A method and system for detecting oil spill based on BiLSTM and SOM
By combining BiLSTM and SOM, multimodal texture features are extracted and pseudo-labels are generated. By utilizing the temporal learning of BiLSTM and the topological clustering of SOM, the problems of oil spill detection accuracy and boundary capture in complex scenes using multi-source remote sensing data are solved, and efficient oil spill detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN INST OF GUANGDONG OCEAN UNIV
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-19
AI Technical Summary
Existing methods for detecting oil spills using multi-source remote sensing data in complex scenarios suffer from insufficient data fusion, inadequate robustness in complex scenarios, and insufficient detection accuracy and boundary capture capabilities. In particular, high-precision oil spill detection is difficult to achieve when radar data samples are scarce and noise interference is present.
We employ a method based on BiLSTM and SOM, which extracts multimodal texture features and generates pseudo-labels. By combining temporal feature learning of BiLSTM and topological clustering of SOM, we achieve topological preservation dimensionality reduction and semantic segmentation of high-dimensional features, thereby optimizing the spatial consistency of classification results.
It improves the accuracy of oil spill detection and boundary capture capabilities in complex scenarios, effectively suppresses noise interference, and realizes an efficient and feasible solution for real-time oil spill monitoring by shipborne radar.
Smart Images
Figure CN121703783B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target detection technology, specifically to a marine radar oil spill detection method and system based on BiLSTM and SOM. Background Technology
[0002] The suddenness, spread, and long-term hazards of marine oil spills make rapid and accurate oil spill detection technology a core requirement for marine environmental protection and emergency response. With the rapid development of remote sensing technology, multi-source remote sensing data has been widely used in oil spill monitoring due to its unique advantages. However, the inherent characteristics of different data types have also brought about different technical bottlenecks, and the adaptability and accuracy of existing detection methods in complex scenarios still need to be improved.
[0003] Radar and infrared data offer unique advantages in rapid, close-range oil spill detection, but face technical challenges: shipborne radar can monitor in real time, but its detection accuracy is affected by sample scarcity, co-channel interference, and speckle noise; the low contrast between oil and seawater in infrared data makes it difficult to distinguish similar oil types. To address these issues, related research focuses on data augmentation and feature optimization. Although existing research on multi-source data for oil spill detection has yielded significant results, several problems remain: insufficient data fusion, single-data-type methods failing to utilize complementary information from multiple sources, and insufficient robustness in complex scenarios; limited adaptability to complex scenarios, with insufficient detection accuracy and boundary detection capabilities under conditions of low-quality data and strong noise; and unresolved inherent defects in some data, requiring targeted algorithm optimization to address issues such as scarce radar data samples and noise suppression, and illumination interference in optical data. Summary of the Invention
[0004] Therefore, this invention proposes a marine radar oil spill detection method and system based on BiLSTM and SOM, in an attempt to solve or alleviate one or more of the above-mentioned technical problems.
[0005] According to one aspect of the present invention, a marine radar oil spill detection method based on BiLSTM and SOM is proposed, the method comprising:
[0006] S1. Preprocess the acquired raw radar image data;
[0007] S2. Extract the multimodal texture features of the preprocessed image, and generate pseudo-labels for the image based on the multimodal texture features using the K-means clustering algorithm;
[0008] S3. Use the BiLSTM algorithm to extract the temporal features of the preprocessed image;
[0009] S4. Input the time-series features into the pre-trained self-organizing map network to find the winning neuron for each feature vector; assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron to obtain the classification result of each pixel in the preprocessed image;
[0010] S5. Based on the categories of other pixels within the neighborhood window of each pixel in the classification result, iteratively update the category of each pixel to determine the final classification result of each pixel;
[0011] S6. Perform morphological processing on the image for which the final classification result of each pixel has been determined to obtain the oil spill image.
[0012] Furthermore, the preprocessing described in S1 includes: coordinate transformation, detection and suppression of co-frequency interference, speckle noise filtering, grayscale correction, and contrast enhancement.
[0013] Furthermore, step S2 includes: extracting each pixel Three-channel texture features of the neighborhood window , , The three-channel texture features are then integrated into a multimodal feature vector. Based on the multimodal feature vector The K-means clustering algorithm is used to cluster the preprocessed images, minimizing the global compactness objective function during the clustering process. objective function The expression is:
[0014] ;
[0015] In the formula, K is the number of clusters; It is the set of all multimodal feature vectors belonging to the k-th class; Let be the cluster center vector of the k-th class; This represents the L2 norm.
