to exclude the background area and make final decisions. The extracted target area is shown in Figure 6.

2.2. Stage ii: CNN-Based Sunken Oil Detection

2.2.1. Brief Description of Convolutional Neural Network

Recently, deep learning methods have been successfully applied to increasingly complex practical engineering problems, including target detection [21], fault diagnosis [22], and condition monitoring [23]. In addition, convolutional neural network algorithms provide an end-to-end learning model and are an “excellent representative” of deep learning methods, which are widely used in tasks such as image classification [24] and object detection [25]. Through multiple convolution layers and pooling operations, CNNs automatically learn features from input images; they then input these features into the fully connected neural network to achieve image identification and classification. The structure of a convolutional neural network usually consists of convolution layers, pool layers, fully connected layers, and output layers.

The convolution layer convolves the input data using convolution filters. According to [26], the convolution operation for the n-th feature map in the i-th convolutional layer is shown in Equation (3).

$$\begin{array}{c}{H}_n^i=f\left({\displaystyle \sum _{m=1}^M}{H}_m^{i-1}\otimes W_{n,m}^i+b_n^i\right)\end{array}$$

(3)

where $W_{n,m}^i$ is the weight kernel connecting the m-th feature map in layer (i − 1) to the n-th feature map in layer i, $H_m^{i-1}$ is the m-th feature map from the previous layer (i − 1), and $H_n^i$ is the n-th feature map in the i-th layer. The two-dimensional convolution operation of the convolution filter of the i-th layer and the feature map of the last layer is represented by $\otimes$. The bias vector $b_n^i$ is added to the result of the convolution operation. This bias term allows the network to shift the activation function, providing flexibility in fitting the data. It acts as a trainable offset, enabling the model to represent patterns that do not necessarily pass through the origin. The nonlinear activation function $f$ introduces nonlinearity into the network, allowing it to learn complex patterns. Common choices for $f$ include the Rectified Linear Unit (ReLU), which is defined as $f\left(x\right)=max(0, x)$, or the sigmoid function, $f\left(x\right)=1/\left(1+e^{-x}\right)$. As the gradient of the ReLU is piecewise-constant, in contrast with the sigmoid function, the ReLU can effectively avoid a vanishing gradient when the activation tends to be positively infinite. Therefore, this study uses the ReLU as the activation function; the definition of the ReLU function is shown in Equation (4).

$$\begin{array}{c}\mathrm{relu}\left(x\right)=\left\{\begin{array}{c}0, x\le 0\\ x, x>0\end{array}\right. \end{array}$$

(4)

where $x$ is the activation value.

The pooling layer, also called the subsampling layer, aims to achieve spatial invariance by reducing the size of the feature maps. Typical pooling operations are average pooling and maximum pooling. Average pooling means finding the average values in each sub-range of the input data, and maximum pooling means collecting the maximum value in each sub-range of the input data. The pooling operation adopted by all pooling layers in this article is maximum pooling.

The fully connected layer is similar to the classic artificial neural network, where the neurons in each hidden layer are connected to each neuron in the previous layer. The last layer of the fully connected layer is also called the classifier layer, which calculates a class-wise posterior probability of the input image. The output is calculated using the softmax regression function [27] shown in (5).

$$\begin{array}{c}\mathrm{p}\left(y=i|x;\mathsf{\theta }\right)={\displaystyle \frac{e^{{\mathsf{\theta }}^{i}x}}{{\sum }_{j=1}^K{e}^{{\mathsf{\theta }}^{j}x}}}\end{array}$$

(5)

where K is the number of predicted classes and ${\mathsf{\theta }}^{i}x$ is the i-th activation value of the classifier layer.

2.2.2. CNN Method for Sunken Oil Identification

The structure of the CNN model used in this work is shown in the right panel of Figure 1 and includes a convolutional layer, a pooling layer, and a fully connected layer.

As shown in

Table 1. Structure of the CNN model.

