CN114897814B - Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network - Google Patents
Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network Download PDFInfo
- Publication number
- CN114897814B CN114897814B CN202210494278.3A CN202210494278A CN114897814B CN 114897814 B CN114897814 B CN 114897814B CN 202210494278 A CN202210494278 A CN 202210494278A CN 114897814 B CN114897814 B CN 114897814B
- Authority
- CN
- China
- Prior art keywords
- network
- oil spill
- training
- parameters
- hyperspectral image
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10036—Multispectral image; Hyperspectral image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20064—Wavelet transform [DWT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
Abstract
The invention provides a hyperspectral image oil spill detection method based on a multistage wavelet decomposition close-connected network, which comprises three steps of image preprocessing, network model training and oil spill detection test, wherein the acquired hyperspectral image data is preprocessed to avoid the influence of abnormal spectral bands on the oil spill detection performance, wavelet transformation is fused into a convolutional neural network structure, so that an MLWBDN network has the functions of frequency domain decomposition and multi-resolution analysis, the method has the advantages that the fine recognition capability of a network on a complex oil spilling area is enhanced, an improved SADFB module is provided by combining wavelet transformation characteristics and a classical Dense Block structure, the SADFB realizes wavelet multi-frequency component fusion by utilizing close connection operation, network parameters are reduced through a continuous attachment strategy, gradient disappearance is avoided through feature multiplexing, a multi-level feature joint decision mechanism of an MLWBDN network is benefited, and detection tasks of different scales can be dealt with by adjusting the number of branch structures.
Description
Technical Field
The invention relates to the technical field of hyperspectral oil spill detection, in particular to a hyperspectral image oil spill detection method based on a multistage wavelet decomposition close-contact network.
Background
The marine oil spill is marine environmental pollution caused by the reasons of pipeline rupture, oil tanker collision, drilling platform explosion and the like, has the characteristics of high occurrence frequency, large hazard degree and wide distribution area, is valued by various countries in the world, and has the petroleum pollution exceeding 10000 tons in each year in the world since 1970, wherein the catastrophic consequences brought by serious marine oil spill pollution events lead people to be eye-catching, not only seriously damage the balance of a marine ecosystem, but also damage the production and life of coastal residents, and is particularly important for effectively making oil spill emergency decisions and carrying out accident treatment, rapidly acquiring the oil spill information on the ocean surface and accurately carrying out oil spill detection. In the put-in monitoring system, satellite remote sensing is one of the most important and effective means, plays an irreplaceable role in offshore oil spill monitoring and emergency response, and at present, the most commonly used satellite remote sensing sensors comprise visible light, infrared, ultraviolet, microwave sensors, hyperspectral sensors and the like, wherein hyperspectral has great potential in the oil spill detection field due to the advantages of integration of 'graph and spectrum', high spectral resolution and wide spectral information range;
although hyperspectral images have rich spectral bands, the problems of increased data volume, high correlation among bands, reduced classification precision, reduced efficiency and the like are brought about along with the increase of the dimension of hyperspectral images, the current method for detecting hyperspectral oil spill at home and abroad mainly comprises a spectral index algorithm, a maximum likelihood classifier, a classification regression tree (CART), an Extreme Learning Machine (ELM), a Support Vector Machine (SVM), a Random Forest (RF) and the like, the method is mainly based on the spectral characteristics of the hyperspectral images, is easy to fall into the problem of dimension disaster, an ideal detection result is difficult to obtain, and is mainly based on a manually selected characteristic shallow model, only low-layer information is utilized, and in recent years, a deep learning method is paid more and more attention as a tool capable of extracting different layers of characteristics.
However, the existing spectrum index algorithm is simple in principle and easy to implement, but in terms of actual operation, the situation of oil spill and sea water misclassification is commonly existed, and is easily influenced by various factors, most of the machine learning methods only use the spectrum characteristics of data, the manual extraction characteristics are complex and time-consuming, the extracted characteristics are possibly not optimal characteristics required by a designed classifier, the existing deep learning-based method is relatively less, better results are obtained compared with the traditional method, but because of the complex environment of oil film diffusion and sea surface, the oil film in the remote sensing image always presents irregular shape distribution, the existing deep learning method is more suitable for the ground object in a regular area in extraction of space information, and suffers from the problem of network gradient disappearance, the number of layers of the existing network is relatively less, the shallow network cannot fully exert the advantages of the deep learning algorithm, and therefore the performance of the deep learning algorithm is not ideal on the overflow oil detection task, and the hyperspectral image oil detection method based on the multistage wavelet decomposition close-contact network is provided to solve the problems existing in the prior art.
