Radar backscattering coefficient image super-resolution reconstruction method, device and medium based on conditional diffusion model

Abstract

The application discloses a radar backscattering coefficient image super-resolution reconstruction method and device based on a conditional diffusion model, and a medium, relates to the field of radar remote sensing image processing, and the method comprises the following steps: processing a high-resolution reference image to generate a pseudo-high-resolution image, and forming a training sample set; based on the training sample set, training a conditional diffusion super-resolution model; performing distribution alignment processing on the numerical distribution difference between the collected long-time sequence low-resolution radar backscattering coefficient image and the pseudo-high-resolution image, and generating a fused radar image; inputting the fused radar image, random noise and time step encoding information into the trained conditional diffusion model, performing an inverse diffusion denoising process, obtaining a residual reconstruction result, then superimposing the residual reconstruction result with the fused radar image, and generating a final high-resolution radar backscattering coefficient image. The application solves the problems of low spatial resolution, blurred texture and difficulty in reflecting regional scale surface heterogeneity of long-time sequence scatterometer radar products.

Description

Technical Field

[0001] This application relates to the field of radar remote sensing image processing, and in particular to a method, device and medium for super-resolution reconstruction of radar backscattering coefficient images based on a conditional diffusion model. Background Technology

[0002] Radar backscattering coefficient data plays a crucial role in long-term environmental monitoring tasks such as vegetation dynamics monitoring, soil moisture retrieval, and climate change analysis due to its all-weather, day-and-night imaging capabilities. Compared to optical remote sensing, radar systems can penetrate clouds and acquire observational data at night, making them more suitable for building long-term, continuous monitoring products of surface environmental changes. However, due to limitations in sensor payload and observation systems, long-term operational scatterometer radar products typically have low spatial resolution, and their spatial detail information is insufficient to meet the needs of regional or local-scale ecological and hydrological modeling applications.

[0003] With the development of remote sensing technology, high-resolution synthetic aperture radar (SAR) imagery (such as Sentinel-1) has been widely used for extracting surface structure information, providing detailed spatial textures and structural features. However, the acquisition time of this type of high-resolution data is relatively short, making it difficult to cover long-term timescales, thus limiting its application value in long-term sequence analysis and historical process reconstruction.

[0004] To enhance spatial details in low-resolution radar backscattering data, previous studies have attempted downscaling reconstruction using multi-source surface information or structural priors. However, these methods are highly dependent on external auxiliary information and parameter settings, and in areas with high surface heterogeneity or complex spatial structures, they are prone to problems such as texture blurring and unnatural transitions, making it difficult to stably characterize the true spatial distribution features of radar backscattering coefficients. With the increasing application of deep learning techniques in remote sensing image super-resolution reconstruction tasks, spatial detail restoration has improved to some extent. However, existing deep learning methods generally rely on paired high-resolution supervised samples and suffer from limited generalization ability and unstable reconstruction results in complex surface scenes, making it difficult to simultaneously maintain spatial texture detail and consistency in physical quantity distribution. Summary of the Invention

[0005] The purpose of this application is to provide a method, device and medium for super-resolution reconstruction of radar backscattering coefficient images based on a conditional diffusion model, which solves the problems of low spatial resolution, blurred texture and difficulty in reflecting regional-scale surface heterogeneity in long-time-series scatterometer radar products.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for super-resolution reconstruction of radar backscattering coefficient images based on a conditional diffusion model, including: A conditional diffusion super-resolution model is constructed by combining forward diffusion with noise addition and backward denoising reconstruction; in the backward denoising reconstruction process, a noise prediction network is used. The high-resolution reference image is processed to generate a pseudo-high-resolution image, and the high-resolution reference image and the pseudo-high-resolution image are combined to form a pair of training samples; multiple pairs of training samples constitute a training sample set. Based on the training sample set, the conditional diffusion super-resolution model is trained to obtain the trained conditional diffusion model. The numerical distribution differences between the acquired long-time low-resolution radar backscattering coefficient image and the pseudo-high-resolution image are processed to perform distribution alignment to generate a fused radar image. The fused radar image, random noise, and time step encoding information are input into the trained conditional diffusion model to perform inverse diffusion denoising and obtain residual reconstruction results. The residual reconstruction results are then superimposed on the fused radar image to generate the final high-resolution radar backscattering coefficient image.

[0007] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a radar backscattering coefficient image super-resolution reconstruction method based on a conditional diffusion model.

[0008] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a radar backscattering coefficient image super-resolution reconstruction method based on a conditional diffusion model.

