Sea fog hyperspectral remote sensing image recognition method based on CNN
By introducing an attention mechanism into the CNN model, combining spectral and spatial features, the problem of insufficient feature extraction in sea fog hyperspectral remote sensing image recognition is solved, and the recognition accuracy and robustness are improved.
Patent Information
- Application Number
- CN202411922585.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-25
- Publication Date
- 2025-04-18
AI Technical Summary
When processing sea fog images, existing hyperspectral remote sensing image recognition methods cannot effectively extract the characteristics of complex land objects, ignore key spectral information, are noise-sensitive and improperly processed redundant features, resulting in poor recognition effect.
Using a CNN-based method, an attention mechanism is introduced to dynamically adjust the feature layer weights during the feature extraction stage, combining spectral and spatial characteristics, and enhancing the model's ability to identify sea fog areas through channel and spatial attention mechanisms.
The model's identification accuracy and robustness of sea fog areas are improved, noise interference is reduced, and classification accuracy and reliability are enhanced.
Smart Images
Figure CN119851152A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the technical field of image recognition, and in particular to a method for recognizing sea fog hyperspectral remote sensing images based on CNN. Background Art
[0002] Traditional feature extraction methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), may not be able to fully capture complex spectral features when processing high-dimensional data, resulting in information loss; especially when processing complex objects, such as sea fog hyperspectral remote sensing images, traditional methods are usually unable to adaptively extract features related to specific tasks and may ignore some key spectral information.
[0003] Traditional methods are often sensitive to noise and redundant information when processing hyperspectral remote sensing images, which may lead to instability and decreased accuracy of classification results; and existing methods cannot effectively distinguish important features from redundant features, thus affecting the overall performance of the model.
[0004] Many traditional hyperspectral remote sensing image recognition methods tend to only focus on spectral information when extracting features, while ignoring the use of spatial context information; this causes the model to perform poorly when dealing with objects with complex spatial distribution.
[0005] In summary, the existing hyperspectral remote sensing image recognition methods have certain shortcomings in feature extraction capability, noise processing, spatial context utilization, etc. These shortcomings limit the effectiveness of existing methods in identifying sea fog hyperspectral remote sensing images. Summary of the invention
[0006] In order to solve the problems existing in the above-mentioned prior art, the present disclosure aims to provide a sea fog hyperspectral remote sensing image recognition method based on CNN to solve the effectiveness problem of the prior art in recognizing sea fog hyperspectral remote sensing images.
[0007] The present invention discloses a method for recognizing sea fog hyperspectral remote sensing images based on CNN, comprising the following steps:
[0008] Sea fog hyperspectral remote sensing image preprocessing: loading sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; in the process of obtaining the feature parameters, an attention mechanism is introduced, and the attention mechanism is used to improve the expression ability of the feature parameters;
[0009] Constructing a data set: combining the obtained feature parameters with the corresponding label data as sample data, and dividing the sample data into a training data set and a test data set;
[0010] Training model: using the training data set to train the CNN model, and using the test data set to test the classification results of the CNN model until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model;
[0011] A sea fog hyperspectral remote sensing image to be identified is obtained and recorded as a target image, and the target image is input into the sea fog hyperspectral remote sensing image recognition model to obtain a classification result of the target image.
[0012] Preferably, the characteristic parameters include spectral characteristics and spatial characteristics, and the spectral characteristics are obtained by the following formula:
[0013]
[0014] Among them, the R i represents the reflectance value of the selected wavelength, and N represents the number of samples of the selected wavelength.
[0015] Preferably, the selected wavelengths include: 1.4 μm, 1.9 μm and at least two wavelengths of 0.45 μm to 0.75 μm.
[0016] Preferably, a channel attention mechanism is introduced in the process of obtaining the feature parameters, and the channel attention mechanism comprises the following steps:
[0017] Calculate the global average pooling and global maximum pooling of sea fog hyperspectral remote sensing images;
[0018] Generate channel weights through a fully connected layer;
[0019] Applying channel weights to the sea fog hyperspectral remote sensing image; obtaining a first weighted sea fog hyperspectral remote sensing image.
