Optical flow-deep learning satellite multi-channel data extrapolation method

Abstract

The application relates to the technical field of satellite meteorological prediction, and discloses an optical flow-depth learning satellite multi-channel data extrapolation method. The method comprises the following steps: acquiring and preprocessing time-series satellite multi-channel data; a two-dimensional motion field is estimated by using a variational optical flow algorithm, preliminary space-time extrapolation is carried out in combination with a semi-Lagrangian algorithm and a mass conservation constraint; a deep learning model integrating space-time convolution and a recurrent neural network is constructed, and the preliminary extrapolation result and original deep features are jointly trained and predicted; two types of results are dynamically fused through an adaptive weight fusion mechanism; finally, the fusion result is corrected based on atmospheric dynamics and radiation transmission physical constraints, and a physically consistent multi-channel extrapolation field is generated. The method combines the advantages of physical motion estimation and data-driven learning, and improves the accuracy and rationality of satellite observation extrapolation from the minute level to the hour level.

Description

Technical Field

[0001] This invention relates to the field of satellite weather forecasting technology, specifically to an optical flow-deep learning method for extrapolating multi-channel satellite data. Background Technology

[0002] Currently, extrapolation methods based on time-series satellite multi-channel observation data are mainly divided into two categories. Traditional optical flow or cross-correlation methods estimate the motion field and perform linear extrapolation by calculating displacement vectors between consecutive frames. However, when dealing with complex cloud evolution, cloud formation, dissipation, and other nonlinear processes, these methods are prone to error accumulation in displacement estimation. Furthermore, the simple motion extrapolation lacks physical constraints, leading to physically unreasonable extrapolation results, easily resulting in deformation distortion or violations of fundamental physical laws such as mass conservation. Pure deep learning methods, on the other hand, directly learn spatiotemporal evolution patterns from historical data using convolutional neural networks or recurrent neural networks. While these methods can capture complex nonlinear characteristics, their prediction process has weak interpretability. The "motion" learned by the model lacks clear physical meaning, and without the guidance of physical laws, it may produce spurious structures that violate atmospheric dynamics principles, resulting in insufficient stability in long-term predictions.

[0003] Existing technologies struggle to effectively integrate data-driven complex pattern recognition capabilities while ensuring the accuracy and physical plausibility of extrapolated motion fields. A key challenge is how to deeply combine explicit physical motion estimation with powerful deep learning feature learning capabilities to achieve multi-channel data extrapolation that combines high accuracy with physical consistency. Summary of the Invention

[0004] The purpose of this invention is to provide an optical flow-deep learning satellite multi-channel data extrapolation method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an optical flow-deep learning satellite multi-channel data extrapolation method, the method comprising:

[0006] Acquire time-series satellite multi-channel observation data, preprocess and filter the time-series satellite multi-channel observation data to obtain optimized multi-channel time-series data;

[0007] Multi-channel feature matching and displacement vector calculation are performed on continuous time frames in the optimized multi-channel time series data. The two-dimensional motion field is accurately estimated based on the variational optical flow algorithm. The two-dimensional motion field is preliminarily extrapolated by combining the semi-Lagrange algorithm and mass conservation constraints to generate the preliminary optical flow extrapolation result.

[0008] A deep learning extrapolation model integrating spatiotemporal convolutional networks and recurrent neural networks is constructed. The preliminary extrapolation result of the optical flow is jointly input into the deep multi-channel features in the optimized multi-channel time series data. The deep learning extrapolation model is then trained to generate deep learning prediction results.

[0009] An adaptive weight fusion mechanism is designed to dynamically allocate contribution weights to the preliminary extrapolation result of the optical flow and the prediction result of the deep learning, thereby generating a fused extrapolation result.

[0010] Based on atmospheric dynamics equations and radiative transfer physical constraints, the fused extrapolation results are subjected to non-physical bias correction processing to generate a corrected multi-channel extrapolation field.

[0011] The final output is a high-precision multi-channel observation field extrapolation product ranging from minute to hour.

[0012] Preferably, the step of acquiring time-series satellite multi-channel observation data, and preprocessing and filtering the time-series satellite multi-channel observation data to obtain optimized multi-channel time-series data includes:

[0013] The time-series satellite multi-channel observation data are sequentially subjected to radiometric calibration, geometric correction and atmospheric correction to generate preprocessed satellite multi-channel data;

[0014] The preprocessed satellite multi-channel data undergoes adaptive filtering to eliminate noise interference and generate optimized multi-channel time-series data, specifically including:

[0015] The preprocessed satellite multi-channel data is subjected to channel separation processing to generate multiple independent single-channel time-series data sequences;

[0016] Calculate the noise statistical characteristics of each of the single-channel time-series data sequences, wherein the noise statistical characteristics include noise variance and noise spatial correlation;

[0017] Based on the noise statistical characteristics corresponding to each channel, a filter kernel function that matches the noise characteristics of the channel is dynamically selected;

[0018] Each single-channel time-series data sequence is subjected to adaptive spatial filtering using the selected filtering kernel function to generate denoised single-channel data.

[0019] The denoised single-channel data of all channels are merged in the original channel order to generate the optimized multi-channel time-series data.

[0020] Preferably, the accurate estimation of the two-dimensional motion field based on the variational optical flow algorithm includes:

[0021] Extract two consecutive frames of multi-channel image data from the optimized multi-channel time-series data;

[0022] A multi-channel data fidelity term based on constant brightness and constant gradient is constructed, and combined with a smoothing constraint term, a variational optical flow energy functional is formed.

[0023] The variational optical flow energy functional is minimized using a numerical optimization algorithm to calculate the motion displacement vector of each pixel between two consecutive frames.

[0024] The motion displacement vectors are integrated in space to generate the two-dimensional motion field, which describes the velocity and direction of the observation field in two-dimensional space.

[0025] Preferably, the preliminary spatiotemporal extrapolation of the two-dimensional motion field by combining the semi-Lagrange algorithm with mass conservation constraints to generate preliminary optical flow extrapolation results includes:

[0026] Using the current observation field as the initial field, and based on the two-dimensional motion field, a semi-Lagrange algorithm is used to track the position of each grid point in the future.