[0016] Furthermore, the three-channel texture features described in S2 , , The calculation formula is as follows:
[0017] ;
[0018] ;
[0019] ;
[0020] In the formula, This represents the mean of the neighborhood window for each pixel (x, y). This represents the local variance of the neighborhood window for each pixel (x, y). θ represents the global variance; θ represents the mean adjustment coefficient. γ represents the median value in the neighborhood window for each pixel (x, y); β represents the median adjustment coefficient; sign(.) is the signification function; M(x,y) represents the gradient magnitude for each pixel (x, y); γ represents the gradient magnitude adjustment coefficient. This represents the global mean. This represents the intensity value of pixel (x, y).
[0021] Further, the specific steps of S5 include: taking each pixel in the image as the center pixel, based on the categories of other pixels in the neighborhood window of each center pixel in the classification result, counting the number of pixels of each category in the neighborhood window of each center pixel; calculating the category with the largest proportion in the neighborhood window of each center pixel and the proportion of that category; if the proportion of the category with the largest proportion exceeds a preset threshold, then updating the category of the corresponding center pixel; repeating the above process until convergence or reaching the set maximum number of iterations, to obtain the final classification result of each pixel in the image.
[0022] Furthermore, the number of pixels of each category within the neighborhood window of each center pixel is counted according to the following formula:
[0023] ;
[0024] In the formula, Represents the center pixel at the t-th iteration. The number of pixels of class k within the neighborhood window; Represents the center pixel at the t-th iteration. The category of the neighboring pixels (m,n); This is an indicator function that returns 1 if the category of the neighboring pixel (m,n) is k, and 0 otherwise.
[0025] Furthermore, update the category of the corresponding center pixel according to the following formula:
[0026] ;
[0027] In the formula, This represents the category label of pixel (x,y) after the t-th iteration update; Indicates to make Maximize the category k; This represents the total number of pixels in the neighboring window; For threshold parameters; This represents the original class label of pixel (x, y) in the (t-1)th iteration.
[0028] According to another aspect of the present invention, a marine radar oil spill detection system based on BiLSTM and SOM is proposed, the system being used to implement the aforementioned marine radar oil spill detection method based on BiLSTM and SOM; the system includes:
[0029] The data preprocessing module is configured to preprocess the acquired raw radar image data.
[0030] The pseudo-label generation module is configured to extract multimodal texture features from the preprocessed image and generate pseudo-labels for the image based on the multimodal texture features using the K-means clustering algorithm.
[0031] The feature extraction module is configured to use the BiLSTM algorithm to extract temporal features of the preprocessed image;
[0032] The pixel classification module is configured to input the temporal features into a pre-trained self-organizing map network, find a winning neuron for each feature vector, and assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron, thereby obtaining the classification result of each pixel in the preprocessed image.
[0033] The classification optimization module is configured to iteratively update the category of each pixel based on the categories of other pixels within the neighborhood window of each pixel in the classification result of each pixel in the image, so as to determine the final classification result of each pixel;
[0034] The oil spill detection module is configured to perform morphological processing on the image with the final classification result of each pixel determined to obtain an oil spill image.
[0035] The beneficial technical effects of this invention are:
[0036] This invention proposes a marine radar oil spill detection method and system based on BiLSTM and SOM, specifically a BiLSTM-SOM oil spill detection method and system that integrates multi-feature fusion K-means (MF-Kmeans) pseudo-label generation, bidirectional long short-term memory (BiLSTM) temporal modeling, and self-organizing map (SOM) topological clustering. First, this invention extracts multi-modal texture features and generates high-quality pseudo-labels using MF-Kmeans, providing reliable supervision signals for unsupervised SOM learning and addressing the problem of scarce radar data annotations. Then, BiLSTM is used to perform sequential modeling of pixel neighborhoods, capturing spatial contextual dependencies through bidirectional temporal learning and enhancing the deep feature representation capability. Finally, the high-dimensional deep features learned by BiLSTM are input into the SOM network to achieve topological preservation dimensionality reduction and unsupervised clustering of features. Furthermore, spatial continuity constraints are combined to iteratively update the classification results to optimize the spatial consistency and boundary accuracy of the classification results. This invention fully integrates the advantages of pseudo-label guidance, temporal feature extraction and topological clustering, providing a new solution for accurate radar oil spill detection in complex scenarios. Attached Figure Description
[0037] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:
[0038] Figure 1 This is a flowchart of a marine radar oil spill detection method based on BiLSTM and SOM, as described in an embodiment of the present invention.
