Deep learning-based underwater hyperspectral image classification method and device, computer equipment and storage medium

Abstract

The application relates to a deep learning-based underwater hyperspectral image classification method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining an underwater hyperspectral image and performing image preprocessing to obtain a first underwater hyperspectral image; simulating light propagation results through an IOP model and an RTE model according to underwater data information, combining the first underwater hyperspectral image, and reconstructing the underwater hyperspectral image through a GAN model to obtain a second underwater hyperspectral image; inputting the first underwater hyperspectral image and the second underwater hyperspectral image into an asymmetric residual network added with an SE module to perform model training, obtaining an underwater hyperspectral image classification model, and thereby classifying the underwater hyperspectral image. The IOP model, the RTE model and the GAN model are used to reconstruct high-quality images, the asymmetric residual network added with the SE module is used to further extract and optimize the features in the images, and the accuracy of the underwater hyperspectral image classification can be improved.

Description

Technical Field

[0001] This invention relates to the field of image classification technology, and in particular to a method, apparatus, computer device, and storage medium for underwater hyperspectral image classification based on deep learning. Background Technology

[0002] The underwater environment is a complex and crucial research field, significant for marine science, water quality monitoring, and aquatic life conservation. Hyperspectral classification within the underwater environment is a key task, aiding in a better understanding of physical, chemical, and biological processes in water bodies and providing crucial information about the state of the underwater environment. Underwater hyperspectral classification technology leverages the rich information in spectral data to identify and analyze underwater objects, water quality parameters, and ecosystem characteristics. Accurate classification and identification of underwater targets enable the monitoring, assessment, and management of the underwater environment, thereby improving the protection and utilization of marine resources and addressing the challenges of environmental change. Currently, methods for underwater hyperspectral classification mainly include those based on spectral reflectance, those based on feature extraction and machine learning, and those based on optical physical models. Spectral reflectance methods classify underwater objects by measuring their spectral reflectance at different wavelengths. By comparing the spectral characteristics of the target object with those of known objects in a reference library, target identification can be achieved. Feature extraction and machine learning methods extract effective features from underwater hyperspectral data and use machine learning algorithms for classification. Commonly used feature extraction methods include Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). Machine learning algorithms can be traditional classification algorithms, such as Support Vector Machines (SVM) and Random Forests, or deep learning algorithms, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Optical physics models are based on underwater optical physics processes and models, modeling and simulating the propagation and interaction of light in water. These models can describe the scattering, absorption, and reflection of light in water, thereby enabling the identification and classification of underwater objects and water quality parameters. IOP models and RTE models are commonly used optical physics models. By comparing measured data with model simulation results, underwater targets can be classified.

[0003] While various methods can demonstrate good performance in underwater hyperspectral image classification, single model structures still have certain limitations. Underwater hyperspectral images possess spatial, spectral, and semantic features, among others. Classification methods based on spectral reflectance face challenges in complex underwater environments due to the influence of depth, scattering, and absorption on the spectrum. Classification methods using feature extraction and machine learning struggle to account for the complexity of underwater optical features and require substantial training data and computational resources. Inherent Optical Properties (IOP) models or Radiative Transfer Equations (RTE) models can overcome the complexity of underwater optical features and reduce the influence of depth, scattering, and absorption on the spectrum; however, the optical-physical models for underwater hyperspectral image classification exhibit low generalization, stability, and robustness.

[0004] Therefore, traditional underwater hyperspectral image classification methods suffer from low classification accuracy. Summary of the Invention

[0005] Therefore, in order to solve the above-mentioned technical problems, a method, apparatus, computer equipment and storage medium for underwater hyperspectral image classification based on deep learning are provided, which can improve the accuracy of underwater spectral image classification.

[0006] A deep learning-based underwater hyperspectral image classification method, the method comprising:

[0007] Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain a first underwater hyperspectral image;

[0008] Based on the underwater data information, the light propagation results are simulated using the IOP model and RTE model, and combined with the first underwater hyperspectral image, the underwater hyperspectral image is reconstructed using the GAN model to obtain the second underwater hyperspectral image.

[0009] The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets respectively, and the image datasets are used as input to an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0010] Underwater hyperspectral images are classified using the aforementioned underwater hyperspectral image classification model.

[0011] In one embodiment, the step of preprocessing the underwater hyperspectral image to obtain a first underwater hyperspectral image includes:

[0012] The underwater atmospheric transmission model was used to remove the brightness shift caused by atmospheric light in the underwater hyperspectral image, and noise was removed by median filtering to obtain a preliminary underwater hyperspectral image.

[0013] The first underwater hyperspectral image is obtained by using a histogram equalization-based method to enhance image details and compensate for illumination changes in the preliminary underwater hyperspectral image.

[0014] In one embodiment, based on the underwater data information, light propagation results are simulated using the IOP model and RTE model, and combined with the first underwater hyperspectral image, an underwater hyperspectral image is reconstructed using the GAN model to obtain a second underwater hyperspectral image, including:

[0015] The IOP model is used to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information, and the inherent optical property parameters of the water body are calculated.

[0016] The inherent optical properties of the water body are input into the RTE model to simulate the propagation path and intensity distribution of light in the water, and the light propagation results are obtained.

[0017] Combining the light propagation results with the first underwater hyperspectral image, the underwater hyperspectral image is reconstructed using a GAN model to obtain the second underwater hyperspectral image.