[0027] During the tracking process, mass conservation constraints are applied to the physical quantities of each grid point to ensure that the total mass remains constant during the motion.

[0028] Based on the tracked future time grid point positions, the physical quantities of the initial field are mapped onto the prediction grid of the future time rules through interpolation methods to generate the preliminary extrapolation field for the future time.

[0029] The process of tracking, applying constraints, mapping, and generating a preliminary extrapolation field is repeated for multiple consecutive future moments to generate the preliminary extrapolation result of the optical flow in time series form.

[0030] Preferably, the construction and training of the deep learning extrapolation model includes:

[0031] A model architecture is constructed that includes a parallel spatiotemporal convolution branch and a recurrent network branch. The spatiotemporal convolution branch is used to extract spatial features and short-term features of multi-channel data, and the recurrent network branch is used to extract long-term dependency features.

[0032] The model architecture incorporates channel attention and spatial attention mechanisms, enabling the model to adaptively focus on features of key channels and key spatial regions.

[0033] The optimized multi-channel time-series data is input into the spatiotemporal convolution branch to extract deep multi-channel features. The deep multi-channel features are then concatenated with the preliminary extrapolation result of the optical flow along the feature dimension to form joint features.

[0034] The joint features are input into the recurrent network branch to further learn and predict the temporal evolution pattern, and the deep learning prediction result is output.

[0035] Using historical satellite multi-channel data and their corresponding future time-based ground truth data, the deep learning extrapolation model is trained end-to-end with the goal of minimizing prediction error.

[0036] Preferably, the design and application of the adaptive weight fusion mechanism includes:

[0037] Calculate the motion stability index of the preliminary extrapolation result of the optical flow during the target prediction period;

[0038] Based on the pre-set forecast task requirements, determine the intensity of the requirement for physical consistency of forecast results and the intensity of the requirement for capturing complex nonlinear features.

[0039] Using the motion stability index, the physical consistency requirement intensity, and the complex nonlinear feature capture requirement intensity as inputs, a learnable weight allocation network dynamically generates weight values ​​assigned to the preliminary extrapolation result of the optical flow and weight values ​​assigned to the deep learning prediction result.

[0040] The generated weight values ​​are used to perform a weighted average fusion of the preliminary extrapolation result of the optical flow and the prediction result of the deep learning to generate the fused extrapolation result.

[0041] Preferably, the calculation of the motion stability index of the preliminary extrapolation result of the optical flow during the target prediction period includes:

[0042] Within the forecast period corresponding to the preliminary extrapolation result of the optical flow, the variation amplitude of the two-dimensional motion field between adjacent extrapolated fields is calculated;

[0043] The mean and variance of the aforementioned variation amplitude over the entire spatial region are statistically analyzed and used as a measure of motion consistency.

[0044] Analyze the temporal continuity of the direction of the two-dimensional motion field and calculate the standard deviation of the direction angle;

[0045] The motion stability index is obtained by combining the mean and variance of the variation amplitude and the standard deviation of the direction angle through linear combination.

[0046] Preferably, the non-physical bias correction processing of the fused extrapolation results based on atmospheric dynamics equations and radiative transfer physical constraints includes:

[0047] Extract multiple physical quantities such as temperature, humidity, water vapor, and cloud water from the fusion extrapolation results;

[0048] Substitute the extracted multi-channel physical quantities into a simplified set of atmospheric dynamics equations and check whether the spatiotemporal evolution of the multi-channel physical quantities satisfies the basic dynamic and thermodynamic relationships described by the set of equations.

[0049] Substitute the aforementioned multi-channel physical quantities into the radiative transfer equation to calculate the simulated satellite channel radiation values;

[0050] By comparing the simulated satellite channel radiation values ​​with the physically reasonable theoretical radiation value range, grid points and channels with non-physical deviations are identified.

[0051] The physical quantities of the grid points and channels with non-physical deviations are adjusted to satisfy the constraints of the atmospheric dynamics equation and the radiative transfer equation, while minimizing the modification of the original fused extrapolation results, thus completing the non-physical deviation correction process and generating the corrected multi-channel extrapolation field.

[0052] Preferably, adjusting the physical quantity values ​​of the identified grid points and channels with non-physical deviations includes:

[0053] For each grid point with non-physical deviation, a local optimization problem is constructed with the satisfaction of physical constraints as the optimization objective and the adjustment of physical quantities as the optimization variables.

[0054] The physical constraints include local equilibrium constraints derived from the atmospheric dynamics equations and reasonable range constraints for channel radiation values ​​defined by the radiative transfer equations.

[0055] The local optimization problem is solved by an iterative optimization algorithm to obtain the physical quantity correction value that maximizes the satisfaction of physical constraints and minimizes the adjustment amount.

[0056] Replace the corresponding grid point and channel values ​​in the original fusion extrapolation result with the physical quantity correction value to complete the correction of a single grid point;

[0057] Traverse all grid points with non-physical deviations, perform corrections sequentially, and generate the corrected multi-channel extrapolation field that satisfies global physical consistency.

[0058] Preferably, the step of performing adaptive spatial filtering on each of the single-channel time-series data sequences using the selected filtering kernel function to generate denoised single-channel data includes:

[0059] For each single-channel time-series data sequence, the size parameters of the filtering kernel function are dynamically determined based on the noise variance in the noise statistics.

[0060] The selected filtering kernel function is convolved with the neighborhood of each pixel in the current single-channel time-series data sequence to calculate the filtered pixel value.

[0061] The weight coefficients of the filtering kernel function are adaptively adjusted based on the local noise variance and image gradient information to preserve image edge features while denoising.

