An unsupervised adaptive hyperspectral remote sensing image denoising method
By employing an unsupervised adaptive hyperspectral remote sensing image denoising method, which utilizes convolutional neural networks and the Huber loss function, the problem of poor generalization ability in existing technologies is solved, achieving efficient denoising and image quality improvement in real-world noisy scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHUHAI ORBIT SATELLITE BIG DATA CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-07-21
AI Technical Summary
Existing hyperspectral remote sensing image denoising methods have poor generalization ability in real noise scenarios, are difficult to adapt to complex noise distributions, and require paired training data.
An unsupervised adaptive hyperspectral remote sensing image denoising method is adopted. By constructing a convolutional neural network denoising model, combining the Adam optimization algorithm and Huber loss function, and utilizing maximum a posteriori probability estimation and orthogonal attention mechanism, the method adaptively processes real noise, achieving efficient denoising without the need for paired training data.
It effectively improves image quality and the accuracy of subsequent applications, can adaptively process in real noisy scenes, and does not require additional clean samples as supervision data.
Smart Images
Figure CN121437314B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image processing technology, and in particular to an unsupervised adaptive hyperspectral remote sensing image denoising method, which can be applied to software development systems for remote sensing image processing. Background Technology
[0002] Hyperspectral remote sensing imagery presents as a three-dimensional data cube, integrating information from the two-dimensional spatial and spectral domains, possessing the characteristic of "image-spectrum integration." It is widely used in fields such as land cover classification, environmental protection, and monitoring of sensitive targets. However, hyperspectral imaging is susceptible to atmospheric interference, photon effects, and other natural conditions, as well as inherent sensor defects, during acquisition. This results in various interferences such as Gaussian noise, sparse noise, and fringe noise, leading to image quality degradation and affecting applications such as classification, super-resolution, and unmixing.
[0003] Existing hyperspectral image denoising methods mainly include traditional filtering and deep learning-based methods. Traditional methods are limited by noise type assumptions and linear models, resulting in limited denoising performance. In recent years, supervised denoising methods based on deep learning have gradually become a research hotspot. These methods achieve adaptive feature learning and nonlinear modeling through neural networks, outperforming traditional methods in denoising performance. However, these methods rely heavily on large-scale clean training samples, while it is difficult to obtain paired data of real noise and clean images in actual hyperspectral images. Usually, training sets need to be constructed by simulating noise. Due to the difference between simulated noise and real noise, the network model has poor generalization ability in practical applications and is difficult to adapt to complex noise distributions.
[0004] Therefore, there is an urgent need for a hyperspectral image denoising method that does not rely on paired training data and can adaptively process real noise in order to improve image quality and the accuracy of subsequent applications. Summary of the Invention
[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art, and to address the problem that the hyperspectral remote sensing image denoising method has poor generalization ability in real noise scenes. This invention proposes an unsupervised adaptive hyperspectral remote sensing image denoising method.
[0006] The technical solution adopted in this invention is as follows: This invention includes the following steps:
[0007] Step S1. Image preprocessing; Input remote sensing image X, normalize the input remote sensing image X so that the pixel values are mapped to the [0,1] interval, and obtain a normalized hyperspectral remote sensing image. ;
[0008] Step S2. Construct an unsupervised denoising framework and denoising model. Step S2 includes the following steps: S21. Construct an unsupervised denoising framework by inputting the remote sensing image X into the degradation model and using a posterior probability estimation method to obtain a clean image from the remote sensing image X. S22. Construct a convolutional neural network denoising model. The convolutional neural network denoising model consists of three main parts: an encoder, a decoder, and a skip connection layer. The encoder and decoder are connected through the skip connection layer.
[0009] Step S3. The network parameters are iteratively updated using the Adam optimization algorithm combined with the backpropagation strategy, and a stopping mechanism is set. The stopping mechanism is triggered when the change of the preset evaluation index is less than the set threshold in a number of consecutive iterations, and the iteration process is terminated.