Disclosure of Invention
Aiming at the problems, the invention aims to provide a hyperspectral image oil spill detection method based on a multistage wavelet decomposition close-contact network, which has the advantages of constructing a multiscale branch structure by utilizing multistage wavelet decomposition, extracting a multi-frequency domain and multi-resolution characteristic for fine oil spill detection and solving the problem that the prior art is not ideal in oil spill detection task.
In order to achieve the purpose of the invention, the invention is realized by the following technical scheme: a hyperspectral image oil spill detection method based on a multistage wavelet decomposition close-contact network comprises the following steps:
step one, image preprocessing
Acquiring original oil spill hyperspectral image data, carrying out normalization and dimension reduction on the acquired original oil spill hyperspectral image data to obtain processed oil spill hyperspectral image sample data, and then carrying out format conversion on the obtained sample data to convert the sample data into a format input by a model;
step two, training a network model
Firstly, determining structural parameters of a network, constructing an MLWBDN network model according to the structural parameters, wherein the structural parameters comprise a network input Patch size, a network branch number, a classification class number, a compression rate, a Dropout value and a growth rate, secondly, determining super parameters of the network, wherein the super parameters comprise model training configuration and an input data format, initializing the parameters of the established network, training the network model according to the set super parameters after the initialization is finished, inputting samples in a training set into the network in a batch mode in the training process, obtaining a predicted oil spill detection result through network forward calculation, inputting the predicted result and an actual label into a loss function calculation loss together, and then updating the network parameters through reverse propagation, wherein all training samples participate in training to finish a complete iteration;
step three, oil spill detection test
And (3) applying the optimal network parameters obtained by training in the step (II) to the oil spill data to be detected, inputting unknown types of samples to be detected into the trained network model, detecting the oil spill, and outputting a detection result.
The further improvement is that: in the first step, the abnormal spectrum band is avoided from affecting the oil spill detection performance, and all hyperspectral sample data are processed by adopting a maximum value normalization method, wherein the formula is as follows:
wherein I is norm Representing normalized spectral band values, I representing the original spectral sequence, I min And I max Representing the minimum and maximum values, respectively, in the sample spectral sequence.
The further improvement is that: in the first step, a spilled oil data set is formed by processed spilled oil hyperspectral image sample data, and the spilled oil data set is divided into a training set, a verification set and a test set.
The further improvement is that: in the second step, the structural parameters include the size of the network input Patch, the number of network branches, the number of classification categories, the compression rate, the Dropout value and the growth rate, and the super parameters include the learning rate, the iteration times, the optimizer selection, the batch processing number and the loss function, wherein the loss function is a cross entropy loss function of classification.
The further improvement is that: in the second step, the network model updates the parameters towards the direction of minimizing the loss by using a random batch gradient descent algorithm, so that the network loss is continuously reduced until the loss converges or the maximum iteration number is reached, and the training is finished.
The further improvement is that: in the second step, the MLWBDN network is a multi-branch joint classification network as a whole, and includes a multi-frequency feature decomposition module MFFDB, a continuous attachment fusion module SADFB, and a multi-stage feature joint decision mechanism.
The further improvement is that: in the second step, after every other iteration, the network parameters are fixed and stored, then the verification set is used for testing the network oil spill detection precision, and after training is finished, the training model parameters with the best performance in the verification set are the optimal network parameters.
The further improvement is that: in the third step, in the verification experiment, the trained network model is tested through the test set, so that performance evaluation is performed on the network through the known label.
The beneficial effects of the invention are as follows: according to the hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-connected network, wavelet transformation is integrated into a convolutional neural network structure, so that an MLWBDN network has frequency domain decomposition and multi-resolution analysis capability, the fine recognition capability of the network on a complex oil spill region is enhanced, an improved SADFB module is provided by combining wavelet transformation characteristics and a classical Dense Block structure, the SADFB realizes wavelet multi-frequency component fusion by utilizing close-connected operation, network parameters are reduced through a continuous attachment strategy, gradient disappearance is avoided through feature multiplexing, and the multistage feature joint decision mechanism of the MLWBDN network is benefited.