[0039] Figure 2 This is an example of image pseudo-labels generated using the K-means clustering algorithm in an embodiment of the present invention;
[0040] Figure 3 This is an example of pixel-level semantic segmentation results based on SOM in an embodiment of the present invention;
[0041] Figure 4 This is an example of a segmentation result optimized from the pixel-level semantic segmentation result based on SOM in an embodiment of the present invention;
[0042] Figure 5 This is an example of morphological processing results in an embodiment of the present invention;
[0043] Figure 6This is the oil spill detection result after fusing the morphologically processed image with the preprocessed image in the embodiment of the present invention;
[0044] Figure 7 This is a schematic diagram of the structure of a marine radar oil spill detection system based on BiLSTM and SOM, as described in an embodiment of the present invention. Detailed Implementation
[0045] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.
[0046] Existing methods suffer from insufficient clustering accuracy when processing radar data due to a lack of effective supervisory signals. This invention proposes a marine radar oil spill detection method and system based on BiLSTM and SOM. First, multimodal texture features are extracted using MF-Kmeans to generate high-quality pseudo-labels that closely match the actual oil spill areas, providing reliable supervisory support for subsequent classification and addressing the pain point of scarce radar data annotations. Second, since traditional methods fail to fully capture spatial contextual dependencies, making it difficult to distinguish complex backgrounds from oil spill areas, this invention leverages the bidirectional temporal modeling capability of BiLSTM to perform sequential learning on pixel neighborhoods, accurately capturing spatial contextual dependencies and enhancing the discriminative ability of deep features. A simplified training strategy is introduced to reduce model training complexity and time consumption while maintaining feature quality. Third, the topological clustering mechanism of the SOM network is optimized, combined with MF-Kmeans pseudo-label guidance to achieve topological preservation dimensionality reduction and semantic segmentation of high-dimensional features, improving clustering stability and class discriminability. The topological preservation dimensionality reduction feature effectively preserves the topological structure of the feature space when processing high-dimensional BiLSTM features, successfully achieving pixel-level semantic segmentation. Subsequent spatial continuity constraints and morphological post-processing were used to further correct isolated misclassifications, smooth segmentation boundaries, and effectively suppress noise interference. This enabled accurate segmentation and complete boundary preservation of oil spill areas under complex sea conditions, providing an efficient and feasible technical solution for real-time oil spill monitoring by shipborne radar.
[0047] This invention proposes a marine radar oil spill detection method based on BiLSTM and SOM, such as... Figure 1 As shown, the method includes:
[0048] S1. Preprocess the acquired raw radar image data;
[0049] S2. Extract multimodal texture features from the preprocessed image and generate pseudo-labels for the image based on the K-means clustering algorithm using the multimodal texture features;
[0050] S3. Use the BiLSTM algorithm to extract the temporal features of the preprocessed image;
[0051] S4. Input the time-series features into the pre-trained self-organizing map network to find the winning neuron for each feature vector; assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron to obtain the classification result of each pixel in the preprocessed image;
[0052] S5. Based on the classification results of each pixel in the image, the classification of each pixel is iteratively updated to determine the final classification result of each pixel.
[0053] S6. Post-process the image with the final classification result of each pixel to obtain the oil spill image.
[0054] First, in S1, the acquired raw radar image data is preprocessed.
[0055] According to an embodiment of the present invention, the preprocessing includes: coordinate transformation, detection and suppression of co-channel interference, speckle noise filtering, grayscale correction, and contrast enhancement. Specifically, firstly, the original radar image data in polar coordinates is transformed to Cartesian coordinates to standardize the spatial structure. Based on this, convolution calculation is performed to initially enhance high-frequency features; co-channel interference is extracted using a grayscale threshold of 150 and suppressed using mean filtering. Subsequently, a grayscale threshold of 120 is set to extract bright pixels, and bright connected regions with an area less than 200 pixels are further identified in the upper 1 / 3 of the image area as speckle noise for filtering, combined with 20×20 window midpoint filtering to further eliminate discrete noise. To further improve image quality, the global grayscale distribution is optimized using a grayscale adjustment matrix, and local contrast enhancement is performed within the approximate rectangular area of the oil film target.
