CN116206203B - Oil spill detection method based on SAR and Dual-EndNet - Google Patents

Oil spill detection method based on SAR and Dual-EndNet Download PDF

Info

Publication number
CN116206203B
CN116206203B CN202310217615.9A CN202310217615A CN116206203B CN 116206203 B CN116206203 B CN 116206203B CN 202310217615 A CN202310217615 A CN 202310217615A CN 116206203 B CN116206203 B CN 116206203B
Authority
CN
China
Prior art keywords
polarization
endnet
dual
oil spill
scattering
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202310217615.9A
Other languages
Chinese (zh)
Other versions
CN116206203A (en
Inventor
宋冬梅
王明月
王斌
张�杰
刘善伟
王大伟
高晗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China University of Petroleum East China
Original Assignee
China University of Petroleum East China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China University of Petroleum East China filed Critical China University of Petroleum East China
Priority to CN202310217615.9A priority Critical patent/CN116206203B/en
Publication of CN116206203A publication Critical patent/CN116206203A/en
Application granted granted Critical
Publication of CN116206203B publication Critical patent/CN116206203B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A20/00—Water conservation; Efficient water supply; Efficient water use
    • Y02A20/20—Controlling water pollution; Waste water treatment
    • Y02A20/204—Keeping clear the surface of open water from oil spills

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses an oil spill detection method based on SAR and Dual-EndNet, which comprises the following steps: acquiring a full polarization SAR image of a target sea surface; pauli decomposition is carried out on the full-polarization SAR image to obtain scattering energy of odd scattering and even scattering, and the PauliRGB image is synthesized by utilizing the scattering energy; extracting polarization characteristics of oil spill detection from the full-polarization SAR image, and selecting the polarization characteristics by utilizing a random forest algorithm to obtain a polarization characteristic image; based on PauliRGB image and polarization characteristic image, obtaining geometric space sample data set, polarization characteristic data set and sample label set, and dividing into training set and test set; constructing a Dual-EndNet network model; training the Dual-EndNet network model by using a training set; and (3) completing the oil spill detection of the target sea surface by using the test set and the trained Dual-EndNet network model.

