Small-scale water area water surface oil spill information identification method

Abstract

The invention relates to the technical field of offshore oil spill, in particular to a small-scale water area water surface oil spill information identification method, which comprises the following steps: collecting an SAR image of the water surface of a target water area; the SAR image is input into a preset recognition model, segmented images containing oil spill areas and non-oil-spill areas are obtained, the recognition model is obtained by training a U-Net model through a training set, and the training set comprises a plurality of oil spill images marked with the oil spill areas and the non-oil-spill areas; and refining the segmented image to obtain a final segmented image, and completing oil spill information identification of the water surface of the target water area. According to the invention, oil spill area detection is carried out on the SAR image by using the deep learning U-Net model, the oil spill area in the SAR image can be accurately identified, and segmentation of the oil spill area in the small-scale water area SAR image is realized.

Description

Technical Field

[0001] The invention relates to the technical field of offshore oil spills, and in particular to a method for identifying oil spill information on a surface of a small-scale water area. Background Art

[0002] As an important precious resource, oil is mainly extracted by offshore drilling platforms and produced by several major oil-producing countries in the world. At present, the main mode of transportation of oil is marine shipping. With the development of the times, the demand for oil in modern industrial society is increasing, and the oil trade volume is gradually increasing, which has greatly promoted the development of marine transportation. In the past few decades, marine transportation has been vigorously developed due to its stability and low transportation cost. At the same time as the development of the marine transportation industry, offshore oil exploitation is also developing. In order to meet the needs of economic development, countries around the world have invested more and more offshore drilling platforms to exploit oil resources in the ocean. But at the same time, it is also accompanied by frequent oil spills at sea.

[0003] Oil spills at sea are a serious threat to the marine environment. It is crucial to improve preventive measures and strengthen emergency response capabilities after oil spills occur, which has a significant effect on reducing the impact of accidents. Once an oil spill occurs, quickly and accurately identifying oil spill information on the water surface is the first step in implementing efficient emergency measures, which is of decisive significance for quickly controlling the development of the accident. At present, the commonly used methods for monitoring oil spills at sea include satellite remote sensing monitoring, aircraft patrol monitoring, patrol ship monitoring, fixed point monitoring and buoy monitoring. Satellite remote sensing monitoring has the advantages of all-weather, large range and high precision, and is a powerful way to monitor the environment; aircraft patrol monitoring has the advantages of strong real-time and high accuracy, but it requires more manpower and has a small range and strong regionality; patrol ship monitoring has the advantages of strong mobility, but is greatly affected by weather and environmental conditions; fixed point monitoring has the advantages of sensitive response and all-weather monitoring, but the range is small, the equipment is easily damaged and the mobility is poor; ocean buoy monitoring can effectively, automatically and all-day collect, mark and send marine environmental data, but it cannot intuitively display the on-site situation of the oil spill area.

[0004] Therefore, in recent years, researchers at home and abroad have been trying various new methods for oil spill information identification, using different methods such as infrared, visible light, hyperspectral and microwave to detect oil spills, so that better oil spill information identification performance can be achieved. Synthetic aperture radar (SAR) technology is valued because of its ability to work under any weather conditions and has become an important tool for oil spill monitoring. Oil spill identification technology utilizes the continuous evolution of SAR image technology. Its method has turned from the threshold judgment of physical spectrum attributes in the early stage to the application of texture analysis technology, and has developed into a neural network method widely used at present. However, the method that traditionally relies on artificial design features is not only inefficient when processing these images, but also requires professionals to have a higher technical level, making the extraction of deep features particularly difficult. Therefore, in order to more effectively utilize these rich SAR image information and overcome the limitations of traditional methods, it is necessary to find more advanced and automated technologies to process and analyze these images. Therefore, the present invention proposes a method for identifying oil spill information on the surface of small-scale waters. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying oil spill information on the surface of small-scale waters, which uses a deep learning U-Net model to detect oil spill areas in SAR images, can accurately identify oil spill areas in SAR images, and realize the segmentation of oil spill areas in small-scale water SAR images.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for identifying oil spill information on a small-scale water surface, comprising:

[0008] Collect SAR images of the target water surface;

[0009] Inputting the SAR image into a preset recognition model to obtain a segmented image containing an oil spill area and a non-oil spill area, wherein the recognition model is obtained by training a U-Net model with a training set, and the training set includes a number of oil spill images marked with oil spill areas and non-oil spill areas;

[0010] The segmented image is refined to obtain a final segmented image, thereby completing the oil spill information recognition on the target water area.

