A method for fast and intelligent identification and classification of offshore oil spill risk sources from high-resolution remote sensing images

Abstract

A kind of offshore oil spill risk source fast intelligent identification classification method of high-resolution remote sensing image belongs to the cross technical field of marine remote sensing and artificial intelligence, including: 1) oil platform and ship as the main risk source of offshore oil spill, with the characteristics of pixel value higher than surrounding seawater, according to the method of pixel value gradient and spot detection, a squeezing excitation residual network combining Gaussian Laplacian operator and connected domain controller is designed;2) using prior knowledge such as pixel gradient and target size, suspected offshore oil spill risk source identification is carried out, the center position and size of the target are obtained using connected domain controller on multiple response points around the target, 3) finally, the suspected offshore oil spill risk source identified is classified by deep learning classification network.The present application only classifies the suspected offshore oil spill risk source target picture in the image, reduces the number of pictures, improves the running time, and the training set and the actual prediction picture size are consistent, only small-scale pictures are needed for training without using the whole remote sensing image, which is easier to train and saves time.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of marine remote sensing and artificial intelligence, and provides a rapid and intelligent identification and classification method for marine oil spill risk sources from high-resolution remote sensing images. Background Technology

[0002] The construction and exploitation of oil platforms and maritime shipping are important components of human activities at sea and are prone to oil spills, representing a major source of risk in marine oil spill incidents. Timely and accurate acquisition of the locations of oil platforms and vessels not only provides data support for the rapid identification and classification of oil spill risk sources but also contributes to the formulation of marine development strategies and the strengthening of marine infrastructure. Remote sensing imagery offers large coverage areas and is easily accessible, making it faster and more efficient for identifying and classifying marine oil spill risk sources. In recent years, with the continuous development of remote sensing satellite technology, more and more high-resolution satellites have been successfully launched, and optical imagery possesses multi-band characteristics and low noise interference, giving it a significant advantage in research on the rapid identification and classification of marine oil spill risk sources using optical satellite imagery as a data source.

[0003] Existing methods for detecting oil platforms in optical imagery commonly rely on multi-temporal image comparison to determine their location based on the platform's positional and size invariance. However, this method requires massive amounts of data, extensive preprocessing, and is time-consuming. Different imaging conditions can also affect the extraction results, and it cannot promptly obtain the location information of newly built oil platforms. Target detection algorithms for marine environments, based on target detection frameworks, consume significant hardware resources when processing large-scale remote sensing images. They often require sliding window prediction of the entire image, wasting considerable time in targetless marine environments. This can lead to misclassifying numerous marine disturbances as vessels, reducing accuracy and extending computation time. To address these issues, this paper presents a rapid and intelligent identification and classification method for marine oil spill risk sources using high-resolution remote sensing imagery. First, suspected targets of marine oil spill risk sources are extracted. Then, a rapid identification and classification method for marine oil spill risk sources is used to classify these suspected targets, quickly identifying oil platforms and vessels within the marine oil spill risk sources. This method achieves rapid identification and classification of marine oil spill risk sources using a single image, reducing unnecessary calculations on large marine areas in remote sensing images, and realizing the goal of rapid and high-precision identification and classification of marine oil spill risk sources. Summary of the Invention

[0004] This invention primarily addresses the challenges of effectively identifying and classifying oil spill risk sources using deep learning, particularly the excessive computation time required for processing large areas of invalid regions. It proposes an intelligent identification and classification method for marine oil spill risk sources using high-resolution remote sensing imagery. First, an oil spill risk source sample dataset is created for training and testing. In the target detection stage, image preprocessing is performed, including linear stretching to improve image quality, downsampling to reduce image size and speed up processing, and threshold segmentation and masking to remove large areas of cloud and land interference. Subsequently, a Laplacian of Gaussian operator connected domain controller squeeze excitation residual network (LCSE-ResNet) is used to extract suspected targets and screen oil platforms and vessels.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A rapid and intelligent identification and classification method for marine oil spill risk sources using high-resolution remote sensing imagery includes the following steps:

[0007] The first step is to preprocess the obtained remote sensing images to be classified, removing interference from clouds and land, which will facilitate subsequent steps in classifying oil platforms and ships in the processed images.

[0008] The preprocessing includes linear stretching, downsampling, threshold segmentation, and mask processing in sequence.