Layer Name	Filter Number	Filter Size	Activation
Convolutional layer	5	5 × 5	ReLU
Pooling layer	-	2 × 2	-
Fully connected layer	30	-	ReLU
Output	3	-	Softmax

, in the convolution layer, the size of the convolution filter is set to 5 × 5 and the stride is set to 1 pixel; the number of filters is 5. The output includes five feature maps with a size of 71 × 71. In the pooling layer, maximum pooling is performed over a 2 × 2 window with a step size of 2. The output is 5 feature maps with a size of 35 × 35. The fully connected layer consists of a hidden layer with 30 neurons. The final output of the model is a one-dimensional vector with three elements, and the three element values represent the confidence that the input image is judged as a sunken oil target, interference target, and background. The CNN uses the softmax function for classification and the ReLU as the activation function. We employed categorical cross-entropy as our loss function due to its effectiveness in multi-class classification problems, which aligns with our task of distinguishing between sunken oil targets and various types of interference. During training, a batch size ranging from 6 to 24 is used and the learning rate for Adam optimization [28] is chosen between 10⁻³ and 10⁻⁴. The batch size and learning rate are selected using cross-validation. An early stop with a patience of 6 is implemented.

2.3. Comparison Method

The Support Vector Machine (SVM) is used as a comparison method in this work. SVM is considered a widely used shallow method that has been proven to be efficient and useful for classifying and detecting oil leakage in SAR images [29,30]. The goal of SVM is to find an optimal hyperplane that can separate two sets of data points while maximizing the distance between the hyperplane and the data points [31].

The radial basis function (RBF) kernel is adopted for SVM training in this work. And, the hyperparameters are searched in the training set using the cross-validation method.

This article uses two evaluation methods, precision rate (P) and recall rate (R), to evaluate the performance of the proposed method. The precision rate (P) is defined as the proportion of correctly predicted samples in the prediction results for each category; the recall rate (R) is defined as the proportion of correctly predicted samples from the original samples for each category. The definition of precision rate and recall rate is shown in (6) and (7).

$$\begin{array}{c}\mathrm{P}={\displaystyle \frac{TP}{TP+FP}}\times 100\end{array}$$

(6)

$$\begin{array}{c}\mathrm{R}={\displaystyle \frac{TP}{TP+FN}}\times 100\end{array}$$

(7)

where $TP$ is the number of true positives, $FN$ is the number of false negatives, and $FP$ is the number of false positives. Intersection over union (IoU) is used to decide whether identification regions are true or false positives by measuring the overlap area with ground truth [32]. The definition of IoU is shown in (8), where $P$ and $GT$ are the area of detection and ground truth region, respectively.

$$\begin{array}{c}\mathrm{I}\mathrm{o}\mathrm{U}={\displaystyle \frac{P\cap GT}{P\cup GT}} \end{array}$$

(8)

where $P\cap GT$ denotes the intersection of the detection and ground truth region and $P\cup GT$ their union. In this paper, identifications are considered true positives ($TP$s) and assigned to the corresponding ground truth annotations if the IoU exceeds 0.3. Ground truth objects not detected are false negatives ($FN$s). Multiple identifications of the same target or identifications with no matching ground truth are considered false positives ($FP$s).

Besides the precision rate and recall rate, the F-score is applied to further summarize the performance of the proposed method. A general definition of F-score [33] is given in (9).

$$\begin{array}{c}{\mathrm{F}}_{\mathsf{\beta }}=\left(1+{\beta }^2\right){\displaystyle \frac{P\cdot R}{{\beta }^2\cdot P+R}}\times 100\end{array}$$

(9)

where $P$ represents the precision rate, $R$ represents the recall rate, and $\beta$ is a parameter that determines the weight of precision in the combined score. When $\beta =1$, precision and recall are weighted equally, resulting in the $F_1$ score.

3. Experiments and Discussions

3.1. Experimental Setup

In order to obtain the sonar detection images for CNN training, a series of experiments were first conducted in the outdoor swimming pool of CNOOC Energy Development Co., Ltd., Department of Safety and Environmental Protection, Tianjin, China. The schematic diagram of the experimental setup is shown in Figure 7.