[0020] Preferably, after introducing the channel attention mechanism in the process of obtaining the feature parameters, the method further includes: introducing a spatial attention mechanism, wherein the introducing the spatial attention mechanism comprises the following steps:
[0021] Calculating global average pooling and global maximum pooling of the first weighted sea fog hyperspectral remote sensing image;
[0022] Generate spatial weights through a convolutional layer;
[0023] The spatial weight is applied to the first weighted sea fog hyperspectral remote sensing image to obtain a second weighted sea fog hyperspectral remote sensing image.
[0024] Preferably, the CNN model comprises two convolutional layers, two maximum pooling layers and two fully connected layers cascaded in sequence;
[0025] Among them, the two convolution layers are respectively the first convolution layer and the second convolution layer, the two maximum pooling layers include the first maximum pooling layer and the second maximum pooling layer, and two fully connected layers are arranged after the second maximum pooling layer. The two fully connected layers include the first fully connected layer and the second fully connected layer. The first fully connected layer is used to integrate the high-dimensional features extracted by the first convolution layer, the second convolution layer and the maximum pooling layer; the second fully connected layer is used to generate classification results.
[0026] Preferably, the input channel of the first convolutional layer is 1, the output channel is 64, the convolution kernel size is 3, and the padding is 1;
[0027] The pooling window size of the maximum pooling layer is 2, the stride is 2, and the padding is 1;
[0028] The second convolutional layer has 64 input channels, 128 output channels, a convolution kernel size of 3, and a padding of 1.
[0029] The input dimension of the first fully connected layer is 128, and the output dimension is 256;
[0030] The input dimension of the second fully connected layer is 256.
[0031] Preferably, a batch normalization layer and an activation layer are arranged between the first convolution layer and the first maximum pooling layer, and between the second convolution layer and the second maximum pooling layer; and the activation layer uses "ReLU" as the activation function.
[0032] The present disclosure also provides a sea fog hyperspectral remote sensing image recognition system, which is characterized by comprising:
[0033] A preprocessing module, which is used to load sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; an attention mechanism is introduced in the process of obtaining the feature parameters, and the attention mechanism is used to improve the expression ability of the feature parameters;
[0034] A data set construction module, which is used to combine the obtained feature parameters with the corresponding label data as sample data, and divide the sample data into a training data set and a test data set;
[0035] A model training module, which is used to train the CNN model using the training data set, and to test the classification result of the CNN model using the test data set, until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model;
[0036] The recognition and classification module is used to obtain the sea fog hyperspectral remote sensing image that needs to be recognized as a target image, input the target image into the sea fog hyperspectral remote sensing image recognition model, and obtain the classification result of the target image.
[0037] The present disclosure also provides a non-temporary computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CNN-based sea fog hyperspectral remote sensing image recognition as described above.
[0038] The advantages of the sea fog hyperspectral remote sensing image recognition method based on CNN disclosed in the present invention are:
[0039] The present disclosure introduces an attention mechanism in the feature extraction stage, which can dynamically adjust the weights of each feature layer and enhance the model's recognition ability of sea fog areas. This mechanism enables the model to pay more attention to important features in the image, suppress irrelevant or redundant information, reduce the interference of noise on feature extraction, and thus improve the accuracy of classification; and by focusing on key features, the attention mechanism can improve the robustness of the model and reduce the probability of misclassification.
[0040] While introducing the attention mechanism, combined with the special scene of sea fog hyperspectral remote sensing images, the texture features and spectral features of sea fog hyperspectral remote sensing images are integrated, and the rich information in hyperspectral remote sensing images is fully utilized. By fusing different types of features, the model can understand the image content more comprehensively, thereby improving the classification accuracy and enhancing the recognition ability of the model.