[0009] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application uses long-term low-resolution radar backscattering coefficient images as the downscaling object and high-resolution reference images as training data. By constructing a conditional diffusion super-resolution model, the model is guided to learn the mapping relationship from low resolution to high resolution, realizing the modeling of residual information between high-resolution images and structurally degraded images. This allows the model to gradually recover high-resolution detail information under multi-time-step noise perturbation mechanisms. Addressing the problem of inconsistent value distribution between the two types of data, this application generates a fused radar image through distribution alignment processing. This image simultaneously possesses the spatial texture features of the high-resolution reference image and the radiometric statistical characteristics of the long-term low-resolution scatterometer data, serving as the conditional input for the diffusion model's inference stage, thereby achieving spatial detail reconstruction and preservation of physical consistency. Finally, without needing to acquire long-term pixel-level high-resolution ground truth data, a high spatial resolution long-term radar backscattering coefficient reconstructed image is output, achieving spatial resolution enhancement of long-term low-resolution radar backscattering coefficient images and improving the structural continuity and numerical consistency of the reconstruction results. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is an application environment diagram of the radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model in one embodiment of this application.

[0012] Figure 2 This is a flowchart illustrating a radar backscattering coefficient image super-resolution reconstruction method based on a conditional diffusion model in one embodiment of this application.

[0013] Figure 3 This is a structural diagram of the noise prediction network in the super-resolution model.

[0014] Figure 4 A flowchart of the steps for building the training dataset and training the super-resolution model.

[0015] Figure 5 This is a flowchart of the distributed migration strategy and image super-resolution reconstruction.

[0016] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] As a probabilistic generative model, the diffusion model demonstrates strong texture restoration and structure fidelity in image generation and super-resolution reconstruction, and can alleviate the shortcomings of traditional deep learning models in detail representation and uncertainty modeling to some extent. However, applying the diffusion model to the spatial downscaling reconstruction of long-term radar backscattering products is still in the exploratory stage. How to maintain the consistency of radar physical quantity distribution characteristics while improving spatial resolution requires further research.

[0019] To address the shortcomings of existing technologies in enhancing the spatial resolution of long-term low-resolution radar backscatter data, which cannot yet meet the practical needs of high-precision remote sensing interpretation and refined analysis of surface processes, this application proposes a conditional diffusion model that integrates high-resolution spatial structure information with the physical distribution characteristics of long-term radar data for super-resolution reconstruction of remote sensing images. This aims to achieve a synergistic improvement in both spatial resolution and physical consistency of long-term radar backscatter products.

[0020] Specifically, this application uses long-term, low-resolution scatterometer data as the downscaling object and Sentinel-1A high-resolution radar imagery as training data. A conditional diffusion model training framework is constructed to guide the model in learning the mapping relationship from low resolution to high resolution. To address the issue of inconsistent value distribution between the two types of data, a distribution alignment strategy is proposed to generate a fused radar image that simultaneously possesses the spatial texture features of Sentinel-1 and the radiometric statistical characteristics of long-term, low-resolution scatterometer data. This image serves as the conditional input for the diffusion model's inference stage, thereby achieving the reconstruction of spatial details and the preservation of physical consistency. Finally, a high spatial resolution long-term radar backscattering coefficient dataset is output. This application effectively solves the problem of low spatial resolution in long-term radar data under complex land cover, providing high-quality data support for ecological monitoring and hydrological simulation.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server. Terminal 101 can send high-resolution reference images and acquired long-term low-resolution radar backscattering coefficient images to server 102. After receiving the data, server 102 constructs a conditional diffusion super-resolution model, processes the high-resolution reference images to obtain a training sample set, and trains the conditional diffusion super-resolution model based on the training sample set. It then performs distribution alignment processing on the numerical distribution differences between the acquired long-term low-resolution radar backscattering coefficient images and pseudo-high-resolution images to generate a fused radar image. The fused radar image is input into the trained conditional diffusion model, and an inverse diffusion denoising process is performed to obtain a residual reconstruction result. The residual reconstruction result is superimposed on the fused radar image to generate the final high-resolution radar backscattering coefficient reconstructed image. Server 102 can feed back the obtained final high-resolution radar backscattering coefficient reconstructed image to terminal 101. Furthermore, in some embodiments, the radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model can also be implemented separately by the server 102 or the terminal 101.

[0023] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 102 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0024] In one exemplary embodiment, such as Figure 2 As shown, a method for super-resolution reconstruction of radar backscattering coefficient images based on a conditional diffusion model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps 201 to 205.