[0010] Step S4. Hyperparameter setting; Based on the noise characteristics of the remote sensing image X, set the hyperparameters of the convolutional neural network denoising model;
[0011] Step S5. Network inference and result output, the hyperspectral remote sensing image after preprocessing in step S1. The hyperparameters set in step S4 are input into the convolutional neural network denoising model constructed in step S2 until the stopping condition set in step S3 is met, and the denoised hyperspectral image is finally output.
[0012] Furthermore, the degradation model in step S21 is represented by formula (1):
[0013]
[0014] Where X represents remote sensing imagery, Represents the true clean image in the degradation model, It is mixed noise;
[0015] The maximum a posteriori probability estimation method is used to recover a clean image from a noisy image X. The optimization objective can be expressed as formula (2):
[0016]
[0017] in, The estimated variables for a clean image are denoted as and the optimal solution is denoted as . Used to approximate R(Y) represents the prior regularization term, used to constrain the spectral consistency of hyperspectral images. The reconstruction error term is represented by the Huber loss function, as defined in formula (3):
[0018]
[0019] in, Representing remote sensing images The reconstruction residuals between the clean image estimate variable Y and the actual reconstruction residuals. This is a robust threshold used to suppress anomalous noise and preserve edge features;
[0020] Denoising model using convolutional neural networks Replacing R(Y), equation (2) is equivalently transformed into equation (4):
[0021]
[0022] in, Indicates parameters A convolutional neural network denoising model, where Z is a randomly generated noise matrix. The optimal network parameters are obtained through Huber loss and unsupervised adaptive optimization. This represents the optimal parameter convolutional neural network denoising model obtained after applying the Huber loss function and unsupervised adaptive optimization. The remote sensing image X is used as the unsupervised loss constraint, enabling the network to gradually approximate the clean image output during the iterative optimization process. , This is also known as denoised hyperspectral remote sensing imagery.
[0023] Furthermore, the encoder is composed of a first DS Block module, a second DS Block module, a third DS Block module, a fourth DS Block module, a fifth DS Block module, an ortho module, and a sixth DS Block module connected sequentially; the decoder is composed of a first US Block module, a second US Block module, a third US Block module, a fourth USBlock module, a fifth US Block module, and a sixth US Block module connected sequentially; the jump connection layer is composed of a first SC Block module, a second SC Block module, a third SC Block module, a fourth SC Block module, a fifth SCBlock module, and a sixth SC Block module. The first SC Block module is connected between the first DS Block module and the first US Block module; the second SC Block module is connected between the second DS Block module and the second US Block module; the third SC Block module is connected between the third DS Block module and the third US Block module; the fourth SC Block module is connected between the fourth DS Block module and the fourth US Block module; the fifth SC Block module is connected between the fifth DS Block module and the fifth US Block module; and the sixth SC Block module is connected between the sixth DS Block module and the sixth US Block module.
[0024] Furthermore, each of the first to sixth DS Block modules consists of seven layers connected sequentially: the first Conv unit, the downsample unit, the first BN layer, the first Leaky ReLU activation function, the second Conv unit, the second BN layer, and the second Leaky ReLU activation function; the ortho module consists of seven major components: Conv Blocks, RNG, Gram-Schmid, Orthogonal Filters, Squeezed, Excitation, and Attention; each of the first to sixth US Block modules consists of eight layers connected sequentially: the third BN layer, the third Conv unit, the fourth BN layer, the third Leaky ReLU activation function, the fourth Conv unit, the fifth BN layer, the fourth Leaky ReLU activation function, and the upsample unit; each of the first to sixth SC Block modules consists of three components: the fifth Conv unit, the sixth BN layer, and the fifth Leaky ReLU activation function.
[0025] Furthermore, the ortho module employs a positive channel attention mechanism, and the channel compression process of the ortho module is shown in formula (6):
[0026]
[0027] The channel attention weights are obtained through the activation function, and the fusion features are finally input using the residual method. The ortho module is represented by formula (7):
[0028]
[0029] in, This represents the feature map tensor output by the preceding network layer. H represents the number of feature channels, and H and W represent the height and width of the feature map, respectively. This represents the feature response value of the feature map at the c-th channel and spatial location (h, w); The orthogonal weighting coefficients corresponding to the c-th channel in the ortho module are used to perform weighted projection of channel features in the spatial dimension. The h and w above represent the spatial height index and spatial width index of the feature map, respectively. This represents the weighted result of the feature map of the c-th channel in the spatial dimension; This represents a set composed of compressed feature vectors from each channel. The constructed channel-level global description vector, i.e. ; This represents the activation function, used to nonlinearly map channel compression features to generate channel attention weights; This represents the orthogonal channel attention weight vector output by the ortho module, used to adjust the feature map. Channel-by-channel weighting is performed, and feature fusion is achieved through residual connection.