[0009] The masking process specifically involves the following steps: For the detection of offshore oil platforms and vessels, since the brightness values ​​of remote sensing images are generally between 0 and 65535, a large bit depth leads to excessively long computation times, and a large number of pixel values ​​in the image are often distributed in a small brightness area. Therefore, a linear stretching method is used to stretch the image to improve contrast and distribute the brightness values ​​between 0 and 255 to speed up the processing time. Secondly, remote sensing images have very large sizes, making it necessary to use downsampling to compress the image size and save resources. Furthermore, clouds and land in the image are significant interference factors. Therefore, a maximum inter-class variance thresholding method is used to segment brighter clouds and land from the darker background, and then masking is used to remove cloud and land interference from the downsampled image.

[0010] The second step is the response of suspected targets from marine oil spill risk sources. This is implemented using the Laplacian of Gaussian (LoG) operator in LCSE-ResNet. LoG uses Gaussian filtering to reduce image noise and then uses the second-order Laplacian gradient operator to obtain locations in the image where brightness changes are significant. Based on the distinct brightness characteristics of ships and oil platforms in the marine environment, gradient change response points are obtained around suspected targets, thus initially identifying the suspected targets. The specific steps are as follows:

[0011] 2.1) Since the grayscale values ​​of oil platforms and ship targets are significantly higher than those of the surrounding marine environment, differential operations are performed on the remote sensing images after the first step of preprocessing to obtain pixels with drastic gradient changes. Since differential operations are sensitive to noise in remote sensing images, and the remote sensing images are blurred while calculating the gradient, a large number of response points can be obtained around suspected oil platforms and ship targets through the above methods.

[0012] 2.2) After step 2.1, there are a large number of response points around each suspected target. The value of the response point reflects the gradient change at that location. Due to the different sizes of the targets at sea and the changes in the marine environment, response points with response values ​​between 5 and 25 are retained as suspected targets. These response points are initially identified as suspected targets.

[0013] The third step is to locate the suspected target.

[0014] Because actual oil platforms and ships are large targets, multiple response points will appear around them. Directly using these response points would lead to repeated calculations of the target, increasing the computational load and making it impossible to accurately locate the target and determine its size. This invention uses a connected domain controller to aggregate and calculate multiple response points around a suspected target, obtaining the center position and size of the suspected target, and then trimming it to obtain a more accurately located suspected target. Specifically:

[0015] 3.1) Calculate the distance between any two points among all response points around each suspected target, and consider all points within a certain region that are within 30 pixels apart and have more than 50 response points as the same connected component, thus obtaining M connected components.

[0016] 3.2) Calculate the average x and y coordinates of all response points in each connected component and the distance between any two response points. Take the maximum distance and use the point with the average x and y coordinates and the maximum distance as the center point and size of the suspected target pointed to by the connected component.

[0017] 3.3) Finally, based on the center point position and size of the suspected target, the suspected target is cropped from the original remote sensing image before preprocessing, resulting in M ​​cropped images of the suspected target. The sizes of the multiple suspected target images can be the same or different, depending on the actual situation. The suspected target image is square, and the center point position obtained in a certain connected component is the center point of the suspected target image, with the size being the maximum distance between any two response points.

[0018] The fourth step is to classify the sources of oil spill risk.

[0019] An oil spill risk source classification network was used to classify the M suspected target images after the third step of cropping, retaining the main risk sources of offshore oil spills: oil platforms and vessels. The specific steps are as follows:

[0020] 4.1) Since the M suspected target images obtained in step 3.3) are not the same size, the size of all images is adjusted to obtain M suspected target images of the same size.

[0021] 4.2) Obtain the trained oil spill risk source classification network

[0022] 4.2.1) First, the network structure of the oil spill risk source classification network will be explained.

[0023] LCSE-ResNet comprises the target localization part (containing the LoG operator (second step) and the connected component controller (third step)) and the oil spill risk source classification network part, as described above. The oil spill risk source classification network contains 50 convolutional layers, with each pair of convolutional layers forming a basic network block. After two convolutional operations, a squeeze excitation operation is used to redistribute the weights of different channels in the output. Then, a shorting method is used to add the input before the two convolutional operations to the output after weight redistribution, which serves as the input to the next network block.

[0024] 4.2.2) By cropping the panchromatic images of the Gaofen-2 optical satellite that have identified ships, oil platforms, clouds, other man-made targets, sea ripples and other interference information, a dataset consisting of multiple images is obtained. Each image is a single-channel grayscale image with a spatial resolution of 1 meter and an image size of 201×201. In addition to the sea surface, the images contain only one of the following: ships, oil platforms, clouds, other man-made targets, ripples or noise interference.