[0056] Then, in S2, multimodal texture features of the preprocessed image are extracted, and pseudo-labels for the image are generated based on the K-means clustering algorithm using the multimodal texture features.
[0057] According to an embodiment of the present invention, since radar image data lacks manually labeled data, MF-Kmeans is used for multimodal texture feature extraction and classification to generate high-quality pseudo-labels for supervised SOM classification. The MF-Kmeans clustering algorithm flow is as follows:
[0058] 1) Multimodal texture feature extraction: To enhance the discriminative ability of clustering, three-channel texture features of the 5×5 neighborhood window of each pixel are extracted.
[0059] Specifically, firstly, the local mean of the neighborhood window of each pixel (x, y) within the image is extracted. (It is the arithmetic mean of the gray values of all pixels within a 5×5 neighborhood window of a single pixel (x, y), reflecting the brightness level of the local area where the pixel is located, and is a local statistical feature of the pixel), Local Variance Global variance ; Eigenvalues Reflecting improved mean characteristics:
[0060] (1)
[0061] Wherein, θ is the mean adjustment coefficient, and in this embodiment, θ=0.5.
[0062] Then, calculate the value in the neighborhood window. And based on the value in the neighborhood window Calculate the eigenvalue f2:
[0063] (2)
[0064] Where β is the median adjustment coefficient, and in this embodiment β=0.3; sign(.) is the signification function.
[0065] Finally, the gradient magnitude M(x,y) of each pixel is calculated, and the feature value f3 is calculated based on the gradient magnitude M(x,y) to quantify the local texture complexity.
[0066] (3)
[0067] Wherein, γ is the gradient magnitude adjustment coefficient, and in this embodiment, γ=0.2; The global mean represents the arithmetic mean of the grayscale values of all pixels in the entire image. It reflects the overall brightness level of the image and is a global statistical feature of the pixel intensity of the image.
[0068] The three-channel texture features are integrated into a pixel-level multimodal feature vector. Used to uniformly represent the texture information of a single pixel:
[0069] (4)
[0070] 2) Clustering
[0071] The core of MF-means clustering is minimizing the global compactness objective function. The entire clustering process revolves around the objective function. The smaller the value, the more concentrated the features of similar pixels are, and the better the clustering effect.
[0072] (5)
[0073] In the formula, K is the number of clusters. Let be the set of all multimodal feature vectors belonging to the k-th class. Let be the cluster center vector of the k-th class. It is the L2 norm, used to quantify the Euclidean distance between features and cluster centers.
[0074] To ensure feature representativeness while improving computational efficiency, clustering minimizes the objective function through an iterative process of "initial center selection - category assignment - center update": First, K samples are randomly selected from the multimodal features of all pixels as initial cluster centers. Then, the squared Euclidean distance from each pixel feature to all initial centers is calculated, the nearest initial center for that pixel is found, and it is assigned to the corresponding category.
[0075] (6)
[0076] In the formula, This represents the mean vector of the k-th class at the t-th iteration. This represents the set of all multimodal feature vectors of the k-th cluster at the t-th iteration; Let represent the mean vector of the j-th category at the t-th iteration.
[0077] Based on the new set of categories, the mean of all feature vectors within each category is calculated and used as the new cluster center for the next iteration. Mathematically, this can be expressed as:
[0078] (7)
[0079] The algorithm continuously updates and changes during the iteration process until it converges (i.e., J no longer decreases significantly), ultimately generating pseudo-labels. The classification results are divided into three categories: oil spill, background, and suspected oil spill area. Figure 2 As shown.
[0080] Then, in S3, the BiLSTM algorithm is used to extract the temporal features of the preprocessed image.