[0030] Furthermore, step S3 specifically involves: calculating the reconstruction error during the iteration process based on the Huber loss function in step S2, using the Adam optimization algorithm combined with a backpropagation strategy to iteratively update the network parameters, and setting a maximum number of iterations. As the upper limit of model training, during the iteration process, self Starting from the next iteration, the peak signal-to-noise ratio (PSNR), mean square error (MSE), and structural similarity between the network output image and the original noisy image in each band are calculated at fixed intervals. The PSNR, MSE, and structural similarity are used as evaluation indicators in the model iteration process, and a stopping mechanism is set. The stopping mechanism is triggered when the changes of the above evaluation indicators are less than a set threshold in several consecutive iterations, and the iteration process is terminated.
[0031] Furthermore, the model hyperparameters in step S4 include the robustness threshold in the Huber loss function. Maximum number of iterations And the thresholds for changes in peak signal-to-noise ratio, mean square error, and structural similarity in the stopping mechanism, as well as the detection interval.
[0032] Furthermore, the input size of the remote sensing image X is ,in , , These represent the number of rows, columns, and bands of the remote sensing image X, respectively.
[0033] The beneficial effects of this invention are: 1. This invention does not require additional clean samples as supervision data, and achieves effective denoising of hyperspectral images through network structure design and robust loss constraints; 2. It does not rely on paired training data, can adaptively process real noise, and can effectively improve image quality and the accuracy of subsequent applications. Attached Figure Description
[0034] Figure 1 This is a flowchart of the algorithm framework of the present invention;
[0035] Figure 2 This is a diagram of the structure of a convolutional neural network denoising model;
[0036] Figure 3 This is a schematic diagram illustrating the specific structure of DS Blocks, US Blocks, and SC Blocks in a convolutional neural network denoising model.
[0037] Figure 4 This is a schematic diagram of the ortho module structure in a convolutional neural network denoising model;
[0038] Figure 5 This is a schematic diagram of the noise reduction effect of each band in the hyperspectral OHS satellite remote sensing image. Detailed Implementation
[0039] like Figures 1 to 5 As shown, this embodiment uses OHS hyperspectral satellite imagery (400-1000nm, 32 bands) as the hyperspectral remote sensing image to be denoised, but the present invention is not limited to this data type and band range. The specific steps are as follows:
[0040] Step S1. Image preprocessing; Input size is The remote sensing image X, in which , , The number of rows, columns, and bands of the remote sensing image X are given, respectively. The input remote sensing image X is then normalized to map pixel values to the [0,1] interval, resulting in a normalized hyperspectral remote sensing image. ;
[0041] Step S2. Construct an unsupervised denoising framework and a denoising model. Step S2 includes the following steps: S21. Construct an unsupervised denoising framework, representing the remote sensing image X degradation model as formula (1):
[0042]
[0043] Represents the true clean image in the degradation model, It is mixed noise.
[0044] The maximum a posteriori probability estimation method is used to estimate the probability from noisy images. The objective of restoring a clean image is expressed as formula (2):
[0045]
[0046] in, The estimated variables for a clean image are denoted as and the optimal solution is denoted as . Used to approximate R(Y) represents the prior regularization term, used to constrain the spectral consistency of hyperspectral images. The reconstruction error term is represented by the Huber loss function, as defined in formula (3):
[0047]
[0048] in, Noise image The reconstruction residuals between the clean image estimate variable Y and the actual reconstruction residuals. A robustness threshold is used to suppress anomalous noise and preserve edge features; when | When |≤δ: The residual is small, belonging to normal pixels (edges / details), and the squared loss term is used. Preserve details;
[0049] When | |>δ: Large residuals indicate anomalous noise (such as salt-and-pepper noise or stripe noise), and a linear loss term is used. This limits the dominant loss due to anomalous noise, achieving robust suppression.