[0025] Step 201: Combine forward diffusion with noise addition and reverse denoising reconstruction to construct a conditional diffusion super-resolution model; wherein, a noise prediction network is used in the reverse denoising reconstruction process.

[0026] In a specific application, the conditional diffusion super-resolution model includes a forward diffusion modeling stage and a reverse denoising and reconstruction stage. In the reverse denoising and reconstruction stage, a noise prediction network is used to model the noise components in the diffusion process, thereby achieving the gradual recovery of residual details.

[0027] In the forward diffusion modeling stage, the residual image between the conditional input image and the high-resolution image is used as the diffusion object. By constructing a diffusion modeling mechanism based on residual learning, the conditional diffusion super-resolution model focuses on learning high-frequency detail compensation information. Gaussian noise is gradually injected into the diffusion object at multiple time steps, causing the residual image to gradually degenerate into an approximately random noise distribution, thereby obtaining a noisy residual image at multiple time steps.

[0028] During the model training phase, the conditional input images and high-resolution images are pseudo-high-resolution images and high-resolution reference images, respectively; during the model inference and application phase, only the conditional input image needs to be fused radar image.

[0029] In the inverse denoising and reconstruction stage, the noisy residual images from multiple time steps, the conditional input image, and the time step encoding information are input into the noise prediction network. The noise components corresponding to each time step are estimated, and the residual information is gradually recovered to reconstruct the high-resolution radar backscattering coefficient image. Specifically, noise prediction is performed on the image at any time step, and then an inverse denoising process is executed. The residual information is gradually recovered from the random noise, and the high-resolution radar backscattering coefficient image is reconstructed under the constraints of the conditional input image.

[0030] Through the above two stages of processing, a super-resolution reconstruction framework for radar backscattering coefficient images based on the conditional diffusion model was constructed.

[0031] like Figure 3 As shown, the noise prediction network includes a downsampling group, an intermediate processing block, and an upsampling group; the downsampling group, the intermediate processing block, and the upsampling group are sequentially cascaded and form a U-Net-based encoder-decoder symmetric structure, and feature transfer between the encoder and decoder is realized through a skip connection structure.

[0032] The downsampling group comprises several cascaded structures, specifically including multiple downsampling blocks and a final-level downsampling block. Each downsampling block includes multiple cascaded residual blocks and a downsampling layer. The final-level downsampling block includes multiple cascaded residual blocks but does not have a downsampling layer to maintain the continuity of the feature map's spatial scale. The downsampling blocks and the final-level downsampling block are used to extract multi-scale feature representations from the received image level by level. Specifically, the downsampling group may include three downsampling blocks and a final-level downsampling block, with each downsampling block containing two cascaded residual blocks and a downsampling layer.

[0033] The intermediate processing module is located between the encoder (downsampling group) and the decoder (upsampling group). The intermediate processing block includes two residual blocks and an attention mechanism block is embedded between the two residual blocks. The intermediate processing block is used to model global context information on the received feature map. The attention mechanism block is used to adaptively weight features at different spatial locations and channel dimensions in the received feature map, thereby enhancing the network's ability to model global context information and complex semantic features.

[0034] The upsampling and downsampling groups are structurally symmetrically distributed and used to progressively restore the spatial resolution of the feature map during the decoding stage. Each upsampling group includes multiple upsampling blocks and a final-level upsampling block; each upsampling block includes multiple cascaded residual blocks and an upsampling layer; the final-level upsampling block includes multiple cascaded residual blocks and does not have an upsampling layer to ensure that the spatial size of the output feature map remains consistent with the network input. The upsampling blocks and the final-level upsampling block are used to progressively restore the spatial resolution and reconstruct detailed information from the received feature map. Specifically, the upsampling group may include three upsampling blocks and one final-level upsampling block, with each upsampling block containing two cascaded residual blocks and an upsampling layer. The feature map size output by the upsampling group remains consistent with the network input.

[0035] The residual block comprises multiple convolutional blocks; each convolutional block includes a normalization layer, a nonlinear activation function layer, and a convolutional layer arranged sequentially. That is, each residual block employs a two-layer convolutional structure, and a nonlinear activation function is introduced between convolutional operations to enhance the network's ability to nonlinearly represent complex textures and edge features.

[0036] By setting up a symmetrical encoder-decoder structure that includes multi-level downsampling, intermediate processing modules, and multi-level upsampling, the noise prediction network can acquire spatial information at different scales during the feature extraction stage and restore spatial resolution step by step during the decoding stage. By introducing an attention mechanism in the intermediate processing module, the feature information is adaptively weighted, enhancing the modeling ability of high-frequency texture information and structural details. Overall, the structure preservation ability and detail reconstruction accuracy of radar backscattering coefficient images during the resolution enhancement process are improved.