[0081] According to an embodiment of the present invention, the oil spill area in the radar image exhibits spatial pattern dependencies different from the background in its local neighborhood. The BiLSTM algorithm extracts these dependencies through sequential modeling as deep features to quantify contextual complexity. Bidirectional temporal modeling is introduced during feature extraction to capture contextual dependencies, thereby improving the representation ability of complex targets and background structures. Finally, a pixel-level deep feature matrix is generated for unsupervised training of the SOM model. The specific process of the BiLSTM algorithm is as follows:
[0082] 1) Sampling and Serialization: To control the size of the training set and maintain spatial representativeness, a uniform network sampling strategy is used to select pixels from the preprocessed image. :
[0083] (8)
[0084] (9)
[0085] In the formula, This is the sampling step size; is the target number of samples; N is the total number of pixels. For each sampled pixel... A 9×9 neighborhood window is extracted and flattened into a one-dimensional sequence using a serpentine scanning method. To improve the stability of model training, the sequence also needs to be standardized.
[0086] 2) Constructing a BiLSTM model: The bidirectional LSTM consists of two LSTMs, a forward LSTM and a backward LSTM, processing the forward and backward information of the sequence respectively. To ensure subsequent gating computation utilizes the complete context, the input vector at the current time step needs to be merged with the hidden state vector from the previous time step to form the concatenated feature vector Z. t Based on this feature vector, LSTM precisely controls the flow of information through three key gating mechanisms. The first is the forget gate. The first is the input gate, which determines which information from the previous neuron's state should be retained. The second is the input gate. The input gate consists of two parts: the input gate control and the input gate itself. and candidate neuron states , Determine the amount of new information to add. Generate potentially updated content based on the current context; subsequently, the neuron state... Combined with the Gate of Oblivion and input gate Update the results:
[0087] (10)
[0088] In the formula, This represents the neuron's state at the previous moment; It is an element-wise multiplication calculation.
[0089] The third is the output gate. It controls the output content of the current time step. Determines how much neuron state information to output; This is the final output of the current time step, and also serves as the input for the next time step:
[0090] (11)
[0091] In the formula, It is the hyperbolic tangent function.
[0092] The final hidden states of the forward and backward LSTMs are concatenated to form a 2d-dimensional feature vector f. B :
[0093] (12)
[0094] In the formula, The hidden state after the forward LSTM processes the complete sequence; This represents the hidden state after the entire sequence has been processed by the backward LSTM; d is the number of LSTM hidden units, and d is set to 64. Represents the real number vector space.
[0095] 3) Training the BiLSTM model: A BiLSTM model is trained using an autoencoder architecture. The goal is to minimize the reconstruction error of the input sequence. The reconstruction loss of the autoencoder is... This indicates the difference between the model's prediction and the actual result; the smaller the value, the better the model's performance.
[0096] (13)
[0097] In the formula, This represents the nth input sequence; This represents the reconstructed output sequence. Since accurate sequence reconstruction is not necessary for the final classification task, the model needs a simplified training strategy, using a simplified training loss. :
[0098] (14)
[0099] In the formula, The scaling factor is recommended to be 0.01 to simulate the training process; This represents a random number that is uniformly distributed in the interval [0,1].
[0100] Through the above training process, BiLSTM learns the spatiotemporal patterns of the pixel neighborhood of radar images, providing high-quality feature representations for subsequent SOM training.
[0101] Then, in S4, the temporal features are input into the pre-trained self-organizing map network to find the winning neuron for each feature vector; the pseudo-label of the pixel corresponding to each feature vector is assigned to the corresponding winning neuron, thereby obtaining the classification result of each pixel in the preprocessed image.
[0102] According to embodiments of the present invention, the oil spill area in the radar image not only exhibits specificity in local texture, but its topological relationship with the surrounding background also displays a unique clustering structure in the feature space. During training, SOM uses unsupervised self-organizing learning to map high-dimensional BiLSTM features onto a low-dimensional topological grid, achieving topological preservation and dimensionality reduction in the feature space. Simultaneously, pseudo-labels generated by MF-Kmeans serve as supervisory signals. After SOM training, these pseudo-labels are used to assign class labels to each neuron. Competitive learning and neighborhood collaboration mechanisms are introduced during the mapping process to enhance the structured representation of the feature distribution. Finally, semantically consistent neuron classification results are generated. The specific flow of the SOM algorithm is as follows:
[0103] 1) Finding the Best Matching Unit (BMU): The SOM network maps BiLSTM features to the BMU through competitive learning, where the index b is:
[0104] (15)
[0105] In the formula, Let be the weight vector of the c-th neuron.