[0018] In one embodiment, the IOP model is used to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information, and the inherent optical property parameters of the water body are calculated, including:

[0019] The IOP model uses the Morel model to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information;

[0020] Among them, the inherent optical properties of water bodies include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters.

[0021] In one embodiment, the step of inputting the inherent optical properties of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water includes:

[0022] The inherent optical properties of the water body are input into the RTE model, which uses the Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in the water.

[0023] In one embodiment, the first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets, including:

[0024] Principal component analysis was used to perform dimensionality reduction on the first and second underwater hyperspectral images to obtain the dimensionality-reduced images.

[0025] The dimensionality-reduced image is converted into several image blocks, and each image block is used as a dataset.

[0026] The dataset is format-converted based on the input requirements of the convolutional neural network.

[0027] In one embodiment, the GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; the step of reconstructing an underwater hyperspectral image using the GAN model to obtain a second underwater hyperspectral image includes:

[0028] The light propagation result is input into the generator;

[0029] The encoder uses convolutional layers to extract and encode features from the light propagation results to obtain encoded features.

[0030] The decoder uses transposed convolutional layers to perform progressive upsampling and reconstruction operations, restoring the encoded features to the reconstructed image, thus obtaining the second underwater hyperspectral image.

[0031] A deep learning-based underwater hyperspectral image classification device, the device comprising:

[0032] The first underwater hyperspectral image acquisition module is used to acquire underwater hyperspectral images and underwater data information, and to perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image.

[0033] The second underwater hyperspectral image acquisition module is used to simulate the light propagation results through the IOP model and RTE model based on the underwater data information, and reconstruct the underwater hyperspectral image through the GAN model in combination with the first underwater hyperspectral image to obtain the second underwater hyperspectral image.

[0034] The model training module is used to process the first underwater hyperspectral image and the second underwater hyperspectral image into image datasets respectively, and input the image datasets into an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0035] An image classification module is used to classify underwater hyperspectral images using the underwater hyperspectral image classification model.

[0036] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0037] Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain a first underwater hyperspectral image;

[0038] Based on the underwater data information, the light propagation results are simulated using the IOP model and RTE model, and combined with the first underwater hyperspectral image, the underwater hyperspectral image is reconstructed using the GAN model to obtain the second underwater hyperspectral image.

[0039] The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets respectively, and the image datasets are used as input to an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0040] Underwater hyperspectral images are classified using the aforementioned underwater hyperspectral image classification model.

[0041] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0042] Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain a first underwater hyperspectral image;

[0043] Based on the underwater data information, the light propagation results are simulated using the IOP model and RTE model, and combined with the first underwater hyperspectral image, the underwater hyperspectral image is reconstructed using the GAN model to obtain the second underwater hyperspectral image.

[0044] The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets respectively, and the image datasets are used as input to an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0045] Underwater hyperspectral images are classified using the aforementioned underwater hyperspectral image classification model.

[0046] The aforementioned deep learning-based underwater hyperspectral image classification method, device, computer equipment, and storage medium use the IOP model, RTE model, and GAN model to reconstruct high-quality images. By adding an asymmetric residual network with an SE module, features in the images are further extracted and optimized to make the images more suitable for classification tasks, thereby further improving the accuracy and stability of underwater hyperspectral image classification. Attached Figure Description

[0047] Figure 1 This is an application environment diagram of a deep learning-based underwater hyperspectral image classification method in one embodiment.

[0048] Figure 2 This is a flowchart illustrating a deep learning-based underwater hyperspectral image classification method in one embodiment.

[0049] Figure 3 This is a schematic diagram of the network structure of an asymmetric residual network in one embodiment;

[0050] Figure 4 This is a schematic diagram of the SE module in one embodiment;

[0051] Figure 5 This is a schematic diagram of the process of reconstructing an image using the IOP+RTE model combined with GAN in one embodiment.

[0052] Figure 6 This is a schematic diagram illustrating the calculation process of the Morel model in the IOP model in one embodiment;

[0053] Figure 7 This is a schematic diagram of the Monte Carlo simulation method in an RTE model in one embodiment;

[0054] Figure 8 This is a schematic diagram of the GAN model structure in one embodiment;

[0055] Figure 9 This is a structural block diagram of a deep learning-based underwater hyperspectral image classification device in one embodiment.

[0056] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] It is understood that the terms "first," "second," etc., used herein may be used to describe underwater hyperspectral images, but these underwater hyperspectral images are not limited by these terms. These terms are only used to distinguish one underwater hyperspectral image from another. For example, without departing from the scope of this application, the first underwater hyperspectral image may be referred to as the second underwater hyperspectral image, and similarly, the second underwater hyperspectral image may be referred to as the first underwater hyperspectral image. Both the first and second underwater hyperspectral images are underwater hyperspectral images, but they are not the same underwater hyperspectral image.

[0059] The underwater hyperspectral image classification method based on deep learning provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes a computer device 110. The computer device 110 can acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain a first underwater hyperspectral image. Based on the underwater data information, the computer device 110 can simulate light propagation results using an IOP model and a RTE model, and combine the first underwater hyperspectral image with a GAN model to reconstruct the underwater hyperspectral image, obtaining a second underwater hyperspectral image. The computer device 110 can process the first and second underwater hyperspectral images into image datasets, and use these image datasets as input to an asymmetric residual network with an added SE module for model training, obtaining an underwater hyperspectral image classification model. The computer device 110 can perform underwater hyperspectral image classification using this underwater hyperspectral image classification model. The computer device 110 can be, but is not limited to, various personal computers, laptops, robots, unmanned aerial vehicles, tablets, and other devices.