[0041] In summary, the present invention integrates texture features and spectral features and adds texture features and spectral features to the feature extraction step of the model through the attention mechanism, so that this method has obvious advantages in the complex sea fog area recognition task, and can provide a more accurate and reliable solution for sea fog hyperspectral remote sensing image recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 is a flow chart of the sea fog hyperspectral remote sensing image recognition method based on CNN described in the present disclosure;
[0043] Figure 2 It is a flow chart for constructing the sea fog hyperspectral remote sensing image recognition model;
[0044] Figure 3 It is a schematic diagram of the architecture of the CNN model. DETAILED DESCRIPTION
[0045] like Figure 1 As shown, the CNN-based sea fog hyperspectral remote sensing image recognition method disclosed in the present invention comprises the following steps:
[0046] Sea fog hyperspectral remote sensing image preprocessing: loading sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; in the process of obtaining the feature parameters, an attention mechanism is introduced, and the attention mechanism is used to improve the expression ability of the feature parameters;
[0047] Constructing a data set: combining the obtained feature parameters with the corresponding label data as sample data, and dividing the sample data into a training data set and a test data set;
[0048] Training model: using the training data set to train the CNN model, and using the test data set to test the classification results of the CNN model until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model;
[0049] A sea fog hyperspectral remote sensing image to be identified is obtained and recorded as a target image, and the target image is input into the sea fog hyperspectral remote sensing image recognition model to obtain a classification result of the target image.
[0050] Specifically, in the training of the CNN model, since the CNN model needs to be trained and learned through the training data set, the association between the data set and the label can be obtained; in the present disclosure, for the sea fog hyperspectral remote sensing image, the water droplets in the sea fog have a significant scattering effect on the light, resulting in blurred edges of objects in the image and reduced visibility. This makes the details and color information in the image unclear. If only the initial sea fog image is simply feature extracted, it is easy to cause low classification accuracy. Therefore, the attention mechanism is introduced in the feature extraction stage to convert the initial sea fog hyperspectral remote sensing image into a weighted feature map. The attention mechanism can improve the model's sensitivity to the relevant features of the target object "sea fog", expand the weight of the relevant features of "sea fog" in the sea fog hyperspectral remote sensing image, and thus improve the model's accuracy in recognizing sea fog images and classification accuracy.
[0051] In the training of the CNN model, cross entropy loss and Adam optimizer are used to train n training cycles. Cross entropy loss is used to measure the difference between the model output and the true label, and Adam optimizer is used to update the model parameters.
[0052] After training for n training cycles, the model is evaluated using the test data set to calculate the confusion matrix, OA (overall accuracy), AA (average accuracy) and Kappa coefficient. The confusion matrix is used to represent the accuracy of classification, OA and AA describe the overall and each category classification accuracy respectively, and the Kappa coefficient quantitatively evaluates the effectiveness of the classifier. When the above indicators meet the standards, the model training is considered complete.
[0053] After the model training is completed, a sea fog hyperspectral remote sensing image recognition model is obtained. When the sea fog hyperspectral remote sensing image recognition model is used to identify and classify the target image, the target image still needs to undergo a preprocessing step to convert the target image into a weighted feature map that can be recognized by the sea fog hyperspectral remote sensing image recognition model, and then be classified after being recognized by the sea fog hyperspectral remote sensing image recognition model described in the present invention.
[0054] Exemplarily, the classification result can be a simple "yes" or "no", that is, "determined to be a sea fog image" or determined to be a "non-sea fog image"; it can also be classified according to ground objects, such as "sea fog", "ocean" or "background".
[0055] like Figure 2 The specific process of model construction is as follows:
[0056] Due to the characteristics of hyperspectral remote sensing images, the characteristic parameters extracted from the sea fog hyperspectral remote sensing images disclosed in the present invention include spectral characteristics and spatial characteristics. Specifically, the spectral characteristics are obtained by the following formula:
[0057]
[0058] Among them, the R i is the reflectance value at the selected wavelength and N is the number of samples at that wavelength.