[0050] The core input variable of the loss function is the residual / error term, which essentially represents the degree of deviation between the model's predicted value and the true value (or reference value). It is the core variable for measuring the "magnitude of deviation." Huber loss measures this deviation by... The segmented processing achieves robust optimization by "preserving details with small errors and suppressing noise with large errors";
[0051] The prior term R(Y) is difficult to model explicitly, so a convolutional neural network denoising model is used. Replacing R(Y), equation (2) is equivalently transformed into equation (4):
[0052]
[0053] in, Indicates parameters A convolutional neural network denoising model, where Z is a randomly generated noise matrix. The optimal network parameters are obtained through Huber loss and unsupervised adaptive optimization. This represents the optimal parameter convolutional neural network denoising model obtained after applying the Huber loss function and unsupervised adaptive optimization. The remote sensing image X is used as the unsupervised loss constraint, enabling the network to gradually approximate the clean image output during the iterative optimization process. .
[0054] S22. Construct a convolutional neural network denoising model. The convolutional neural network denoising model is the UOrth-Net model with an encoder-decoder structure. Its network output corresponds to formula (4) in step S21 and is expressed as formula (5):
[0055]
[0056] Where Z is a randomly generated input noise matrix. The network parameters are obtained through Huber loss and unsupervised adaptive optimization. This represents the optimal parameter convolutional neural network denoising model obtained after Huber loss function and unsupervised adaptive optimization; the convolutional neural network denoising model consists of three main parts: encoder, decoder and skip connection layer, and the encoder and decoder are connected through skip connection layer;
[0057] The encoder is composed of a first DS Block module, a second DS Block module, a third DS Block module, a fourth DS Block module, a fifth DS Block module, and a sixth DS Block module connected in sequence. It is used to extract multi-scale spatial-spectral features of hyperspectral images. An ortho module is set between the fifth DS Block module and the sixth DS Block module. The ortho module, also known as the orthogonal channel attention mechanism, is used for orthogonalization of band features to enhance band differences and highlight noise features.
[0058] The first to sixth DS Block modules are each composed of seven layers connected in sequence: the first Conv unit, the downsample unit, the first BN layer, the first Leaky ReLU activation function, the second Conv unit, the second BN layer, and the second Leaky ReLU activation function; the ortho module consists of seven major components: Conv Blocks, RNG, Gram-Schmid, Orthogonal Filters, Squeezed, Excitation, and Attention.
[0059] The channel compression process of the ortho module is shown in equation (6). By introducing a set of weighting coefficients that are orthogonal in spatial dimensions, the channels of the input feature map are weighted and converged to obtain a channel-level global description. Specifically, for the input feature map... Configure each channel with orthogonal weighting coefficients consistent with its spatial dimensions. And the channel features are weighted and summed in the spatial dimension to achieve channel compression. Unlike the traditional channel compression method based on average pooling, this method reduces the correlation redundancy between different channel compressed representations through orthogonal constraints and improves the discriminativeness of channel feature representation. After channel compression is completed, the obtained channel-level global description vector is input into the activation function, and the corresponding channel attention weights are generated through nonlinear mapping. The orthogonal channel attention mechanism is shown in formula (7). Finally, the channel attention weights are applied to the input features in a channel-by-channel weighted manner and fused with the original features through residual connection, thereby enhancing the responsiveness to key information channels while maintaining feature stability. The calculation process of the c-th channel compressed feature is shown in formula 6:
[0060]
[0061] The channel attention weights are obtained through the activation function, and the fusion features are finally input using the residual method. The ortho module is represented by formula (7):
[0062]
[0063] in, This represents the feature map tensor output by the preceding network layer. H represents the number of feature channels, and H and W represent the height and width of the feature map, respectively. This represents the feature response value of the feature map at the c-th channel and spatial location (h, w); The orthogonal weighting coefficients corresponding to the c-th channel in the ortho module are used to perform weighted projection of the channel features in the spatial dimension. The h and w above represent the spatial height index and spatial width index of the feature map, respectively. This represents the weighted result of the feature map of the c-th channel in the spatial dimension; This represents a set composed of compressed feature vectors from each channel. The constructed channel-level global description vector, i.e. ; This represents the activation function, used to nonlinearly map channel compression features to generate channel attention weights; This represents the orthogonal channel attention weight vector output by the ortho module, used to adjust the feature map. Channel-by-channel weighting is performed, and feature fusion is achieved through residual connection.