[0060] In one embodiment, such as Figure 2 As shown, a deep learning-based underwater hyperspectral image classification method is provided, including the following steps:

[0061] Step 202: Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image.

[0062] Computer equipment can acquire underwater hyperspectral images and underwater data. The underwater hyperspectral images can be acquired using image acquisition equipment, while the underwater data can be acquired using detection devices or sensors. The underwater data can include auxiliary data such as water quality parameters, depth information, underwater environment, turbidity, and salinity.

[0063] Next, the computer equipment can perform image preprocessing on the underwater hyperspectral image, removing atmospheric light, noise and other data from the underwater hyperspectral image, and compensating for changes in illumination, and use the processed image as the first underwater hyperspectral image.

[0064] Step 204: Based on the underwater data information, simulate the light propagation results using the IOP model and RTE model, and combine the first underwater hyperspectral image with the GAN model to reconstruct the underwater hyperspectral image, thereby obtaining the second underwater hyperspectral image.

[0065] Computer equipment can use the IOP+RTE physical model combined with pre-processed hyperspectral images to reconstruct underwater hyperspectral images through a Generative Adversarial Network (GAN). Specifically, the computing equipment can first use the IOP model to estimate parameters based on underwater data, and then input the estimated parameters into the RTE model to simulate light propagation results. Next, the computing equipment can combine the light propagation results with the first underwater hyperspectral image and use the GAN model to complete image reconstruction, thereby obtaining the second underwater hyperspectral image.

[0066] Step 206: The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets, and the image datasets are used as input to an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0067] The computer equipment can further preprocess the first and second underwater hyperspectral images to make the processed data suitable for the input and output of a convolutional neural network. Specifically, the computer equipment can process the first and second underwater hyperspectral images into image datasets suitable for the input and output of a convolutional neural network, and use an asymmetric residual network (SE-ARN) with an added SE module for image classification.

[0068] In the asymmetric residual network (SE-ARN), the number of convolutional kernels in the two sub-networks differs from the network structure. An SE module is introduced to implement a channel attention mechanism, and residual connections are added to achieve skip connections. The asymmetric residual network (SE-ARN) with the added SE module mainly includes the Squeeze-and-Excitation Module (SE module) in the attention mechanism, the first sub-network, the second sub-network, residual connection operations, feature concatenation operations, and the output layer.

[0069] In this embodiment, the SE module implements a channel attention mechanism through global average pooling (GlobalAveragePooling2D), shape transformation (Reshape), fully connected layers (FC), and multiplication operations (Multiply). Both the first and second sub-networks include convolutional layers (CONV2D), max pooling layers (MaxPooling2D), and the SE module, using ReLU as the activation function. A convolutional layer is added to the first sub-network, and a convolutional layer and an average pooling layer are added to the second sub-network. The feature map obtained from this convolutional or pooling layer is then concatenated with the outputs of the SE modules in the first and second sub-networks to complete residual connections and achieve cross-layer information transfer. In this embodiment, the feature concatenation operation is performed along the feature channel dimension to generate a higher-dimensional feature map.

[0070] Specifically, the asymmetric residual network structure and the SE module structure are as follows: Figure 3 , Figure 4 As shown, where, Figure 3 The network structure is an asymmetric residual network. Figure 4 This describes the structure of the SE module. The first and second sub-networks in the asymmetric residual network both include convolutional layers, max pooling layers, average pooling layers, the SE module, Add layers, Concatenate layers, Flatten layers, and Dense layers, with ReLU activation functions.

[0071] The number of convolutional kernels and network structures differ between the two sub-networks. The second sub-network has one more average pooling layer than the first, ensuring that the Add and Concatenate layers can perform addition and concatenation operations correctly. The SE module consists of a global average pooling layer (GAP), a Reshape layer, and two fully connected layers, using ReLU as the activation function. The GAP layer primarily obtains the mean Z of the feature values ​​for each channel, calculated using the following formula: Where H and W represent the height and width of the input tensor, respectively, c is the number of channels, and i and j are the indices of H and W, respectively.

[0072] The first fully connected layer (called the Squeeze operation) maps z to a small feature vector f, calculated as: f = ReLU(FC(z)). The second fully connected layer (called the Excitation operation) maps the feature vector f to a channel weight vector s, where each element of s is scaled to the range [0, 1], calculated as: s = Sigmoid(FC(f)). Finally, the SE module produces the output by performing a weighted summation on each channel of the input feature tensor. The calculation formula is: Where Xj represents the j-th channel of the input feature tensor X, s j It is the j-th element of the channel weight vector s.

[0073] In the first sub-network, the residual connection adds the output of the second max-pooling layer (referred to as `residual_a`) to the output of the convolutional layer (referred to as `se_block_a`) and the output of the SE module (referred to as `se_block_a`) element-wise. The corresponding operation is: `res_block_a = residual_a + se_block_a`.