[0059] Exemplarily, the process of obtaining the reflectivity value generally includes the following steps:
[0060] Data Collection:
[0061] When acquiring hyperspectral remote sensing images, the hyperspectral sensor captures reflected light information at multiple wavelengths. The wavelength selection can be to include all wavelengths in a band or to select some specific wavelengths in a band.
[0062] Data processing:
[0063] The hyperspectral remote sensing image is subjected to radiation correction. The steps of radiation correction include:
[0064] Geometric correction: Ensure that the geometric shape of the image is consistent with the actual object and eliminate errors caused by shooting angle and position.
[0065] Radiation calibration: The calibration coefficient is obtained through laboratory measurement or data provided by the sensor manufacturer, and the digital value (DN value) measured by the sensor is converted into reflectance through the calibration coefficient.
[0066] Reflectivity calculation:
[0067] The calculation formula for reflectivity is:
[0068] R(λ)=L(λ) / E(λ)
[0069] in:
[0070] R(λ) is the reflectance value.
[0071] L(λ) is the radiance measured by the sensor.
[0072] E(λ) is the radiance of the incident light (solar radiation), which can be obtained through solar radiation models or measured data.
[0073] Furthermore, in order to improve the sensitivity of the model to the feature of “sea fog”, when acquiring the spectral features of the hyperspectral image, some wavelengths are selected to obtain the reflectivity of specific wavelengths, and then the spectral features of the “sea fog” at specific wavelengths are obtained.
[0074] Specifically, the selected wavelengths include: 1.4 μm, 1.9 μm, and at least two wavelengths of 0.45 μm to 0.75 μm.
[0075] Among them, 0.45μm to 0.75μm is the visible light band. Since sea fog is formed by small water droplets suspended in the air, water droplets will scatter the light in the visible light band. The scattering phenomenon makes the spectral characteristics of the sea fog area show a higher reflectivity in the spectrum graph. Based on this, the spectral characteristics of sea fog can be reflected by the reflectivity of light in the visible light band.
[0076] 1.4μm and 1.9μm belong to the near-infrared band, and at these two wavelengths, water vapor has significant absorption characteristics, that is, light with wavelengths of 1.4μm and 1.9μm will be absorbed when passing through water vapor, resulting in reduced reflectivity; and sea fog is usually formed on the sea surface, especially in warm and humid climate conditions, when water vapor cools and condenses; that is, the concentration of water vapor directly affects the formation and development of sea fog. Therefore, in the sea fog area, the distribution and concentration changes of water vapor are closely related to the characteristics of sea fog. The absorption characteristics at wavelengths of 1.4μm and 1.9μm can reflect the concentration and distribution of water vapor, and thus reflect the characteristics of sea fog.
[0077] In summary, by utilizing the fact that light in the visible light band has a higher reflectivity in the sea fog area and light in the near-infrared band has a lower reflectivity in the sea fog environment, the sensitivity of the model to the spectral characteristics of sea fog is improved, thereby improving the recognition accuracy of the trained model for sea fog hyperspectral remote sensing images.
[0078] Exemplarily, wavelengths of 0.55 μm, 0.65 μm, 1.4 μm and 1.9 μm may be selected to obtain the spectral characteristics of the sea fog hyperspectral remote sensing image.
[0079] Furthermore, a channel attention mechanism is introduced in the process of obtaining feature parameters;
[0080] The channel attention mechanism includes the following steps:
[0081] Calculate the global average pooling and global maximum pooling of sea fog hyperspectral remote sensing images;
[0082] Generate channel weights through a fully connected layer;
[0083] Applying channel weights to the sea fog hyperspectral remote sensing image; obtaining a first weighted sea fog hyperspectral remote sensing image.
[0084] Exemplarily, assume that the input sea fog hyperspectral remote sensing image is F, and the dimension is C*H*W, where C is the number of channels, H is the height of the feature map, and W is the width of the feature map.