Drawings
FIG. 1 is a schematic flow chart of the steps of the present invention.
Fig. 2 is a schematic diagram of the overall flow of oil spill detection according to the present invention.
Fig. 3 is a schematic diagram of the MLWBDN network model of the present invention.
FIG. 4 is a schematic diagram of classical Dense Block (right) and SADFB (left) structures of the present invention.
FIG. 5 is a schematic diagram of a three-level feature joint decision structure of the present invention.
Detailed Description
The present invention will be further described in detail with reference to the following examples, which are only for the purpose of illustrating the invention and are not to be construed as limiting the scope of the invention.
According to the embodiments shown in fig. 1 to 5, a hyperspectral image oil spill detection method based on a multistage wavelet decomposition close-contact network is provided, which comprises the following steps:
step one, image preprocessing
The method comprises the steps of obtaining original oil spill hyperspectral image data, carrying out normalization and dimension reduction on the obtained original oil spill hyperspectral image data to obtain processed oil spill hyperspectral image sample data, carrying out format conversion on the obtained sample data, converting the obtained sample data into a format input by a model, avoiding the influence of abnormal spectrum wave bands on the oil spill detection performance, and processing all hyperspectral sample data by adopting a maximum value normalization method, wherein the formula is as follows:
wherein I is norm Representing normalized spectral band values, I representing the original spectral sequence, I min And I max Representing the minimum value and the maximum value in a sample spectrum sequence respectively, simultaneously, carrying out spectrum dimension characteristic Principal Component Analysis (PCA) on normalized data, converting the original high-correlation upper hundred dimension spectrum characteristic into mutually uncorrelated characteristic by the PCA through orthogonal transformation, and retaining the partial component which contains the original information most, namely the principal component, so as to achieve the purpose of reducing dimension but not reducing the data expression capability, wherein the hyperspectral data after dimension reduction cannot be directly input into a network, because, in order to comprehensively extract space-spectrum information, a sample needs to be input in a Patch form, namely, a single pixel is used as a center, space cube data with a specified size is used as a supervised learning method, the input sample of the network model in the embodiment is Patch cube data with the pixel as the center, the corresponding label is the category of the center pixel, and in addition, the edge sample without enough space neighborhood needs to be subjected to data filling;
before formally training a network model, forming an oil spill data set by processed oil spill hyperspectral image sample data, and dividing the oil spill data set into a training set, a verification set and a test set, wherein the training set is used for training neural network parameters, the verification set is used for selecting optimal network parameters, the test set is a sample to be classified and used for evaluating the actual performance of a network, and the ratio of the three data sets is 1:1:8;
step two, training a network model
Firstly, determining structural parameters of a network, constructing an MLWBDN network model according to the structural parameters, wherein the structural parameters comprise a network input Patch size, a network branch number, a classification category number, a compression rate, a Dropout value and a growth rate, secondly, determining structural parameters and super parameters of the network model, wherein the super parameters comprise a learning rate, iteration times, an optimizer selection, a batch processing number and a loss function, the loss function is a classified cross entropy loss function, and the specific formula of the cross entropy loss function is as follows:
wherein N is the number of test samples, M is the number of classes, y ic As a sign function, if the true class of the sample i is c, y ic 1, otherwise 0.P is p ic The probability of the model prediction sample i category being c is represented, the cross entropy can measure the difference degree of two different probability distributions in the same random variable, and the difference is represented as the difference between the true probability distribution and the prediction probability distribution in the classification task. The smaller the cross entropy loss value is, the better the prediction classification result of the network is;
after the network construction is completed, initializing network parameters, wherein the parameters are set to be random numbers between 0 or 0 and 1, training a network model after the initialization is completed, in the training process, inputting samples in a training set into the network in batches in a Patch mode, obtaining a predicted oil spill detection result through network forward calculation, inputting the predicted result and an actual tag into a loss function to calculate loss, updating the network parameters through reverse propagation, wherein all training samples participate in training to complete one complete iteration, updating the parameters towards the direction of minimizing the loss by a random batch gradient descent algorithm, enabling the network loss to be continuously reduced until the loss converges or the maximum iteration number is reached, simultaneously, in order to obtain the optimal network parameters, fixing and storing the network parameters after every other iteration, testing the network oil spill detection precision by using a verification set, and after the training is completed, obtaining the training model parameters with the best performance in the verification set as the optimal network parameters;
step three, oil spill detection test
And (3) applying the optimal network parameters obtained by training in the step (II) to the oil spill data to be detected, inputting unknown types of samples to be detected into a trained network model, detecting the oil spill, outputting a detection result, and testing the trained network model through a test set in a verification experiment so as to evaluate the performance of the network through a known label.