Description

Oil spill detection method based on SAR and Dual-EndNet
Technical Field
The invention belongs to the technical field of ocean, and particularly relates to an oil spill detection method based on SAR and Dual-EndNet.
Background
With the rapid development of marine economy, activities such as marine petroleum resource development, marine transportation and the like are increasingly prosperous, and marine oil spill accidents also occur frequently. The marine oil spill pollution has serious influence on marine ecological environment, ecological resources and marine economy, and has important significance in timely developing oil spill remote sensing detection. Among them, synthetic aperture radar (Synthetic Aperture Radar, SAR) has become one of the most important means for detecting oil spills on the sea at present due to its advantages of all-day, all-weather and wide-range observation. Compared with the traditional single polarization SAR, the polarization SAR (polar SAR) can acquire different polarization information of a target, and can record the phase difference of echo signals under different polarization state combinations while measuring the ground object echo amplitude so as to fully reveal the physical scattering mechanism of the target. Therefore, the advantages of polarized SAR images in terms of oil spill detection are increasingly prominent. As an advanced imaging radar technology, polSAR has become one of the main means of marine surface oil leak detection. Meanwhile, deep learning has been a significant research hotspot in the remote sensing field because of the ability to automatically mine deep features of images. The advent of deep learning provides a new opportunity for polarized SAR target detection in the big data age.
The deep learning framework provides a solution for detecting the oil spill of the polarized SAR image, but still faces the following problems. First, the polarized SAR image is a special microwave image, and features having the same gray scale value do not necessarily have the same optical characteristics. Therefore, the existing deep learning framework for optical image construction cannot be directly applied to the field of polarized SAR. Secondly, the existing method for detecting the offshore oil spill by using the polarized SAR image mostly only pays attention to the polarization characteristics of the image, and ignores the spatial characteristics of the image. The PauliRGB image contains rich features such as outlines, textures and the like, and is very matched with a real ground scene, so that the space features of the polarized SAR image can be effectively extracted. Therefore, the invention provides an oil spill detection method based on SAR and Dual-EndNet.
Disclosure of Invention
The invention aims to solve the defects of the prior art, and provides an oil spill detection method based on SAR and Dual-EndNet, wherein a Dual-channel coding-decoding network suitable for polarized SAR oil spill detection is constructed, and PauliRGB images and polarized characteristic images after characteristic optimization are respectively input into two channels, so that space information and polarization information in polarized SAR data can be fully extracted; the method has stronger oil spill detection capability, can not only distinguish sea water from oil spill, but also effectively improve the detection capability of different oil films of crude oil, emulsified oil and biological oil films.
In order to achieve the above object, the present invention provides the following solutions:
the oil spill detection method based on SAR and Dual-EndNet comprises the following steps:
acquiring a full polarization SAR image of a target sea surface;
pauli decomposition is carried out on the full-polarization SAR image, so that scattering energy of odd scattering and even scattering is obtained, and the PauliRGB image is synthesized by utilizing the scattering energy of the odd scattering and even scattering;
extracting the polarization characteristics of oil spill detection from the full-polarization SAR image, and selecting the polarization characteristics by utilizing a random forest algorithm to obtain a polarization characteristic image;
based on the PauliRGB image and the polarization feature image, a geometric space sample data set, a polarization feature data set and a sample label set are obtained, and the geometric space sample data set, the polarization feature data set and the sample label set are divided into a training set and a testing set;
constructing a Dual-EndNet network model;
training the Dual-EndNet network model by using the training set;
and finishing the oil spill detection of the target sea surface by using the test set and the trained Dual-EndNet network model.
Preferably, the method for Pauli decomposition of the full polarization SAR image comprises the following steps:
wherein S represents a polarized scattering matrix, and element S in the matrix ij Representing the complex scattering coefficients transmitted in polarization i and received in polarization j, respectively, each a, b, c, d being complex,
preferably, the even scattering includes: secondary scattering at a 0 ° direction angle and secondary scattering at a 45 ° direction angle.
Preferably, the method for selecting the polarization characteristic by using a random forest algorithm comprises the following steps:
and analyzing the capacity of distinguishing oil spill, oil film like and sea water of each oil spill detection polarization characteristic based on a random forest feature selection method, and selecting optimal features from the capacity to form a feature combination.
Preferably, the specific process of selecting the polarization feature by using a random forest algorithm includes:
assuming that the number of decision trees in a random forest algorithm is T, D features exist in a sample, and an original sample set is D= { (x) 1 ,y 1 ),(x 2 ,y 2 ),…,(x n ,y n ) Sample set capacity n;
extracting T sample sets from the sample set D by using a Bootstrap self-help sampling method, wherein the number of samples in each sample set is n;
respectively training T decision tree models by using the extracted T training sets with the size of n;
by calculating features X j The feature X is calculated by the average variance of the out-of-bag data errors of d twice before and after noise perturbation on all decision tree models, j=1, 2 j Importance score of (2);
and accumulating the importance scores of the features obtained by each iteration through multiple iterations, finally obtaining importance scores of the features after multiple iterations, sorting all the features according to the importance scores after the iterations, selecting a plurality of features with the importance scores sorted in front, and completing feature optimization.
Preferably, by calculating the characteristic X j The feature X is calculated by the average variance of the out-of-bag data errors of d twice before and after noise perturbation on all decision tree models, j=1, 2 j The method of importance score of (2) comprises:
s1: calculating OOB (out-of-bag data of ith decision tree) i Out-of-bag data error ErrOOB i ；