[0025] 4.2.3) Divide the dataset obtained in step 4.2.2) into a training set and a test set. Use the training set to train the oil spill risk source classification network from step 4.1), and use the test set to validate it. Specifically: A classification network model is constructed by combining a squeeze-excitation network module and a residual network. The cross-entropy loss function is used to compare the model's output with the true label, measuring the difference between them. The network model is trained using the training dataset, and backpropagation and gradient descent are used to update the model's weights and biases to minimize the loss function. During training, the classification model learns the mapping relationship from the input image to the output category. Through backpropagation and optimization algorithms, the model adjusts its weights and biases so that, given an input image, it can produce an output that matches the true label. In other words, the classification model learns the mapping relationship from the input image to the output category during training. Through backpropagation and optimization algorithms, the model adjusts its weights and biases so that, given an input image, it can produce an output that matches the true label.

[0026] 4.3) The trained oil spill risk source classification network is used to classify the suspected targets from step 4.3). This network outputs a label corresponding to the input image based on the correspondence between the training set images and labels, thus determining the actual category of the suspected target. Targets classified as ships and oil platforms are retained from the M suspected targets. In other words, the image is fed into the trained classification network to obtain the category corresponding to that image.

[0027] Marine targets exhibit a high degree of similarity, and in cropped images, marine oil spill risk source targets almost occupy the center of the image. Each image consists almost entirely of targets with concentrated brightness and similar size, along with their surrounding marine environment. The oil spill risk source classification network (also a deep learning classification model) can combine extracted channel information and spatial information to allow the model to learn the relationships between different channels, enabling rapid identification and classification of marine oil spill risk source targets. In the shallow layers of the oil spill risk source classification network, it exhibits class-independent characteristics between different classes, improving the quality of low-level representations. As the number of layers increases, class-related features emerge, enhancing the more important features between each channel and making the internal and contour features of ships and oil platforms more distinct from other marine disturbance factors.

[0028] The beneficial effects of this invention are as follows:

[0029] This invention provides a rapid and intelligent identification and classification method for marine oil spill risk sources based on high-resolution optical satellite marine remote sensing imagery. The algorithm utilizes prior knowledge such as pixel gradient and target size to detect suspected targets, uses a connected component controller to obtain the target's center position and size from multiple response points around the target, and finally classifies and filters the targets.

[0030] This invention only classifies suspected target images within the imagery, significantly reducing the number of images and improving runtime. Furthermore, the training set is approximately the same size as the actual predicted images, requiring only a small number of images for training instead of the entire remote sensing image, making the algorithm easier to train and more time-efficient. Attached Figure Description

[0031] Figure 1 A block diagram of a rapid and intelligent identification and classification method for marine oil spill risk sources based on high-resolution remote sensing images; Figure 1 In the diagram, (a) represents linear stretching; (b) represents downsampling; (c) and (d) represent masking; and (e) represents the oil spill risk source classification network.

[0032] Figure 2 The image shows the results of rapid identification and classification of marine oil spill risk sources from images. Circular boxes represent oil platforms, and square boxes represent ships. Figure 2 (a) shows the identification and classification results of oil platforms and vessels when there is a lot of fog interference at sea; Figure 2 (b) shows the identification and classification results of oil platforms and vessels when there is a lot of cloud interference; Figure 2 (c) and (d) represent the identification and classification results of oil platforms and vessels when there is a lot of sea ripple interference; Figure 2 (e) and (f) represent the identification and classification results of oil platforms and vessels when the sea conditions are good and the distribution of oil platforms and vessels is dense. Detailed Implementation

[0033] To make the problem solved by the present invention, the method adopted, and the effect achieved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings, not all of them.

[0034] like Figure 1 As shown in the embodiment of the present invention, a rapid and intelligent identification and classification method for marine oil spill risk sources based on high-resolution remote sensing imagery includes:

[0035] The first step is remote sensing image preprocessing.