In this embodiment, as shown in fig. 2, the MLWBDN network is a multi-branch joint classification network as a whole, which includes a multi-frequency feature decomposition module MFFDB, a continuous attachment fusion module SADFB and a multi-stage feature joint decision mechanism, two-dimensional wavelet transformation is introduced into the network through the multi-frequency feature decomposition module MFFDB, and the extracted spatial features are transformed into the wavelet domain, then the multi-frequency spatial features and the spectral features are fused through the continuous attachment fusion module SADFB, the multi-stage decomposition structure and the multi-stage feature joint decision mechanism are adopted by the model to complete joint oil spill detection by integrating the multi-stage decomposition structure and the multi-stage feature joint decision mechanism, and the number of branches of the MLWBDN network is an adjustable structural parameter corresponding to the multi-stage analysis capability of the network. Therefore, the model can adapt to various requirements under actual conditions.
For the Multi-frequency feature decomposition module (Multi-Frequency Feature DecompositionBlock) MFFDB, the function of the Multi-frequency feature decomposition module is to perform channel-by-channel two-dimensional discrete wavelet decomposition on the input or the feature map extracted in the previous stage, and perform feature fusion or compression on the decomposed wavelet features respectively. Taking a characteristic diagram X with the size of 2W multiplied by 2H multiplied by C as input, obtaining four sub-band components of which the high-low frequency information is mutually combined after the X passes through 2D-DWT, wherein the four sub-band components are respectively a low-frequency approximate sub-band LL, three high-frequency detail characteristic sub-bands HL, LH and HH, the sizes of the four sub-band characteristic diagrams are half of the original characteristic diagram, and the number of channels is unchanged, namely W multiplied by H multiplied by C. The four frequency combination features are then each fed into a transmission layer. The Transition layer is composed of 1×1 convolution, and is mainly used for carrying out spectral dimension feature fusion on the decomposed feature map, and controlling the dimension of the output feature channel through the compression rate lambda, as shown in a multi-frequency feature decomposition module in fig. 3. After the Transition layer, the number of { LL, HL, LH, HH } channels changes from C to λC. In addition λ at the primary MFFDB is set to 1 to avoid information loss caused by recompression of the reduced-size data.
For a continuous attachment fusion module (Successively Attached Dense Fusion Block) SADFB, the function of the module is to fuse the space frequency characteristics and the spectrum characteristics of the decomposed wavelet components, extract deep characteristics for classification decision, realize characteristic multiplexing through close-fitting operation, ensure characteristic diversity, slow down gradient disappearance, and assume that the size of { LL, HL, LH, HH } obtained after MFFDB is W×H×K 0 Dense Block requires that they be concatenated as input, so that the input becomes WXH 4K 0 In order to reduce the parameter surge caused by the channel dimension increase, a continuous attachment strategy is used, since the approximation component LL retains most of the information of the original image, we use LL as the initial input instead of all components after concatenation, and in each subsequent bonding process, in addition to the output of each convolution layer before addition, additional wavelet components are added successively in HL, LH, HH order, classical Dense Block and SADFB Block pair of module structures such as shown in FIG. 4. Each of the close-coupled layer packets in the SADFBThe method comprises a bottleneck layer and a convolution layer, wherein the sizes of the bottleneck layer and the convolution layer are respectively 1 multiplied by 4K and 3 multiplied by K, the convolution is formed by connecting batch normalization layer BN and LeakyRelu activation functions, K represents growth rate parameters (growth rate), and the number of input channels of the close-coupled layers in SADFB and Dense Block is as follows:
wherein the method comprises the steps ofInput channel number representing the ith bonding layer of SADFB module,/for the bonding layer>The input channel number of the ith bonding layer of the Dense Block module is represented, and since the SADFB comprises 4 bonding layers, the strategy can reduce 6K 0 The number of channels, experiments show that in practical cases the continuous attachment strategy can reduce the parameter amount of a single Dense Block by 6% to 21%, while in theory, this approach can reduce the parameter by more than 30% when K0 > K.