S2: for out-of-bag data OOB i Features X in j Adding noise disturbance to obtain OOB' i Recalculating out-of-bag data OOB' i Out-of-bag data error ErrOOB' i ；
S3: repeating said S1 and said S2 until { ErrOOB is obtained i I=1, 2, …, T } and { ErrOOB' i ,i＝1,2,…,T}；
S4: computing all decision tree features X j Mean value of the change in the data error outside the bag before and after substitution:
in the formula, VI (X) j ) Is characteristic X j Importance score of (c).
Preferably, the Dual-EndNet network model consists of two-channel encoder-decoder, including two channels of F10-EndNet and P-EndNet, extracting polarization information and spatial features from input data respectively; the extracted polarization information and spatial features are then stacked together and passed through a convolution and Softmax classifier to achieve the final prediction.
Preferably, at the encoder stage, a total of 6 convolutional layers and 2 pooling layers are provided:
for the first four convolution layers, adding one maximum pooling layer into one group for every two convolution layers, and dividing the two groups, wherein the convolution kernel size of each convolution layer is 3 multiplied by 3, and the step length is 1; the parameter settings of the step length and the convolution kernel size of the last two convolution layers are consistent with those of the first four convolution layers;
the core size of the maximum pooling layer is 2 x 2 with a step size of 2.
Preferably, at the decoder stage, 2 upsampling layers and 4 convolutional layers are designed in total, symmetrically to the encoder stage:
each up-sampling layer is connected with two convolution layers, and relevant parameters are the same as the arrangement of symmetrical parts of the convolution stage;
the up-sampled features are stacked together with the feature maps before the pooling layer by using a jump cascade and input into the convolutional layer together.
Compared with the prior art, the invention has the beneficial effects that:
according to the method, 30 polarization features are extracted from the full-polarization SAR image to form a polarization feature image, and data dimension reduction is performed by utilizing a random forest, so that polarization information of the image is fully utilized, and information redundancy is avoided.
The PauliRGB image provided by the invention contains rich features such as contours, textures and the like, is very consistent with a real ground scene, and can be used for effectively extracting the spatial features of the polarized SAR image.
The invention takes the Encoder-decoder as a basic framework to construct a new Dual-EndNet, designs the PauliRGB images and the optimized 10 polarized characteristic images which are respectively input into the two branches, simultaneously extracts the space information and the polarized information in the oil spill polarized SAR images, and fuses the two branch information so as to highlight the important characteristics of the oil spill target and improve the performance of the oil spill detection method.
The invention utilizes advanced technology in the field of artificial intelligence such as deep learning, and research on a polarized SAR oil spill detection method based on deep learning is developed by means of polarized SAR data according to the imaging principle and the image characteristics of sea surface SAR images. The method aims at expanding the application field of deep learning, discussing the capability of the deep learning on oil spill detection work, improving the precision of marine pollution monitoring and providing a certain research foundation for subsequent researches.
Drawings
In order to more clearly illustrate the technical solutions of the present invention, the drawings that are needed in the embodiments are briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic flow chart of an oil spill detection method based on SAR and Dual-EndNet;
FIG. 2 is a schematic diagram of a random forest algorithm of the present invention;
FIG. 3 is a flow chart of a polarized SAR oil spill detection experiment based on Dual-EndNet of the present invention;
FIG. 4 is a diagram of the Dual-EndNet network of the present invention;
FIG. 5 is PauliRGB image of two-view polarized SAR spilled oil image of the present invention, wherein FIG. (a) is a schematic view of spilled oil image obtained in the gulf of Mexico, and FIG. (b) is a schematic view of Radarsat-2 image obtained in the offshore spilled oil experiment in North sea of Europe;
FIG. 6 is a schematic diagram of the visual results of oil spill detection according to the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
Example 1
As shown in FIG. 1, the invention provides an oil spill detection method based on SAR and Dual-EndNet, which comprises the following steps:
acquiring a full polarization SAR image of a target sea surface;
pauli decomposition is carried out on the full-polarization SAR image to obtain scattering energy of odd scattering and even scattering, and the PauliRGB image is synthesized by utilizing the scattering energy of the odd scattering and even scattering;
extracting polarization characteristics of oil spill detection from the full-polarization SAR image, and selecting the polarization characteristics by utilizing a random forest algorithm to obtain a polarization characteristic image;
based on PauliRGB images and polarization feature images, a geometric space sample data set, a polarization feature data set and a sample label set are obtained, and the geometric space sample data set, the polarization feature data set and the sample label set are divided into a training set and a testing set;
constructing a Dual-EndNet network model;
training the Dual-EndNet network model by using a training set;
and (3) completing the oil spill detection of the target sea surface by using the test set and the trained Dual-EndNet network model.
In the embodiment, two-scene Radarsat-2 full-polarization SAR oil spill data are selected for experiments. Firstly, clipping a polarized SAR image and denoising the image based on the Lee method.
In this embodiment, the PauliRGB image is a pseudo-color image synthesized by scattering energy corresponding to the first three scattering mechanisms decomposed by Pauli, and contains abundant features such as contours and textures, which are very consistent with a real ground scene, so that the spatial features of the polarized SAR image can be effectively extracted.
Pauli decomposition is a decomposition method that decomposes the polarized scattering matrix into four basis matrices and corresponds to four scattering mechanisms, and then represents the overall scattering process of the target with the four basis scattering mechanisms. The Pauli decomposition formula is shown in formula (1).
Wherein S represents a polarized scattering matrix, and the elements S in the matrix ij The complex scattering coefficients, a, b, c, d, representing the transmission in polarization i and the reception in polarization j, respectively, are complex numbers. Their values are given by formula (2):