[0062] The convolution operation results are normalized to generate denoised single-channel data;

[0063] The denoised single-channel data is recombined in time series order to form a denoised single-channel time series data sequence.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] A variational optical flow algorithm is employed for two-dimensional motion field estimation, and a semi-Lagrange algorithm and mass conservation constraints from numerical weather prediction are introduced for preliminary extrapolation. This method embeds the guidance of physical laws into the data-driven motion estimation stage, ensuring that the motion trajectory and field structure obtained from the preliminary extrapolation follow the basic principles of mass transport and conservation. This overcomes the distortion caused by the simple translation of traditional optical flow, providing a more physically reasonable initial prediction field with better spatiotemporal structure preservation for subsequent processing, thus improving the stability and rationality of long-term extrapolation.

[0066] A deep learning model integrating spatiotemporal convolutional networks and recurrent neural networks is constructed, and the preliminary extrapolation results of optical flow are jointly input with the deep multi-channel features of the original data. The spatiotemporal convolutional network effectively extracts multi-scale features and local motion correlations in the spatial neighborhood, while the recurrent neural network models the temporal evolution dependence of the features. This joint input mechanism enables the model to not only learn the deep evolution patterns of the original data but also to fuse and optimize the preliminary motion field optimized by physical constraints as strong prior information. This architectural design allows the final prediction result to possess the ability to learn complex nonlinear processes from massive amounts of data, while the generated structure is constrained by physically reasonable motion trajectories, thus achieving synergistic enhancement in prediction accuracy and physical consistency. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the working principle of the optical flow-deep learning satellite multi-channel data extrapolation method described in this invention.

[0068] Figure 2 The flowchart shows the data preprocessing and adaptive filtering process.

[0069] Figure 3 This is a flowchart for estimating a two-dimensional motion field based on the variational optical flow algorithm;

[0070] Figure 4 A comparison chart of motion stability indicators and components for different satellite channels;

[0071] Figure 5The curves show the changes in the objective function loss value and the physical constraint residual during the local optimization iteration process. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figure 1 This invention provides an optical flow-deep learning satellite multi-channel data extrapolation method. The method includes: acquiring time-series satellite multi-channel observation data; preprocessing and filtering the time-series satellite multi-channel observation data to obtain optimized multi-channel time-series data; processing continuous time-series frames in the optimized multi-channel time-series data; accurately estimating a two-dimensional motion field based on a variational optical flow algorithm; and performing preliminary spatiotemporal extrapolation processing on the two-dimensional motion field by combining a semi-Lagrange algorithm and mass conservation constraints to generate preliminary optical flow extrapolation results; constructing a deep learning extrapolation model integrating a spatiotemporal convolutional network and a recurrent neural network; jointly inputting the preliminary optical flow extrapolation results with deep multi-channel features extracted from the optimized multi-channel time-series data; training the model to generate deep learning prediction results; designing an adaptive weight fusion mechanism to dynamically allocate contribution weights between the preliminary optical flow extrapolation results and the deep learning prediction results to generate fused extrapolation results; and performing non-physical bias correction processing on the fused extrapolation results based on atmospheric dynamics equations and radiative transfer physical constraints to generate a corrected multi-channel extrapolation field. The final output is a high-precision multi-channel observation field extrapolation product ranging from minute to hour.

[0074] Example 1: See Figure 2Radiometric calibration, geometric correction, and atmospheric correction are sequentially performed on time-series satellite multi-channel observation data to generate preprocessed satellite multi-channel data. Channel separation processing is then applied to the preprocessed satellite multi-channel data to generate multiple independent single-channel time-series data sequences. The noise statistical characteristics of each single-channel time-series data sequence are calculated, including noise variance and noise spatial correlation. Based on the noise statistical characteristics corresponding to each channel, a filter kernel function matching the channel's noise characteristics is dynamically selected. For each single-channel time-series data sequence, the size parameters of the filter kernel function are dynamically determined based on the noise variance in the noise statistical characteristics. The selected filter kernel function is convolved with the neighborhood of each pixel in the current single-channel time-series data sequence to calculate the filtered pixel value. The weight coefficients of the filter kernel function are adaptively adjusted based on local noise variance and image gradient information. The convolution result is normalized to generate denoised single-channel data. The denoised single-channel data is then recombined in time-series order to form a denoised single-channel time-series data sequence. The denoised single-channel data from all channels are merged in the original channel order to generate optimized multi-channel time-series data.

[0075] In practice, continuous time-series observation images from the infrared and visible light channels of the Himawari-8 satellite are acquired. Each frame contains, for example, observation values ​​from the 10.4-micron infrared channel and the 0.64-micron visible light channel. Processing steps are performed on the time-series satellite multi-channel observation data. Radiometric calibration, geometric correction, and atmospheric correction are sequentially performed on the time-series satellite multi-channel observation data to convert the raw satellite count values ​​into atmospheric parameters with clear physical meaning. All frame data are then reprojected onto a unified geographic coordinate grid to generate preprocessed satellite multi-channel data.

[0076] The preprocessed satellite multi-channel data undergoes channel separation processing, decomposing the multi-channel data cube along the channel dimension to generate independent time-series data sequences for the infrared and visible light channels. The noise statistical characteristics of each single-channel time-series data sequence are calculated. In practice, for the infrared channel time-series data sequence, the noise variance is estimated by calculating the variance of pixel values ​​in a local uniform region of the image, and the spatial correlation of noise is estimated by analyzing the correlation of noise between adjacent pixels. The calculation of the noise statistical characteristics of the visible light channel time-series data sequence follows the same principle.

[0077] Based on the noise statistics of each channel, a filter kernel function that matches the channel noise characteristics is dynamically selected. In some embodiments, for infrared channel data sequences with large noise variance and high spatial correlation, a larger Gaussian filter kernel function is selected; for visible light channel data sequences with small noise variance and weak spatial correlation, a smaller bilateral filter kernel function with better edge preservation characteristics is selected. For each single-channel time-series data sequence, the size parameter of the filter kernel function is dynamically determined based on the noise variance in the noise statistics. In a specific implementation, when the calculated noise variance of the infrared channel is 4.2, the size parameter of the corresponding Gaussian filter kernel function is set to 5x5 pixels; when the calculated noise variance of the visible light channel is 1.1, the spatial domain kernel size parameter of the corresponding bilateral filter kernel function is set to 3x3 pixels.