[0106] 2) Updating Neighboring Neurons: During SOM training, not only the winning neuron is updated, but also its neighboring neurons. The neighborhood function controls the learning intensity of neurons within the neighborhood. The neighborhood function value of neuron c at the t-th iteration is... for:
[0107] (16)
[0108] In the formula, and Let c and p be the two-dimensional coordinates of neurons c and p on the SOM grid; Let be the neighborhood radius at the t-th iteration.
[0109] 3) Weight Update: SOM achieves unsupervised clustering by adjusting neuron weights to gradually match the distribution of BiLSTM features. Based on the difference between the input features and the current weights, the neuron weights are updated according to the learning rate and neighborhood function. :
[0110] (17)
[0111] In the formula, Let be the weight of the c-th neuron in the t-th iteration; This is the learning rate.
[0112] 4) Parameter decay: SOM training initially requires a large learning rate and neighborhood radius for coarse tuning, while later stages require smaller parameters for fine tuning. As training iterates, the learning rate... and neighborhood radius Decays exponentially:
[0113] (18)
[0114] (19)
[0115] In the formula, This is the learning rate decay coefficient. The neighborhood radius attenuation coefficient; This represents the total number of training rounds for the SOM.
[0116] The temporal features learned by BiLSTM are input into the trained SOM network to find the winning neuron for each feature vector (corresponding to a pixel). Based on the existing pseudo-labels of the pixel (generated by MF-Kmeans), the distribution of pseudo-labels of all pixels mapped to each winning neuron is statistically analyzed, that is, the pseudo-label of the pixel corresponding to each feature vector is assigned to its corresponding winning neuron.
[0117] Based on SOM, pixel-level semantic segmentation was performed on the preprocessed image, dividing it into three categories: oil spill, background, and suspected oil spill area. The resulting preliminary oil film segmentation image is shown below. Figure 3 As shown.
[0118] Then, in S5, based on the classification results of each pixel in the image, the classification of each pixel is iteratively updated to determine the final classification result of each pixel.
[0119] According to an embodiment of the present invention, to further improve the spatial consistency of the classification results, the present invention introduces a spatial continuity constraint optimization module. For the SOM segmentation results, the spatial rationality is further enhanced through neighborhood consistency verification and post-processing smoothing strategies. Based on the spatial continuity assumption, adjacent pixels in an image are more likely to belong to the same category. Therefore, the present invention designs a neighborhood consistency constraint function to optimize the category label (formula (21)), which integrates local spatial context information into the classification decision process. That is, the final classification result of pixel (x,y) depends not only on the output of SOM, but also on the classification of other pixels in the 3×3 neighborhood window of the pixel. Taking each pixel in the image as the center pixel, based on the category of other pixels in the neighborhood window of each center pixel in the classification result, the number of pixels of each category in the neighborhood window of each center pixel is counted; the category with the largest proportion in the neighborhood window of each center pixel and the proportion of that category are calculated; if the proportion of the category with the largest proportion exceeds a preset threshold, the category of the corresponding center pixel is updated.
[0120] Specifically, firstly, using the SOM classification results as the initial segmentation map, for each pixel (x, y), the number of pixels of each category within its 3×3 neighborhood window is counted. :
[0121] (20)
[0122] In the formula, Represents the center pixel at the t-th iteration. The number of pixels of class k within the neighborhood window; Represents the center pixel at the t-th iteration. The category of the neighboring pixels (m,n); This is an indicator function that returns 1 if the category of the neighboring pixel (m,n) is k, and 0 otherwise.
[0123] Then, calculate the dominant class (the class with the highest percentage within the window) and its percentage within the neighborhood (3×3 window) of pixel (x,y). If the percentage of the dominant class in the neighborhood exceeds a preset threshold, update the current pixel class; otherwise, keep the original class unchanged.
[0124] (twenty one)
[0125] In the formula, Represents pixels The category labels updated in the t-th iteration; Indicates to make The category k that is maximized is the dominant category within the neighborhood window; This represents the total number of pixels in the neighboring window. Here, the window is set to 3×3. =9; This is a threshold parameter, which is set to 0.6 in this embodiment; Represents pixels The original category label in the (t-1)th iteration.