according to the description of the basic properties of Pauli base matrices by the theory of polarization base transformation of electromagnetic waves, the four base matrices respectively represent a scattering mechanism, and the four base matrices are respectively: (1) odd scattering; (2) secondary scattering at a 0 ° direction angle; (3) secondary scattering at 45 ° direction angle; (4) an asymmetric component in the scattering matrix S. Pauli decomposition theory is simple and easy to understand, and the orthogonal base is adopted for decomposition, so that the Pauli decomposition theory has certain anti-interference capability.
In this embodiment, the full polarization SAR has four polarization channels, and can acquire more abundant polarization scattering information through different polarization channels. Compared with the traditional unipolar SAR, the full-polarization SAR can improve the oil spill detection precision. Currently, marine oil spill detection using polarized SAR images is mostly performed by extracting polarization features. Different polarization characteristic construction mechanisms are different, have differences in oil spill characterization capability and change with the change of marine environment. In order to more comprehensively describe the oil spill characteristics, 30 oil spill detection polarization characteristics commonly used at present are summarized and combed, the capability of distinguishing oil spill, oil film like and sea water of each characteristic is analyzed by utilizing a random forest-based characteristic selection method, the optimal characteristic is selected from the characteristics, and a characteristic combination is formed, so that data preparation is carried out for subsequent oil spill detection based on deep learning.
a. Polarization feature extraction
The present invention provides a comprehensive combing summary of the currently commonly used 30 spilled oil detection polarization characteristics, as shown in table 1.
TABLE 1
b. Feature optimization based on random forests
The feature selection is performed based on random forests, mainly by using classification accuracy as a measurement standard, and the feature with larger contribution rate is selected by calculating the importance score of each feature, so that the schematic diagram is shown in fig. 2. Specifically, testing the data outside the bag by using a trained random forest model, and calculating a classification error; and then replacing a certain feature by random noise, generating new data outside the bag, recalculating the classification error, and calculating the importance degree of the feature according to the classification error calculated twice. The more important the feature is, the more the result of random forest model prediction changes, and conversely, the less the change. In this way, the importance of all features in each decision tree can be calculated. If a feature calculates a plurality of importance scores in a plurality of decision trees, taking the average value of the scores as the final importance score of the feature. The method comprises the specific steps of obtaining a sample set, constructing a random forest model, calculating feature importance, and selecting features according to feature importance scores.
Assuming that the number of decision trees in a random forest algorithm is T, D features exist in a sample, and an original sample set is D= { (x) 1 ,y 1 ),(x 2 ,y 2 ),…,(x n ,y n ) And if the sample set capacity is n, performing feature selection by using a random forest, wherein the specific steps are as follows:
step 1: and (3) acquiring a sample set. And extracting T sample sets from the sample set D by using a Bootstrap self-help sampling method, wherein the number of samples in each sample set is n. During sampling, some samples are not sampled and thus do not appear in any sample set, and such samples are called Out-Of-Bag (OOB) data, the presence Of which can be used as a test set to verify the performance Of the random forest model.
Step 2: and (5) constructing a random forest model. And respectively training T decision tree models by using the extracted T sampling sets with the size of n. In the process of generating the decision tree, k attributes are randomly selected when nodes are divided, and the optimal attribute is selected from the k attributes by using a characteristic evaluation method to divide the nodes. The RF algorithm directly integrates the generated T decision tree models into the modelTogether, the final result is obtained using majority voting principles or averaging. Where the parameter k is used to control the degree of randomness, typically k=log 2 d。
Step 3: and calculating the importance of the features. By calculating features X j (j=1, 2, …, d) calculating the feature X by averaging the variance of the out-of-bag data errors twice before and after noise perturbation on all decision trees j Importance score of (c).
The specific process is as follows:
s1: calculating OOB (out-of-bag data of ith decision tree) i Out-of-bag data error ErrOOB i 。
S2: to OOB while keeping other features unchanged i Features X in j Adding noise disturbance to obtain OOB' i Recalculating out-of-bag data OOB' i Out-of-bag data error ErrOOB' i 。
S3: repeating the steps S1 and S2 until { ErrOOB is obtained i I=1, 2, …, T } and { ErrOOB' i ,i＝1,2,…,T}。
S4: computing all decision tree features X j Mean value of the change in the data error outside the bag before and after substitution:
in the formula, VI (X) j ) Is characteristic X j Importance score of (c).
Step 4: and (5) feature selection. And accumulating the importance scores of the features obtained by each iteration through multiple iterations, and finally obtaining the importance scores of the features after multiple iterations. And sorting all the features according to the scores, and selecting the first few features with higher scores to finish feature preference.
In this embodiment, the invention provides a Dual-channel encoding-decoding network (Dual-EndNet) full-polarization SAR image marine oil spill detection method. The method consists of two channels of encoder-decoders: one channel is mainly used for extracting polarization information (F10-EndNet) in 10 polarization feature images, and the other channel is mainly used for extracting spatial information (P-EndNet) of PauliRGB images. The features extracted by the two channels are input into a convolution layer together for mutual fusion, and then input into a classifier to finish classification work. The specific flow of the method is shown in figure 3.
The invention uses the encoder-decoder as a basic framework to build a Dual-channel encoding-decoding network (Dual-EndNet). The Dual-EndNet comprises two channels of F10-EndNet and P-EndNet, and polarization information and spatial characteristics are extracted from input data respectively; the extracted polarization information and spatial features are then stacked together and predicted by convolution and Softmax classifiers. The Dual-EndNet network firstly utilizes rolling and pooling to extract deep features of the image; then gradually restoring the size of the image by utilizing up-sampling and convolution operation; the feature maps of the two channels are then fused to achieve the final prediction, the schematic of which is shown in fig. 4.
