Sea surface oil spill detection method based on decision fusion of multi-core classification models

Abstract

Embodiments of the present disclosure disclose a sea surface oil spill detection method based on decision fusion of multi-core classification models. The method includes: obtaining a multi-spectral remote sensing image of a target sea area as an image to be detected; extracting features from the image to be detected to obtain feature information; constructing input data based on the feature information; inputting the input data into classification models of multiple single-core functions respectively to obtain multiple classification results; and obtaining a decision fusion result based on the multiple classification results to distinguish the oil spill area in the image to be detected. The sea surface oil spill detection method provided by the present invention inputs the input data into classification models of multiple single-core functions respectively to obtain multiple classification results, and then performs decision fusion on the multiple classification results to obtain a decision fusion result. The multi-core decision fusion algorithm can complement the advantages of each single core, improve the classification accuracy, effectively detect oil spills under a complex water color background, and further improve the accuracy of sea surface oil spill detection.

Description

Technical Field

[0001] The present invention relates to the field of marine exploration, and particularly to a method for detecting oil spills on the sea surface based on decision fusion of a multi-core classification model. Background Art

[0002] With the development of the marine transportation industry, oil spill accidents occur frequently. The spread of oil products in the ocean will cause losses and hazards to aspects such as the ecological environment, economic construction, marine industry, and the lives of coastal residents. Most oil spill accidents caused by either natural factors or human reasons are sudden and dynamic, which makes it more difficult for relevant departments to analyze the nature of the incident and deploy emergency measures. Remote sensing technology can play an important role in this regard.

[0003] Support Vector Machine (SVM) has good computational performance in remote sensing image research and can detect targets with high precision using only a small number of samples. The selection of the SVM kernel function and other parameters will affect the detection effect, while the traditional single-core SVM needs to set the kernel function according to the characteristics of the research target, which has limitations. Summary of the Invention

[0004] In view of this, the present invention provides a method for detecting oil spills on the sea surface based on decision fusion of a multi-core classification model to solve the problem in the prior art that the single-core classification model needs to set the kernel function according to the characteristics of the research target and has limitations.

[0005] The technical solution of a method for detecting oil spills on the sea surface based on decision fusion of a multi-core classification model provided by the present invention is as follows:

[0006] A method for detecting oil spills on the sea surface based on decision fusion of a multi-core classification model, the method comprising:

[0007] Obtain a multi-spectral remote sensing image of the target sea area as the image to be detected;

[0008] Extract feature information from the image to be detected;

[0009] Construct input data based on the feature information;

[0010] Input the input data into classification models of multiple single-core functions respectively to obtain multiple classification results;

[0011] Obtain a decision fusion result based on the multiple classification results to distinguish the oil spill area in the image to be detected.

[0012] In some embodiments, the feature information includes spectral features and texture features.

[0013] In some embodiments, the method for extracting the spectral features includes:

[0014] Select a plurality of initial data of preset bands in the image to be detected;

[0015] Preprocess the plurality of initial data of preset bands to obtain a plurality of processed data of preset bands;

[0016] Based on the optimal index method, select a predetermined number of band combinations from the plurality of preset band data;

[0017] Determine the plurality of preset band data and the predetermined number of band combinations as the spectral features.

[0018] In some embodiments, the method for extracting the texture features includes:

[0019] Perform principal component analysis on the image to be detected to extract the first principal component;

[0020] Determine the texture features based on the first principal component.

[0021] In some embodiments, based on the gray-level co-occurrence matrix, obtain a plurality of features of the first principal component in multiple directions, and take the average value of the multiple directions of each feature to obtain the texture features.

[0022] In some embodiments, calculate the gray-level co-occurrence matrix by using windows of multiple different sizes respectively to extract the texture features.

[0023] In some embodiments, the method for constructing the input data based on the feature information includes:

[0024] Calculate the oil-water Bhattacharyya distance of each feature in the feature information;

[0025] Select the features with larger oil-water Bhattacharyya distance as the best oil spill spatial spectrum feature dataset;

[0026] Determine the best oil spill spatial spectrum feature dataset as the input data.

[0027] In some embodiments, the method for obtaining the decision fusion result based on the plurality of classification results includes:

[0028] Based on the fuzzy membership degree, determine the category to which each pixel belongs after fusion pixel by pixel.