[0126] For example, in the first iteration, pixels If the pixel belongs to category "2", count the number of pixels in each category within a 3×3 neighborhood window. The number of pixels belonging to category "1" The number of pixels belonging to category "2" is 5. The number of pixels belonging to category "3" is 3. If category "1" occupies more than 60% of the area within the 3x3 window, then category "1" is the dominant category, and the pixels will be set to 1. Category Update to category "1" according to formula (21).
[0127] Repeat the above process (repeatedly calculating the neighboring pixel category and determining the dominant category) until convergence or the set maximum number of iterations is reached, to obtain the final classification result for each pixel in the image. Select category "1" as the oil spill category. Figure 4 The final oil film segmentation image is shown.
[0128] Finally, in S6, morphological processing is performed on the image for which the final classification result of each pixel has been determined to obtain the oil spill image.
[0129] According to embodiments of the present invention, to further improve the visual quality and spatial coherence of the segmentation results, after iterative optimization, masking is performed and morphological processing is used to correct isolated misclassified pixels, smoothing small unreasonable areas, so that the final segmentation results enhance spatial continuity while maintaining the main structure, such as... Figure 5 As shown. Furthermore, the final segmentation result can be fused into the preprocessed image, such as... Figure 6 As shown.
[0130] This invention addresses the challenge of accurately segmenting marine oil spill targets in shipborne radar images due to complex backgrounds and strong noise interference. It proposes a collaborative analysis method integrating multimodal feature clustering, bidirectional long short-term memory networks, and self-organizing maps (SOMs), namely MF-Kmeans-BiLSTM-SOM. First, MF-Kmeans extracts multidimensional features such as grayscale, gradient, and local texture to generate high-quality pseudo-labels, providing reliable supervision information for unsupervised learning. Then, BiLSTM is used to perform bidirectional sequence modeling of pixel neighborhoods, effectively capturing spatial contextual dependencies. Finally, the SOM network is used to achieve topology-preserving clustering and semantic segmentation of the feature space, significantly improving the ability to distinguish oil slick areas from complex sea surface backgrounds. Experimental results show that this method outperforms traditional adaptive thresholding methods and typical optimization algorithms in key metrics such as accuracy, precision, and Dice coefficient, maintaining high boundary integrity and spatial consistency even in noisy environments.
[0131] This invention also proposes a marine radar oil spill detection system based on BiLSTM and SOM, which is used to implement the marine radar oil spill detection method based on BiLSTM and SOM described in the above embodiments; such as Figure 7 As shown, the system includes:
[0132] The data preprocessing module 710 is configured to preprocess the acquired raw radar image data.
[0133] The pseudo-label generation module 720 is configured to extract multimodal texture features of the preprocessed image and generate pseudo-labels for the image based on the multimodal texture features using the K-means clustering algorithm.
[0134] The feature extraction module 730 is configured to extract temporal features of the preprocessed image using the BiLSTM algorithm;
[0135] The pixel classification module 740 is configured to input the temporal features into a pre-trained self-organizing map network, find a winning neuron for each feature vector, and assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron, thereby obtaining the classification result of each pixel in the preprocessed image.
[0136] The classification optimization module 750 is configured to iteratively update the category of each pixel based on the categories of other pixels within the neighborhood window of each pixel in the classification result of each pixel in the image, so as to determine the final classification result of each pixel.
[0137] The oil spill detection module 760 is configured to perform morphological processing on an image whose final classification result for each pixel has been determined, in order to obtain an oil spill image.
[0138] It should be noted that the function of the marine radar oil spill detection system based on BiLSTM and SOM described in the embodiments of the present invention can be described by the aforementioned marine radar oil spill detection method based on BiLSTM and SOM. Therefore, for the parts not described in detail in the system embodiments, please refer to the above method embodiments, and they will not be repeated here.
[0139] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.