For a multi-Level feature joint decision mechanism, a multi-Level branch structure is established in a network to process different scale features extracted after multiple times of wavelet decomposition, and inter-Level feature fusion and classification are performed at the same time, as shown in fig. 5, a Level1, a Level2 and a Level3 feature map extracted through an MFFDB and an SADFB enter respective branched classifiers to complete classification processes of corresponding scales. The classifier consists of a 3x3 size depth separable convolution and a 1 x 1 convolution for depth feature extraction. Wherein the depth separable convolution is to decompose the normal two-dimensional convolution operation into a point-wise convolution (Pointwise Conv) and a channel-wise convolution (Depthwise Conv) to reduce network parameters. Because the feature graphs of different scales are different in size after multi-stage decomposition, the classifier adopts global average pooling (global average pooling) GAP to map the feature graphs into one-dimensional feature vectors after a convolution layer, and Dropout is added to enhance model generalization. And finally, connecting the three branch feature vectors in series, and outputting a classification result after full connection layer and softmax operation. It is noted that in order to enhance information interaction and gradient transfer between different hierarchical features, the network establishes a hierarchical short circuit (level-wise short cut) before inputting the features into the classifier, so that a high-level gradient can be propagated into a low-level structure to realize multi-resolution information exchange, and in addition, by changing the number of branches in the network, an MLWBDN network focusing on different scales can be constructed.
In this embodiment, in the stage of verifying the validity of the algorithm, the data of monitoring the oil spill on the sea in the aviation hyperspectral Dalian of 24 days of 7 months of 2010 are taken as an example. The sensor is a portable airborne spectrum imager AISA eagle produced by Finland, the wavelength is 400-970 nm,258 spectrum bands, and the spatial resolution is 1.41m. The data has been subjected to geometric and radiation corrections of the system. Because of the large data volume, a rectangular area with the size of 350 multiplied by 360 is cut out as the experimental data set. Images are herein classified into thick oil film, thin oil film and sea water according to colors exhibited by the spilled images in the RGB bands (R: 645.13nm, G:554.44nm, B:469.05 nm) and the Bonn protocol. In the data set, each type of data in the data set is randomly divided into a training set, a verification set and a test set according to a ratio of 1:1:8.
The available sample distribution for each class of the spilled data set for training and testing is listed by table one, which is shown below:
table-available sample distribution of oil spill data set
And compared with the following four methods: in order to examine the performance of different methods, random Forest (RF), support Vector Machine (SVM), neural Network (NN) and LeNet-5 (a convolution Network), we use various classification accuracy, average Accuracy (AA), overall Accuracy (OA) and Kappa coefficient as evaluation indexes. In the experiment, the MLWBDN model is set to be 2 in scale, the input Patch is 24 multiplied by 24, the learning rate is 0.002, the weight attenuation is 0.001, the total iteration is 50, the optimizer is Adam, and the batch processing number is 100. Compression λ=0.1, growth rate is 48. The experimental results are shown in a second table, and the average value of various classification precision, AA, OA and Kappa coefficients after each method is operated for 5 times on the oil spill data set is recorded from top to bottom, and the standard deviation is listed in the last three indexes, wherein the second table is as follows:
table two classification results of different methods on oil spill dataset
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (7)
1. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-contact network is characterized by comprising the following steps of: the method comprises the following steps:
step one, image preprocessing
Acquiring original oil spill hyperspectral image data, carrying out normalization and dimension reduction on the acquired original oil spill hyperspectral image data to obtain processed oil spill hyperspectral image sample data, and then carrying out format conversion on the obtained sample data to convert the sample data into a format input by a model;
step two, training a network model