[0078] The selected filtering kernel function is convolved with the neighborhood of each pixel in the current single-channel time-series data sequence to calculate the filtered pixel value. Essentially, for a frame of data in the infrared channel, a Gaussian kernel function with a size parameter of 5 is used to calculate a weighted average of the 5x5 neighborhood of each pixel. The weight coefficients of the filtering kernel function are adaptively adjusted based on the local noise variance and image gradient information. In specific implementations, for visible light channel data, spatial similarity weights are reduced in image edge regions based on larger image gradient values ​​to preserve clear edge features, while smoothing weights are increased in flat regions based on smaller noise variance. The convolution result is normalized to generate denoised single-channel data. The denoised single-channel data is then recombined in time-series order to form denoised infrared channel time-series data sequences and denoised visible light channel time-series data sequences.

[0079] The denoised single-channel data from all channels are merged in the original channel order. In specific implementations, the processed infrared channel time-series data sequence and the visible light channel time-series data sequence are recombinated along the channel dimension to restore their data cube structure, generating optimized multi-channel time-series data. In some embodiments, during the adaptive spatial filtering process, the filtered pixel value is calculated using the following formula:

[0080]

[0081] in: This represents the denoised pixel value at position (x, y) and time t. This represents a local neighborhood window centered at (x, y). This represents the weight coefficient of the adaptive filtering kernel function at position (i,j), which is determined by the standard deviation of the spatial distance. Standard deviation of pixel intensity difference Joint decision, This represents the original input pixel values. It can be understood as the standard deviation of spatial distance. Dynamically adjusted based on noise variance, standard deviation of pixel intensity difference Adjustments are made based on local image gradients. Optionally, in infrared channel data processing, the spatial distance standard deviation is considered. The value is greater than the value in visible light channel data processing. Optionally, in image edge regions, the standard deviation of pixel intensity difference... The value of is reduced to decrease the smoothing intensity.

[0082] Example 2: See Figure 3 Two consecutive frames of multi-channel image data are extracted from the optimized multi-channel time-series data. A multi-channel data fidelity term based on constant brightness and gradient is constructed, combined with a smoothing constraint term, to form a variational optical flow energy functional. The variational optical flow energy functional is minimized using a numerical optimization algorithm to calculate the motion displacement vector of each pixel between two consecutive frames. The motion displacement vectors are integrated spatially to generate a two-dimensional motion field describing the velocity and direction of the observation field in two-dimensional space. Using the observation field at the current moment as the initial field, the position of each grid point at future moments is tracked using a semi-Lagrange algorithm based on the two-dimensional motion field. During the tracking process, mass conservation constraints are applied to the physical quantities of each grid point. Based on the tracked future grid point positions, the physical quantities of the initial field are mapped to the prediction grid of future moment rules using an interpolation method to generate a preliminary extrapolation field for future moments. The process of tracking, applying constraints, mapping, and generating the preliminary extrapolation field is repeated for multiple consecutive future moments to generate a preliminary extrapolation result of optical flow in time-series form.

[0083] In the specific implementation, two consecutive frames of multi-channel image data are extracted from the optimized multi-channel time-series data for processing. Taking the observation images of the 10.8-micron infrared channel and the 0.64-micron visible light channel of the Himawari-8 satellite at time t and t+Δt as examples, where Δt is the satellite's temporal resolution, a multi-channel data fidelity term based on constant brightness and constant gradient is constructed. For the infrared channel, the constant brightness assumption is reflected in the stable brightness temperature value of the same moving object in two consecutive frames, and the constant gradient assumption is reflected in the spatial gradient pattern of the brightness temperature field in the image shifting with motion. The reflectivity data of the visible light channel also follows similar constant constraints. Combined with the smoothing constraint term, a variational optical flow energy functional is formed. A numerical optimization algorithm is used to minimize the variational optical flow energy functional. In the specific implementation, a coarse-to-fine pyramid strategy combined with iterative reweighted least squares method is used to calculate the motion displacement vector of each pixel between two consecutive frames. By integrating the motion displacement vectors in space, a two-dimensional motion field is generated that describes the velocity and direction of the 10.8-micron infrared channel brightness temperature field and the 0.64-micron visible light channel reflectivity field in two-dimensional space. It can be understood that the two-dimensional motion field is a vector field, in which each grid point contains a two-dimensional vector.

[0084] Using the observation field at the current time t as the initial field, and based on the two-dimensional motion field, a semi-Lagrange algorithm is used to track the position of each grid point at future times. In some embodiments, for the grid point position predicted at the future time τ, the origin position of that grid point at the initial time t is traced in reverse. The tracking process is based on the integration of the velocity vector provided by the two-dimensional motion field. During the tracking process, mass conservation constraints are applied to the physical quantities of each grid point. For conserved physical quantities such as the integral water vapor content of water vapor channels, it is ensured that the total mass remains constant from the initial position to the target position during the motion. Based on the tracked future grid point positions, the physical quantities of the initial field are mapped to the predicted grid of future time regularity using an interpolation method. In specific implementations, for the origin point traced in reverse and located between the regular grids of the initial field, the physical quantity value of the origin point is calculated from the values ​​of the four neighboring grid points of the regular grid of the initial field using a bilinear interpolation method. This value is then assigned to the target regular grid point corresponding to the future time, generating the preliminary extrapolation field for the future time. It can be understood that the preliminary extrapolation field contains the estimated brightness temperature or reflectivity values ​​of each channel at the future time. The process involves repeatedly tracking, applying constraints, mapping, and generating preliminary extrapolation fields for multiple consecutive future moments. In specific implementation, extrapolation field sequences of t+10 minutes, t+20 minutes, and up to t+120 minutes are generated sequentially to produce preliminary extrapolation results of optical flow in time series form.