[0074] In the second sub-network, the residual connection adds the output of the second max-pooling layer (residual_b) obtained through convolutional and average pooling layers, and the output of the SE module (se_block_b) element-wise. The corresponding operation is res_block_b = residual_b + se_block_b. The outputs of the two networks, output_1 (res_block_a) and output_2 (res_block_b), are then processed through concatenate, flatten, and dense layers to perform underwater hyperspectral image classification. The SE-ARN network optimizer uses Adam, and the loss function is categorical cross-entropy, which can be expressed as: CategoricalCross-Entropy Loss = -∑(yi*log(p) i Where, Categorical Cross-Entropy Loss is the result of the classification cross-entropy loss function; ∑ represents the summation over all categories; y i p is the i-th element of the true label, representing the probability that the sample belongs to the i-th class; it is a one-hot encoded vector. i It is the i-th element of the model's predicted output, representing the probability that the model predicts the sample belongs to the i-th class, which is obtained through the Softmax function.

[0075] Furthermore, after the network is built, the model is trained and optimized using the image dataset. The trained and optimized SE-ARN model can then be used to obtain the underwater hyperspectral image classification results.

[0076] Step 208: Classify underwater hyperspectral images using an underwater hyperspectral image classification model.

[0077] In this embodiment, the computer device uses the IOP model, RTE model, and GAN model to reconstruct high-quality images. By adding an asymmetric residual network with an SE module, features in the images are further extracted and optimized to make the images more suitable for classification tasks, thereby further improving the accuracy and stability of underwater hyperspectral image classification.

[0078] In one embodiment, a deep learning-based underwater hyperspectral image classification method may further include an image preprocessing process, specifically including: using an underwater atmospheric transmission model to remove brightness shifts caused by atmospheric light in the underwater hyperspectral image, and removing noise through median filtering to obtain a preliminary underwater hyperspectral image; using a histogram equalization-based method to enhance image details in the preliminary underwater hyperspectral image and compensate for illumination changes to obtain a first underwater hyperspectral image.

[0079] Specifically, the computer device can use an underwater atmospheric transmission model to remove brightness shifts caused by atmospheric light and median filtering to remove noise. In this embodiment, both of these steps can be performed using ENVI software. Next, the computer device can continue to use the histogram equalization method in ENVI software to enhance image details and compensate for illumination changes, thereby obtaining the first underwater hyperspectral image.

[0080] like Figure 5 As shown, in one embodiment, a deep learning-based underwater hyperspectral image classification method may further include an image reconstruction process, specifically including:

[0081] Step 502: Using the IOP model, the scattering, absorption and reflection characteristics of the water body are estimated based on underwater data information, and the inherent optical property parameters of the water body are calculated.

[0082] Step 504: Input the inherent optical properties of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water and obtain the light propagation results.

[0083] Step 506: Combining the light propagation results and the first underwater hyperspectral image, the underwater hyperspectral image is reconstructed using a GAN model to obtain the second underwater hyperspectral image.

[0084] Computer equipment can use the IOP model to estimate the scattering, absorption, and reflection characteristics of water bodies based on measurement data, thereby obtaining the intrinsic optical property parameters of the water bodies. Next, the computer equipment can input the estimated parameters into the RTE model to simulate the propagation path and intensity distribution of light in water. Then, combining the light propagation results simulated by the RTE model with the original hyperspectral image, image reconstruction is performed using a GAN model to obtain a second underwater hyperspectral image.

[0085] In one embodiment, a deep learning-based underwater hyperspectral image classification method may further include a process of calculating relevant parameters. The specific process includes: using the Morel model with the IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information; wherein, the inherent optical property parameters of the water body include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters of the water body.

[0086] In this embodiment, the IOP model can use the Morel model, which consists of five processes, such as... Figure 6 As shown, the process includes collecting measurement data, selecting reference data, parameter estimation, model building, and validation and correction.

[0087] The Morel model can estimate the scattering, absorption, and reflection characteristics of water bodies, including the calculation of the total absorption coefficient of water bodies, the calculation of chlorophyll absorption coefficient, and the calculation of chlorophyll spectral absorption curve parameters.

[0088] The formula for calculating the total absorption coefficient of water is: a total (lambda) = a phy (lambda)+a cd (lambda)+a nap (lambda)+a y (lambda)+a sm (lambda). Where lambda is the wavelength, a phy a is the chlorophyll absorption coefficient; cd a is the absorption coefficient of dissolved organic matter; nap a is the absorption coefficient of non-melting particulate matter; y a is the absorption coefficient of xanthine solubility; sm is the absorption coefficient of soluble particulate matter.

[0089] The formula for calculating the chlorophyll absorption coefficient is: a phy (lambda) = Chl*a phi (lambda). Where lambda is the wavelength; Chl is the chlorophyll concentration; a phi (lambda) is the parameter of the chlorophyll spectral absorption curve.

[0090] The formula for calculating the parameters of the chlorophyll spectral absorption curve is: a phi (lambda) = a water (lambda) + beta(lambda) * Chl + Chl 2 *a cs (lambda). Where lambda is the wavelength, a water (lambda) is the pure water absorption coefficient, beta(lambda) is the chlorophyll ratio coefficient, and a cs (lambda) is a non-linear relationship parameter for chlorophyll concentration.

[0091] In one embodiment, a deep learning-based underwater hyperspectral image classification method may further include a process of simulating light in water, specifically including: inputting the inherent optical properties of the water body into an RTE model, wherein the RTE model uses a Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in water.