In the encoder stage, a total of 6 convolutional layers and 2 pooling layers are provided. For the first four convolution layers, each two convolution layers are added with one maximum pooling layer to form a group, and the two groups are divided. Wherein the convolution kernel size of each convolution layer is 3×3, and the step length is 1; the core size of the maximum pooling layer is 2 x 2 with a step size of 2. The parameter settings of the last two convolution layers, the step size and the convolution kernel size are consistent with the first four convolution layers. The number of convolution kernels for each convolution layer is shown in table 2.
TABLE 2
In the decoder stage, 2 up-sampling layers and 4 convolutional layers are designed, essentially symmetrically to the encoder stage. Each up-sampling layer is followed by two convolution layers, and the relevant parameters are the same as the arrangement of the symmetrical parts of the convolution stage. The upsampling is followed by a jump cascade, which is stacked together with the feature map before the pooling layer and is jointly input into the convolutional layer. In particular, all convolution layers in the network contain bulk normalization operations and ReLU activation functions.
In this embodiment, in the network training process, the loss function selects a cross entropy loss function and an Adam optimization algorithm. Meanwhile, in order to solve the over-fitting problem of the network, L2 regularization with a weight decay coefficient of 0.01 is used.
The invention takes PauliRGB image and 10 polarized characteristic images as input data of double channels respectively, cuts the input data, takes the neighborhood of each pixel as a sample for providing polarized characteristic information and space information respectively, and takes the pixel category as a label. Then, the processed geometrical space sample data set and the polarized characteristic data set and the sample label set are divided into a training set and a testing set.
The invention divides the characteristic image into patch input networks with the size of 16 multiplied by 16, and the ratio of the training set to the testing set is 8:2. The data used are two groups of data, the first data is oil spill and sea water, and the second data image contains oil films of crude oil, emulsified oil and vegetable oil of 3 different types, so that the training set of the data 1 is set to be less than the training set of the data 2. The learning rate and the batch size are key parameters for deep learning. The learning rate has great influence on the training effect, and the improper setting of the learning rate can lead to slow divergence or convergence of the network. The invention sets the learning rate to be 1 multiplied by 10 through experimental comparison -4 . The Batchsize refers to the number of training samples used per iteration in the training process. Its size significantly affects the progress of model optimization, and moreover, properly setting the network parameters of the batch size can make the gradient descent direction more accurate. Considering the size of the training set and the GPU platform used in the present invention, the batch size is set to 20. In order to accelerate the convergence rate of the network, the invention adopts an Adam optimizer to optimize the network and update parameters, and sets the maximum iteration number to 100. Parameters such as the learning rate of the network and the batch processing size are all adjusted through experiments so as to obtain the optimal classification performance. The framework version and the hardware and software configuration including the computing platform used in the present invention are shown in table 3.
TABLE 3 Table 3
Embodiment two:
(1) Introduction of data
According to the invention, two-scene Radarsat-2 full-polarization SAR oil spill data are selected for experiment. Wherein data 1 is an image of spilled oil acquired in the gulf of mexico on 5 months 8 days 2015, as shown in fig. 5 (a). The data resolution is 4.7mX4.8m, the coverage area is 32.95km X23.2 km, and the data format is single vision complex data (SLC). Data set 2 is a Radarsat-2 image of an offshore oil spill experiment in north europe taken on month 6 and 8 2011, as shown in fig. 5 (b). The image contains 3 different types of oil films of crude oil, emulsified oil and vegetable oil, the coverage area is 37.59km multiplied by 15.95km, and the resolution is 4.7mmultiplied by 4.8m. The detailed imaging parameters are shown in table 4. Note that, since data 1 only contains seawater and crude oil, and the number of types is smaller than that of data 2, data 1 training set is set to be smaller than data 2.
TABLE 4 Table 4
(2) Analysis of results
In order to test the performance of the polarized SAR oil spill detection algorithm based on Dual-EndNet, the invention adopts the overall accuracy, the average classification accuracy, the Kappa coefficient, the F1-fraction and the average cross-over ratio to evaluate the oil spill detection result. And comparing the new algorithm with SVM, CNN, U-Net and FCN 4 algorithms, wherein each algorithm respectively takes the PualiRGB image, the F10 feature, the F30 feature and the FP feature image after fusion of the F10 feature and PauliRGB as input data for experiments.
a. Dataset 1 results analysis
Table 5 gives the importance scores for the 30 polarization features extracted by dataset 1. The first 10 features with importance scores from large to small are respectively the maximum feature value, the average intensity, the homopolar cross-term real part, the total polarized power, the VV intensity, the geometric intensity, the surface scattering score, the polarized entropy, the Gini coefficient and the h_a12 combination parameter.
TABLE 5
Fig. 6 shows the visual results of the oil spill detection of the oil spill data set 1 using the present method and the comparison algorithm, and its corresponding PauliRGB image and ground truth label map. As is apparent from fig. 6, the method has the advantages of almost no specks on the sea surface, higher purity and greatly reduced probability of sea water misclassification compared with other comparison algorithms in visual effect. The result of the SVM algorithm visually presents many clutter compared to the deep learning algorithm. The reason is that the SVM algorithm does not use the space information of the image, only uses the polarization information, so that a plurality of misclassification phenomena exist on the sea surface, and more speckles appear in the classification result diagram. According to the method, through fusion of the F10-EndNet and the P-EndNet, polarization information specific to the polarized SAR is considered, spatial information of an image is extracted, and the accuracy of oil spill detection is effectively improved.