[0011] Optionally, before training the U-Net model using the training set, the step further includes preprocessing the training set, and the preprocessing includes:

[0012] Representing the oil spill area in the oil spill image in the form of a mask;

[0013] The oil spill image and its mask are converted into a grayscale image, and then the grayscale image is converted into an array and then normalized.

[0014] Optionally, the U-Net model includes an encoder, a decoder and a bottleneck layer, the encoder is used to extract a feature map containing context information of the oil spill image by reducing the resolution layer by layer and increasing the number of feature channels; the bottleneck layer is used to input the feature map into the decoder, the decoder is jump-connected to the encoder, and is used to reduce the number of feature channels by half layer by layer to restore information of the feature map, and classify the feature map at the pixel level, and output a segmented image containing oil spill areas and non-oil spill areas.

[0015] Optionally, the encoder includes four convolution blocks connected in sequence, and each convolution block is connected to a maximum pooling layer, wherein each convolution block includes two first convolution layers with the same structure, and each first convolution layer is initialized using a ReLU activation function and he_normal.

[0016] Optionally, the decoder includes four upsampling layers, each upsampling layer is connected to two second convolutional layers with the same structure, and finally a third convolutional layer is used to output a final image through a sigmoid activation function, and the second convolutional layer and the third convolutional layer have different sizes.

[0017] Optionally, the bottleneck layer includes two fourth convolutional layers with the same structure, and a Dropout layer is added to the bottleneck layer and the last convolutional block of the encoder to reduce overfitting.

[0018] Optionally, the Adam optimizer is used when training the U-Net model through the training set, the binary cross entropy is used as the loss function, and the accuracy is used as the performance evaluation indicator. When the loss value and the accuracy both meet the preset requirements, the training is completed.

[0019] Optionally, performing thinning processing on the segmented image includes:

[0020] The segmented image is expanded, eroded, and smoothed to obtain the final segmented image.

[0021] The beneficial effects of the present invention are:

[0022] The present invention uses a deep learning U-Net model to detect oil spill areas in SAR images, uses an open source SOS dataset as a training set and a test set for the model, and uses accuracy as an evaluation indicator. The accuracy of the model after training reaches 90%. The trained U-Net model is used to identify oil spill areas in SAR images, and the oil spill areas in SAR images can be accurately identified, thereby realizing the segmentation of oil spill areas in SAR images of small-scale waters. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 This is a flow chart of a method for identifying oil spill information on a small-scale water surface according to an embodiment of the present invention;

[0025] Figure 2 Schematic diagram of the U-Net model structure of an embodiment of the present invention;

[0026] Figure 3 The accuracy of the U-Net model of the embodiment of the present invention on the training set and the test set;

[0027] Figure 4 The loss verification results of the U-Net model of the embodiment of the present invention on the training set and the test set;

[0028] Figure 5 The visualization results of the trained U-Net model of an embodiment of the present invention on the PALSAR test set are shown in Figure 1, where (a) is a SAR image, (b) is a predicted image, (c) is a curve-thinned image, and (d) is an original annotated image.

[0029] Figure 6 This is the visualization result of the trained U-Net model of an embodiment of the present invention on the Sentinel-1A test set, where (a) is a SAR image, (b) is a predicted image, (c) is a curve-thinned image, and (d) is the original annotated image. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] This embodiment provides a method for identifying oil spill information on a small-scale water surface. Figure 1 As shown, including:

[0033] Collect SAR images of the target water surface;

[0034] The SAR image is input into a preset recognition model to obtain a segmented image containing the oil spill area and the non-oil spill area, wherein the recognition model is obtained by training a U-Net model with a training set, and the training set includes a number of oil spill images marked with the oil spill area and the non-oil spill area;

[0035] The segmented image is refined to obtain the final segmented image and complete the oil spill information recognition on the target water surface.