[0085] Calculate the global average pooling F of sea fog hyperspectral remote sensing images avg and global maximum pooling F max :
[0086] Global average pooling F avg Calculated by the following formula:
[0087]
[0088] Global maximum pooling F max Calculated by the following formula:
[0089]
[0090] In the above formula, F(c, i, j) represents the value of the cth channel at position (i, j) in the feature map.
[0091] The global average pooling F avg and global maximum pooling F max The results are connected together to form a new feature vector F concat :
[0092]
[0093] Generate channel weights through a fully connected layer:
[0094] The channel weight is calculated as follows:
[0095] Z=σ(W2·δ(W1·F concat )).
[0096] Applying the channel weights to the sea fog hyperspectral remote sensing image; obtaining a first weighted sea fog hyperspectral remote sensing image:
[0097] The generated channel weights are used to weight the input feature map to obtain the first weighted sea fog hyperspectral remote sensing image F':
[0098] F'=Z⊙F;
[0099] Among them, ⊙ represents the element-by-element multiplication operation.
[0100] The first weighted sea fog hyperspectral remote sensing image F' obtained is denoted as F'(c,h,w).
[0101] The channel attention mechanism dynamically adjusts the weights of each channel in the feature map, enhancing the network's attention to important features and improving the performance of the model.
[0102] Furthermore, after introducing the channel attention mechanism in the process of obtaining feature parameters, it also includes: introducing the spatial attention mechanism,
[0103] The spatial attention mechanism consists of the following steps:
[0104] Calculate the global average pooling and global maximum pooling of the first weighted sea fog hyperspectral remote sensing image;
[0105] Generate spatial weights through a convolutional layer;
[0106] Applying spatial weights to the first weighted sea fog hyperspectral remote sensing image; obtaining a second weighted sea fog hyperspectral remote sensing image.
[0107] Exemplarily, the first weighted sea fog hyperspectral remote sensing image F' obtained by the channel attention mechanism is used as input;
[0108] Calculate the global average pooling F′ of the first weighted sea fog hyperspectral remote sensing image avg and global maximum pooling F′ max :
[0109] Global average pooling F′ avg Calculated by the following formula:
[0110]
[0111] Global maximum pooling F′ maxCalculated by the following formula:
[0112]
[0113] In the above formula, F'(c,h,w) represents the value of the cth channel at position (h,w) in the feature map.
[0114] The global average pooling F′ avg and global maximum pooling F′ max The results are connected together to form a new feature vector F′ concat :
[0115]
[0116] Generate spatial attention weights S through convolutional layers:
[0117] S = σ(Conv(F′ concat ));
[0118] Where Conv represents the convolution operation and σ is the sigmoid function.
[0119] Applying the spatial weight to the first weighted sea fog hyperspectral remote sensing image; obtaining a second weighted sea fog hyperspectral remote sensing image:
[0120] Use the generated spatial weights to weight the input feature map to obtain the weighted feature map F":
[0121] F”=S⊙F’;
[0122] Among them, ⊙ represents the element-by-element multiplication operation.
[0123] The spatial attention mechanism is added on the basis of the channel attention mechanism to form a comprehensive attention module that can simultaneously focus on the channel information and spatial information of the feature map. By dynamically adjusting the weights of each channel and spatial position in the feature map, the model can more effectively extract and express important features, thereby improving the accuracy of feature extraction.
[0124] Furthermore, the CNN model includes two convolutional layers, two max pooling layers, and two fully connected layers cascaded in sequence;
[0125] Among them, the two convolution layers are respectively the first convolution layer and the second convolution layer, the two maximum pooling layers include the first maximum pooling layer and the second maximum pooling layer, and two fully connected layers are arranged after the second maximum pooling layer. The two fully connected layers include the first fully connected layer and the second fully connected layer. The first fully connected layer is used to integrate the high-dimensional features extracted by the first convolution layer, the second convolution layer and the maximum pooling layer; the second fully connected layer is used to generate classification results.