[0064] The decoder is composed of a first US Block module, a second US Block module, a third US Block module, a fourth US Block module, a fifth US Block module, and a sixth US Block module connected sequentially, used to restore feature resolution step by step; each of the first to sixth US Block modules is composed of eight layers connected sequentially: a third BN layer, a third Conv unit, a fourth BN layer, a third Leaky ReLU activation function, a fourth Conv unit, a fifth BN layer, a fourth Leaky ReLU activation function, and an Upsample unit, used to restore image and structural detail features layer by layer;
[0065] The skip connection layer consists of a first SC Block module, a second SC Block module, a third SC Block module, a fourth SC Block module, a fifth SC Block module, and a sixth SC Block module. The first SC Block module is connected between the first DS Block module and the first US Block module; the second SC Block module is connected between the second DS Block module and the second US Block module; the third SC Block module is connected between the third DS Block module and the third US Block module; the fourth SC Block module is connected between the fourth DS Block module and the fourth US Block module; the fifth SC Block module is connected between the fifth DS Block module and the fifth US Block module; and the sixth SC Block module is connected between the sixth DS Block module and the sixth US Block module. The skip connection layer is used to transfer features from the encoding stage to the decoding stage. The skip connection layer can achieve the fusion of features at different scales, which helps to preserve edge and texture details in the image and effectively alleviates the gradient vanishing problem in deep network training. Each of the first to sixth SC Block modules consists of three components: a fifth Conv unit, a sixth BN layer, and a fifth Leaky ReLU activation function.
[0066] From the first BN layer to the sixth BN layer, all layers use Batch Normalization.
[0067] Step S3. Model Parameter Update and Stopping Mechanism: During iterative training, the reconstruction error between the network output image and the reference image is calculated based on the Huber loss function described in Step S2. Since the Huber loss function is defined piecewise within different error amplitude ranges, it maintains squared loss characteristics in small error regions and adopts linear loss in large error regions, thereby achieving adaptive suppression of gradients from abnormal noise samples. Based on this, the gradient information of the Huber loss function relative to each trainable parameter of the network is calculated using backpropagation (BP), and the gradient is weighted and updated using an adaptive optimization (AO) strategy. The adaptive optimization strategy estimates the first and second-order statistics of historical gradients and adaptively adjusts the update step size of each parameter. Through this method, while ensuring the continuity of gradient propagation, the convergence stability of the model under non-smooth robust loss constraints is improved, achieving iterative updates of network parameters. A maximum number of iterations is also set. As the upper limit of model training, during the iteration process, self Starting from the next iteration, the peak signal-to-noise ratio (MPSNR), mean square error (MMSE), and structural similarity (MSSIM) of the network output image and the original noisy image are calculated at fixed intervals in each band. The MPSNR, MMSE, and MSSIM are used as evaluation indicators in the model iteration process, and a stopping mechanism is set. The stopping mechanism is triggered when the change of the evaluation indicators is less than a set threshold in a number of consecutive iterations, at which point the iteration process is terminated. The peak signal-to-noise ratio (MPSNR) measures the overall reconstruction quality of the network output image and the original noisy image across various spectral bands. It is obtained by statistically analyzing the MPSNR of each band and reflects the fidelity of the denoising result in the spectral dimension. The mean squared error (MMSE) measures the average level of pixel-level differences between the network output image and the original noisy image across various bands; a smaller MMSE value indicates a lower reconstruction error. The multi-band structural similarity index (MSSIM) comprehensively evaluates the spatial structural consistency between the network output image and the original noisy image from aspects such as brightness, contrast, and structural information; a larger MSSIM value indicates a stronger ability to preserve structure.