Firstly, determining structural parameters of a network, constructing an MLWBDN network model according to the structural parameters, secondly, determining super parameters of the network, wherein the super parameters comprise model training configuration and input data format, initializing the parameters of the established network, training the network model according to the set super parameters after the initialization is completed, inputting samples in a training set into the network in batches in a Patch mode in the training process, obtaining a predicted oil spill detection result through forward calculation of the network, inputting the predicted result and an actual label into a loss function calculation loss together, and updating the parameters of the network through back propagation, wherein all training samples participate in training to complete a complete iteration; in the second step, the MLWBDN network is a multi-branch joint classification network as a whole, and includes a multi-frequency feature decomposition module MFFDB, a continuous attachment fusion module SADFB, and a multi-stage feature joint decision mechanism;
the multi-frequency feature decomposition module MFFDB carries out channel-by-channel two-dimensional discrete wavelet decomposition on the feature map extracted in the input or previous stage, and carries out feature fusion or compression on the decomposed wavelet features respectively;
the continuous attachment fusion module SADFB fuses the space frequency characteristics and the spectrum characteristics of the decomposed wavelet components, extracts deep characteristics for classification decision, realizes characteristic multiplexing through close connection operation, ensures characteristic diversity, and slows down gradient disappearance;
the multi-level feature joint decision mechanism establishes a multi-level branch structure to process different scale features extracted after multiple times of wavelet decomposition, and performs inter-level feature fusion and classification;
step three, oil spill detection test
And (3) applying the optimal network parameters obtained by training in the step (II) to the oil spill data to be detected, inputting unknown types of samples to be detected into the trained network model, detecting the oil spill, and outputting a detection result.
2. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network according to claim 1, wherein the method is characterized by comprising the following steps of: in the first step, the abnormal spectrum band is avoided from affecting the oil spill detection performance, and all hyperspectral sample data are processed by adopting a maximum value normalization method, wherein the formula is as follows:
wherein I is norm Representing normalized spectral band values, I representing the original spectral sequence, I min And I max Representing the minimum and maximum values, respectively, in the sample spectral sequence.
3. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network according to claim 1, wherein the method is characterized by comprising the following steps of: in the first step, a spilled oil data set is formed by processed spilled oil hyperspectral image sample data, and the spilled oil data set is divided into a training set, a verification set and a test set.
4. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network according to claim 1, wherein the method is characterized by comprising the following steps of: in the second step, the structural parameters include the size of the network input Patch, the number of network branches, the number of classification categories, the compression rate, the Dropout value and the growth rate, and the super parameters include the learning rate, the iteration times, the optimizer selection, the batch processing number and the loss function, wherein the loss function is a cross entropy loss function of classification.
5. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network according to claim 1, wherein the method is characterized by comprising the following steps of: in the second step, the network model updates the parameters towards the direction of minimizing the loss by using a random batch gradient descent algorithm, so that the network loss is continuously reduced until the loss converges or the maximum iteration number is reached, and the training is finished.
6. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network according to claim 1, wherein the method is characterized by comprising the following steps of: in the second step, after every other iteration, the network parameters are fixed and stored, then the verification set is used for testing the network oil spill detection precision, and after training is finished, the training model parameters with the best performance in the verification set are the optimal network parameters.