[0037] Step 202: Process the high-resolution reference image to generate a pseudo-high-resolution image, and combine the high-resolution reference image and the pseudo-high-resolution image to form a pair of training samples; multiple pairs of training samples constitute a training sample set.

[0038] In one application, high-resolution reference images (such as high-resolution synthetic aperture radar images) are downsampled and upsampled to generate pseudo-high-resolution images, which are then used to obtain training samples. Correspondingly, such as... Figure 4As shown, step 202 includes the following steps (21)-(23).

[0039] (21) Downsampling is performed on the high-resolution reference image to obtain a low-resolution image. Specifically, Sentinel-1A synthetic aperture radar VV polarization data is selected and aggregated to a spatial resolution of 2km as a high-resolution reference image. Downsampling is performed according to a preset magnification to obtain the corresponding low-resolution image.

[0040] (22) Upsample the low-resolution image to generate a pseudo-high-resolution image (PHR) corresponding to the high-resolution reference image. The pseudo-high-resolution image maintains the overall spatial structure and texture of the original high-resolution SAR image while weakening local high-frequency details, and is used as a structural guidance condition for the subsequent conditional diffusion model.

[0041] (23) The high-resolution reference image and the corresponding pseudo-high-resolution image are spatiotemporally aligned to obtain paired training samples, that is, paired image patch samples, for training the conditional diffusion super-resolution model.

[0042] Step 203: Based on the training sample set, the conditional diffusion super-resolution model is trained to obtain the trained conditional diffusion model. In application, a residual diffusion training mechanism is constructed based on the training sample set, and the parameters of the conditional diffusion super-resolution model are optimized by minimizing the difference between the predicted noise and the real noise. In one embodiment, step 203 includes the following steps (31)-(34).

[0043] (31) For any pair of training samples in the training sample set, calculate the difference between the high-resolution reference image and the pseudo-high-resolution image to obtain a residual image, and use the residual image as the diffusion object.

[0044] (32) Gaussian noise is gradually injected into the residual image at multiple time steps to obtain a noisy residual image at multiple time steps.

[0045] (33) Input the noisy residual image of multiple time steps, the pseudo high-resolution image and time step coding information into the noise prediction network, estimate the noise component corresponding to each time step, and obtain the predicted noise.

[0046] (34) Construct a loss function based on the difference between the predicted noise and the actual noise, and iteratively update the model parameters of the conditional diffusion super-resolution model using the backpropagation algorithm until the model converges to obtain the trained conditional diffusion model; wherein, the actual noise refers to the Gaussian noise. That is, by minimizing the loss function between the predicted noise and the actual noise, iteratively update the model parameters until the model converges to obtain the trained conditional diffusion model.

[0047] This application constructs pseudo-high-resolution images as structural guidance conditions, enabling the conditional diffusion super-resolution model to explicitly learn the residual mapping relationship between high-resolution images and structurally degraded images during training, thereby enhancing its ability to model high-frequency detail compensation information. By constructing a progressively noisy forward diffusion process on the residual information and modeling the multi-time-step noise prediction process under conditional constraints, the model can characterize the recovery rules of complex textures and spatial structural details under different noise levels. Overall, it improves the structural consistency and detail integrity of the high-resolution reconstruction results, providing a reliable model foundation for the high-resolution reconstruction of radar backscattering coefficient images.

[0048] Step 204: Perform distribution alignment processing on the numerical distribution differences between the acquired long-time-series low-resolution radar backscattering coefficient image and the pseudo-high-resolution image to generate a fused radar image. Here, "long-time-series" refers to an image whose temporal values ​​are higher than a preset value.

[0049] In a practical application, step 204 includes the following steps (41)-(43).

[0050] (41) Perform parameterized probability distribution modeling (e.g., Beta distribution modeling) on ​​the pixel values ​​of the pseudo-high resolution image and the pixel values ​​of the long-time low resolution radar backscattering coefficient image respectively, and obtain the corresponding probability distribution parameters.

[0051] (42) Based on two probability distribution parameters, a chain distribution transformation model is constructed. When the parameterized probability distribution is a Beta distribution, the chain distribution mapping transformation process of the chain distribution transformation model includes: mapping the pixel values ​​of the pseudo-high resolution image of the Beta distribution as variables to the Fisher distribution space; using an approximate transformation method to map the variables of the Fisher distribution space to the standard normal distribution space; performing an inverse transformation based on the target Beta distribution parameters corresponding to the low-resolution radar backscattering coefficient image, and sequentially mapping the variables of the standard normal distribution space back to the target Fisher distribution space and the target Beta distribution space.