[0068] Peak signal-to-noise ratio (MPSNR), mean square error (MMSE), and structural similarity (MSSIM) are used as quality evaluation indicators for multi-band hyperspectral images. They are used to comprehensively evaluate the stability and convergence of denoising results from three aspects: spectral consistency, numerical error, and spatial structure preservation.
[0069] Step S4. Hyperparameter setting; Based on the noise characteristics of the remote sensing image X, set the hyperparameters of the convolutional neural network denoising model, including: initial learning rate Lr, robustness threshold in the Huber loss function. Maximum number of iterations The parameters include the threshold values for changes in Peak Signal-to-Noise Ratio (MPSNR), Mean Square Error (MMSE), and Structural Similarity (MSSIM) in the stopping mechanism, as well as the detection interval. These parameters can be adaptively adjusted according to the noise characteristics of different satellite payload images.
[0070] The initial learning rate Lr controls the update step size of network parameters during the initial training phase of the model. Its magnitude affects the convergence speed and stability of the model in the initial iteration stage. Setting a reasonable initial learning rate can avoid oscillations or slow convergence during parameter updates and provides a basis for subsequent dynamic adjustment of the learning rate using adaptive optimization strategies. These hyperparameters can be adaptively adjusted according to the noise characteristics of different satellite payload images to improve the model's denoising robustness and reconstruction performance under different imaging conditions.
[0071] Step S5. Network inference and result output, the hyperspectral remote sensing image after preprocessing in step S1. The hyperparameters set in step S4 are input into the convolutional neural network denoising model constructed in step S2 until the stopping condition set in step S3 is met. The final output is the denoised hyperspectral image, which is the clean image. .
[0072] Although the embodiments of the present invention are described with reference to actual solutions, they do not constitute a limitation on the meaning of the present invention. Modifications to the embodiments and combinations with other solutions based on this specification will be obvious to those skilled in the art.
Claims
1. An unsupervised adaptive hyperspectral remote sensing image denoising method, characterized in that: It includes the following steps: Step S1. Input remote sensing image X, and normalize the input remote sensing image X so that the pixel values are mapped to the [0,1] interval to obtain a normalized hyperspectral remote sensing image. ; Step S2. Construct an unsupervised denoising framework and a denoising model. Step S2 includes the following steps: S21. Construct an unsupervised denoising framework by inputting the remote sensing image X into the degradation model and using a posterior probability estimation method to obtain the true clean image in the degradation model from the remote sensing image X. S22. Construct a convolutional neural network denoising model. The convolutional neural network denoising model consists of three main parts: an encoder, a decoder, and a skip connection layer. The encoder and decoder are connected through the skip connection layer. Step S3. The network parameters are iteratively updated using the Adam optimization algorithm combined with the backpropagation strategy, and a stopping mechanism is set. The stopping mechanism is triggered when the change of the preset evaluation index is less than the set threshold in a number of consecutive iterations, and the iteration process is terminated. Step S4. Based on the noise characteristics of the remote sensing image X, set the hyperparameters of the convolutional neural network denoising model; Step S5. Network inference and result output, the hyperspectral remote sensing image after preprocessing in step S1. The hyperparameters set in step S4 are input into the convolutional neural network denoising model constructed in step S2 until it meets the stopping condition set in step S3, and the final output is the denoised hyperspectral remote sensing image. The degradation model in step S21 is represented by formula (1): ， Where X represents remote sensing imagery, Represents the true clean image in the degradation model, It is mixed noise; A clean image is recovered from remote sensing image X using a posterior probability estimation method. The optimization objective can be expressed as formula (2): ， in, The estimated variables for a clean image are denoted as and the optimal solution is denoted as . Used to approximate R(Y) represents the prior regularization term, used to constrain the spectral consistency of hyperspectral images. The reconstruction error term is represented by the Huber loss function, as defined in formula (3): ， in, Representing remote sensing images The reconstruction residuals between the clean image estimate variable Y and the actual reconstruction residuals. This is a robust threshold used to suppress anomalous noise and preserve edge features; Denoising model using convolutional neural networks Replacing R(Y), equation (2) is equivalently transformed into equation (4): ， in, Indicates having parameters A convolutional neural network denoising model, where Z is a