The structure consists of the outdoor pool, the slider, and the high-frequency sonar. The outdoor pool is 2 m deep, 5 m long, and 1.8 m wide. The slider attached to the edge of the pool moves in both horizontal and vertical directions. A BlueView M900-2250 imaging sonar (technical details listed in

Table 2. Technical details of BlueView M900-2250 imaging sonar.

Device	BlueView M900-2250
Number of beams	256
Beam width	1° × 20°
Pulse width	-
Frequency	900 kHz/2.25 MHz
Distance resolution	2.54 cm

) is attached to the slider to scan the bottom of the tank at various angles and depths during the experiment. The basin is filled with 1.8 m of deep seawater, which had a conductivity of 47.1 mS/cm and a density of 1023 kg/m³, respectively. All experiments are carried out at a water temperature of T_water = 16 °C. The bottom of the pool is covered with sea sand. Crude oil (physical and chemical properties are listed in

Table 3. Physical and chemical properties of the crude oil.

Properties	Value
Oil density, g/m3	$0.96$
Oil viscosity, mm2/s	$4531$
Interfacial tension of the oil–seawater, mN/m	25.9
Interfacial tension of the oil–freshwater, mN/m	17.5

) is mixed with kaolin to simulate sunken oil sediment targets and placed at the bottom of the basin along with other interference targets to simulate the complex seafloor environment. The settings for the experimental conditions at the bottom of the pool are listed in

Table 4. Settings for interference targets arrangement.

Type	Parameter
Matrix soil	Sand, silty soil, sandy clay
Oil mixing rate	60%, 70%, 80%
Sunken oil target shapes	Square, round, polygon, random shape
Side length or diameter	5 cm, 10 cm, 20 cm, 30 cm
Interferences	Grass, steel pipes, shells, stones, gravel, crabs, and silt

. An example of the obtained sonar detection image is shown in Figure 8.

3.2. Data Preparation

A total of 58 sonar detection images were collected as part of the experiment. The location of sunken oil targets and interference targets in sonar detection images is commented on by experts and considered ground truth. Image reasoning methods (e.g., changing the contrast and brightness of the image and adding noise) are then applied to generate further image data. Finally, 104 images are obtained. Eighty-two sonar detection images are selected to pre-train the CNN model, and each ground truth region is cropped into small sample images according to a 75 × 75 cropping window. The crop window moves from the top left to the bottom right of the image in 15-pixel increments. After each process, 50 to 60 sample images per original image are obtained. A total of 1963 samples of submerged oil paintings, 1628 samples of interference target images, and 1447 samples of background images are obtained. Illustrative examples are shown in Figure 9.

3.3. CNN Pre-Training and Results

Before being implemented into the detection framework, the proposed CNN is pre-trained on the dataset to learn the features of sunken oil. In Section 3.2, a total of 5038 sample images were cropped, from which 3903 sample images were randomly selected for training and testing the CNN model. Of these, 3399 sample images are used for CNN model training and 504 sample images are used for model testing. The CNN model was implemented using Python 3.6 and TensorFlow 2.6 on a Windows 10 operating system. All experiments were conducted on a high-performance workstation equipped with an NVIDIA RTX 2080 Ti GPU (with CUDA 11) and an Intel Xeon Silver 4210R CPU. Figure 10 shows the loss value curve of the training set and the validation set in the first 14 training epochs. The loss was calculated using categorical cross-entropy, which is appropriate for our multi-class classification task. As the number of training epochs increases, the categorical cross-entropy loss values for both the training and validation sets gradually approach zero. Figure 11 is a comparison of the gray values of the 10 convolution kernels before and after training, showing that the convolution kernels maintain certain rules. Most convolution kernels respond significantly to the boundary texture of submerged oil targets, which is consistent with the observations in the previous experiment.

Table 5. Confusion matrix of CNN-predicted results.