[0036] Furthermore, before training the U-Net model with the training set, the training set is also preprocessed, and the preprocessing includes:

[0037] The oil spill area in the oil spill image is represented in the form of a mask;

[0038] The oil spill image and its mask are converted into grayscale images, and then the grayscale images are converted into arrays and normalized.

[0039] Furthermore, the U-Net model includes an encoder, a decoder and a bottleneck layer. The encoder is used to extract a feature map containing context information of the oil spill image by reducing the resolution layer by layer and increasing the number of feature channels; the bottleneck layer is used to input the feature map into the decoder. The decoder is jump-connected to the encoder to reduce the number of feature channels by half layer by layer to restore information of the feature map, and classify the feature map at the pixel level to output a segmented image containing oil spill areas and non-oil spill areas.

[0040] The encoder includes four convolution blocks connected in sequence, and each convolution block is connected to a maximum pooling layer, wherein each convolution block contains two first convolution layers with the same structure, and each first convolution layer is initialized using a ReLU activation function and he_normal.

[0041] The decoder consists of four upsampling layers. Each upsampling layer is connected to two second convolutional layers with the same structure. Finally, the third convolutional layer is used to output the final image through the sigmoid activation function. The second and third convolutional layers have different sizes.

[0042] The bottleneck layer includes two fourth convolutional layers with the same structure. A Dropout layer is added to the bottleneck layer and the last convolutional block of the encoder to reduce overfitting.

[0043] Furthermore, the Adam optimizer is used to train the U-Net model through the training set, the binary cross entropy is used as the loss function, and the accuracy is used as the performance evaluation indicator. When the loss value and the accuracy meet the preset requirements, the training is completed.

[0044] Furthermore, performing refinement processing on the segmented image includes: performing expansion, erosion, and smoothing processing on the segmented image to obtain a final segmented image.

[0045] Specifically, the construction, training, and testing of the recognition model in this embodiment include the following:

[0046] (1) Constructing a dataset;

[0047] This embodiment uses the Deep-SAR Oil Spill (SOS) dataset jointly produced and shared by China University of Geosciences (HPSCIL) and Wuhan University (RS-IDEA). The SOS dataset uses SAR images from the ALOS satellite and the Sentinel-1A satellite as the basis. Among them, 14 original SAR images from the ALOS satellite record the sea surface conditions in the Gulf of Mexico from May 21 to July 25 after the explosion of the "Deepwater Horizon" drilling platform in the Gulf of Mexico, USA on April 20, 2010. The SAR images are pre-processed by multiple processes such as map projection and multi-vision processing. The pixel spacing is 12.5 meters, and the polarization mode is horizontal transmission and horizontal reception (HH). The oil spill area is clearly displayed on the HH polarization image; the other 7 original SAR images from the Sentinel-1A satellite recorded the Persian Gulf region from August 5, 2017 to August 14, 2017, with a spatial resolution of 5m×20m, and the polarization modes used vertical transmission vertical reception (VV) and vertical transmission horizontal reception (VH). Considering that the oil spill area appears relatively clear on the VV polarization image, the SAR image with VV polarization mode is selected as the original data and then the image is preprocessed by speckle filtering, radiation calibration, and terrain correction.

[0048] The sample size of 21 original SAR images is too small to be used in the training of convolutional neural networks. Therefore, data augmentation technology is used to expand the original data set. Operations such as cropping, rotation, and adding noise are performed on the original grayscale images. The original 21 SAR images are divided into 6456 oil spill images of size 416 pixels × 416 pixels. In order to balance the quantitative difference between oil spill samples and non-oil spill samples, careful manual screening and cropping operations are performed. Finally, a data set containing 8070 samples is successfully constructed. Each sample is 256 pixels × 256 pixels in size, and is accompanied by clear oil spill area and non-oil spill area annotations. In terms of data set division, it is divided into training set and test set in a ratio of 8:2. Among them, the training set contains 3101 images from the Mexican oil spill area, and the test set includes 776 images of the area. Similarly, the training set also covers 3354 images from the Persian Gulf oil spill area, and the test set contains 839 images of the area.

[0049] (2) Constructing a recognition model;

[0050] The U-Net model includes an encoder, a decoder and a bottleneck layer. The encoder is used to extract a feature map containing context information of the oil spill image by reducing the resolution layer by layer and increasing the number of feature channels. The bottleneck layer is used to input the feature map into the decoder. The decoder is jump-connected with the encoder to reduce the number of feature channels by half layer by layer to restore information of the feature map, classify the feature map at the pixel level, and output a segmented image containing oil spill areas and non-oil spill areas.