[0052] like Figure 5 As shown, let the pseudo-high resolution image be... The corresponding random variable is denoted as Long-time low-resolution radar backscattering coefficient image is The corresponding random variable is denoted as To construct a reversible and parameterized numerical distribution mapping relationship to achieve statistical consistency transformation between different images, probability distribution modeling of image pixel values ​​is performed.

[0053] Considering that the image data, after normalization, exhibits statistical characteristics that match the Beta distribution, this embodiment uses the Beta distribution for parametric modeling. Statistical fitting yields the following: .

[0054] in, , These are two positive parameters of the Beta distribution.

[0055] The objective of this embodiment is: to maintain Without changing the spatial structure, align its pixel value distribution to... Consistency, thereby constructing fused conditional images The specific implementation method is as follows: (1) Beta Distribution to Fisher Distribution: To facilitate subsequent standardization, the Beta distribution variable is mapped to the Fisher distribution variable. The mapping function is defined as follows: .

[0056] Then Mapped to: .

[0057] Under this transformation, Follows Fisher distribution: .

[0058] (2) Mapping from Fisher distribution to standard normal distribution: After completing the mapping from Beta distribution to Fisher distribution, in order to eliminate the influence of the difference in different distribution parameters on the subsequent numerical migration process, a mapping process from Fisher distribution space to standard normal distribution space is further constructed.

[0059] In this embodiment, the Paulson approximation transform method is used to convert the Fisher-distributed random variable into a standard normally distributed random variable. The specific expression is as follows: .

[0060] .

[0061] .

[0062] and The function and Defined as: .

[0063] Under the action of the aforementioned mapping function, the mapping of Fisher-distributed random variables can be completed. To standard normal distribution random variable The conversion is expressed as follows: .

[0064] .

[0065] .

[0066] Through the above transformation, the Fisher-distributed random variable is mapped to the standard normal distribution space, obtaining a unified intermediate standardized representation, which provides a foundation for the subsequent inverse mapping process under the target distribution parameters.

[0067] (3) Inverse mapping from standard normal distribution to Fisher distribution: In the previous steps, the standard normal random variable has been obtained. .

[0068] To reconstruct the standard normal distribution space into the target Fisher distribution space, an inverse approximation transformation process is further constructed. In the application, the Paulson inverse approximation method is used to realize the mapping from the standard normal distribution to the Fisher distribution. This inverse mapping relationship can be represented by the following implicit equation, which is used to characterize the standard normal random variable. Fisher-distributed random variables The correspondence between them: .

[0069] The parameters are defined as follows: .

[0070] .

[0071] And the function and The steps remain consistent with those described above. To facilitate the solution, variable substitution is introduced: .

[0072] The expression can then be transformed into a statement about Standard quadratic equation form: .

[0073] This equation can be solved using the quadratic formula, and the solution can be expressed as: .

[0074] in: ,and Represents the different root branches of a quadratic equation.

[0075] In practical applications, let And using the parameters corresponding to the target Fisher distribution: .

[0076] .

[0077] This yields the corresponding Fisher distribution random variable. Ultimately, via standard normal random variables... Reconstructed random variables Follows Fisher distribution: .

[0078] For each sample point The corresponding Fisher distribution sample According to The symbol is selected to be used for calculation corresponding to the root branch, that is: .

[0079] Through the above inverse transformation process, random variables in the standard normal distribution space can be reconstructed in a statistical sense into target Fisher distribution random variables, thereby completing the closed-loop construction of the "standardization-inverse standardization" link in the chain distribution migration process.

[0080] (4) Fisher distribution to Beta distribution: Further mapping back to the Beta distribution. Define the mapping function: .

[0081] Then we get the aligned Beta variable: .

[0082] thereby In a statistical sense, it is related to the target distribution. Consistent.

[0083] (43) Based on the chain-like distribution transformation model, the pixel value distribution of the pseudo-high-resolution image is transformed into a target statistical distribution consistent with the pixel value distribution of the long-time-series low-resolution radar backscattering coefficient image, thereby generating a fused radar image. Specifically, through the above-mentioned chain-like mapping, the numerical distribution of the pseudo-high-resolution image is migrated to the numerical distribution of the long-time-series low-resolution radar backscattering coefficient image, thereby generating a fused radar image.