[0092] The RTE model uses the Monte Carlo simulation method, which consists of six processes, such as... Figure 7 As shown, this includes parameter settings, photon trajectory generation, light propagation simulation, scattering and absorption calculations, statistical analysis, and output. The formula for the light intensity propagation model in the Monte Carlo simulation method is: Where dl / ds represents the rate of change of light intensity along the propagation path; c is the speed of light; a is the absorption coefficient; s is the scattering coefficient; I is the intensity of the incident light; and B is the radiation term (including scattering and divergence components). The Morel model and Monte Carlo simulation method were implemented using the professional Hydrolight software.

[0093] In one embodiment, the underwater hyperspectral image classification method based on deep learning may further include a process of processing the underwater hyperspectral images. The specific process includes: using principal component analysis to perform dimensionality reduction processing on the first underwater hyperspectral image and the second underwater hyperspectral image to obtain dimensionality-reduced images; converting the dimensionality-reduced images into several image patches, and using each image patch as a dataset; and performing format conversion on the dataset according to the input requirements of the convolutional neural network.

[0094] Specifically, the computer equipment can use Principal Component Analysis (PCA) to perform dimensionality reduction on the first and second underwater hyperspectral images; convert the data into several image patches to train the neural network model; divide the dataset into training and testing sets; increase the weak data by copying samples of the weak category; save the preprocessed dataset 1 (X1_train, X1_test, y1_train, y1_test) corresponding to the first underwater hyperspectral image and the dataset 2 (X2_train, X2_test, y2_train, y2_test) corresponding to the second underwater hyperspectral image, and convert the labels to one-hot encoding for the input required by the convolutional neural network.

[0095] Furthermore, the computer device can convert the two image datasets into a format suitable for the convolutional network, with each dataset having a shape of (batch_size, height, width, channels). Here, batch_size represents the batch size, height and width represent the height and width of the image, respectively, and channels represent the number of input bands, i.e., the number of channels.

[0096] In one embodiment, a deep learning-based underwater hyperspectral image classification method may further include a process of reconstructing an underwater hyperspectral image. The specific process includes: inputting the light propagation result into a generator; the encoder using a convolutional layer to extract and encode features from the light propagation result to obtain encoded features; and the decoder using a transposed convolutional layer to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image, thereby obtaining a second underwater hyperspectral image.

[0097] Among them, such as Figure 8 As shown, the GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers (Conv2DTranspose) and LeakyReLU layers (with an alpha parameter of 0.2).

[0098] The generator takes the light propagation result simulated by RTE (let's call it G_input) as input. The encoder uses a convolutional layer to extract and encode features from G_input to obtain encoded features. The decoder uses a transposed convolutional layer to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image (let's call it G_output).

[0099] The discriminator takes the original hyperspectral image (D_input) and the generated reconstructed image (G_output) as input. It then uses convolutional layers to extract features from D_input and G_output, obtaining feature representations. Finally, convolutional and fully connected layers are used to discriminate these feature representations, outputting a judgment result indicating whether the input is the real image (D_input) or the generated image (G_output).

[0100] In this embodiment, the loss function used in the GAN model is binary cross-entropy, which can be expressed as: L(y,y) pred )=-(y*log(y pred )+(1-y)*log(1-y predAmong them, L(y,y) pred The expression represents the result of the loss function (binary_crossentropy), where y is the true label (0 or 1) and y_pred is the model's predicted probability value (between 0 and 1). When y = 1, the first term y * log(y_pred) represents the sample belonging to the true image; when y = 0, the second term (1-y) * log(1-y_pred) represents the sample belonging to the generated image.

[0101] In this embodiment, the Adam optimizer is used to minimize the loss function during GAN training. The parameters are set as follows: lr = 0.0002, beta_1 = 0.5. The Adam optimization process is as follows: First, initialize the parameters: learning rate α; momentum parameter β1; second-order moment estimation parameter (RMSProp decay rate): β2; initial gradient accumulation m = 0; initial squared gradient accumulation v = 0; iteration count t = 0. Second, in each iteration, calculate the gradient with respect to parameter θ. J(θ) is the gradient of the loss function with respect to the parameter θ; next, update the first moment estimate (momentum): m = β1*m + (1-β1)*g; update the second moment estimate (RMSProp): v = β2*v + (1-β2)*g 2 Then, bias corrections are applied to the estimated first and second moments: Update parameters: Repeat the above steps until the parameters converge and the loss is minimized. Finally, the computer can use the trained GAN to obtain the reconstructed image G_output.

[0102] In one embodiment, to illustrate the implementation process of the asymmetric residual network in the deep learning-based underwater hyperspectral image classification method, the method is used on a hypothetical dataset. It is assumed that the images obtained after image preprocessing and reconstruction are all 150×150 pixels, containing a total of 224 bands and 16 categories. The images are named Image_1 and Image_2, respectively.

[0103] When using the deep learning-based underwater hyperspectral image classification method provided in this embodiment, firstly, PCA is used to reduce the two high-dimensional data to 25 dimensions. Then, Image_1 is divided into blocks using a 19×19 window, and Image_2 is divided into blocks using a 21×21 window. Secondly, weak class data is increased by copying samples of the weak class. Finally, the dataset is divided into 7:3 segments and the data is saved. The final shapes of X1_train, X1_test, y1_train, and y1_test data are [27697, 19, 19, 25],

[27697] , [3075, 19, 19, 25], and

[3075] , respectively. The shapes of X2_train, X2_test, y2_train, and y2_test data are [27697, 21, 21, 25],

[27697] , [3075, 21, 21, 25], and

[3075] , respectively.