randomly generated noise matrix. The optimal network parameters are obtained through Huber loss and unsupervised adaptive optimization. This represents the optimal parameter convolutional neural network denoising model obtained after applying the Huber loss function and unsupervised adaptive optimization. The remote sensing image X is used as the unsupervised loss constraint, enabling the network to gradually approximate the clean image output during the iterative optimization process. , This is also known as denoised hyperspectral remote sensing imagery; The encoder is composed of a first DS Block module, a second DS Block module, a third DS Block module, a fourth DSBlock module, a fifth DS Block module, an ortho module, and a sixth DS Block module connected sequentially; the decoder is composed of a first US Block module, a second US Block module, a third US Block module, a fourth US Block module, a fifth US Block module, and a sixth US Block module connected sequentially; the skip connection layer is composed of a first SC Block module, a second SC Block module, a third SC Block module, a fourth SC Block module, a fifth SC Block module, and a sixth SCBlock module. The first SC Block module is connected between the first DS Block module and the first US Block module; the second SC Block module is connected between the second DS Block module and the second US Block module; the third SCBlock module is connected between the third DS Block module and the third US Block module; the fourth SC Block module is connected between the fourth DS Block module and the fourth US Block module; the fifth SC Block module is connected between the fifth DS Block module and the fifth US Block module; and the sixth SC Block module is connected between the sixth DS Block module and the sixth USBlock module. The first to sixth DS Block modules each consist of seven layers connected sequentially: the first Conv unit, the downsample unit, the first BN layer, the first Leaky ReLU activation function, the second Conv unit, the second BN layer, and the second Leaky ReLU activation function. The ortho module consists of seven major components: Conv Blocks, RNG, Gram-Schmid, Orthogonal Filters, Squeezed, Excitation, and Attention. The first to sixth US Block modules each consist of eight layers connected sequentially: the third BN layer, the third Conv unit, the fourth BN layer, the third Leaky ReLU activation function, the fourth Conv unit, the fifth BN layer, the fourth Leaky ReLU activation function, and the upsample unit. The first to sixth SC Block modules each consist of three components: the fifth Conv unit, the sixth BN layer, and the fifth Leaky ReLU activation function. The ortho module is a positive channel attention mechanism, and the channel compression process of the ortho module is shown in formula (6): ， The channel attention weights are obtained through the activation function, and the fusion features are finally input using the residual method. The ortho module is represented by formula (7): ， in, This represents the feature map tensor output by the preceding network layer. H represents the number of feature channels, and H and W represent the height and width of the feature map, respectively. This represents the feature response value of the feature map at the c-th channel and spatial location (h, w); The orthogonal weighting coefficients corresponding to the c-th channel in the ortho module are used to perform weighted projection of the channel features in the spatial dimension; h and w above represent the spatial height index and spatial width index of the feature map, respectively. This represents the weighted result of the feature map of the c-th channel in the spatial dimension; This represents a set composed of compressed feature vectors from each channel. The constructed channel-level global description vector, i.e. ; This represents the activation function, used to nonlinearly map channel compression features to generate channel attention weights; This represents the orthogonal channel attention weight vector output by the ortho module, used to adjust the feature map. Channel-by-channel weighting is performed, and feature fusion is achieved through residual connection. In step S3, the reconstruction error during the iteration process is calculated based on the Huber loss function from step S2. The network parameters are iteratively updated using the Adam optimization algorithm combined with a backpropagation strategy, and a maximum number of iterations is set. As the upper limit of model training, during the iteration process, self Starting from the next iteration, the peak signal-to-noise ratio (MPSNR), mean square error (MMSE), and structural similarity (MSSIM) of the network output image and the original noisy image are calculated at fixed intervals in each band. The MPSNR, MMSE, and MSSIM are used as evaluation indicators in the model iteration process, and a stopping mechanism is set. The stopping mechanism is triggered when the change of the evaluation indicators is less than a set threshold in a number of consecutive iterations, and the iteration process is terminated.
2. The unsupervised adaptive hyperspectral remote sensing image denoising method according to claim 1, characterized in that: The input size of the remote sensing image X is ,in , , These represent the number of rows, columns, and bands of the remote sensing image X, respectively.