[0085] In practical implementation, the variational optical flux energy functional is constructed to minimize the sum of the data fidelity term and the motion field smoothing term, and its expression is:

[0086]

[0087] in: This represents the total energy functional that needs to be minimized. This represents the total number of channels involved in the calculation. Represents the entire image plane spatial domain. This represents the weight coefficient assigned to the c-th channel. Describe a robust penalty function. Indicates the position of the c-th channel. and time The observed intensity value, Indicates the position to be solved The two-dimensional displacement vector at that point. Indicates through the displacement vector The intensity value obtained by interpolation in the next frame image. It is a regularization parameter that controls the weights of the smoothing term. This represents the penalty function for the magnitude of the gradient of the sports field. and Representing displacement components respectively and The spatial gradient. This can be understood by adjusting the weight coefficients. and regularization parameters This can balance the contributions of different channels with the smoothness of the motion field. In some embodiments, for the infrared channel and the visible light channel, the weighting coefficients are... The settings can be differentiated based on the noise level or importance of the channel. An optional robustness penalty function is also available. The introduction of this approach helps to address situations where the assumption of constant brightness fails locally due to factors such as cloud formation and dissipation, and changes in lighting.

[0088] Example 3: A model architecture comprising a parallel spatiotemporal convolutional branch and a recurrent network branch is constructed. The spatiotemporal convolutional branch is used to extract spatial and short-term features from multi-channel data, while the recurrent network branch is used to extract long-term temporal dependency features. Channel attention and spatial attention mechanisms are embedded in the model architecture. Optimized multi-channel temporal data is input into the spatiotemporal convolutional branch to extract deep multi-channel features. These deep multi-channel features are then concatenated with the preliminary optical flow extrapolation results along the feature dimension to form joint features. These joint features are input into the recurrent network branch for further learning and prediction of temporal evolution patterns, outputting deep learning prediction results. Using historical satellite multi-channel data and their corresponding future ground truth data, the deep learning extrapolation model is trained end-to-end with the goal of minimizing prediction error.

[0089] In the specific implementation, a deep learning extrapolation model is constructed. Its architecture includes a parallel spatiotemporal convolutional branch and a recurrent network branch to process 10.8-micron and 6.2-micron multi-channel time-series data from the Himawari-8 satellite. The spatiotemporal convolutional branch employs a three-dimensional convolutional neural network structure, with its convolutional kernels sliding along a two-dimensional spatial dimension and a one-dimensional temporal dimension. This is used to extract spatial texture features and short-term temporal evolution features between adjacent frames from the multi-channel data. The recurrent network branch is composed of stacked gated recurrent units, used to extract long-term temporal dependencies spanning multiple time steps from longer time series. Specifically, when constructing the deep learning extrapolation model, a parallel branch architecture is used to process the 10.8-micron and 6.2-micron multi-channel time-series data from the Himawari-8 satellite. The spatiotemporal convolutional branch, based on a three-dimensional convolutional neural network structure, performs sliding convolution operations along a two-dimensional spatial dimension and a one-dimensional temporal dimension, extracting spatial texture features and short-term temporal evolution patterns between adjacent frames layer by layer from the multi-channel data. The recurrent network branch consists of stacked multi-layer gated recurrent units, which cyclically process the temporal sequence through a gating mechanism to capture long-term temporal dependencies spanning multiple time steps. The model architecture embeds channel attention and spatial attention mechanisms. The channel attention mechanism adaptively calculates the weights of each channel to focus on key channel features, while the spatial attention mechanism generates a spatial importance map to highlight key regional features. Optimized multi-channel temporal data is input into the spatiotemporal convolutional branch to extract deep multi-channel features. These features are then concatenated with the preliminary extrapolation results of optical flow along the feature dimension to form joint features, which are then input into the recurrent network branch for temporal evolution prediction. Finally, the model is trained end-to-end using historical satellite multi-channel data and future ground truth data. The channel attention mechanism calculates the weights of each channel in the input feature map through a fully connected layer network, enabling the model to adaptively focus on the features of key channels that contribute more to the forecasting task. The spatial attention mechanism calculates the importance map of the spatial location of the feature map, enabling the model to adaptively focus on features of key spatial regions such as cloud development and fronts.

[0090] The optimized multi-channel temporal data is input into the spatiotemporal convolution branch to extract deep multi-channel features. In specific implementations, the input data consists of seven consecutive frames of satellite images from the past hour, with a temporal resolution of 10 minutes. Each frame contains observations from three channels. After several layers of 3D convolution and pooling operations, a tensor containing high-level abstract features is output. The deep multi-channel features are concatenated with the preliminary extrapolation results of optical flow along the feature dimension to form joint features. It can be understood that the preliminary extrapolation results of optical flow are a prediction sequence for future moments. Connecting this with the deep multi-channel features extracted from historical observations along the feature channel dimension allows the subsequent network to simultaneously perceive the preliminary prediction based on physical motion laws and the complex patterns learned from the data. The joint features are input into the recurrent network branch for further learning and prediction of temporal evolution laws. In some embodiments, the concatenated joint feature sequence is input into the gated recurrent unit network step by step. The internal state of the gated recurrent unit network is gradually updated to remember long-term dependencies, and finally, the deep learning prediction results for the next hour are output at 10-minute intervals.

[0091] By utilizing historical satellite multi-channel data and their corresponding ground truth data for future moments, and aiming to minimize prediction error, an end-to-end deep learning extrapolation model is trained. The training data consists of a large number of time-series satellite observation samples from the past period and their corresponding future real observation fields. The loss function used in model training measures the difference between the deep learning prediction results and the future real observation fields, and updates all parameters in the 3D convolutional layers, gated recurrent unit layers, and attention mechanisms of the model through a backpropagation algorithm. In some embodiments, the Huber loss is used, and its expression is:

[0092]

[0093] in: This represents the Huber loss function value. This represents the total number of samples in a training batch. This represents the total number of future time steps that need to be predicted. Indicates the total number of channels. This represents the deep learning prediction result of the model for the nth sample, the τth future time, and the cth channel. This represents the true value data for the corresponding future time. This is the Huber loss function, specifically defined as using squared loss when the absolute error is less than a threshold δ, and linear loss when it is greater than or equal to δ. In essence, the Huber loss combines the advantages of mean squared error and mean absolute error, providing a more stable gradient during training. The training process aims to iteratively optimize the loss function. The value of is constantly decreasing.