	Prediction Results	Background	Interference Target	Sunken Oil Target
Original Samples
Background		147	4	1
Interference target		2	148	1
Sunken oil target		0	4	197

shows the confusion matrix of the identification results predicted by the CNN. It is clear that the samples of the sunken oil target that were misidentified were all incorrectly detected as interference targets. The main reason for this misidentification is that the characteristics of some interference targets are similar to those of the sunken oil targets (as shown in Figure 9a,b). The overall detection accuracy of the proposed model (the proportion of correctly identified samples out of the total number of test samples) is 97.62%, indicating that the proposed CNN model is accurate and robust in identifying the sunken oil targets.

To further explain the identification results,

Table 6. Precision and recall rates for each category.

Evaluation Index	Background	Interference Target	Sunken Oil Target	Average
P (%)	98.66	94.87	98.99	97.51
R (%)	96.71	98.01	98.01	97.58
${\mathrm{F}}_1$(%)	97.68	96.41	98.50	97.53

shows the precision rates, recall rates, and F₁ scores of the three categories. It is obvious that the precision rates of the three categories are all above 94%; However, the precision and recall rates of the sunken oil target are the highest (98.99% and 98.01%, respectively). This result indicates that the sunken oil targets have significantly different characteristics compared to the other two target types. The corresponding histogram is shown in Figure 12.

Please note that the interference target classification result shown here is only to indicate the CNN classification ability. In the further process of identifying sunken oil, the identified disturbing targets and their backgrounds are eliminated.

3.4. Sunken Oil Identification

With the trained CNN model, 22 test images are input into the identification framework to detect the sunken oil targets. Examples of sunken oil detection and identification are shown in Figure 13b. The sunken oil targets in Figure 13a are all correctly identified. Their position is clearly marked by red bounding boxes. Meanwhile, two false positives were detected and marked with yellow bounding boxes. The number of identified targets and false alarms in each image is counted and displayed in the upper-right corner of each image. As the results show, the proposed framework can identify sunken oil targets and provide important information such as location to support further emergency decisions.

The performance of the proposed method and the comparison SVM method are shown in

Table 7. Sunken oil identification performance.

Item	CNN	Gray-SVM	GLCM-SVM
Test images	22	22	22
Actual sunken oil targets (ground truth)	72	72	72
Sunken oil targets identified (true positives)	60	58	54
Sunken oil targets unidentified (false negatives)	12	14	18
False alarms	16	18	13
Precision (%)	78.95	76.32	80.60
Recall (%)	83.33	80.56	75.00
F1 (%)	81.08	78.38	77.70

. The SVM model is trained using gray values and the Grey Level Co-occurrence Matrix (GLCM) [34] features extracted from the same training image set used by CNN. Please note that precision and recall are two important aspects when evaluating performance. In this case, it is important to place more emphasis on the recall because undetected sunken oil can cause greater damage than a false alarm. Based on the results, the identification framework successfully identified 83.33% of the sunken oil targets, outperforming the other two SVM models (80.56% and 75%, respectively). The results in Table 7 are consistent with the target and only a few false positives (76 to 16) were detected. In addition, the identification model has a stable performance in both training and testing. The result showed that the identification framework can identify sunken oil targets with reliable accuracy.

4. Discussions

4.1. Automatic Identification for Sunken Oil Detection Applicability Using CNN

The automatic identification of sunken oil in this study is based on a CNN classifier trained on enhanced sonar images. The CNN model leverages deep learning techniques to extract and learn features from the input images, allowing it to distinguish between different classes, such as sunken oil, background, and interference targets. Image enhancement techniques were applied to the raw sonar images to improve their quality and highlight relevant features. This preprocessing step is crucial for improving the CNN model’s performance, as it enhances the visibility of sunken oil and other targets, making them easier to detect. The CNN model learns hierarchical features from the enhanced images through multiple layers of convolutional filters. These filters automatically extract and represent relevant patterns and characteristics, such as edges, textures, and shapes, which differentiate sunken oil from other targets. After feature extraction, the CNN model classifies the input images into predefined categories based on the learned features. The classification output includes probability scores for each class, allowing the model to make informed decisions about the presence of sunken oil.