[0084] Based on the chain distribution transformation mapping relationship of parameterized distribution, the pixel values ​​of pseudo-high resolution images are mapped from the source distribution space to the intermediate normalized distribution space. According to the target distribution parameters corresponding to the long-term low-resolution radar backscattering coefficient images, the normalized variables are inversely mapped to the target distribution space. Thus, numerical distribution alignment is completed while keeping the spatial texture structure of pseudo-high resolution images unchanged, and fused radar images are generated.

[0085] In other words, to address the difference in numerical distribution between long-term low-resolution (scatterometer) radar backscattering coefficient images and pseudo-high-resolution images, a distribution alignment strategy is constructed. Through statistical distribution modeling and chain-like distribution mapping, the numerical distribution of the pseudo-high-resolution images is transformed into a target statistical distribution consistent with that of the long-term low-resolution radar backscattering coefficient images. Under the premise of maintaining the spatial structure information of the pseudo-high-resolution images, a fusion conditional input image is constructed, thus obtaining the fused radar image.

[0086] Step 205: Input the fused radar image, random noise, and time step encoding information into the trained conditional diffusion model, perform inverse diffusion denoising process, and obtain residual reconstruction result; superimpose the residual reconstruction result with the fused radar image to generate the final high-resolution radar backscattering coefficient image.

[0087] Specifically, the fused radar image is used as a conditional input image and input together with random noise and time step coding information into the noise prediction network of the trained conditional diffusion model; noise is predicted step by step through back diffusion and noise components are removed to restore residual information and obtain residual reconstruction results; the residual reconstruction results are superimposed with the fused radar image to generate the final high-resolution radar backscattering coefficient image.

[0088] The LHScat long-time series low-resolution radar backscattering coefficient image is selected as the input data source. Based on the aforementioned distribution alignment strategy, the numerical distribution of the pseudo-high-resolution image is converted into a statistical distribution consistent with the LHScat data, constructing a fused radar image (FRI). The fused radar image, noise term (i.e., random noise), and time step information (i.e., time step encoding information) are used as conditional inputs and fed into the U-Net network in the trained conditional diffusion model to perform the backdiffusion generation process. During the backdiffusion process, the U-Net network at each diffusion time step... The noise components in the current input state are predicted, and the corresponding denoising results are gradually constructed based on the prediction results. As the diffusion time step decreases progressively from the initial time step, the model gradually weakens the influence of noise components on the results, achieving a gradual evolution from a high-noise state to a low-noise state. When the diffusion time step advances to... When the noise-free condition is met, the prediction output result corresponding to the noise-free condition is obtained.

[0089] Because the U-Net network uses the residual information between the high-resolution reference image and the pseudo-high-resolution image as the learning target during the training phase, it performs a back-diffusion process during the inference phase. The output obtained at each step essentially corresponds to the detailed residual information of the target's high-resolution radar backscattering coefficient image relative to the fused radar image. Furthermore, this residual result is overlaid pixel-by-pixel with the fused radar image to reconstruct the final high-resolution radar backscattering coefficient image.

[0090] By introducing a distribution migration strategy, the consistency of numerical distribution and the prior spatial structure are effectively integrated, thereby satisfying both physical statistical consistency and structural expressiveness in the generation stage. On this basis, a stepwise reverse denoising inference is performed using a post-trained conditional diffusion model, which enables high-resolution details to be gradually restored under conditional constraints, avoiding numerical offset and texture distortion problems caused by direct reconstruction. Overall, the stability and reliability of the reconstruction results in terms of spatial resolution and texture detail preservation are improved.

[0091] This application constructs a conditional diffusion super-resolution reconstruction framework based on residual modeling to model the residual information between high-resolution images and structurally degraded images, enabling the model to gradually recover high-resolution detail information under multi-time-step noise perturbation mechanisms. By introducing a distribution alignment strategy, the fusion conditional input image in the inference stage simultaneously possesses high-resolution spatial structural features and statistical distribution characteristics of long-term low-resolution radar products, thereby completing image reconstruction under both structural constraints and distribution consistency conditions. Finally, without acquiring long-term pixel-level high-resolution ground truth data, the spatial resolution of long-term low-resolution radar backscattering coefficient images is enhanced, improving the structural continuity and numerical consistency of the reconstruction results.

[0092] Based on the same inventive concept, this application also provides a system. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more system embodiments provided below can be found in the limitations of the method above, and will not be repeated here.