[0051] The encoder includes four convolution blocks connected in sequence, and each convolution block is connected to a maximum pooling layer, wherein each convolution block contains two first convolution layers with the same structure, and each first convolution layer is initialized using a ReLU activation function and he_normal.

[0052] The decoder consists of four upsampling layers. Each upsampling layer is connected to two second convolutional layers with the same structure. Finally, the third convolutional layer is used to output the final image through the sigmoid activation function. The second and third convolutional layers have different sizes.

[0053] The bottleneck layer includes two fourth convolutional layers with the same structure. A Dropout layer is added to the bottleneck layer and the last convolutional block of the encoder to reduce overfitting.

[0054] Specifically, Figure 2 As shown in the figure, the U-Net model is a U-shaped symmetrical structure. The contraction path on the left, that is, the encoder captures and extracts image context information, and gradually reduces the feature map resolution and increases the number of feature channels using convolution, ReLU activation and maximum pooling layers; the expansion path on the right, that is, the decoder restores target details through upsampling and convolution operations, and at each level, the high-resolution features of the encoder are fused with the upsampled feature maps of the corresponding level in the decoder through jump connections to retain more spatial information. Finally, at the end of the expansion path, the feature map is converted into an output channel equal to the number of categories through a 1x1 convolution layer, and the category label of each pixel is determined, thereby generating a high-precision semantic segmentation map.

[0055] The U-Net model takes a single-channel 256x256 pixel image as input. The encoder part contains four convolution blocks, each of which contains two 3x3 convolution layers, using the ReLU activation function and he_normal initialization, followed by a 2x2 maximum pooling layer. The function formula of the activation function ReLU is:

[0056] f(x)=max(0,x) (1)

[0057] When the input value c > 0, its reciprocal is always 1, allowing better gradient propagation and faster model training.

[0058] The number of channels starts at 64 and doubles after each convolution block until it reaches 512. A Dropout layer is also added in the fourth convolution block with a dropout rate of 0.5 to reduce overfitting. A convolution layer with 1024 channels is used in the bottleneck layer, and a Dropout layer is also added in between with a dropout rate of 0.5. The decoder part is upsampled by UpSampling2D, and then the number of channels is halved through the convolution layer and spliced ​​with the output of the corresponding encoder part. This process gradually reduces the number of channels from 1024 to 64. After each upsampling, two convolution operations are performed to further refine the features. Finally, after several layers of convolution, the output layer uses a 1x1 convolution kernel to output the final image through the sigmoid activation function. The number of channels is 1, indicating the binary classification prediction of the pixel point. The function formula of the activation function sigmoid is:

[0059]

[0060] The model uses the Adam optimizer, the learning rate is 0.0001, and the loss function is binary cross entropy. The mathematical expression of the binary cross entropy loss function is:

[0061]

[0062] Where L is the loss function, N is the number of samples, and y i is the true value, is the predicted value.

[0063] The performance evaluation indicator is accuracy, and the mathematical expression is:

[0064]

[0065] Among them, TP (True Positives) is the number of positive classes correctly predicted as positive classes, TN (True Negatives) is the number of negative classes correctly predicted as negative classes, FP (False Positives) is the number of negative classes incorrectly predicted as positive classes, and FN (False Negatives) is the number of positive classes incorrectly predicted as negative classes.

[0066] (3) Train and test the recognition model based on the dataset;

[0067] a. Training:

[0068] The oil spill area of ​​the oil spill image in the training set is represented in the form of a mask, and the oil spill image and its mask are converted into a 256x256 grayscale image, which is then converted into an array and normalized to the range of [0,1].

[0069] Use the processed training set to train the model, use the fit method, set the batch size to 32, and the training cycle to 20 times, and save the trained model.

[0070] b. Test:

[0071] Load the trained U-Net model, convert the oil spill image in the test set into grayscale format and resize it to 256x256 pixels, then convert it into an array and normalize it to the range of [0,1], input it into the U-Net model, and get the predicted mask. This mask represents the part that the model identifies as the oil spill area.