7. The hyperspectral image oil spill detection method based on the multistage wavelet decomposition close-coupled network, which is characterized by comprising the following steps of: in the third step, in the verification experiment, the trained network model is tested through the test set, so that performance evaluation is performed on the network through the known label.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210494278.3A CN114897814B (en) | 2022-05-07 | 2022-05-07 | Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210494278.3A CN114897814B (en) | 2022-05-07 | 2022-05-07 | Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114897814A CN114897814A (en) | 2022-08-12 |
| CN114897814B true CN114897814B (en) | 2023-07-25 |
Family
ID=82722409
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210494278.3A Active CN114897814B (en) | 2022-05-07 | 2022-05-07 | Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114897814B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516596A (en) * | 2019-08-27 | 2019-11-29 | 西安电子科技大学 | Empty spectrum attention hyperspectral image classification method based on Octave convolution |
| CN112232280A (en) * | 2020-11-04 | 2021-01-15 | 安徽大学 | Hyperspectral image classification method based on self-encoder and 3D depth residual error network |
| CN113052216A (en) * | 2021-03-15 | 2021-06-29 | 中国石油大学(华东) | Oil spill hyperspectral image detection method based on two-way graph U-NET convolutional network |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105844298A (en) * | 2016-03-23 | 2016-08-10 | 中国石油大学(华东) | High spectral oil overflow image classification method based on Fuzzy ARTMAP neural network |
| CN109919123B (en) * | 2019-03-19 | 2021-05-11 | 自然资源部第一海洋研究所 | Sea surface oil spill detection method based on multi-scale feature deep convolution neural network |
| CN111191736B (en) * | 2020-01-05 | 2022-03-04 | 西安电子科技大学 | Hyperspectral image classification method based on depth feature cross fusion |
| CN111695469B (en) * | 2020-06-01 | 2023-08-11 | 西安电子科技大学 | Hyperspectral image classification method of light-weight depth separable convolution feature fusion network |
| CN112733659B (en) * | 2020-12-30 | 2022-09-20 | 华东师范大学 | Hyperspectral image classification method based on self-learning double-flow multi-scale dense connection network |
-
2022
- 2022-05-07 CN CN202210494278.3A patent/CN114897814B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110516596A (en) * | 2019-08-27 | 2019-11-29 | 西安电子科技大学 | Empty spectrum attention hyperspectral image classification method based on Octave convolution |
| CN112232280A (en) * | 2020-11-04 | 2021-01-15 | 安徽大学 | Hyperspectral image classification method based on self-encoder and 3D depth residual error network |
| CN113052216A (en) * | 2021-03-15 | 2021-06-29 | 中国石油大学(华东) | Oil spill hyperspectral image detection method based on two-way graph U-NET convolutional network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114897814A (en) | 2022-08-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107316013B (en) | Hyperspectral image classification method based on NSCT (non-subsampled Contourlet transform) and DCNN (data-to-neural network) | |
| CN107358260B (en) | Multispectral image classification method based on surface wave CNN | |
| CN111368825B (en) | Pointer positioning method based on semantic segmentation | |
| CN113052216B (en) | Oil spill hyperspectral image detection method based on two-way graph U-NET convolutional network | |
| CN111783884A (en) | Unsupervised hyperspectral image classification method based on deep learning | |
| Chen et al. | A rotation-invariant convolutional neural network for image enhancement forensics | |
| CN103500343A (en) | Hyperspectral image classification method based on MNF (Minimum Noise Fraction) transform in combination with extended attribute filtering | |
| CN111738052B (en) | Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning | |
| CN110689065A (en) | Hyperspectral image classification method based on flat mixed convolution neural network | |
| CN104809471A (en) | Hyperspectral image residual error fusion classification method based on space spectrum information | |
| CN114882368A (en) | Non-equilibrium hyperspectral image classification method | |
| CN116977747B (en) | Small sample hyperspectral classification method based on multipath multi-scale feature twin network | |
| CN114897814B (en) | Hyperspectral image oil spill detection method based on multistage wavelet decomposition close-coupled network | |
| CN116482618B (en) | Radar active interference identification method based on multi-loss characteristic self-calibration network | |
| CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
| CN116188981A (en) | Hyperspectral high-spatial-resolution remote sensing image classification method and device | |
| CN115984632A (en) | Hyperspectral plastic waste material rapid classification method and device and storage medium | |
| Albano et al. | Graph theoretic metrics for spectral imagery with application to change detection | |
| CN107832790A (en) | A kind of semi-supervised hyperspectral image classification method based on local low-rank representation | |
| CN114299382A (en) | Hyperspectral remote sensing image classification method and system | |
| CN114863293B (en) | Hyperspectral oil spill detection method based on double-branch GAN network | |
| CN112949771A (en) | Hyperspectral remote sensing image classification method based on multi-depth multi-scale hierarchical attention fusion mechanism | |
| CN115471677B (en) | Hyperspectral image classification method based on double-channel sparse network | |
| CN116188830B (en) | Hyperspectral image cross-domain classification method based on multi-level feature alignment | |
| CN115861824B (en) | Remote sensing image recognition method based on improved transducer |
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 |