[0036] 1.1) Linear Stretching. Most remote sensing images have a bit depth of 16 bits, with pixel values ​​ranging from 0 to 65535. However, the actual pixel values ​​of each image are often concentrated in a very small range, leading to reduced contrast. Furthermore, since the pixel ranges of different remote sensing images are often inconsistent, this can affect the algorithm results. Additionally, a 16-bit pixel depth results in excessively large data volumes, increasing computational complexity. Therefore, using linear stretching for image enhancement can improve the contrast and visibility of remote sensing images, while simultaneously mapping the pixel range to 0-255, reducing the data size.

[0037] 1.2) Downsampling. Due to the large size of the original remote sensing image data, the computation speed is slow, resulting in reduced real-time performance. Therefore, the image is processed by sampling at a width and height interval of 5 pixels to reduce the image data ratio and improve computation speed.

[0038] 1.3) Thresholding Segmentation and Masking. For the detection of offshore oil platforms and vessels, clouds and land in the imagery are significant interfering factors. Therefore, thresholding segmentation and masking are used to remove cloud and land interference. A thresholding method based on maximum inter-class variance is used to segment brighter clouds and land from a darker background. Then, morphological processing is used to refine the thresholding results, including dilation, erosion, and filling. The morphologically processed binary image is inverted and used as a mask to mask the downsampling results.

[0039] For example, in the GF2_PMS1_E124.9_N29.0_20191007_L1A00042 image, the pixel range is concentrated below 30,000 luminance values, making it difficult to effectively distinguish different objects. Linear stretching improves the contrast, which is beneficial for subsequent localization of suspected targets. The image size is 29200×27632, and after downsampling, the size is 5840×5526, which greatly reduces the data size and improves the processing speed.

[0040] The second step is the response of suspected targets from marine oil spill risk sources. This is implemented using the Laplacian of Gaussian (LoG) operator in LCSE-ResNet. LoG uses Gaussian filtering to reduce image noise and then uses the second-order Laplacian gradient operator to obtain locations in the image where brightness changes are significant. Based on the distinct brightness characteristics of ships and oil platforms in the marine environment, gradient change response points are obtained around suspected targets, thus initially identifying the suspected targets. The specific steps are as follows:

[0041] 2.1) Since the grayscale values ​​of oil platforms and ship targets are significantly higher than those of the surrounding marine environment, differential operations are performed on the remote sensing images after the first step of preprocessing to obtain pixels with drastic gradient changes. Since differential operations are sensitive to noise in remote sensing images, and the remote sensing images are blurred while calculating the gradient, a large number of response points can be obtained around suspected oil platforms and ship targets through the above methods.

[0042] 2.2) After step 2.1, there are a large number of response points around each suspected target. The value of the response point reflects the gradient change at that location. Due to the different sizes of the targets at sea and the changes in the marine environment, response points with response values ​​between 5 and 25 are retained as suspected targets. These response points are initially identified as suspected targets.

[0043] This step yields a large number of response points around targets such as ships, oil platforms, and clouds. For example, thousands of response points were obtained for the GF2_PMS1_E124.9_N29.0_20191007_L1A00042 image.

[0044] The third step is to locate the suspected target.

[0045] Because there are too many response points, directly using these response points would lead to repeated calculations of the target, increasing the computational load and making it impossible to accurately locate the target and determine its size. This invention uses a connected component controller to aggregate and calculate multiple response points around a suspected target, obtaining the center position and size of the suspected target, and then cropping it to obtain a more accurately located suspected target. Specifically:

[0046] 3.1) Calculate the distance between any two points among all response points around each suspected target, and consider all points within a certain region that are within 30 pixels apart and have more than 50 response points as the same connected component, thus obtaining M connected components.

[0047] 3.2) Calculate the average x and y coordinates of all response points in each connected component and the distance between any two response points. Take the maximum distance and use the point with the average x and y coordinates and the maximum distance as the center point and size of the suspected target pointed to by the connected component.

[0048] 3.3) Finally, based on the center point position and size of the suspected target, the suspected target is cropped from the original remote sensing image before preprocessing, resulting in M ​​cropped images of the suspected target. The sizes of the multiple suspected target images can be the same or different, depending on the actual situation. The suspected target image is square, and the center point position obtained in a certain connected component is the center point of the suspected target image, with the size being the maximum distance between any two response points.

[0049] For example, the number of suspected targets in the GF2_PMS1_E124.9_N29.0_20191007_L1A00042 image decreased from several thousand to 141.

[0050] The fourth step is to classify the sources of oil spill risk.