[0093] In one exemplary embodiment, a radar backscattering coefficient image super-resolution reconstruction system based on a conditional diffusion model is provided, comprising: The model building module is used to construct a conditional diffusion super-resolution model by combining forward diffusion with denoising and backward denoising reconstruction; in the backward denoising reconstruction process, a noise prediction network is used.

[0094] The training sample set construction module is used to process the high-resolution reference image to generate a pseudo-high-resolution image, and combine the high-resolution reference image and the pseudo-high-resolution image to form a pair of training samples; multiple pairs of training samples constitute a training sample set.

[0095] The model training module is used to train the conditional diffusion super-resolution model based on the training sample set to obtain the trained conditional diffusion model.

[0096] The distribution alignment strategy processing module is used to perform distribution alignment processing on the numerical distribution differences between the acquired long-time-series low-resolution radar backscattering coefficient image and the pseudo-high-resolution image to generate a fused radar image.

[0097] The high-resolution reconstruction module is used to input the fused radar image, random noise, and time step encoding information into the trained conditional diffusion model, perform inverse diffusion denoising process, and obtain residual reconstruction results; the residual reconstruction results are superimposed on the fused radar image to generate the final high-resolution radar backscattering coefficient image.

[0098] This application utilizes a conditional diffusion model to reconstruct long-time low-resolution radar backscattering coefficient (σ) data with spatial resolution enhancement. By introducing the conditional diffusion model and combining it with a distribution alignment strategy, the spatial resolution of long-term low-resolution radar backscattering products is enhanced while maintaining the consistency between numerical distribution and physical properties. This solves the problems of low spatial resolution and missing spatial structure texture in long-time low-resolution scatterometer radar backscattering coefficient (σ) images.

[0099] This application proposes a super-resolution reconstruction method for radar backscattering coefficients by constructing a fusion conditional diffusion model and a distribution alignment strategy. Under constrained conditions, it models and progressively denoises and reconstructs high-resolution detail residuals, achieving a synergistic improvement in spatial detail enhancement and numerical distribution consistency. Compared to traditional downscaling methods based on interpolation, regularization constraints, or statistical regression, this method does not rely on empirical models or manual weight settings. Instead, it introduces conditional information images to constrain the generation process, enabling the model to effectively learn the spatial structure mapping relationship between low-resolution and high-resolution images, mitigating texture blurring and structural distortion in complex surface areas. A residual diffusion training mechanism is constructed based on training samples, and the parameters of the conditional diffusion model are optimized by minimizing the difference between predicted and actual noise. The distribution alignment strategy achieves statistical distribution matching between the reconstructed results and the original long-term radar products while maintaining spatial texture structure stability, thus avoiding the problem of insufficient physical consistency despite visual enhancement. Ultimately, without requiring strictly paired high- and low-resolution samples, it achieves stable and high-quality spatial resolution enhancement of long-term radar backscattering coefficient products, improving their practicality and reliability in applications such as ecological monitoring and hydrological analysis.

[0100] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a radar backscattering coefficient image super-resolution reconstruction method based on a conditional diffusion model.

[0101] Those skilled in the art will understand that Figure 6 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.

[0102] In one exemplary embodiment, a computer device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the above-described method embodiments.

[0103] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0104] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0106] Those skilled in the art will understand that all or part of the processes in 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. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0107] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0108] 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.

[0109] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for super-resolution reconstruction of radar backscattering coefficient images based on a conditional diffusion model, characterized in that, The method includes: A conditional diffusion super-resolution model is constructed by combining forward diffusion with noise addition and backward denoising reconstruction; in the backward denoising reconstruction process, a noise prediction network is used. The high-resolution reference image is processed to generate a pseudo-high-resolution image, and the high-resolution reference image and the pseudo-high-resolution image are combined to form a pair of training samples; multiple pairs of training samples constitute a training sample set. Based on the training sample set, the conditional diffusion super-resolution model is trained to obtain the trained conditional diffusion model. The numerical distribution differences between the acquired long-time low-resolution radar backscattering coefficient image and the pseudo-high-resolution image are processed to perform distribution alignment to generate a fused radar image. The fused radar image, random noise, and time step encoding information are input into the trained conditional diffusion model to perform inverse diffusion denoising and obtain residual reconstruction results. The residual reconstruction results are then superimposed on the fused radar image to generate the final high-resolution radar backscattering coefficient image.

2. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 1, characterized in that, The conditional diffusion super-resolution model includes a forward diffusion modeling stage and a reverse denoising and reconstruction stage. In the forward diffusion modeling stage, the residual image between the conditional input image and the high-resolution image is used as the diffusion object; Gaussian noise is gradually injected into the diffusion object at multiple time steps to obtain a noisy residual image at multiple time steps. In the reverse denoising and reconstruction stage, the noisy residual images of multiple time steps, the conditional input images, and the time step coding information are input into the noise prediction network to estimate the noise components corresponding to each time step and gradually recover the residual information to reconstruct the high-resolution radar backscattering coefficient image.

3. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 1, characterized in that, The noise prediction network includes a downsampling group, an intermediate processing block, and an upsampling group; the downsampling group, the intermediate processing block, and the upsampling group are sequentially cascaded and form an encoder-decoder symmetrical structure, and feature transfer between the encoder and the decoder is realized through a skip connection structure. The downsampling group includes multiple downsampling blocks and a final-level downsampling block; the downsampling block includes multiple cascaded residual blocks and a downsampling layer; the final-level downsampling block includes multiple cascaded residual blocks; the downsampling blocks and the final-level downsampling block are used to extract multi-scale feature representations from the received image step by step; The intermediate processing block includes two residual blocks, with an attention mechanism block embedded between them; the intermediate processing block is used to model global context information of the received feature map; the attention mechanism block is used to adaptively weight features at different spatial locations and channel dimensions in the received feature map; The upsampling group includes multiple upsampling blocks and a final-level upsampling block; the upsampling block includes multiple cascaded residual blocks and an upsampling layer; the final-level upsampling block includes multiple cascaded residual blocks; the upsampling blocks and the final-level upsampling block are used to recover the spatial resolution and reconstruct the detail information of the received feature map step by step.

4. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 1, characterized in that, The high-resolution reference image is processed to generate a pseudo-high-resolution image, and the high-resolution reference image and the pseudo-high-resolution image are combined to form paired training samples, including: The high-resolution reference image is downsampled to obtain the low-resolution image. The low-resolution image is upsampled to generate a pseudo-high-resolution image corresponding to the high-resolution reference image. The high-resolution reference image and the corresponding pseudo-high-resolution image are spatiotemporally aligned to obtain paired training samples.

5. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 1, characterized in that, Based on the training sample set, the conditional diffusion super-resolution model is trained to obtain a trained conditional diffusion model, including: For any pair of training samples in the training sample set, the difference between the high-resolution reference image and the pseudo-high-resolution image is calculated to obtain the residual image; Gaussian noise is gradually injected into the residual image at multiple time steps to obtain a noisy residual image at multiple time steps; The noisy residual image at multiple time steps, the pseudo-high resolution image, and the time step encoding information are input into the noise prediction network to estimate the noise components corresponding to each time step and obtain the predicted noise. A loss function is constructed based on the difference between the predicted noise and the actual noise, and the model parameters of the conditional diffusion super-resolution model are iteratively updated using the backpropagation algorithm until the model converges, thus obtaining the trained conditional diffusion model; wherein, the actual noise refers to the Gaussian noise.

6. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 1, characterized in that, The numerical distribution differences between the acquired long-time-series low-resolution radar backscattering coefficient image and the pseudo-high-resolution image are aligned to generate a fused radar image, including: Parametric probability distribution models were performed on the pixel values ​​of pseudo-high resolution images and the pixel values ​​of long-time series low-resolution radar backscattering coefficient images to obtain their respective probability distribution parameters. Based on two probability distribution parameters, a chain distribution transformation model is constructed; Based on the chain-like distribution transformation model, the pixel value distribution of the pseudo-high resolution image is transformed into a target statistical distribution that is consistent with the pixel value distribution of the long-time-series low-resolution radar backscattering coefficient image, thereby generating a fused radar image.

7. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 6, characterized in that, The parameterized probability distribution is a Beta distribution; The chain distribution transformation model's chain distribution mapping transformation process includes: The pixel values ​​of the pseudo-high resolution image with Beta distribution are mapped as variables to the Fisher distribution space. An approximate transformation method is used to map the variables in the Fisher distribution space to the standard normal distribution space; Based on the target Beta distribution parameters corresponding to the long-time low-resolution radar backscattering coefficient image, an inverse transformation is performed to map the variables of the standard normal distribution space back to the target Fisher distribution space and the target Beta distribution space in sequence.

8. The radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to claim 3, characterized in that, The residual block includes multiple convolutional blocks; The convolutional block includes a normalization layer, a nonlinear activation function layer, and a convolutional layer arranged sequentially.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the radar backscattering coefficient image super-resolution reconstruction method based on the conditional diffusion model as described in any one of claims 1-8.

Images