[0094] Example 4: Calculating the motion stability index of the preliminary optical flow extrapolation results within the target forecast period. The calculation of the motion stability index of the preliminary optical flow extrapolation results within the target forecast period includes calculating the variation amplitude of the two-dimensional motion field between adjacent extrapolated times within the forecast period corresponding to the preliminary optical flow extrapolation results; statistically analyzing the mean and variance of the variation amplitude over the entire spatial region; analyzing the temporal continuity of the direction of the two-dimensional motion field and calculating the standard deviation of the direction angle; and combining the mean and variance of the variation amplitude and the standard deviation of the direction angle to obtain a quantified motion stability index through linear combination. Based on the preset forecast task requirements, the intensity of the physical consistency requirement for the forecast results and the intensity of the requirement for capturing complex nonlinear features are determined. Using the motion stability index, the intensity of the physical consistency requirement, and the intensity of the requirement for capturing complex nonlinear features as inputs, a learnable weight allocation network dynamically generates weight values ​​assigned to the preliminary optical flow extrapolation results and weight values ​​assigned to the deep learning prediction results. The generated weight values ​​are used to perform a weighted average fusion of the preliminary optical flow extrapolation results and the deep learning prediction results to generate a fused extrapolation result.

[0095] In the specific implementation, an adaptive weight fusion mechanism is designed to fuse the preliminary optical flow extrapolation results with the deep learning prediction results, taking the generation of cloud top brightness and temperature field extrapolation products for the next 2 hours as an example. The motion stability index of the preliminary optical flow extrapolation results within the target forecast period is calculated. Within the next 2-hour forecast period corresponding to the preliminary optical flow extrapolation results, the change amplitude of the two-dimensional motion field between adjacent extrapolation times is calculated. Specifically, for the extrapolation field at every 10-minute interval, the difference in Euclidean distance between its two-dimensional motion field vector and the two-dimensional motion field vector at the previous time point at each grid point is calculated as the change amplitude. The mean and variance of the change amplitude over the entire spatial region are statistically analyzed as a measure of motion consistency. In the specific implementation, for a region containing 100x100 grid points, the mean change amplitude of all grid points at all adjacent times is calculated to be 0.8 pixels / frame, and the standard deviation of the change amplitude is 0.5 pixels / frame. The temporal continuity of the direction of a two-dimensional motion field is analyzed, and the standard deviation of the direction angle is calculated. The direction angle of the two-dimensional motion field is defined by the angle between the vector and the due east direction. The standard deviation of the direction angle at each grid point over the next 2 hours is calculated, and the average standard deviation of the direction angle over the region is 0.26 radians. By combining the mean of the variation amplitude, the standard deviation of the variation amplitude, and the standard deviation of the direction angle, a quantitative motion stability index is calculated through normalized linear combination. The lower the value of the motion stability index, the more stable the two-dimensional motion field.

[0096] Based on the pre-defined forecasting task requirements, the intensity of the requirement for physical consistency of forecast results and the intensity of the requirement for capturing complex nonlinear features are determined, with different intensity settings corresponding to different forecasting task requirements. It can be understood that the intensity of the physical consistency requirement reflects the emphasis on the forecast field adhering to physical laws such as mass conservation and smooth motion, while the intensity of the requirement for capturing complex nonlinear features reflects the emphasis on the forecast field accurately depicting complex nonlinear processes such as cloud formation and dissipation, and thunderstorm outbreaks. See Table 1.

[0097] Table 1: Demand Intensity Setting Table for Two Typical Forecasting Tasks

[0098]

[0099] Using motion stability indices, physical consistency requirements, and the requirements for capturing complex nonlinear features as inputs, a learnable weight allocation network dynamically generates weight values ​​assigned to the preliminary extrapolation results of the optical flow and weight values ​​assigned to the deep learning prediction results. In some embodiments, the weight allocation network is a feedforward neural network with two fully connected layers. Its output is activated by a Sigmoid function to obtain weight values ​​for the preliminary extrapolation results of the optical flow, which are between 0 and 1. The weight values ​​for the deep learning prediction results are obtained by subtracting the weight values ​​for the preliminary extrapolation results of the optical flow from 1. The generated weight values ​​are used to perform a weighted average fusion of the preliminary optical flow extrapolation results and the deep learning prediction results to generate a fused extrapolation result. In specific implementation, for the advection forecasting task of large-scale layered cloud systems, the motion stability index is calculated to be 0.2, the physical consistency requirement intensity is 0.9, and the complex nonlinear feature capture requirement intensity is 0.3. The weight value of the preliminary optical flow extrapolation result output by the weight allocation network is 0.75, and the weight value of the deep learning prediction result is 0.25. The fused extrapolated brightness temperature value of a certain grid point at a future time is calculated by multiplying the preliminary optical flow extrapolation brightness temperature value by 0.75 and adding the deep learning prediction brightness temperature value multiplied by 0.25. The motion stability index is calculated using the following formula:

[0100]

[0101] in: This represents the calculated dimensionless motion stability index. This represents the mean of the variation over the entire spatial region and over all adjacent time points. It is a reference average value of change, used for... Normalize it to make it dimensionless. The standard deviation represents the magnitude of change. It is a reference standard deviation of the range of variation, used for... Normalize, The average value of the region representing the standard deviation of the orientation angle. It is a reference orientation angle standard deviation, used for... Normalize, , and These are dimensionless weighting coefficients used to balance the three contributions. In some embodiments, the reference value... , , This can be obtained statistically from historical extrapolation datasets. Optionally, before deployment, the weight allocation network has been trained offline using historical samples containing various weather scenarios and expert-defined optimal fusion weights as supervision information, enabling the network to learn the mapping relationship from input to optimal weights. Optionally, for different satellite channels, their corresponding motion stability indices can be calculated independently and fusion weights assigned. For example, the stability indices for the water vapor channel and the infrared channel can be calculated separately, and different fusion weights can be assigned to the preliminary optical flow extrapolation results and deep learning prediction results for each channel.