[0126] Furthermore, the input channel of the first convolutional layer is 1, the output channel is 64, the convolution kernel size is 3, and the padding is 1;
[0127] The pooling window size of the maximum pooling layer is 2, the stride is 2, and the padding is 1;
[0128] The second convolutional layer has 64 input channels, 128 output channels, a convolution kernel size of 3, and a padding of 1.
[0129] The input dimension of the first fully connected layer is 128, and the output dimension is 256;
[0130] The input dimension of the second fully connected layer is 256.
[0131] Further, such as Figure 3 As shown, a batch normalization layer and an activation layer are arranged between the first convolution layer and the first maximum pooling layer, and between the second convolution layer and the second maximum pooling layer; the activation layer uses "ReLU" as the activation function.
[0132] The present disclosure also provides a sea fog hyperspectral remote sensing image recognition system, comprising:
[0133] A preprocessing module, which is used to load sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; an attention mechanism is introduced in the process of obtaining the feature parameters, and the attention mechanism is used to improve the expression ability of the feature parameters;
[0134] A data set construction module, which is used to combine the obtained feature parameters with the corresponding label data as sample data, and divide the sample data into a training data set and a test data set;
[0135] A model training module, which is used to train the CNN model using the training data set, and to test the classification result of the CNN model using the test data set, until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model;
[0136] The recognition and classification module is used to obtain the sea fog hyperspectral remote sensing image that needs to be recognized as a target image, input the target image into the sea fog hyperspectral remote sensing image recognition model, and obtain the classification result of the target image.
[0137] The present disclosure also provides a non-temporary computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CNN-based sea fog hyperspectral remote sensing image recognition as described above.
[0138] The present disclosure introduces an attention mechanism in the feature extraction stage, which can dynamically adjust the weights of each feature layer and enhance the model's recognition ability of sea fog areas. This mechanism enables the model to pay more attention to important features in the image, suppress irrelevant or redundant information, reduce the interference of noise on feature extraction, and thus improve the accuracy of classification; and by focusing on key features, the attention mechanism can improve the robustness of the model and reduce the probability of misclassification.
[0139] While introducing the attention mechanism, combined with the special scene of sea fog hyperspectral remote sensing images, the texture features and spectral features of sea fog hyperspectral remote sensing images are integrated, and the rich information in hyperspectral remote sensing images is fully utilized. By fusing different types of features, the model can understand the image content more comprehensively, thereby improving the classification accuracy and enhancing the recognition ability of the model.
[0140] In summary, the present invention integrates texture features and spectral features and adds texture features and spectral features to the feature extraction step of the model through the attention mechanism, so that this method has obvious advantages in the complex sea fog area recognition task, and can provide a more accurate and reliable solution for sea fog hyperspectral remote sensing image recognition.
[0141] In the description of the present disclosure, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present disclosure and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the scope of protection of the present disclosure.
[0142] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present disclosure.
Claims
1. A sea fog hyperspectral remote sensing image recognition method based on CNN, characterized in that: The following steps are involved: Sea fog hyperspectral remote sensing image preprocessing: loading sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; in the process of obtaining the feature parameters, an attention mechanism is introduced, and the attention mechanism is used to improve the expression ability of the feature parameters; Constructing a data set: combining the obtained feature parameters with the corresponding label data as sample data, and dividing the sample data into a training data set and a test data set; Training model: using the training data set to train the CNN model, and using the test data set to test the classification results of the CNN model until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model; A sea fog hyperspectral remote sensing image to be identified is obtained and recorded as a target image, and the target image is input into the sea fog hyperspectral remote sensing image recognition model to obtain a classification result of the target image.
2. The sea fog hyperspectral remote sensing image recognition method based on CNN according to claim 1 is characterized in that: The characteristic parameters include spectral characteristics and spatial characteristics, and the spectral characteristics are obtained by the following formula: Among them, R i represents the reflectance value of the selected wavelength, and N represents the number of samples of the selected wavelength.