For quantitative comparison of the ability of the present invention to detect spilled oil, table 6 lists the accuracy of the spilled oil detection of the present invention and the comparative method on data set 1. As can be seen from the table, the invention is higher than other algorithms in the five indexes of OA, AA, kappa coefficient, F1 fraction and MIoU, namely 96.36%, 96.38%, 0.9272, 0.9636 and 0.9297 respectively. The OA is improved by 4.77% as compared with other algorithms, the AA is improved by 4.78% as compared with other algorithms, the F1 fraction is improved by 4.77% as compared with other algorithms, the MIoU is improved by 8.47%, and the Kappa coefficient is improved by 9.53% as compared with other algorithms. The oil film detection precision of the method provided by the invention reaches 95.78%, the seawater precision reaches 96.98%, and the method is superior to other comparison algorithms. In addition, in all comparison algorithms, the F10 characteristic obtains better experimental results than the F30 characteristic, and the preferred effectiveness of the characteristic of the invention is proved. And the performance of Dual-EndNet is superior to the experimental results of FP fusion characteristics of other algorithms, and the design of the invention as a Dual-channel network model is proved to be feasible.
TABLE 6
b. Dataset 2 results analysis
A greater variety of oil films are contained in dataset 2 than in dataset 1: crude oil, emulsified oil and vegetable oil, and various types of oil films are also more difficult to distinguish. The experiment utilizes vegetable oil to simulate a biological oil film so as to test the detection capability of the algorithm to the biological oil film. The experiment is mainly used for verifying the distinguishing capability of Dual-EndNet on crude oil, emulsified oil and biological oil films.
Table 7 gives the importance scores for the 30 polarization features extracted for data 2. It can be seen from the table that the top 10 features of the importance score of data 2 are the maximum feature value, the average intensity, the VV intensity, the SERD, the surface scattering score, the real part of the homopolar cross term, the polarization entropy, the geometric intensity, the Gini coefficient and the substrate height, respectively.
TABLE 7
Data 2 oil spill test results are shown in fig. 6. As can be clearly seen from the figure, the sea level in the experimental result of the method provided by the invention is relatively pure, can obviously distinguish crude oil, vegetable oil and emulsified oil, and has high detection precision. The result of using the SVM method shows that the overall visual effect of the classification result is poor, a plurality of specks exist on the sea surface, crude oil and emulsified oil cannot be effectively distinguished, and a plurality of crude oil and emulsified oil are mistakenly separated into biological oil films. Compared with SVM, the traditional deep learning algorithm has obvious distinction between three oil films, solves the problem that crude oil and emulsified oil are mistakenly separated into biological oil films to a certain extent, but still has more miscellaneous spots. The F10-FCN can better distinguish three oil films due to the input of multi-channel polarization characteristics, but more miscellaneous spots exist on the sea surface, so that the accuracy of oil spill detection is affected. The Dual-EndNet algorithm has the advantages of double channels, not only can better identify three types of oil films, but also has fewer mixed spots on the whole and highest visual purity, better solves the problems, and further improves the accuracy of oil spill detection.
Table 8 lists the accuracy of the algorithm of the present invention and the comparison method for oil spill detection on data 2. As can be seen from the table, the algorithm of the present invention still performed best on data 2, where OA was 98.76%, AA was 95.34%, kappa coefficient was 0.8913, F1 score was 0.8788, MIoU was 0.7952. Of the comparison algorithms, the F10-FCN performed best and the P-SVM performed worst. Compared with F10-FCN, OA is improved by 0.6 percentage points, AA is improved by 4.3 percentage points, kappa coefficient and F1 percentage are respectively improved by 5 percentage points and 6.2 percentage points, MIoU is improved most obviously, and 8 percentage points are improved. Compared with the P-SVM, the OA of the algorithm is improved by 9.76%, the AA is improved by 40.26%, the F1 fraction is improved by 44.67%, the MIoU is improved by 43.35%, and the Kappa coefficient is improved by 49.24% to the maximum extent. Further examining the classification precision of sea water and various oil films, the classification precision of the method on the sea water reaches 99.04%, the crude oil is 94.32%, and the emulsified oil and the vegetable oil are 95.71% and 92.28%, respectively. Compared with other algorithms, the algorithm of the invention achieves the optimal performance and has obvious promotion. In conclusion, the method of the invention not only greatly reduces the misjudgment probability of the seawater, but also improves the distinguishing capability of crude oil, emulsified oil and biological oil film.
TABLE 8
In summary, the invention firstly carries out Pauli decomposition on polarized SAR data, and synthesizes PauliRGB images by utilizing the scattering energy of odd scattering and two even scattering obtained after decomposition. Meanwhile, extracting 30 polarization features commonly used for oil spill detection, and selecting 10 polarization features with higher importance for oil spill detection through a random forest-based feature selection algorithm. And respectively taking the PauliRGB image and the 10 polarized characteristic images as input data of the double channels, cutting the input data, taking the neighborhood of each pixel as a sample for providing polarized characteristic information and spatial information, and taking the pixel type as a label. Then dividing the processed geometric space sample data set and the polarized characteristic data set and the sample label set into a training set and a testing set; finally, the space information and the polarization information of the polarized SAR data are respectively extracted by utilizing the constructed Dual-EndNet, the performance of oil spill detection is improved by fusing the two information, and the network model is trained and optimized by utilizing the data; and finally, detecting the test set by using the trained model. Experimental results show that the algorithm has strong oil spill detection capability, can distinguish sea water from oil spill, and can effectively improve the detection capability of different oil films of crude oil, emulsified oil and biological oil films.
The above embodiments are merely illustrative of the preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, but various modifications and improvements made by those skilled in the art to which the present invention pertains are made without departing from the spirit of the present invention, and all modifications and improvements fall within the scope of the present invention as defined in the appended claims.