[0102] See Figure 4 In the stability analysis phase of the optical flow extrapolation results, a comparison of the statistical values ​​of motion stability indices and their constituent elements (mean of variation, standard deviation of variation, and standard deviation of orientation angle) for the infrared, water vapor, and visible light satellite channels was presented. Specifically, the mean of variation (blue bars, unit: pixels / frame), standard deviation of variation (orange bars, unit: pixels / frame), and standard deviation of orientation angle (green bars, unit: radians) for each channel provide the basic data for calculating the motion stability index, while the motion stability index (red line, dimensionless) is a quantitative result combining the three. From the data characteristics: the mean of variation in the water vapor channel is significantly higher than that in the infrared and visible light channels, and its standard deviation of variation and orientation angle are also at a high level, corresponding to its highest motion stability index among the three (approximately 1.1); the mean, standard deviation, and standard deviation of variation in the infrared channel are all at a medium level, with a motion stability index of approximately 0.5; the statistical values ​​of all constituent elements in the visible light channel are the lowest among the three, corresponding to a motion stability index of approximately 0.1. This figure visually reflects the differences in motion stability of the optical flow extrapolation results for different satellite channels, providing data support for channel-level weight allocation in the subsequent adaptive weight fusion mechanism.

[0103] Example 5: Extracting multi-channel physical quantities of temperature, humidity, water vapor, and cloud water from the fusion extrapolation results. Substituting the extracted multi-channel physical quantities into a simplified set of atmospheric dynamics equations, the spatiotemporal evolution of these quantities is checked to ensure it satisfies the fundamental dynamic and thermodynamic relationships described by the equations. Substituting these multi-channel physical quantities into the radiative transfer equation, simulated satellite channel radiative values ​​are calculated. The simulated satellite channel radiative values ​​are compared with the physically reasonable theoretical radiative value range to identify grid points and channels with non-physical deviations. The physical quantity values ​​of the identified grid points and channels with non-physical deviations are adjusted. For each grid point with non-physical deviations, a local optimization problem is constructed with physical constraint satisfaction as the optimization objective and physical quantity adjustment as the optimization variable. The physical constraints include local equilibrium constraints derived from the atmospheric dynamics equations and reasonable range constraints for channel radiative values ​​defined by the radiative transfer equation. An iterative optimization algorithm is used to solve the local optimization problem to obtain the physical quantity correction value that maximizes physical constraint satisfaction and minimizes the adjustment amount. This correction value replaces the corresponding grid point and channel values ​​in the original fusion extrapolation results. Iterate through all grid points with non-physical deviations, perform corrections sequentially, and generate a corrected multi-channel extrapolation field that satisfies global physical consistency.

[0104] In practical implementation, non-physical bias corrections are applied to the fusion extrapolation results based on atmospheric dynamics equations and radiative transfer physical constraints. Taking the fusion extrapolation results of the 6.2-micron water vapor channel and 10.8-micron infrared channel of the Himawari-8 satellite as an example, multi-channel physical quantities such as temperature, humidity, water vapor, and cloud water are extracted from the fusion extrapolation results. In practice, using the established physical inversion relationship between satellite brightness temperature and atmospheric temperature and humidity profiles and water vapor content, the future fusion extrapolated brightness temperature field of the 6.2-micron and 10.8-micron channels is converted into a mid-atmosphere water vapor content field and a cloud top temperature field. The extracted multi-channel physical quantities are substituted into a simplified set of atmospheric dynamics equations to check whether the spatiotemporal evolution of the multi-channel physical quantities satisfies the basic dynamic and thermodynamic relationships described by the equations. In some embodiments, the omega equation under the quasi-geostrophic approximation is used to check the consistency between the vertical velocity field and the water vapor and temperature fields, identifying physically contradictory grid points where water vapor converges strongly but vertical motion is downward. Multi-channel physical quantities are substituted into the radiative transfer equation to calculate simulated satellite channel radiative values. Using a fast radiative transfer model, the extrapolated temperature field, humidity field, and cloud water field before correction are used as inputs to calculate the simulated brightness temperature values ​​for the corresponding 6.2-micron and 10.8-micron channels in a forward simulation. The simulated satellite channel radiative values ​​are compared with the physically reasonable theoretical radiative value range to identify grid points and channels with non-physical deviations. In practice, the physically reasonable theoretical radiative value range is determined jointly by historical observation statistics and radiative transfer theory. For example, under clear sky conditions, the brightness temperature of the 10.8-micron channel should not be lower than the surface temperature minus a reasonable deviation threshold. It was identified that the simulated brightness temperature value of a cloud region in the fused extrapolation results was lower than the theoretical reasonable lower limit, indicating that the cloud top temperature in that region was incorrectly corrected too low during the extrapolation process.

[0105] The physical quantity values ​​of grid points and channels with identified non-physical deviations are adjusted. For each grid point with non-physical deviations, a local optimization problem is constructed with physical constraint satisfaction as the optimization objective and the physical quantity adjustment amount as the optimization variable. In specific implementation, for the aforementioned grid points with excessively low cloud top temperatures, the local optimization problem is defined as finding the minimum adjustment amount for the temperature and water vapor content of that grid point, so that the brightness temperature value calculated after substituting the adjusted physical quantities into the radiative transfer model returns to a reasonable range, and satisfies the simplified thermodynamic equation constraints with the temperature and humidity fields of surrounding grid points. Physical constraints include local equilibrium constraints derived from atmospheric dynamic equations and reasonable range constraints for channel radiation values ​​defined by the radiative transfer equation. Local equilibrium constraints can be expressed as a simplified form of the relationship between the water vapor change rate and horizontal advection and vertical transport. An iterative optimization algorithm is used to solve the local optimization problem to obtain the physical quantity correction value that maximizes physical constraint satisfaction and minimizes the adjustment amount. In some embodiments, the Lagrange multiplier method combined with gradient descent is used for iterative solution, with each iteration fine-tuning the physical quantity values ​​and recalculating the constraint satisfaction until the objective function converges. The correction of a single grid point is completed by replacing the corresponding grid point and channel values ​​in the original fusion extrapolation result with the physical quantity correction value. It can be understood that the correction is only applied to specific physical quantities that are identified as having non-physical deviations, such as only correcting excessively low cloud top temperatures and the associated water vapor content.