[0104] The data shape conforms to the input requirements of the SE-ARN network and does not need to be reshaped. The shapes of Input1 and Input2 are [None, 19, 19, 25] and [None, 21, 21, 25], respectively.

[0105] In the asymmetric residual network, the first, second, and third convolutional layers of the first subnetwork use 64, 128, and 256 3×3 convolutional kernels, respectively, with a stride of 1. The max-pooling layer uses a 2×2 pooling window with a stride of 1. In the SE module, the global pooling layer pools the input shape to [None, 256], and then the Reshape layer transforms it into the shape [None, 1, 1, 256]. The two fully connected layers have 16 and 256 neurons, respectively, and the output is multiplied by the result of the third max-pooling layer. The residual connection uses a convolutional layer with 256 1×1 convolutional kernels and a stride of 2. The residual connection inputs the output from the second max-pooling layer to the output of the convolutional layer and the output from the SE module, and then performs an element-wise addition operation to obtain the output of the first subnetwork, output_1([None, 2, 2, 256]).

[0106] In the second subnetwork, the first, second, and third convolutional layers use 32, 64, and 128 3×3 kernels respectively, with a stride of 1. The max-pooling layer uses a 2×2 pooling window with a stride of 1. In the SE module, the global pooling layer pools the input shape to [None, 128], and then the reshape layer reshapes the input to [None, 1, 1, 256]. The two fully connected layers have 8 and 128 neurons respectively, and the output is multiplied by the result of the third max-pooling layer. The residual connection convolutional layer uses 128 3×3 kernels with a stride of 2, and the average pooling layer has a pooling window of 2 with a stride of 1. The residual connection adds the output of the second subnetwork, output_2([None, 2, 2, 128]), obtained by passing the output of the second max-pooling layer through the convolutional layer and the average pooling layer.

[0107] The outputs of the first sub-network (output_1) and the second sub-network (output_2) pass through a concatenate layer, a flatten layer, and three dense layers to obtain the output result, Output. The network has a total of 1,553,832 parameters and occupies approximately 5.93MB of memory, classifying it as a lightweight convolutional neural network.

[0108] When compiling the model, the Adam optimizer is used, the loss function is Categorical_Crossentropy, batch_size=32, epochs=50, and validation_split=0.1.

[0109] The model performance is evaluated using four metrics: accuracy (Acc), precision (Pre), recall, and F1 score.

[0110] This embodiment provides a deep learning-based underwater hyperspectral image classification method. In the initial raw image preprocessing stage, atmospheric light is removed using software, noise is removed through median filtering, and histogram equalization is used to compensate for illumination variations. Next, using auxiliary data (such as water quality parameters, depth information, underwater ambient temperature, turbidity, salinity, etc.) and the processed underwater hyperspectral image, the underwater hyperspectral image is reconstructed using an IOP+RTE physical model and a GAN. These two images are then input into an asymmetric residual network (SE-ARN) to complete the underwater hyperspectral image classification task. The advantages of this approach are: firstly, it improves upon traditional single models or networks by using an underwater-suitable optical physics model (IOP) combined with an RTE model and a GAN to reconstruct high-quality images; secondly, the residual network further extracts and optimizes features in the image, making it more suitable for classification tasks and improving classification accuracy and network performance.

[0111] Furthermore, during preprocessing, PCA is used to reduce data dimensionality, remove redundant feature information, and shorten computation time, converting the data into several image patches for training the neural network model. The residual network uses an asymmetric residual network, adapted to both underwater hyperspectral images and the reconstructed images. An SE module is added to fully utilize channel relationships and model parameter efficiency. Residual connections use skip connections to address the gradient vanishing problem during deep network training, enabling SE-ARN to better extract key features in image classification tasks, further improving network performance and classification accuracy.

[0112] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0113] In one embodiment, such as Figure 9 As shown, a deep learning-based underwater hyperspectral image classification device is provided, comprising: a first underwater hyperspectral image acquisition module 910, a second underwater hyperspectral image acquisition module 920, a model training module 930, and an image classification module 940, wherein:

[0114] The first underwater hyperspectral image acquisition module 910 is used to acquire underwater hyperspectral images and underwater data information, and to perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image.

[0115] The second underwater hyperspectral image acquisition module 920 is used to simulate the light propagation results through the IOP model and RTE model based on underwater data information, and reconstruct the underwater hyperspectral image through the GAN model by combining the first underwater hyperspectral image to obtain the second underwater hyperspectral image.

[0116] The model training module 930 is used to process the first underwater hyperspectral image and the second underwater hyperspectral image into image datasets respectively, and input the image datasets into an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model.

[0117] Image classification module 940 is used to classify underwater hyperspectral images using an underwater hyperspectral image classification model.

[0118] In one embodiment, the first underwater hyperspectral image acquisition module 910 is further configured to use an underwater atmospheric transmission model to remove the brightness shift caused by atmospheric light in the underwater hyperspectral image, and remove noise by median filtering to obtain a preliminary underwater hyperspectral image; and use a histogram equalization-based method to enhance the image details in the preliminary underwater hyperspectral image and compensate for illumination changes to obtain the first underwater hyperspectral image.

[0119] In one embodiment, the second underwater hyperspectral image acquisition module 920 is further configured to use the IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information, and calculate the inherent optical property parameters of the water body; input the inherent optical property parameters of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water, and obtain the light propagation results; combine the light propagation results with the first underwater hyperspectral image, and reconstruct the underwater hyperspectral image through the GAN model to obtain the second underwater hyperspectral image.