[0072] On the training set and test set, the accuracy and loss verification results of the U-Net model are as follows: Figure 3 , Figure 4 After training on the training set, the accuracy of the U-Net model on the test set can reach more than 90%, and the value of the binary cross entropy loss function is around 0.2, indicating that the model can accurately identify the oil spill area in the SAR image and has good oil spill area recognition ability.

[0073] The predicted mask is further binarized with a threshold of 0.5 and morphological operations, including dilation and erosion, are performed to remove prediction noise and improve the quality of the mask. In addition, Gaussian blur is applied to smooth the mask edges. The original image, the mask predicted by the model, and the mask after refinement are visualized for comparison, as shown in Figure 2. Figure 5 , Figure 6 shown. Figure 5 The visualization results of the trained U-Net model on the PALSAR test set, where (a) is the SAR image, (b) is the predicted image, (c) is the curve-thinned image, and (d) is the original annotation image; Figure 6 The visualization results of the trained U-Net model on the Sentinel-1A test set, where (a) is the SAR image, (b) is the predicted image, (c) is the curve-refined image, and (d) is the original annotation image.

[0074] It can be seen that the predicted image after refinement has a high degree of overlap with the original annotated image. The white area in the image is the oil spill area. The U-Net model can accurately identify the oil spill area on the PALSAR test set and the Sentinel-1 test set, and can accurately segment the oil spill area from the SAR image. The visualization results show that the U-Net model has good oil spill area recognition capabilities.

[0075] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying oil spill information on small-scale water surfaces. It is characterized in that include: Collect SAR images of the target water surface; Inputting the SAR image into a preset recognition model to obtain a segmented image containing an oil spill area and a non-oil spill area, wherein the recognition model is obtained by training a U-Net model with a training set, and the training set includes a number of oil spill images marked with oil spill areas and non-oil spill areas; The segmented image is refined to obtain a final segmented image, thereby completing the oil spill information recognition on the target water area.

2. The method for identifying oil spill information on a small-scale water surface according to claim 1, It is characterized in that Before training the U-Net model with the training set, the training set is preprocessed, and the preprocessing includes: Representing the oil spill area in the oil spill image in the form of a mask; The oil spill image and its mask are converted into a grayscale image, and then the grayscale image is converted into an array and then normalized.

3. The method for identifying oil spill information on a small-scale water surface according to claim 1, It is characterized in that The U-Net model includes an encoder, a decoder and a bottleneck layer. The encoder is used to extract a feature map containing context information of the oil spill image by reducing the resolution layer by layer and increasing the number of feature channels. The bottleneck layer is used to input the feature map into the decoder. The decoder is jump-connected with the encoder to reduce the number of feature channels by half layer by layer to restore information of the feature map, and to classify the feature map at the pixel level to output a segmented image containing oil spill areas and non-oil spill areas.

4. The method for identifying oil spill information on a small-scale water surface according to claim 3, It is characterized in that The encoder includes four convolution blocks connected in sequence, and each convolution block is connected to a maximum pooling layer, wherein each convolution block includes two first convolution layers with the same structure, and each first convolution layer is initialized using a ReLU activation function and he_normal.

5. The method for identifying oil spill information on a small-scale water surface according to claim 3, It is characterized in that The decoder includes four upsampling layers, each of which is connected to two second convolutional layers with the same structure, and finally a third convolutional layer is used to output a final image through a sigmoid activation function, and the second convolutional layer and the third convolutional layer have different sizes.

6. The method for identifying oil spill information on a small-scale water surface according to claim 3, It is characterized in that The bottleneck layer includes two fourth convolutional layers with the same structure, and a Dropout layer is added to the bottleneck layer and the last convolutional block of the encoder to reduce overfitting.

7. The method for identifying oil spill information on a small-scale water surface according to claim 1, It is characterized in that The Adam optimizer is used to train the U-Net model through the training set, the binary cross entropy is used as the loss function, and the accuracy is used as the performance evaluation indicator. When the loss value and the accuracy both meet the preset requirements, the training is completed.

8. The method for identifying oil spill information on a small-scale water surface according to claim 1, It is characterized in that The thinning process of the segmented image includes: The segmented image is expanded, eroded, and smoothed to obtain the final segmented image.

Images