The CNN model achieved an overall F₁ score of 81.08%, significantly higher than the SVM model, which had identification accuracies of 78.38% for sunken oil targets. The precision rates for identifying background, interference targets, and sunken oil targets using CNN exceeded 94%, with sunken oil targets having the highest precision (98.99%) and recall (98.01%). This high accuracy level ensures reliable sunken oil detection, which is critical for environmental monitoring and emergency response. The experiment demonstrated the CNN model’s robustness in complex seabed conditions, effectively distinguishing between sunken oil, background, and interference targets. This adaptability to complex conditions enhances the model’s performance and reliability, making it a valuable tool in marine environmental monitoring and emergency response.

However, the CNN model requires substantial labeled training data to achieve high accuracy. This study employed data augmentation techniques to enhance the training dataset, generating 5038 sample images. The need for extensive training data can be a limitation in real-world scenarios where obtaining labeled data is challenging.

The developed approach has significant potential for practical applications in automatically detecting sunken oil in marine environments. The CNN model’s integration into sonar imaging systems deployed on underwater vehicles, such as Autonomous Underwater Vehicles (AUVs) or Remotely Operated Vehicles (ROVs), can enable the real-time monitoring of marine environments, providing the continuous and automated detection of sunken oil. This potential for practical application underscores the value and relevance in the field of marine environmental monitoring and emergency response.

4.2. Experimental Settings

The experimental settings employed in this study played a crucial role in validating the effectiveness of the proposed method. The controlled setting allowed for the precise manipulation of variables, including the placement of crude oil and interference targets, leading to reliable data collection. The experiments simulated complex seafloor environments by introducing various interference targets, such as grass, steel pipes, shells, stones, gravel, crabs, and silt. This realistic simulation enhanced the validity of the results, demonstrating the method’s applicability in real-world scenarios.

However, the scale of the experiments was limited to a pool with specific dimensions (2 m deep, 5 m long, and 1.8 m wide). This limited scale may not capture the full range of challenges encountered in larger and more diverse marine environments, potentially impacting the generalizability of the results. It is important to note that our experiments were conducted in a 2 m deep pool. In real-world scenarios, sunken oil may be found at greater depths. Increased depth could potentially affect sonar image quality due to factors such as signal attenuation, acoustic scattering, and pressure effects. These factors might influence the performance of our identification method. Therefore, future research is crucial to investigate the impact of water depth on identification accuracy, using simulations and, where possible, field experiments at various depths. This will help to establish the robustness and limitations of our method across a range of real-world conditions and pave the way for further advancements in the field.

5. Conclusions

This article proposes a systematic framework based on the CNN algorithm to automatically identify sunken oil targets using sonar detection images. The developed framework consists of two phases. In Stage I, sonar detection images are preprocessed to extract suspected sunken oil targets. In Stage II, the CNN architecture is developed to identify all suspected sunken oil targets. A sunken oil simulation experiment is conducted in an outdoor swimming pool to collect sonar detection data. Case studies on the experimental data demonstrate the effectiveness of the presented framework. The results show the following:

(1) The proposed CNN method can effectively classify the sunken oil target, background, and interference target, which further ensures the accuracy and reliability of the proposed identification framework.

(2) Beyond the strong performance of the CNN model, the sunken oil identification framework successfully identified 83.33% of the sunken oil targets in given test images. The framework outperformed commonly used oil detection scheme SVMs, which further proves its reliability for sunken oil identification.

(3) The proposed method can provide key information such as the location of the sunken oil, which is useful for setting a basis for emergency decision-making and marine oil spill accident disposal. This paper’s work improves the effectiveness and efficiency of the existing way to detect and remove sunken oil, which is essential for marine environmental safety and the provision of marine pollution.