[0120] In one embodiment, the second underwater hyperspectral image acquisition module 920 is also used to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information using the Morel model of the IOP model; wherein, the inherent optical property parameters of the water body include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters of the water body.

[0121] In one embodiment, the second underwater hyperspectral image acquisition module 920 is also used to input the inherent optical property parameters of the water body into the RTE model, which uses the Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in water.

[0122] In one embodiment, the model training module 930 is further configured to perform dimensionality reduction processing on the first underwater hyperspectral image and the second underwater hyperspectral image using principal component analysis to obtain dimensionality-reduced images; convert the dimensionality-reduced images into several image blocks, and use each image block as a dataset; and perform format conversion on the dataset according to the input requirements information of the convolutional neural network.

[0123] In one embodiment, the GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; the second underwater hyperspectral image acquisition module 920 is also used to input the light propagation result into the generator; the encoder uses convolutional layers to extract and encode features from the light propagation result to obtain encoded features; the decoder uses transposed convolutional layers to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image, thereby obtaining the second underwater hyperspectral image.

[0124] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a deep learning-based underwater hyperspectral image classification method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0125] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0127] Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image;

[0128] Based on underwater data, the light propagation results were simulated using the IOP model and RTE model. Combined with the first underwater hyperspectral image, the underwater hyperspectral image was reconstructed using the GAN model to obtain the second underwater hyperspectral image.

[0129] The first and second underwater hyperspectral images were processed into image datasets, and the image datasets were used as input to an asymmetric residual network with an added SE module for model training, resulting in an underwater hyperspectral image classification model.

[0130] Underwater hyperspectral images are classified using an underwater hyperspectral image classification model.

[0131] In one embodiment, when the processor executes the computer program, it further performs the following steps: using an underwater atmospheric transmission model to remove the brightness shift caused by atmospheric light in the underwater hyperspectral image, and removing noise by median filtering to obtain a preliminary underwater hyperspectral image; using a histogram equalization-based method to enhance image details in the preliminary underwater hyperspectral image and compensate for illumination changes to obtain a first underwater hyperspectral image.

[0132] In one embodiment, when the processor executes the computer program, it further performs the following steps: using the IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information, and calculating the inherent optical property parameters of the water body; inputting the inherent optical property parameters of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water, and obtaining the light propagation results; combining the light propagation results and the first underwater hyperspectral image, reconstructing the underwater hyperspectral image through the GAN model, and obtaining the second underwater hyperspectral image.

[0133] In one embodiment, when the processor executes the computer program, it further implements the following steps: the IOP model uses the Morel model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information; wherein, the inherent optical property parameters of the water body include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters of the water body.

[0134] In one embodiment, the processor, when executing the computer program, also performs the following steps: inputting the inherent optical properties of the water body into the RTE model, which uses the Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in the water.

[0135] In one embodiment, when the processor executes the computer program, it further performs the following steps: using principal component analysis to perform dimensionality reduction processing on the first underwater hyperspectral image and the second underwater hyperspectral image to obtain dimensionality-reduced images; converting the dimensionality-reduced images into several image blocks, and using each image block as a dataset; and performing format conversion on the dataset according to the input requirements information of the convolutional neural network.

[0136] In one embodiment, the GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; when the processor executes the computer program, it also implements the following steps: the light propagation result is input into the generator; the encoder uses a convolutional layer to extract and encode features from the light propagation result to obtain encoded features; the decoder uses a transposed convolutional layer to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image, thereby obtaining a second underwater hyperspectral image.

[0137] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0138] Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image;

[0139] Based on underwater data, the light propagation results were simulated using the IOP model and RTE model. Combined with the first underwater hyperspectral image, the underwater hyperspectral image was reconstructed using the GAN model to obtain the second underwater hyperspectral image.

[0140] The first and second underwater hyperspectral images were processed into image datasets, and the image datasets were used as input to an asymmetric residual network with an added SE module for model training, resulting in an underwater hyperspectral image classification model.

[0141] Underwater hyperspectral images are classified using an underwater hyperspectral image classification model.

[0142] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: using an underwater atmospheric transmission model to remove the brightness shift caused by atmospheric light in the underwater hyperspectral image, and removing noise by median filtering to obtain a preliminary underwater hyperspectral image; using a histogram equalization-based method to enhance image details in the preliminary underwater hyperspectral image and compensate for illumination changes to obtain a first underwater hyperspectral image.

[0143] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: using the IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information, and calculating the inherent optical property parameters of the water body; inputting the inherent optical property parameters of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water, and obtaining the light propagation results; combining the light propagation results and the first underwater hyperspectral image, reconstructing the underwater hyperspectral image through the GAN model, and obtaining the second underwater hyperspectral image.

[0144] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: the IOP model uses the Morel model to estimate the scattering, absorption, and reflection characteristics of the water body based on underwater data information; wherein, the inherent optical property parameters of the water body include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters of the water body.

[0145] In one embodiment, when the computer program is executed by the processor, it also performs the following steps: inputting the inherent optical properties of the water body into the RTE model, which uses the Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in the water.

[0146] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: using principal component analysis to perform dimensionality reduction processing on the first underwater hyperspectral image and the second underwater hyperspectral image to obtain dimensionality-reduced images; converting the dimensionality-reduced images into several image blocks, and using each image block as a dataset; and performing format conversion on the dataset according to the input requirements information of the convolutional neural network.