[0051] An oil spill risk source classification network was used to classify the M suspected target images after the third step of cropping, retaining the main risk sources of offshore oil spills: oil platforms and vessels. The specific steps are as follows:

[0052] 4.1) Since the M suspected target images obtained in step 3.3) are not the same size, the size of all images is adjusted to obtain M suspected target images of the same size.

[0053] 4.2) Obtain the trained oil spill risk source classification network

[0054] 4.2.1) First, the network structure of the oil spill risk source classification network will be explained.

[0055] LCSE-ResNet comprises the target localization part consisting of the LoG operator and the connected component controller, and the oil spill risk source classification network part. The oil spill risk source classification network contains 50 convolutional layers, with each pair of convolutional layers forming a basic network block. After two convolutional operations, a squeezing excitation operation is used to redistribute the weights of the different channels of the output. Then, a shorting method is used to add the input before the two convolutional operations to the output after weight redistribution, which serves as the input to the next network block.

[0056] 4.2.2) By cropping the panchromatic images of the Gaofen-2 optical satellite that have identified ships, oil platforms, clouds, other man-made targets, sea ripples and other interference information, a dataset consisting of multiple images is obtained. Each image is a single-channel grayscale image with a spatial resolution of 1 meter and an image size of 201×201. In addition to the sea surface, the images contain only one of the following: ships, oil platforms, clouds, other man-made targets, ripples or noise interference.

[0057] 4.2.3) Divide the dataset obtained in step 4.2.2) into a training set and a test set. Use the training set to train the oil spill risk source classification network from step 4.1), and use the test set to validate it. Specifically: A classification network model is constructed by combining a squeeze-excitation network module and a residual network. The cross-entropy loss function is used to compare the model's output with the true label, measuring the difference between them. The network model is trained using the training dataset, and backpropagation and gradient descent are used to update the model's weights and biases to minimize the loss function. During training, the classification model learns the mapping relationship from the input image to the output category. Through backpropagation and optimization algorithms, the model adjusts its weights and biases so that, given an input image, it can produce an output that matches the true label.

[0058] 4.4) The trained oil spill risk source classification network is used to classify the suspected targets in step 4.3). The classification network will output the corresponding label to the input image based on the correspondence between the training set images and labels during training, so as to obtain the actual category of the suspected target and retain the targets classified as ships and oil platforms among the M suspected targets.

[0059] In the GF2_PMS1_E124.9_N29.0_20191007_L1A0004292462 image, 9 ship targets and 3 oil platform targets were successfully classified out from 141 suspected targets, which is consistent with the actual number.

[0060] Marine targets exhibit high similarity, and in cropped images, targets almost occupy the center of the image. Each image is almost entirely composed of targets of similar brightness and size, along with their surrounding marine environment. For traditional convolutional operations, most improvements have focused on increasing the receptive field, fusing more spatial features, or extracting multi-scale spatial information. When performing feature fusion across multiple channels, convolutional operations essentially fuse all channels of the input feature map. However, the importance of channels is intertwined with the spatial correlations captured by the convolutional layers, failing to effectively utilize the importance of features from different channels. For ships, oil platforms, and other similar targets, spatial correlation alone is insufficient for effective differentiation. Oil spill risk source classification networks can mix extracted channel and spatial information, allowing the model to learn relationships between different channels. In shallower layers, this improves the quality of lower-level representations; as the number of layers increases, it enhances the more important features between each channel, making the internal and contour features of ships and oil platforms more distinct from other marine disturbances.

[0061] The oil spill risk source classification network uses a squeezing operation to compress the global spatial information of each channel into a channel descriptor to obtain the global characteristics of the channel. This description uses global average pooling to obtain the descriptive information of a single channel. After obtaining the channel information, the network uses an excitation operation to learn the relationships between each channel. This operation includes two fully connected layers.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the method solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the method solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the method features, do not cause the essence of the corresponding method solutions to deviate from the scope of the method solutions of the embodiments of the present invention.