[0029] In some embodiments, for the pixels at the same position in the plurality of classification results, the method for determining the category to which the pixel belongs after fusion includes:

[0030] If the categories to which the pixels belong in the plurality of classification results are the same, assign this category to the decision fusion result of the pixel;

[0031] If the classification results of multiple categories are different for the cell's belonging category, obtain the category with the largest number in the predetermined window around the cell for the multiple classification results. If the belonging category of a classification result for the cell is the same as the category with the largest number in the predetermined window for the other classification results, assign this category to the decision fusion result of the cell. Otherwise, calculate the membership degrees of the cell belonging to each category in each classification result, and use the category with the highest membership degree as the decision fusion result of the cell.

[0032] In some embodiments, the classification models of the multiple single-kernel functions include an SVM classification model with an RBF kernel function, an SVM classification model with a Linear kernel function, an SVM classification model with a Polynomial kernel function, and an SVM classification model with a Sigmoid kernel function.

[0033] The method for detecting sea surface oil spills based on decision fusion of a multi-kernel classification model provided by the present invention inputs input data into classification models of multiple single-kernel functions respectively to obtain multiple classification results, and then performs decision fusion on the multiple classification results to obtain a decision fusion result. The multi-kernel decision fusion algorithm can complement the advantages of each single kernel, improve the classification accuracy, effectively detect oil spills under a complex water color background, and further improve the accuracy of sea surface oil spill detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings are used to better understand the present invention and do not unduly limit the present invention. Among them:

[0035] Figure 1 is a flowchart of the method for detecting sea surface oil spills based on decision fusion of a multi-kernel classification model provided by a specific embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of the principle of obtaining a decision fusion result based on multiple classification results provided by a specific embodiment of the present invention;

[0037] Figure 3 is a true color composite image provided by a specific embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of the principle of the method for detecting sea surface oil spills based on decision fusion of a multi-kernel classification model provided by a specific embodiment of the present invention;

[0039] Figures 5a to 5e Successively show the classification result diagrams of the oil spill detection of the SVM classification model with an RBF kernel function, the SVM classification model with a Linear kernel function, the SVM classification model with a Polynomial kernel function, the SVM classification model with a Sigmoid kernel function, and the multi-kernel classification model;

[0040] Figure 6 Comparison diagram of multi-kernel decision fusion and the highest / lowest single-kernel accuracy. Detailed implementation manners

[0041] The present invention will be further described below in conjunction with embodiments.

[0042] Aiming at the problem that the single-core classification model in the prior art needs to set the kernel function according to the characteristics of the research target and has limitations, this embodiment provides a sea surface oil spill detection method based on decision fusion of a multi-core classification model, as Figure 1 shown, which includes the following steps:

[0043] S100. Obtain a multi-spectral remote sensing image of the target sea area as the image to be detected.

[0044] As an example, a multi-spectral remote sensing image of the target sea area collected by the Sentinel-2 high-resolution multi-spectral imaging satellite can be used as the image to be detected. There are currently two satellites, Sentinel-2A and Sentinel-2B, covering 13 spectral bands, and the ground resolution can reach up to 10m at most.

[0045] S200. Extract features from the image to be detected to obtain feature information.

[0046] As an example, the feature information is spectral information. In a preferred embodiment, the feature information includes spectral features and texture features, so as to consider the target features more comprehensively, further improve the classification accuracy, and effectively detect oil spills under a complex water color background.

[0047] In one embodiment, the method for extracting spectral features includes:

[0048] S210. Select multiple preset band initial data in the image to be detected.

[0049] As an example, six bands, namely blue, green, red, near-infrared (broad), short-wave infrared 2, and short-wave infrared 3 bands, are selected as the preset band initial data.

[0050] S220. Preprocess the multiple preset band initial data to obtain the processed multiple preset band data.

[0051] As an example, the preprocessing includes atmospheric correction, resampling, reflectance normalization, etc.

[0052] S230. Select a predetermined number of band combinations from the multiple preset band data based on the optimal index method.

[0053] As an example, the following formula is used to calculate the OIF (Optimum Index Factor) index of each band combination:

[0054]

[0055] Among them, S k represents the standard deviation of the k-th band;

[0056] r k,l represents the correlation coefficient between the k-th band and the l-th band.

[0057] The larger the standard deviation and the smaller the correlation coefficient, the larger the OIF index, and the more information the band combination contains. In this embodiment, the top 4 groups of bands with the highest OIF index are combined to obtain spectral feature data with stronger oil-water spectral differences.