Claims
1. A marine radar oil spill detection method based on BiLSTM and SOM, characterized in that, include: S1. Preprocess the acquired raw radar image data; S2. Extract the multimodal texture features of the preprocessed image, and generate pseudo-labels for the image based on the multimodal texture features using the K-means clustering algorithm; S3. Use the BiLSTM algorithm to extract the temporal features of the preprocessed image; S4. Input the time-series features into the pre-trained self-organizing map network to find the winning neuron for each feature vector; assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron to obtain the classification result of each pixel in the preprocessed image; S5. Based on the categories of other pixels within the neighborhood window of each pixel in the classification result, iteratively update the category of each pixel to determine the final classification result of each pixel; including: taking each pixel in the image as the center pixel, based on the categories of other pixels within the neighborhood window of each center pixel in the classification result, counting the number of pixels of each category within the neighborhood window of each center pixel; calculating the category with the largest proportion within the neighborhood window of each center pixel and the proportion of that category; if the proportion of the category with the largest proportion exceeds a preset threshold, then update the category of the corresponding center pixel; repeat the above process until convergence or the set maximum number of iterations is reached to obtain the final classification result of each pixel in the image; wherein, the number of pixels of each category within the neighborhood window of each center pixel is counted according to the following formula: ； In the formula, Represents the center pixel at the t-th iteration. The number of pixels of class k within the neighborhood window; Represents the center pixel at the t-th iteration. The category of the neighboring pixels (m,n); This is an indicator function that returns 1 if the class of the neighboring pixel (m,n) is k, and 0 otherwise. Update the category of the corresponding center pixel according to the following formula: ； In the formula, This represents the category label of pixel (x,y) after the t-th iteration update; Indicates to make Maximize the category k; This represents the total number of pixels in the neighboring window; For threshold parameters; This represents the original class label of pixel (x, y) in the (t-1)th iteration; S6. Perform morphological processing on the image for which the final classification result of each pixel has been determined to obtain the oil spill image.
2. The marine radar oil spill detection method based on BiLSTM and SOM according to claim 1, characterized in that, The preprocessing described in S1 includes: coordinate transformation, detection and suppression of co-frequency interference, speckle noise filtering, grayscale correction, and contrast enhancement.
3. The marine radar oil spill detection method based on BiLSTM and SOM according to claim 1, characterized in that, The steps in S2 include: extracting each pixel. Three-channel texture features of the neighborhood window , , The three-channel texture features are then integrated into a multimodal feature vector. Based on the multimodal feature vector The K-means clustering algorithm is used to cluster the preprocessed images, minimizing the global compactness objective function during the clustering process. objective function The expression is: ； In the formula, K is the number of clusters; It is the set of all multimodal feature vectors belonging to the k-th class; Let be the cluster center vector of the k-th class; This represents the L2 norm.
4. The marine radar oil spill detection method based on BiLSTM and SOM according to claim 3, characterized in that, The three-channel texture features described in S2 , , The calculation formula is as follows: ； ； ； In the formula, This represents the mean of the neighborhood window for each pixel (x, y). This represents the local variance of the neighborhood window for each pixel (x, y). θ represents the global variance; θ represents the mean adjustment coefficient. γ represents the median value in the neighborhood window for each pixel (x, y); β represents the median adjustment coefficient; sign(.) is the signification function; M(x,y) represents the gradient magnitude for each pixel (x, y); γ represents the gradient magnitude adjustment coefficient. This represents the global mean. This represents the intensity value of pixel (x, y).
5. A marine radar oil spill detection system based on BiLSTM and SOM, characterized in that, The system is used to implement the marine radar oil spill detection method based on BiLSTM and SOM as described in any one of claims 1-4; the system includes: The data preprocessing module is configured to preprocess the acquired raw radar image data. The pseudo-label generation module is configured to extract multimodal texture features from the preprocessed image and generate pseudo-labels for the image based on the multimodal texture features using the K-means clustering algorithm. The feature extraction module is configured to use the BiLSTM algorithm to extract temporal features of the preprocessed image; The pixel classification module is configured to input the temporal features into a pre-trained self-organizing map network, find a winning neuron for each feature vector, and assign the pseudo-label of the pixel corresponding to each feature vector to the corresponding winning neuron, thereby obtaining the classification result of each pixel in the preprocessed image. The classification optimization module is configured to iteratively update the category of each pixel based on the categories of other pixels within the neighborhood window of each pixel in the classification result of each pixel in the image, so as to determine the final classification result of each pixel; The oil spill detection module is configured to perform morphological processing on the image with the final classification result of each pixel determined to obtain an oil spill image.