Claims

1. A rapid and intelligent identification and classification method for marine oil spill risk sources from high-resolution remote sensing imagery, characterized in that, It includes the following steps: The first step is to preprocess the obtained remote sensing images to be classified, removing interference from clouds and land. The second step is to respond to suspected sources of marine oil spill risk. The method is implemented using the Gaussian Laplacian operator LoG in LCSE-ResNet. LoG uses Gaussian filtering to reduce noise in the image and then uses the second-order Laplacian gradient operator to obtain the locations in the image where the brightness changes significantly. Based on the obvious brightness characteristics of ships and oil platforms in the marine environment, gradient change response points are obtained around the suspected targets, thus initially identifying the suspected targets. The third step is to locate the suspected target. The connected domain controller is used to aggregate and calculate multiple response points around the suspected target to obtain the center position and size of the suspected target, and then cropped to obtain the accurately located suspected target; The fourth step is to classify the sources of oil spill risk. An oil spill risk source classification network was used to classify the M suspected target images after the third step of cropping, retaining the main risk sources of offshore oil spills: oil platforms and vessels. The specific steps are as follows: 4.1) Since the M suspected target images obtained in step 3.3) are not the same size, the size of all images is adjusted to obtain M suspected target images of the same size; 4.2) Obtain the trained oil spill risk source classification network 4.2.1) First, the network structure of the oil spill risk source classification network will be explained. LCSE-ResNet includes the target localization part composed of the LoG operator and the connected component controller, and the oil spill risk source classification network part. The oil spill risk source classification network contains 50 convolutional layers, with each two convolutional layers forming a basic network block. After two convolutional operations, a squeezing excitation operation is used to redistribute the weights of the different channels of the output. Then, a shorting method is used to add the input before the two convolutional operations and the output after weight redistribution as the input of the next network block. 4.2.2) By cropping the panchromatic images of the Gaofen-2 optical satellite that have identified ships, oil platforms, clouds, other man-made targets, sea ripples and other interference information, a dataset consisting of multiple images is obtained. Each image is a single-channel grayscale image, and the images contain only one of the following, in addition to the sea surface: ships, oil platforms, clouds, other man-made targets, ripples or noise interference. 4.2.3) Divide the dataset obtained in step 4.2.2) into a training set and a test set; use the training set to train the oil spill risk source classification network in step 4.1), and use the test set to validate it; the specific process is as follows: use the squeezing excitation network module and the residual network to build a classification network model; use the cross-entropy loss function to compare the output of the model with the true label and measure the difference between them; The network model is trained using a training dataset, and backpropagation and gradient descent are used to update the model's weights and biases to minimize the loss function. During training, the classification model learns the mapping relationship from the input image to the output class. Through backpropagation and optimization algorithms, the model adjusts its own weights and biases so that it can produce an output that matches the true label given an input image. 4.3) The trained oil spill risk source classification network is used to classify the suspected targets in step 4.3). The classification network will output the corresponding label of the input image based on the correspondence between the training set image and the label during training, so as to obtain the actual category of the suspected target and retain the targets classified as ships and oil platforms among the M suspected targets.

2. The rapid and intelligent identification and classification method for marine oil spill risk sources based on high-resolution remote sensing imagery according to claim 1, characterized in that, The first step of the preprocessing process includes linear stretching, downsampling, threshold segmentation, and mask processing in sequence.

3. The rapid and intelligent identification and classification method for marine oil spill risk sources from high-resolution remote sensing imagery according to claim 1, characterized in that, The specific steps of the second step are as follows: 2.1) Perform differential operations on the remote sensing image after the first step of preprocessing to obtain pixels with drastic gradient changes. Since differential operations are sensitive to noise in remote sensing images, and the remote sensing image is blurred while calculating the gradient, a large number of response points can be obtained around the suspected oil platform and ship targets. 2.2) After step 2.1, there are a large number of response points around each suspected target. The response points with response values ​​between 5 and 25 are retained and these response points are initially identified as suspected targets.

4. The rapid and intelligent identification and classification method for marine oil spill risk sources from high-resolution remote sensing imagery according to claim 1, characterized in that, The specific steps of the third step are as follows: 3.1) Calculate the distance between any two points among all response points around each suspected target, and consider all points within a certain region that are within 30 pixels apart and have more than 50 response points as the same connected component, thus obtaining M connected components; 3.2) Calculate the average x and y coordinates of all response points in each connected component and the distance between any two response points. Take the maximum distance and use the point with the average x and y coordinates and the maximum distance as the center point and size of the suspected target pointed to by the connected component. 3.3) Based on the center point location and size of the suspected target, the suspected target is cropped from the original remote sensing image before preprocessing, resulting in M ​​cropped images of the suspected target. The sizes of the multiple suspected target images can be the same or different, depending on the actual situation. The suspected target image is a square, and the center point obtained in a certain connected component is the center point of the suspected target image. The size is the maximum distance between any two response points.

Images