[0147] In one embodiment, the GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; when the computer program is executed by the processor, it also performs the following steps: the light propagation result is input into the generator; the encoder uses a convolutional layer to extract and encode features from the light propagation result to obtain encoded features; the decoder uses a transposed convolutional layer to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image, thereby obtaining a second underwater hyperspectral image.

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0149] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A deep learning-based underwater hyperspectral image classification method, characterized in that, The method includes: Acquire underwater hyperspectral images and underwater data information, and perform image preprocessing on the underwater hyperspectral images to obtain a first underwater hyperspectral image; Based on the underwater data, light propagation results are simulated using the IOP model and RTE model. Combined with the first underwater hyperspectral image, a second underwater hyperspectral image is reconstructed using a GAN model. This process includes: using the IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data, and calculating the inherent optical property parameters of the water body; inputting the inherent optical property parameters of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water, obtaining the light propagation results; and combining the light propagation results with the first underwater hyperspectral image, reconstructing the underwater hyperspectral image using a GAN model to obtain the second underwater hyperspectral image. The GAN model includes a generator and a discriminator; the generator includes an input layer, an encoder, a decoder, and an output layer; the discriminator includes an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; the process of reconstructing an underwater hyperspectral image using the GAN model to obtain a second underwater hyperspectral image includes: the light propagation result being input into the generator; the encoder using a convolutional layer to extract and encode features from the light propagation result to obtain encoded features; and the decoder using a transposed convolutional layer to perform progressive upsampling and reconstruction operations, restoring the encoded features to the reconstructed image to obtain the second underwater hyperspectral image. The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets respectively, and the image datasets are used as input to an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model. Underwater hyperspectral images are classified using the aforementioned underwater hyperspectral image classification model.

2. The underwater hyperspectral image classification method based on deep learning according to claim 1, characterized in that, The step of preprocessing the underwater hyperspectral image to obtain a first underwater hyperspectral image includes: The underwater atmospheric transmission model was used to remove the brightness shift caused by atmospheric light in the underwater hyperspectral image, and noise was removed by median filtering to obtain a preliminary underwater hyperspectral image. The first underwater hyperspectral image is obtained by using a histogram equalization-based method to enhance image details and compensate for illumination changes in the preliminary underwater hyperspectral image.

3. The underwater hyperspectral image classification method based on deep learning according to claim 1, characterized in that, The IOP model is used to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information, and the inherent optical property parameters of the water body are calculated, including: The IOP model uses the Morel model to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information; Among them, the inherent optical properties of water bodies include the total absorption coefficient, chlorophyll absorption coefficient, and chlorophyll spectral absorption curve parameters.

4. The underwater hyperspectral image classification method based on deep learning according to claim 1, characterized in that, The process of inputting the inherent optical properties of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water includes: The inherent optical properties of the water body are input into the RTE model, which uses the Monte Carlo simulation method to simulate the propagation path and intensity distribution of light in the water.

5. The underwater hyperspectral image classification method based on deep learning according to claim 1, characterized in that, The first underwater hyperspectral image and the second underwater hyperspectral image are processed into image datasets, including: Principal component analysis was used to perform dimensionality reduction on the first and second underwater hyperspectral images to obtain the dimensionality-reduced images. The dimensionality-reduced image is converted into several image blocks, and each image block is used as a dataset. The dataset is format-converted based on the input requirements of the convolutional neural network.

6. A deep learning-based underwater hyperspectral image classification device, characterized in that, The device includes: The first underwater hyperspectral image acquisition module is used to acquire underwater hyperspectral images and underwater data information, and to perform image preprocessing on the underwater hyperspectral images to obtain the first underwater hyperspectral image. The second underwater hyperspectral image acquisition module is used to simulate light propagation results using an IOP model and an RTE model based on the underwater data information, and reconstruct an underwater hyperspectral image using a GAN model in conjunction with the first underwater hyperspectral image to obtain a second underwater hyperspectral image. This includes: using an IOP model to estimate the scattering, absorption, and reflection characteristics of the water body based on the underwater data information, and calculating the inherent optical property parameters of the water body; inputting the inherent optical property parameters of the water body into the RTE model to simulate the propagation path and intensity distribution of light in the water to obtain the light propagation results; and combining the light propagation results with the first underwater hyperspectral image to reconstruct the underwater hyperspectral image using a GAN model to obtain the second underwater hyperspectral image. The GAN model comprises a generator and a discriminator; the generator comprises an input layer, an encoder, a decoder, and an output layer; the discriminator comprises an input layer, a feature extraction layer, and a discrimination layer; and both the generator and the discriminator are composed of transposed convolutional layers and LeakyReLU layers; the process of reconstructing an underwater hyperspectral image using the GAN model to obtain a second underwater hyperspectral image includes: the light propagation result being input into the generator; the encoder using a convolutional layer to extract and encode features from the light propagation result to obtain encoded features; and the decoder using a transposed convolutional layer to perform progressive upsampling and reconstruction operations to restore the encoded features to the reconstructed image, thereby obtaining the second underwater hyperspectral image. The model training module is used to process the first underwater hyperspectral image and the second underwater hyperspectral image into image datasets respectively, and input the image datasets into an asymmetric residual network with an added SE module for model training to obtain an underwater hyperspectral image classification model. An image classification module is used to classify underwater hyperspectral images using the underwater hyperspectral image classification model.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images