Claims (6)

1. The oil spill detection method based on SAR and Dual-EndNet is characterized by comprising the following steps:
acquiring a full polarization SAR image of a target sea surface;
pauli decomposition is carried out on the full-polarization SAR image, so that scattering energy of odd scattering and even scattering is obtained, and the PauliRGB image is synthesized by utilizing the scattering energy of the odd scattering and even scattering;
extracting the polarization characteristics of oil spill detection from the full-polarization SAR image, and selecting the polarization characteristics by utilizing a random forest algorithm to obtain a polarization characteristic image;
based on the PauliRGB image and the polarization feature image, a geometric space sample data set, a polarization feature data set and a sample label set are obtained, and the geometric space sample data set, the polarization feature data set and the sample label set are divided into a training set and a testing set;
constructing a Dual-EndNet network model;
the Dual-EndNet network model consists of two channels of encoding-decoding devices, wherein the encoding-decoding devices comprise two channels of F10-EndNet and P-EndNet, 10 polarized characteristic images and PauliRGB images are respectively used as input data of the two channels, and polarized information and spatial characteristics are respectively extracted from the input data; then stacking the extracted polarization information and the spatial features together, and performing convolution and Softmax classifier to realize final prediction;
in the encoder stage, a total of 6 convolutional layers and 2 pooling layers are provided:
for the first four convolution layers, adding one maximum pooling layer into one group for every two convolution layers, and dividing the two groups, wherein the convolution kernel size of each convolution layer is 3 multiplied by 3, and the step length is 1; the parameter settings of the step length and the convolution kernel size of the last two convolution layers are consistent with those of the first four convolution layers;
the core size of the maximum pooling layer is 2×2, and the step size is 2;
in the decoder stage, 2 upsampling layers and 4 convolution layers are designed in total, symmetrically to the encoder stage:
each up-sampling layer is connected with two convolution layers, and relevant parameters are the same as the arrangement of symmetrical parts of the convolution stage;
stacking the feature images before the pooling layer together by utilizing jump cascade after up-sampling and inputting the feature images into a convolution layer together;
training the Dual-EndNet network model by using the training set;
and finishing the oil spill detection of the target sea surface by using the test set and the trained Dual-EndNet network model.
2. The SAR and Dual-EndNet based oil spill detection method according to claim 1, wherein the method of Pauli decomposition of the fully polarized SAR image comprises:
wherein ,Srepresenting a polarized scattering matrix, elements in the matrix +.>Representing the respective polarization modes->Emission and polarization mode->The received complex scattering coefficients, a, b, c, d are all complex, -/-, and>。
3. the SAR and Dual-end based oil spill detection method of claim 1, wherein the even scattering comprises: secondary scattering at a 0 ° direction angle and secondary scattering at a 45 ° direction angle.
4. The SAR and Dual-EndNet based oil spill detection method of claim 1, wherein the method of selecting the polarization feature using a random forest algorithm comprises:
and analyzing the capacity of distinguishing oil spill, oil film like and sea water of each oil spill detection polarization characteristic based on a random forest feature selection method, and selecting optimal features from the capacity to form a feature combination.
5. The SAR and Dual-EndNet based oil spill detection method according to claim 1, wherein the specific process of selecting the polarization feature using a random forest algorithm comprises:
assuming that the number of decision trees in a random forest algorithm is T, d features are in samples, and an original sample set isThe sample set capacity is n;
extracting T sample sets from the sample set D by using a Bootstrap self-help sampling method, wherein the number of samples in each sample set is n;
respectively training T decision tree models by using the extracted T training sets with the size of n;
by calculating characteristics，j=1，2...，dCalculating the feature +.>Importance score of (2);
and accumulating the importance scores of the features obtained by each iteration through multiple iterations, finally obtaining importance scores of the features after multiple iterations, sorting all the features according to the importance scores after the iterations, and selecting a plurality of features with the importance scores sorted in front.
6. The SAR and Dual-EndNet based oil spill detection method according to claim 5, wherein the characteristic is calculated，j=1，2...，dCalculating the feature +.>The method of importance score of (2) comprises:
s1: calculate the firstiOut-of-bag data of decision treeOut-of-bag data error->；
S2: for out-of-bag dataInternal features->Adding noise disturbance to obtain->Recalculating the out-of-bag data->Out-of-bag data error->；
S3: repeating said S1 and said S2 until obtaining and />；
S4: computing all decision tree featuresMean value of the change in the data error outside the bag before and after substitution:
in the formula ,/>Is characterized by->Importance score of (c).
CN202310217615.9A 2023-03-08 2023-03-08 Oil spill detection method based on SAR and Dual-EndNet Active CN116206203B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310217615.9A CN116206203B (en) 2023-03-08 2023-03-08 Oil spill detection method based on SAR and Dual-EndNet