3. The sea fog hyperspectral remote sensing image recognition method based on CNN according to claim 2 is characterized in that: The selected wavelengths include: 1.4 μm, 1.9 μm, and at least two wavelengths of 0.45 μm to 0.75 μm.
4. The sea fog hyperspectral remote sensing image recognition method based on CNN according to claim 2 is characterized in that: A channel attention mechanism is introduced in the process of obtaining feature parameters, and the channel attention mechanism includes the following steps: Calculate the global average pooling and global maximum pooling of sea fog hyperspectral remote sensing images; Generate channel weights through a fully connected layer; The channel weights are applied to the sea fog hyperspectral remote sensing image to obtain a first weighted sea fog hyperspectral remote sensing image.
5. The method for recognizing sea fog hyperspectral remote sensing images based on CNN according to claim 4, characterized in that: After the channel attention mechanism is introduced in the process of obtaining the feature parameters, the process further includes: introducing a spatial attention mechanism, wherein the introducing the spatial attention mechanism includes the following steps: Calculating global average pooling and global maximum pooling of the first weighted sea fog hyperspectral remote sensing image; Generate spatial weights through a convolutional layer; The spatial weight is applied to the first weighted sea fog hyperspectral remote sensing image to obtain a second weighted sea fog hyperspectral remote sensing image.
6. The method for recognizing sea fog hyperspectral remote sensing images based on CNN according to claim 1, characterized in that: The CNN model includes two convolutional layers, two maximum pooling layers and two fully connected layers cascaded in sequence; Among them, the two convolution layers are respectively the first convolution layer and the second convolution layer, the two maximum pooling layers include the first maximum pooling layer and the second maximum pooling layer, and two fully connected layers are arranged after the second maximum pooling layer. The two fully connected layers include the first fully connected layer and the second fully connected layer. The first fully connected layer is used to integrate the high-dimensional features extracted by the first convolution layer, the second convolution layer and the maximum pooling layer; the second fully connected layer is used to generate classification results.
7. The method for recognizing sea fog hyperspectral remote sensing images based on CNN according to claim 6, characterized in that: A batch normalization layer and an activation layer are arranged between the first convolution layer and the first maximum pooling layer, and between the second convolution layer and the second maximum pooling layer; and the activation layer uses "ReLU" as the activation function.
8. The method for recognizing sea fog hyperspectral remote sensing images based on CNN according to claim 6, characterized in that: The first convolutional layer has an input channel of 1, an output channel of 64, a convolution kernel size of 3, and a padding of 1; The pooling window size of the maximum pooling layer is 2, the stride is 2, and the padding is 1; The second convolutional layer has 64 input channels, 128 output channels, a convolution kernel size of 3, and a padding of 1. The input dimension of the first fully connected layer is 128, and the output dimension is 256; The input dimension of the second fully connected layer is 256.
9. A sea fog hyperspectral remote sensing image recognition system, characterized in that: include: A preprocessing module, which is used to load sea fog hyperspectral remote sensing image data and corresponding label data; the loading of sea fog hyperspectral remote sensing image data includes extracting features of the sea fog hyperspectral remote sensing image data to obtain feature parameters; an attention mechanism is introduced in the process of obtaining the feature parameters, and the attention mechanism is used to improve the expression ability of the feature parameters; A data set construction module, which is used to combine the obtained feature parameters with the corresponding label data as sample data, and divide the sample data into a training data set and a test data set; A model training module, which is used to train the CNN model using the training data set, and to test the classification result of the CNN model using the test data set, until the number of iterations or classification accuracy requirements of the CNN model are met, thereby obtaining a sea fog hyperspectral remote sensing image recognition model; The recognition and classification module is used to obtain the sea fog hyperspectral remote sensing image that needs to be recognized as a target image, input the target image into the sea fog hyperspectral remote sensing image recognition model, and obtain the classification result of the target image.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the CNN-based sea fog hyperspectral remote sensing image recognition as described in any one of claims 1 to 8 is realized.