[0058] S240. Determine multiple preset band data and the predetermined number of band combinations as spectral features.

[0059] In this embodiment, combining a single spectral band with a multi-band combination to obtain spectral features can obtain more oil-water information.

[0060] In one embodiment, the method for extracting texture features includes:

[0061] S201. Perform principal component analysis on the image to be detected, and extract the first principal component;

[0062] S202. Determine the texture features based on the first principal component.

[0063] Extracting the first principal component through principal component analysis to obtain texture features reduces the data dimension and noise interference.

[0064] Further, step S202 is specifically: based on the Gray-level Co-occurrence Matrix (GLCM), obtain multiple features of the first principal component in multiple directions, and take the average of the multiple directions of each feature to obtain the texture features.

[0065] As an example, based on the Gray-level Co-occurrence Matrix (GLCM), obtain 8 features (mean, variance, co-occurrence, contrast, dissimilarity, entropy, angular second moment, and correlation) of the first principal component in 8 directions, and take the average of the 8 directions of each feature.

[0066] Since the window size determines the authenticity of the texture and the calculation speed, based on this, calculate the Gray-level Co-occurrence Matrix using multiple different window sizes. As an example, two sets of calculations with window sizes of 3×3 and 5×5 are performed, and a total of 16 texture features are obtained.

[0067] In this embodiment, the obtained spectral feature data is combined with spatial texture data to construct a more effective feature dataset for oil spill detection.

[0068] S300. Construct input data based on feature information.

[0069] However, it is not the case that the more features, the more beneficial it is for model detection. Feature redundancy not only causes excessive computational load and reduces computational efficiency, but also easily leads to model overfitting and affects detection accuracy.

[0070] Based on this, preferably, step S300 further includes:

[0071] S310. Calculate the oil-water Bhattacharyya distance of each feature in the feature information.

[0072] The Bhattacharyya Distance (BD) is used to evaluate the separability between classes in classification, and its formula is as follows:

[0073]

[0074] where μ i represents the mean of the i-th class of samples;

[0075] σ i represents the standard deviation of the i-th class of samples.

[0076] S320. Select the features with larger oil-water Bhattacharyya distances as the best oil spill spatial-spectral feature dataset;

[0077] S330. Determine the best oil spill spatial-spectral feature dataset as the input data.

[0078] The number of features selected in step S320 is not limited. In order to determine the number of features in the best spatial-spectral feature dataset, this embodiment conducts oil spill detection comparison experiments with different feature amounts. The top 5, 10, 15, 20, and 25 feature amounts are selected in descending order of BD for superposition. 10% of the total samples are selected for training, and 90% are used for testing. SVM with RBF as the kernel function is used as the classification model. Through accuracy comparison, it is found that the detection result is the best when the feature amount is 20. Therefore, this embodiment finally selects the top 20 feature amounts arranged in descending order of BD as the best oil spill spatial-spectral feature dataset.

[0079] In this embodiment, SVM with RBF as the kernel function is used as the classification model. The single-band spectral features, spectral index features, and texture features are respectively superimposed and combined with each other. A total of 7 feature datasets covering different types and numbers of features are obtained for oil spill detection respectively. The obtained accuracy results are shown in the following table. The experimental results show that the OA, Kappa coefficient, and F1 score of oil spill detection using the best spatial-spectral feature dataset established by the present invention are higher than those of other feature datasets, demonstrating higher accuracy and consistency. This also verifies that it is not the case that the more features, the better the classification effect.

[0080]

[0081] S400. Input the input data into the classification models of multiple single-core functions respectively to obtain multiple classification results;

[0082] S500. Based on the multiple classification results, obtain a decision fusion result to distinguish the oil spill area in the image to be detected.

[0083] As an example, such as Figure 2 shown, the classification models of multiple single-core functions include the SVM classification model with RBF kernel function, the SVM classification model with Linear kernel function, the SVM classification model with Polynomial kernel function, and the SVM classification model with Sigmoid kernel function. The expressions of each kernel function are shown in the following table:

[0084]

[0085] Based on the fuzzy membership degree, determine the category to which each pixel belongs after fusion pixel by pixel, that is, calculate the membership degree of each pixel, determine the category of each pixel according to the membership degree of each pixel, and finally merge the categories to which each pixel belongs after determination to obtain the final decision fusion result map.

[0086] Before performing the sea surface oil spill detection method, first train the classification models of each single-core function. For example, based on the L1C-level product taken by Sentinel-2 in the study area on May 18, 2021, perform preprocessing such as atmospheric correction, resampling, and reflectance normalization on the 6 band data suitable for oil spill detection. The blue, green, red, near-infrared (broad), short-wave infrared 2, and short-wave infrared 3 bands of Sentinel-2 are conducive to distinguishing oil-water characteristics, so these 6 bands are selected for spectral analysis. Figure 3 It is a true-color composite image, and the red, green, and blue channels correspond to 664.5nm, 560nm, and 496.6nm respectively. A total of 16,325 pixels are selected for training data, and 152,837 pixels are selected for test data, with a ratio of approximately 1:9, including two categories of oil film and sea water.

[0087] For the pixels at the same position in the multiple classification results, the methods for determining the category to which the pixel belongs after fusion include:

[0088] If the categories to which the multiple classification results belong to the pixel are the same, then assign this category to the decision fusion result of the pixel;

[0089] If the categories to which multiple classification results belong to a pixel are different, obtain the category with the largest number in a predetermined window (such as a 3×3 window around the pixel) around the pixel. If the category to which a classification result belongs to the pixel is the same as the category with the largest number in the predetermined window of other classification results, assign this category to the decision fusion result of the pixel. Otherwise (that is, if the above situation is not met), calculate the membership degree of the pixel belonging to each category in each classification result, and use the category with a high membership degree as the decision fusion result of the pixel.

[0090] As an example, the calculation formula for the membership degree is as follows:

[0091]

[0092] Among them, P i is the membership degree of the i-th category;

[0093] W j is the weight of the target feature;

[0094] is the membership degree of the m-th pixel in the single-core classification map belonging to the n-th category.

[0095] Figure 4 Illustrates the schematic diagram of the sea surface oil spill detection method based on decision fusion of multi-core classification models. As Figure 4 shown, first, construct a spectral feature index based on the Optimum Index Factor (OIF) to extract the spectral features of sea surface oil spills, extract the texture features of sea surface oil spills based on the Gray Level Co-occurrence Matrix (GLCM), use the Bhattacharyya distance method for feature selection, construct a sea surface oil spill spatial-spectral feature dataset, use a single-core SVM as the classification model to detect oil spills in the study area, and finally, based on the decision fusion algorithm, realize the multi-core fusion of the sea surface oil spill detection results. The experimental results show that the sea surface oil spill detection method based on decision fusion of multi-core classification models proposed in this embodiment can consider target features more comprehensively, improve the classification accuracy, and effectively detect oil spills under complex water color backgrounds. After comparative analysis, the OA, Kappa coefficient, and F 1 score of multi-core decision fusion are all higher than those of the single-core model. OA has increased by 1.0 - 8.2%, the Kappa coefficient has increased by 0.013 - 0.122, and the F 1 score has increased by 0.009 - 0.097. The multi-core decision fusion algorithm complements the advantages of each single-core, further improving the accuracy of sea surface oil spill detection.

[0096] Next, the performance of the sea surface oil spill detection method based on decision fusion of multi-core classification models provided by the present invention is verified.

[0097] Apply the determined optimal spatial-spectral feature dataset to different single-core SVM models, and introduce multi-core SVM decision fusion for oil spill detection.Figures 5a to 5e It is a classification result diagram of single-core and multi-core SVM oil spill detection. Figure 6 It is a comparison diagram of multi-core decision fusion and the highest / lowest precision of single-core, and the precision evaluation is shown in the following table:

[0098]

[0099] As Figures 5a to 5e shown, the classification results obtained by different kernel functions are all inconsistent. Generally speaking, the RBF, Linear and Polynomial kernel functions are not very different, and the differences are mainly in the upper left and right edges of the study area. There are fewer fragmented patches in Polynomial, but it does not handle well the area with complex oil-water distribution in the upper left. The Sigmoid kernel function has more classification errors. Preliminary analysis shows that due to factors such as lack of detailed parameter tuning, the existence of mixed pixels in the training set, and overfitting, the Sigmoid kernel function fails to correctly distinguish some easily confused pixels in certain texture features. The image after multi-core SVM decision fusion does not leave the problem of misclassification in Sigmoid, overcomes the noise problem of RBF, and retains the classification details of the area with more oil spills in the center of the study area. From the above table and Figure 6 the precision data of single-core and multi-core decision fusion, it can be seen that in single-core SVM oil spill detection, RBF, Linear and Polynomial all have good classification performance. Among them, the OA of RBF is the highest, reaching 89.25%, while the F1 score of Polynomial is the highest, reaching 0.890; the evaluation indexes of Sigmoid are all the lowest, which is consistent with the effect of the classification diagram; the OA, Kappa coefficient and F1 score of multi-core decision fusion are all higher than those of the single-core model. The OA has increased by 1.0 - 8.2%, the Kappa coefficient has increased by 0.013 - 0.122, and the F1 score has increased by 0.009 - 0.097. The multi-core decision fusion algorithm proposed by the present invention complements the advantages of each single-core, and further improves the accuracy of sea surface oil spill detection.

[0100] Those skilled in the art should understand that the embodiments of the present disclosure can be provided as methods, systems, or computer program products. Therefore, the present disclosure can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Moreover, the present disclosure can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program code.

[0101] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce a means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 and / or one or more of the blocks.

[0102] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction means that implements the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 and / or one or more of the blocks.

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 and / or one or more of the blocks.

[0104] Those of ordinary skill in the art can understand that all or part of the processes in the above-described method embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-described method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0105] Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A method for detecting sea surface oil spills based on decision fusion of multi-core classification models, characterized in that, the method includes: Obtain multi-spectral remote sensing images of the target sea area as the images to be detected; Extract feature information from the images to be detected; Construct input data based on the feature information; Input the input data into the classification models of multiple single-core functions respectively to obtain multiple classification results; Based on the multiple classification results, obtain a decision fusion result to distinguish the oil spill area in the images to be detected; For the pixels at the same position in the multiple classification results, the method for determining the category to which the pixel belongs after fusion includes: If the categories to which the multiple classification results belong to the pixel are the same, then assign this category to the decision fusion result of the pixel; If the categories to which the multiple classification results belong to the pixel are different, then obtain the category with the largest number in the predetermined window around the pixel in the multiple classification results. If the category to which a classification result belongs to the pixel is the same as the category with the largest number in the predetermined window of the other classification results, then assign this category to the decision fusion result of the pixel. Otherwise, calculate the membership degree of the pixel belonging to each category in each classification result, and use the category with the highest membership degree as the decision fusion result of the pixel.

2. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to claim 1, characterized in that, the feature information includes spectral features and texture features.

3. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to claim 2, characterized in that, the method for extracting the spectral features includes: Select multiple initial data of preset bands in the images to be detected; Preprocess the multiple initial data of preset bands to obtain multiple processed preset band data; Based on the optimal index method, select a predetermined number of band combinations from the multiple preset band data; Determine the multiple preset band data and the predetermined number of band combinations as the spectral features.

4. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to claim 2, characterized in that, the method for extracting the texture features includes: Perform principal component analysis on the images to be detected and extract the first principal component; Determine the texture features based on the first principal component.

5. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to claim 4, characterized in that, Based on the gray-level co-occurrence matrix, obtain multiple features of the first principal component in multiple directions, and take the average value of each feature in multiple directions to obtain the texture features.

6. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to claim 4, characterized in that, Calculate the gray-level co-occurrence matrix using windows of multiple different sizes respectively to extract the texture features.

7. The method for detecting sea surface oil spills based on decision fusion of multi-core classification models according to any one of claims 1 to 6, characterized in that, the method for constructing input data based on the feature information includes: Calculate the oil-water Bhattacharyya distance of each feature in the feature information; The features with larger oil-water Baccarat distance were selected as the best oil spill spatial spectrum feature dataset; The optimal oil spill spatial spectrum feature data set is determined as the input data.

8. The method for detecting oil spill on sea surface based on decision fusion of multi-core classification model according to any one of claims 1 to 6, It is characterized in that Based on the multiple classification results, a method for obtaining a decision fusion result includes: Based on the fuzzy membership, the category of each pixel after fusion is determined pixel by pixel.

9. The method for detecting oil spill on sea surface based on decision fusion of multi-core classification model according to any one of claims 1 to 6, It is characterized in that The classification models of the multiple single kernel functions include an SVM classification model of an RBF kernel function, an SVM classification model of a Linear kernel function, an SVM classification model of a Polynomial kernel function, and an SVM classification model of a Sigmoid kernel function.

Images