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310217615.9A CN116206203B (en) 2023-03-08 2023-03-08 Oil spill detection method based on SAR and Dual-EndNet

Publications (2)

Publication Number Publication Date
CN116206203A CN116206203A (en) 2023-06-02
CN116206203B true CN116206203B (en) 2023-08-18

Family

ID=86511062

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310217615.9A Active CN116206203B (en) 2023-03-08 2023-03-08 Oil spill detection method based on SAR and Dual-EndNet

Country Status (1)

Country Link
CN (1) CN116206203B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116894999B (en) * 2023-07-18 2024-05-03 中国石油大学(华东) Method and device for detecting oil spill polarization SAR based on condition parameterized convolution

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105718957A (en) * 2016-01-26 2016-06-29 西安电子科技大学 Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network
CN107180434A (en) * 2017-05-23 2017-09-19 中国地质大学（武汉） Polarization SAR image segmentation method based on super-pixel and fractal net work evolution algorithmic
CN108846338A (en) * 2018-05-29 2018-11-20 南京林业大学 Polarization characteristic selection and classification method based on object-oriented random forest
CN110110813A (en) * 2019-05-20 2019-08-09 西安邮电大学 Classification of Polarimetric SAR Image method based on binary channels convolutional network
CN111025291A (en) * 2019-11-06 2020-04-17 中国石油大学(华东) Ocean oil spill detection method based on new characteristics of fully-polarized SAR
CN114241402A (en) * 2021-11-15 2022-03-25 齐鲁空天信息研究院 Sea surface oil spill detection method and device, electronic equipment and storage medium
CN114882371A (en) * 2022-07-08 2022-08-09 中国石油大学(华东) Sea surface oil spill detection model construction method based on full-polarization SAR image

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105718957A (en) * 2016-01-26 2016-06-29 西安电子科技大学 Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network
CN107180434A (en) * 2017-05-23 2017-09-19 中国地质大学（武汉） Polarization SAR image segmentation method based on super-pixel and fractal net work evolution algorithmic
CN108846338A (en) * 2018-05-29 2018-11-20 南京林业大学 Polarization characteristic selection and classification method based on object-oriented random forest
CN110110813A (en) * 2019-05-20 2019-08-09 西安邮电大学 Classification of Polarimetric SAR Image method based on binary channels convolutional network
CN111025291A (en) * 2019-11-06 2020-04-17 中国石油大学(华东) Ocean oil spill detection method based on new characteristics of fully-polarized SAR
CN114241402A (en) * 2021-11-15 2022-03-25 齐鲁空天信息研究院 Sea surface oil spill detection method and device, electronic equipment and storage medium
CN114882371A (en) * 2022-07-08 2022-08-09 中国石油大学(华东) Sea surface oil spill detection model construction method based on full-polarization SAR image

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于改进FCN 的多极化SAR 影像海上溢油检测;杜艳玲等;《激光与光电子学进展》;第59卷(第4期);第321-328页 *

Also Published As

Publication number Publication date
CN116206203A (en) 2023-06-02

Similar Documents

Publication Publication Date Title
CN110533631B (en) SAR image change detection method based on pyramid pooling twin network
CN112700429B (en) Airport pavement underground structure disease automatic detection method based on deep learning
CN106248801B (en) A kind of Rail crack detection method based on more acoustie emission event probability
Xu et al. A comparative study of different classification techniques for marine oil spill identification using RADARSAT-1 imagery
Mishra et al. Land cover classification of PALSAR images by knowledge based decision tree classifier and supervised classifiers based on SAR observables
CN110766051A (en) Lung nodule morphological classification method based on neural network
CN110555841B (en) SAR image change detection method based on self-attention image fusion and DEC
CN101853509A (en) SAR (Synthetic Aperture Radar) image segmentation method based on Treelets and fuzzy C-means clustering
CN105718942A (en) Hyperspectral image imbalance classification method based on mean value drifting and oversampling
CN116206203B (en) Oil spill detection method based on SAR and Dual-EndNet
CN115035361A (en) Target detection method and system based on attention mechanism and feature cross fusion
CN112699717A (en) SAR image generation method and generation device based on GAN network
Nurmasari et al. Oil palm plantation detection in Indonesia using Sentinel-2 and Landsat-8 optical satellite imagery (case study: Rokan Hulu regency, Riau Province)
CN106097290A (en) SAR image change detection based on NMF image co-registration
CN109766926A (en) A kind of radar emitter signal intrapulse feature comprehensive estimation method and system
CN107609577B (en) Method for extracting polarized SAR sea surface oil film by using random forest
CN115601572A (en) Ultrasonic phased array image optimization reconstruction method and system based on semi-supervised CycleGan network
Ge et al. Azimuth-Sensitive Object Detection in Sar Images Using Improved Yolo V5 Model
Mogushi et al. NNETFIX: an artificial neural network-based denoising engine for gravitational-wave signals
Wu et al. Object detection and X-ray security imaging: A survey
CN112989940B (en) Raft culture area extraction method based on high-resolution third satellite SAR image
Chen et al. Regional classification of urban land use based on fuzzy rough set in remote sensing images
CN105046286A (en) Supervision multi-view feature selection method based on automatic generation of view and unit with l1 and l2 norm minimization
CN113298007A (en) Small sample SAR image target identification method
CN115456957B (en) Method for detecting change of remote sensing image by full-scale feature aggregation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant