A Super-Resolution Reconstruction Method for Typhoon Cloud Image Sequences Based on Fluid Dynamics Constraints

Abstract

This invention discloses a super-resolution reconstruction method for typhoon cloud image sequences based on fluid dynamics constraints, belonging to the field of image super-resolution reconstruction. The method includes: acquiring the original typhoon cloud image sequence of the satellite observation area; preprocessing the typhoon cloud image sequence to obtain a low-resolution image sequence; constructing a super-resolution reconstruction network model based on fluid dynamics constraints; training and optimizing the super-resolution reconstruction network model using a training set; and inputting the low-resolution image sequence from the validation set into the trained super-resolution reconstruction network model to obtain the reconstructed high-resolution image. This invention can effectively maintain the structural coherence and dynamic continuity of typhoon cloud images, and is suitable for the generation and analysis of high-precision cloud image sequences in fields such as weather forecasting and satellite remote sensing, possessing strong practical value and promising prospects for application.

Description

Technical Field

[0001] This invention relates to the field of image super-resolution reconstruction, and in particular to a method for super-resolution reconstruction of typhoon cloud image sequences based on hydrodynamic constraints. Background Technology

[0002] Typhoons are large-scale meteorological phenomena that typically require timely and effective forecasting to minimize their potential impacts. Sequenced satellite images of cloud systems during typhoon formation and development provide extensive spatial coverage and high-frequency temporal resolution, making them crucial for analyzing typhoon formation, progression, and evolution. By analyzing cloud morphology and structure in satellite cloud images, meteorologists can assess typhoon intensity and identify key features, playing a vital role in predicting typhoon tracks and revealing factors influencing typhoon evolution.

[0003] With the rapid development of artificial intelligence technology, super-resolution (SR) models have been widely used in meteorological image processing in recent years to obtain clearer and higher-resolution typhoon cloud image resources. Super-resolution technology effectively improves the spatial resolution of images by recovering more detailed information from low-resolution images. Meanwhile, research shows that there is a close correlation between the temporal and spatial resolution of images; by utilizing multi-frame image sequences, it is possible to recover the missing detailed information in a single frame, thereby achieving higher-quality image reconstruction.

[0004] Optical flow estimation, as one of the key technologies in multi-frame super-resolution, calculates the relative motion between consecutive frames to achieve motion compensation, aligning images at different time points and greatly improving image fusion results. Although optical flow technology has made significant progress in the field of image super-resolution, current mainstream optical flow algorithms still face many challenges in handling complex fluid dynamics and are difficult to meet the special requirements of non-rigid deformation and rapid flow in typhoon cloud images.

[0005] Traditional optical flow algorithms are typically based on the assumption of constant brightness, assuming that image grayscale values ​​remain unchanged between adjacent frames. However, in scenarios with strong non-rigid deformation and fluid dynamics, such as typhoon cloud images, complex fluid boundaries, and ever-changing motion characteristics, this assumption often fails, thus affecting the accuracy of optical flow estimation. The resulting motion estimation errors directly limit the effectiveness and accuracy of super-resolution reconstruction. Summary of the Invention

[0006] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a super-resolution reconstruction method for typhoon cloud image sequences based on fluid dynamics constraints. By constructing an optical flow module tailored to the dynamic characteristics of fluids, the limitations of the traditional brightness uniformity assumption are overcome, and the complex motion characteristics of clouds are accurately captured.

[0007] Technical solution: The present invention provides a super-resolution reconstruction method for typhoon cloud image sequences based on fluid dynamics constraints, comprising the following steps:

[0008] The original typhoon cloud image sequence of the satellite observation area is obtained, the typhoon cloud image sequence is preprocessed to obtain a low-resolution image sequence, the original typhoon cloud image sequence is used as label data, and the low-resolution image sequence is divided into training set and validation set.

[0009] A super-resolution reconstruction network model based on fluid dynamics constraints was constructed, and the super-resolution reconstruction network model was trained and optimized using a training set.

[0010] The low-resolution image sequences in the validation set are input into the trained super-resolution reconstruction network model to obtain the reconstructed high-resolution images.

[0011] Furthermore, the super-resolution reconstruction network model is a combination of... A multi-step bidirectional grid data propagation architecture with parallel optical flow units. Each optical flow unit receives typhoon cloud image data at one time point. Each optical flow unit includes four sequentially arranged optical flow modules. Each optical flow module receives time information of two step sizes, and the typhoon cloud image data is propagated back and forth in time in an alternating manner. It is a natural number greater than 4.

[0012] Furthermore, the steps for training and optimizing the super-resolution reconstruction network model using the training set include:

[0013] For any optical flow unit, the typhoon cloud image data at the current time t is denoted as... Typhoon cloud image data Image features are extracted using residual modules composed of convolutions to obtain fluid image features. , representing the two-dimensional coordinates at time t. Image intensity at that location;

[0014] Fluid image features As an input item to the optical flow unit, it is input into the first-layer optical flow module, where Features representing fluid image propagation in a single step. The fluid image features representing two-step propagation are obtained through two O-transforms in the first-layer optical flow module. and Optical flow between and image features and Optical flow between The o-transform represents the calculation of the optical flow transformation between two sets of features;

[0015] The obtained optical flow , and fluid image features at the current moment Input them together into the spatiotemporal alignment module to obtain alignment features. ;in The transformation represents the alignment operation between the two sets of optical flow features calculated by the o-transform and the features at the current time. The spatiotemporal alignment module consists of three convolutions.

[0016] Align features and fluid image features Data features are concatenated and then fused using convolution to obtain the first layer of fused motion information features. ,in and Features belonging to the same type;

[0017] Features After being processed sequentially by the second, third, and fourth optical flow modules, the features of the fourth layer fused motion information are obtained. ;

[0018] Features Pixel-Shuffle upsampling was performed, along with data from typhoon cloud imagery. After performing bilinear interpolation upsampling on the feature map, a reconstructed high-resolution image is generated. The formula is:

[0019] ,

[0020] In the formula, Up represents the Pixel-Shuffle upsampling operation, and B represents the bilinear interpolation upsampling operation.

[0021] Furthermore, the implementation process of the o-transform includes:

[0022] S21, Construct the optical flow estimation prediction equation, expressed as:

[0023] ,

[0024] In the formula, Represents the velocity vector of optical flow. This represents the partial derivative of the image feature intensity in the x-direction. This represents the partial derivative of the image feature intensity in the y-direction. This represents the time partial derivative of the image feature intensity along the image time axis.

[0025] S22, Discretize the optical flow estimation prediction equation onto the computational grid, with a time step of [time step value missing]. pixel coordinates are Image intensity and speed Defined as a discretized quantity, the spatial and temporal derivative terms are discretized using the finite difference method. This discretization yields an update equation, where the time term... Represented as:

[0026] ,

[0027] In the formula, Discretization to a computational grid The time partial derivative in the t direction, Discretization to a computational grid The optical flow velocity vector u at that point Discretization to a computational grid The optical flow velocity vector v at that point Discretization to a computational grid The partial derivative in the x-direction, Discretization to a computational grid The partial derivative in the y-direction;

[0028] S23, for the time term It can also be expressed as a discretization based on forward difference, that is:

[0029] ,

[0030] In the formula, Discretized representation of image feature intensity at the previous time step. Discretized representation of image feature intensity at the next time step;

[0031] Finally, the following relation is obtained:

[0032] ;

[0033] S24. Performing a matrix transformation on the relation, we obtain the following equation:

[0034] ;

[0035] S25, the equation is solved using the local least squares method:

[0036] First, define: , , where a is Sliding window area, The side length of the sliding window area;

[0037] The optical flow within window region a is then solved using the least squares method as follows:

[0038] ,

[0039] Expanding the expression, the specific solution for the optical flow within window region a is obtained as follows:

[0040] ,

[0041] in, ;

[0042] The estimated optical flow from each local region is combined to obtain the local optical flow field of the entire feature region, which is represented as:

[0043] ,

[0044] In the formula, The vector u represents the optical flow velocity of the entire image, which is composed of the various local regions. This represents the optical flow velocity vector of the entire image, composed of the various local regions. ;

[0045] S26, introducing a gradient-based prior, the final optical flow velocity vector u is expressed as:

[0046] ,

[0047] in, The weighting factor is used to adjust the contribution ratio between local estimation and global gradient prior; thus, two directional components are obtained. and Together, they describe the optical flow estimation results of fluid motion in image features;

[0048] The Sobel operator is used to perform convolution operations on the image to estimate the image gradient, thus obtaining... and ;

[0049] S27. To achieve optical flow estimation between image features, the hydrodynamically constrained optical flow module performs matrix operations based on all known variables to complete the calculation of the final optical flow estimation result.

[0050] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0051] 1. This invention constructs a fluid dynamics-constrained optical flow module, which overcomes the limitations of traditional optical flow methods in handling non-rigid fluid motion, and can more accurately model the complex dynamic behavior in typhoon cloud images, effectively improving the motion estimation accuracy;

[0052] 2. By combining the designed data propagation structure, multi-frame alignment and feature fusion strategy, this invention can fully mine the temporal context information in the image sequence, and enhance the spatial detail reconstruction capability and temporal continuity performance of the image.

[0053] 3. The super-resolution reconstruction framework proposed in this invention not only improves the texture clarity and structural integrity of the reconstructed image, but also has stronger adaptability and robustness.

[0054] 4. This invention is applicable to various practical application scenarios such as typhoon remote sensing image enhancement, weather forecasting and disaster monitoring, and has good engineering practicality and promotional value. Attached Figure Description

[0055] Figure 1 This is a flowchart of the present invention;

[0056] Figure 2 This is a structural block diagram of the super-resolution reconstruction network model of the present invention;

[0057] Figure 3 A schematic diagram of data flow for feature alignment;

[0058] Figure 4 A schematic diagram of the structure of an optical flow module constrained by fluid dynamics;

[0059] Figure 5 Experimental results of super-resolution reconstruction of cloud images for different typhoons using different methods;

[0060] Figure 6 Experimental results of super-resolution reconstruction of cloud images at different stages of typhoons using different methods;

[0061] Figure 7 Experimental results for different optical flow modules. Detailed Implementation

[0062] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0063] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0064] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0065] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0066] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0067] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0068] Combination Figure 1 The super-resolution reconstruction method for typhoon cloud image sequences based on fluid dynamics constraints described in this embodiment includes the following steps:

[0069] Step 1: Obtain the original typhoon cloud image sequence of the satellite observation area, preprocess the typhoon cloud image sequence to obtain a low-resolution image sequence, use the original typhoon cloud image sequence as label data, and divide the low-resolution image sequence into training set and validation set.

[0070] In one embodiment, the Digital Typhoon Dataset, developed and maintained by the National Institute of Informatics (NII) of Japan, was used as the original typhoon cloud image sequence. Typhoon data samples from the western North Pacific basin between 2010 and 2022 were selected as the satellite observation area. The data were presented as 512×512 pixel infrared cloud images, which were used as a high-resolution image sequence with a spatial resolution of 5 km and a temporal resolution of 1 hour. Based on the model's structural design, the cloud images were downsampled to 128×128 pixels as a low-resolution image sequence. During the selection process, typhoon data with missing time steps or excessively short total durations were removed, ultimately selecting 50 sets of typhoon data, containing a total of 12,383 typhoon cloud images. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio, with 40 sets of typhoon data used as the training set, 5 sets as the validation set, and 5 sets as the test set. The input data is a single-channel grayscale image, which is normalized in the [0,1] interval to avoid overfitting of the model.

[0071] Step 2: Build a super-resolution reconstruction network model based on fluid dynamics constraints, and train and optimize the super-resolution reconstruction network model using the training set.

[0072] Combination Figure 2 As shown, the super-resolution reconstruction network model is a system that includes... A multi-step bidirectional grid data propagation architecture with parallel optical flow units, where each optical flow unit receives typhoon cloud image data at a single time point, such as the typhoon cloud image data at the current time t. Each optical flow unit includes four sequentially arranged optical flow modules. Each optical flow module receives time information in two steps and propagates typhoon cloud image data back and forth in time in an alternating manner. It is a natural number greater than 4.

[0073] In this example, n is 5, meaning the super-resolution reconstruction network model includes 5 parallel optical flow units. Each optical flow unit includes a residual module, 4 optical flow modules, and an image reconstruction module, processing the typhoon cloud image data... The input is fed into a super-resolution reconstruction network model, where image features are first extracted through a residual module to obtain fluid image features. Then fluid image features The features of the first layer of fused motion information are obtained through processing by the first layer optical flow module. ,feature The second-layer optical flow module processes the data to obtain the features of the fused motion information in the second layer. ,feature The features of the third-layer fused motion information are obtained through processing by the third-layer optical flow module. ,feature The features of the fourth-layer fused motion information are obtained through processing by the fourth-layer optical flow module. , will feature The image is input into the image reconstruction module to obtain a reconstructed high-resolution image. .

[0074] Furthermore, the steps for training and optimizing the super-resolution reconstruction network model using the training set include:

[0075] For any optical flow unit, the typhoon cloud image data at the current time t is denoted as... Typhoon cloud image data Image features are extracted using residual modules composed of convolutions to obtain fluid image features. , representing the two-dimensional coordinates at time t. Image intensity at that location;

[0076] Fluid image features As an input item to the optical flow unit, it is input into the first-layer optical flow module, where Features representing fluid image propagation in a single step. Representing fluid image features of two-step propagation, combined with Figure 3 In the first optical flow module, image features are obtained through two O-transforms. and Optical flow between and image features and Optical flow between The o-transform represents the calculation of the optical flow transformation between two sets of features;

[0077] The obtained optical flow , and fluid image features at the current moment Input them together into the spatiotemporal alignment module to obtain alignment features. ;in The transformation represents the alignment operation between the two sets of optical flow features calculated by the o-transform and the features at the current time. The spatiotemporal alignment module consists of three convolutions.

[0078] Align features and fluid image features Data features are concatenated and then fused using convolution to obtain the first layer of fused motion information features. ,in and Features belonging to the same type;

[0079] Features After being processed sequentially by the second, third, and fourth optical flow modules, the features of the fourth layer fused motion information are obtained. ;

[0080] Features Pixel-Shuffle upsampling was performed, along with data from typhoon cloud imagery. After performing bilinear interpolation upsampling on the feature map, a reconstructed high-resolution image is generated. The formula is:

[0081] ,

[0082] In the formula, Up represents the Pixel-Shuffle upsampling operation, and B represents the bilinear interpolation upsampling operation.

[0083] Combination Figure 4 Furthermore, the implementation process of the o-transformation includes the following steps:

[0084] S21. To correlate optical flow with image intensity, a time-dependent partial differential equation (PDE) from fluid dynamics, specifically the advection equation, is used. This equation is often used to simulate changes in mass, having a mass (density) variable and a velocity sub-variable. In this task, mass is replaced by image intensity, and the velocity sub-variable is the optical flow vector to be determined. Therefore, the optical flow estimation prediction equation is constructed as follows:

[0085] ,

[0086] In the formula, Represents the velocity vector of optical flow. This represents the partial derivative of the image feature intensity in the x-direction. This represents the partial derivative of the image feature intensity in the y-direction. This represents the time partial derivative of the image feature intensity in the t direction.

[0087] S22, In order to use this equation to calculate the optical flow of the image features at the current time step and the image features at the previous time step, the optical flow estimation prediction equation is discretized into a computational grid with a time step of [value missing]. pixel coordinates are Image intensity and speed Defined as a discretized quantity, the spatial and temporal derivative terms are discretized using the finite difference method (FDM). Discretization yields the update equation, where the time term... Represented as:

[0088] ,

[0089] In the formula, Discretization to a computational grid The time partial derivative in the t direction, Discretization to a computational grid The optical flow velocity vector u at that point Discretization to a computational grid The optical flow velocity vector v at that point Discretization to a computational grid The partial derivative in the x-direction, Discretization to a computational grid The partial derivative in the y-direction.

[0090] S23, for the time term It can also be expressed as a discretization based on forward difference, that is:

[0091] ,

[0092] In the formula, Discretized representation of image feature intensity at the previous time step. Discretized representation of image feature intensity at the next time step;

[0093] Finally, the following relation is obtained:

[0094] .

[0095] S24. Performing a matrix transformation on the relation, we obtain the following equation:

[0096] .

[0097] S25, the equation is solved using the local least squares method (Lucas–Kanade):

[0098] First, define: , , where a is Sliding window area, The side length of the sliding window area;

[0099] The optical flow within window region a is then solved using the least squares method as follows:

[0100] ,

[0101] Expanding the expression, the specific solution for the optical flow within window region a is obtained as follows:

[0102] ,

[0103] in, ;

[0104] The estimated optical flow from each local region is combined to obtain the local optical flow field of the entire feature region, which is represented as:

[0105] ,

[0106] In the formula, The vector u represents the optical flow velocity of the entire image, which is composed of the various local regions. This represents the optical flow velocity vector of the entire image, composed of the various local regions. .

[0107] Local least squares methods based on the advection equation have the ability to adapt to local non-rigid deformations and can extract local motion features with fine granularity. However, relying solely on local information is insufficient to fully characterize the global evolution of the cloud field. Therefore, this example introduces a gradient-based prior to roughly reflect the overall motion trend of the image, thereby compensating for the shortcomings of local methods.

[0108] S26, introducing a gradient-based prior, the final optical flow velocity vector u is expressed as:

[0109] ,

[0110] in, The weighting factor is used to adjust the contribution ratio between local estimation and global gradient prior; thus, two directional components are obtained. and Together, they describe the optical flow estimation results of fluid motion in image features;

[0111] The Sobel operator is used to perform convolution operations on the image to estimate the image gradient, thus obtaining... and .

[0112] S27, to achieve optical flow estimation between image features, the hydrodynamically constrained optical flow module performs matrix operations based on all known variables to calculate the final optical flow estimation result. For ease of subsequent description, this transformation operation is defined as the o-transformation, denoted as: .

[0113] In one example, a system of constraint equations can be constructed using a 3×3 sliding local window, a method that can capture cloud patterns within a fine-scale region of motion and evolution.

[0114] Step 3: Input the low-resolution image sequence from the validation set into the trained super-resolution reconstruction network model to obtain the reconstructed high-resolution image.

[0115] To further demonstrate the superiority of the super-resolution reconstruction method for typhoon cloud image sequences based on hydrodynamic constraints described in this invention, the following examples illustrate the advantages. The super-resolution reconstruction network model constructed in this invention is compared with existing super-resolution reconstruction models for typhoon cloud image sequences. The comparison results are shown in Table 1, demonstrating the image reconstruction performance of different benchmark models under a 4x super-resolution task. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as evaluation metrics. As can be seen from Table 1, this invention achieves ideal PSNR and SSIM values ​​on the test sequences, indicating its superiority in detail recovery and structure preservation. Compared to the high-performance BasicVSR series architecture, the introduction of physical information further improves the reconstruction quality of the model on multiple sequences, especially demonstrating outstanding performance on the 202214 and 202220 sequences.

[0116] Table 1. Comparison of super-resolution metrics for cloud images of different models under different typhoons.

[0117]

[0118] To visually demonstrate the reconstruction performance of different models under a 4x super-resolution task, Figure 5 Visual comparisons of the various models on different test sequences are presented. From the image details, it can be seen that traditional methods such as TDAN and EDVR exhibit some blurring in the recovery of high-frequency textures, while the BasicVSR series shows improvement in structure preservation. In contrast, the present invention, by introducing fluid equations, allows the model to further optimize cloud detail textures and edge information, resulting in reconstructed images that more closely resemble real high-resolution images and exhibit better visual quality.

[0119] To further demonstrate the super-resolution performance of the model constructed in this invention across different intensities of typhoons, typhoon number 202201 from the Digital Typhoon Dataset was selected. This typhoon developed for 283 hours and has 283 consecutive time-series images, covering five different intensity ranges from 2 to 6. Time-series images of the typhoon within five different intensity ranges were selected to showcase the super-resolution performance of the model constructed in this invention across these ranges. The typhoon cloud image indexes for each intensity level are shown in Table 2.

[0120] Table 2. Typhoon cloud image indexes at various intensity levels

[0121]

[0122] Table 2 shows the super-resolution reconstruction performance of different models across different intensities of Typhoon 202201. The PSNR and SSIM indices of each model were evaluated at different time points for five different intensities (2-6) during the typhoon's development. The results show that this invention achieves relatively ideal reconstruction results across all intensities. Compared to the BasicVSR series models, the introduction of the fluid equation results in better performance in restoring complex cloud structures and preserving texture details.

[0123] Figure 6 This paper showcases the super-resolution reconstruction performance of different models across various intensities of Typhoon 202201. It can be seen that traditional methods such as TDAN and EDVR exhibit relatively blurry details at cloud edges, while BasicVSR and its improved versions show improvement in overall structure preservation, though some detail loss still exists. Under complex multi-cloud structures, this invention can further enhance cloud texture details, resulting in clearer, more refined reconstructed images and demonstrating superior super-resolution reconstruction capabilities.

[0124] To verify the effectiveness of the hydrodynamically constrained optical flow module in this invention, a comparative experiment was designed with two representative methods: the traditional Farneback algorithm and the deep learning-based SPyNet method. The Farneback algorithm is a classic dense optical flow estimation algorithm, widely used in traditional motion analysis due to its efficiency in estimating smooth motion fields, and often used as a baseline. SPyNet, on the other hand, is a deep learning-based optical flow method that uses a spatial pyramid structure to capture motion information from multiple scales, demonstrating high accuracy in general optical flow tasks. It should be noted that this evaluation focuses only on optical flow estimation itself and does not integrate these methods into downstream tasks such as super-resolution.

[0125] Comparative experiments were conducted on a pair of consecutive typhoon cloud images to evaluate the performance of the proposed hydrodynamically constrained optical flow module. Experiments included comparisons with the traditional Farneback algorithm and the deep learning method SPyNet. To ensure fairness and reproducibility, the SPyNet model used in the experiments was the officially released version and pre-trained on the Vimeo90K dataset based on the MMagic framework. Experimental results are as follows: Figure 7 As shown, this invention demonstrates significant advantages in capturing complex, non-rigid motion patterns in atmospheric images.

[0126] In summary, the results demonstrate that the super-resolution reconstruction network model constructed in this invention possesses outstanding capabilities in generating high-quality, accurate cloud images, effectively capturing the dynamic evolution of typhoon patterns, and enhancing structural consistency and detail preservation. The proposed hydrodynamically constrained optical flow module exhibits significant advantages in capturing non-rigid fluid motion in typhoon cloud images. By combining local least squares estimation with global prior information based on image gradients, this invention can not only accurately reconstruct small-scale local dynamics but also better reflect the overall cloud cluster evolution trend. Furthermore, this invention demonstrates higher robustness and expressive power when processing cloud image sequences with strong deformation and complex boundaries. Experimental results verify the effectiveness of this module in atmospheric image flow field modeling, providing a reliable optical flow estimation foundation for subsequent tasks such as image alignment and super-resolution reconstruction.

Claims

1. A super-resolution reconstruction method for typhoon cloud image sequences based on fluid dynamics constraints, characterized in that, Includes the following steps: The original typhoon cloud image sequence of the satellite observation area is obtained, the typhoon cloud image sequence is preprocessed to obtain a low-resolution image sequence, the original typhoon cloud image sequence is used as label data, and the low-resolution image sequence is divided into training set and validation set. A super-resolution reconstruction network model based on fluid dynamics constraints was constructed, and the super-resolution reconstruction network model was trained and optimized using a training set. The low-resolution image sequence in the validation set is input into the trained super-resolution reconstruction network model to obtain the reconstructed high-resolution image; The super-resolution reconstruction network model is a system that includes... A multi-step bidirectional grid data propagation architecture with parallel optical flow units. Each optical flow unit receives typhoon cloud image data at one time point. Each optical flow unit includes four sequentially arranged optical flow modules. Each optical flow module receives time information of two step sizes, and the typhoon cloud image data is propagated back and forth in time in an alternating manner. It is a natural number greater than 4; The steps for training and optimizing a super-resolution reconstruction network model using a training set include: For any optical flow unit, the typhoon cloud image data at the current time t is denoted as... Typhoon cloud image data Image features are extracted using residual modules composed of convolutions to obtain fluid image features. , representing the two-dimensional coordinates at time t. Image intensity at that location; Fluid image features As an input item to the optical flow unit, it is input into the first-layer optical flow module, where Features representing fluid image propagation in a single step. The fluid image features representing two-step propagation are obtained through two O-transforms in the first-layer optical flow module. and Optical flow between and image features and Optical flow between The o-transform represents the calculation of the optical flow transformation between two sets of features; The obtained optical flow , and fluid image features at the current moment Input them together into the spatiotemporal alignment module to obtain alignment features. ;in The transformation represents the alignment operation between the two sets of optical flow features calculated by the o-transform and the features at the current time. The spatiotemporal alignment module consists of three convolutions. Align features and fluid image features Data features are concatenated and then fused using convolution to obtain the first layer of fused motion information features. ,in and Features belonging to the same type; Features After being processed sequentially by the second, third, and fourth optical flow modules, the features of the fourth layer fused motion information are obtained. ; Features Pixel-Shuffle upsampling was performed, along with data from typhoon cloud imagery. After performing bilinear interpolation upsampling on the feature map, a reconstructed high-resolution image is generated. The formula is: ， In the formula, Up represents the Pixel-Shuffle upsampling operation, and B represents the bilinear interpolation upsampling operation.

2. The method for super-resolution reconstruction of typhoon cloud image sequences based on fluid dynamics constraints according to claim 1, characterized in that, The implementation process of the o-transform includes: S21, Construct the optical flow estimation prediction equation, expressed as: ， In the formula, Represents the velocity vector of optical flow. This represents the partial derivative of the image feature intensity in the x-direction. This represents the partial derivative of the image feature intensity in the y-direction. This represents the time partial derivative of the image feature intensity along the image time axis. S22, Discretize the optical flow estimation prediction equation onto the computational grid, with a time step of [time step value missing]. pixel coordinates are Image intensity and speed Defined as a discretized quantity, the spatial and temporal derivative terms are discretized using the finite difference method. This discretization yields an update equation, where the time term... Represented as: ， In the formula, Discretization to a computational grid The time partial derivative in the t direction, Discretization to a computational grid The optical flow velocity vector u at that point Discretization to a computational grid The optical flow velocity vector v at that point Discretization to a computational grid The partial derivative in the x-direction, Discretization to a computational grid The partial derivative in the y-direction; S23, for the time term It can also be expressed as a discretization based on forward difference, that is: ， In the formula, Discretized representation of image feature intensity at the previous time step. Discretized representation of image feature intensity at the next time step; Finally, the following relation is obtained: ； S24. Performing a matrix transformation on the relation, we obtain the following equation: ； S25, the equation is solved using the local least squares method: First, define: , , where a is Sliding window area, The side length of the sliding window area; The optical flow within window region a is then solved using the least squares method as follows: ， Expanding the expression, the specific solution for the optical flow within window region a is obtained as follows: ， in, ; The estimated optical flow from each local region is combined to obtain the local optical flow field of the entire feature region, which is represented as: ， In the formula, The vector u represents the optical flow velocity of the entire image, which is composed of the various local regions. This represents the optical flow velocity vector of the entire image, composed of the various local regions. ; S26, introducing a gradient-based prior, the final optical flow velocity vector u is expressed as: ， in, The weighting factor is used to adjust the contribution ratio between local estimation and global gradient prior; thus, two directional components are obtained. and Together, they describe the optical flow estimation results of fluid motion in image features; Specifically, the Sobel operator is used to perform convolution operations on the image to estimate the image gradient, thus obtaining... and ; S27. To achieve optical flow estimation between image features, the hydrodynamically constrained optical flow module performs matrix operations based on all known variables to complete the calculation of the final optical flow estimation result.

Images