[0106] In practice, the objective function of the local optimization problem is constructed as a combination of multiple normalized penalty terms to ensure that the dimensions of each term are consistent and additive. Its expression is:

[0107]

[0108] in: This represents the dimensionless objective function value to be minimized. This indicates that the current grid point state vector needs to be changed. The amount of adjustment made, It is a with A reference state vector of the same dimension is used to adjust the amount. Normalize, , and It is a dimensionless coefficient used to balance the weights of various components. This represents the simulated radiation value calculated using the adjusted physical quantities through the radiative transfer equation. This represents the boundary value of the reasonable range of theoretical radiation values. It is a radiation value scaling factor used to normalize radiation residuals. This represents the residual vector obtained by substituting the adjusted physical quantities into the simplified atmospheric dynamics equations. This is a scaling factor used to normalize the residuals of the physical equations. Optional, a reference vector. This can be taken as the climate average or the current regional average. Optional, scale factor. and It can be set based on the statistical variance of historical data or the typical magnitude of physical equations.

[0109] See Figure 5 In the correction of non-physical biases based on atmospheric dynamics and radiative transfer constraints, the evolution characteristics of the objective function loss value and constraint residuals during the iterative solution of a local optimization problem at a single grid point are demonstrated. Specifically, the objective function loss value (blue curve) is a normalized combination term including penalty terms for radiative and dynamic residuals. The radiative transfer residual (red curve) and atmospheric dynamic residual (green curve) correspond to the degree of constraint satisfaction of the radiative transfer equation and atmospheric dynamic equation, respectively. During the iteration process, in the initial stage (iteration number < 5), both the objective function loss value and the two types of residuals decrease rapidly, reflecting the rapid correction of significant non-physical biases by the optimization algorithm. As the iteration progresses (5 < iteration number < 15), the rate of decrease of each curve slows down, indicating that the physical quantity adjustment gradually approaches the optimal solution that satisfies the constraints. In the later stage (iteration number > 15), each curve tends to stabilize, indicating that the objective function has converged, the physical quantity adjustment has reached its minimum, and the radiative transfer and atmospheric dynamic constraints are satisfied. In terms of parameter configuration, the iterative process uses gradient descent combined with the Lagrange multiplier method to solve the problem. The weight coefficients of each penalty term in the objective function (such as the weights of radiation residuals and dynamic residuals) are set according to the constraint priority.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extrapolating multi-channel satellite data in optical flow-deep learning, characterized in that, The method includes: Acquire time-series satellite multi-channel observation data, preprocess and filter the time-series satellite multi-channel observation data to obtain optimized multi-channel time-series data; Multi-channel feature matching and displacement vector calculation are performed on continuous time frames in the optimized multi-channel time series data. The two-dimensional motion field is accurately estimated based on the variational optical flow algorithm. The two-dimensional motion field is preliminarily extrapolated by combining the semi-Lagrange algorithm and mass conservation constraints to generate the preliminary optical flow extrapolation result. A deep learning extrapolation model integrating spatiotemporal convolutional networks and recurrent neural networks is constructed. The preliminary extrapolation result of the optical flow is jointly input into the deep multi-channel features in the optimized multi-channel time series data. The deep learning extrapolation model is then trained to generate deep learning prediction results. An adaptive weight fusion mechanism is designed to dynamically allocate contribution weights to the preliminary extrapolation result of the optical flow and the prediction result of the deep learning, thereby generating a fused extrapolation result. Based on atmospheric dynamics equations and radiative transfer physical constraints, the fused extrapolation results are subjected to non-physical bias correction processing to generate a corrected multi-channel extrapolation field. The final output is a high-precision multi-channel observation field extrapolation product ranging from minute to hourly levels; The construction and training of the deep learning extrapolation model includes: A model architecture is constructed that includes a parallel spatiotemporal convolution branch and a recurrent network branch. The spatiotemporal convolution branch is used to extract spatial features and short-term features of multi-channel data, and the recurrent network branch is used to extract long-term dependency features. The model architecture incorporates channel attention and spatial attention mechanisms, enabling the model to adaptively focus on features of key channels and key spatial regions. The optimized multi-channel time-series data is input into the spatiotemporal convolution branch to extract deep multi-channel features. The deep multi-channel features are then concatenated with the preliminary extrapolation result of the optical flow along the feature dimension to form joint features. The joint features are input into the recurrent network branch to further learn and predict the temporal evolution pattern, and the deep learning prediction result is output. Using historical satellite multi-channel data and their corresponding future time-based ground truth data, the deep learning extrapolation model is trained end-to-end with the goal of minimizing prediction error.

2. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 1, characterized in that, The process of acquiring time-series satellite multi-channel observation data, and preprocessing and filtering the time-series satellite multi-channel observation data to obtain optimized multi-channel time-series data includes: The time-series satellite multi-channel observation data are sequentially subjected to radiometric calibration, geometric correction and atmospheric correction to generate preprocessed satellite multi-channel data; The preprocessed satellite multi-channel data undergoes adaptive filtering to eliminate noise interference and generate optimized multi-channel time-series data, specifically including: The preprocessed satellite multi-channel data is subjected to channel separation processing to generate multiple independent single-channel time-series data sequences; Calculate the noise statistical characteristics of each of the single-channel time-series data sequences, wherein the noise statistical characteristics include noise variance and noise spatial correlation; Based on the noise statistical characteristics corresponding to each channel, a filter kernel function that matches the noise characteristics of the channel is dynamically selected; Each single-channel time-series data sequence is subjected to adaptive spatial filtering using the selected filtering kernel function to generate denoised single-channel data. The denoised single-channel data of all channels are merged in the original channel order to generate the optimized multi-channel time-series data.

3. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 1, characterized in that, The precise estimation of the two-dimensional motion field based on the variational optical flow algorithm includes: Extract two consecutive frames of multi-channel image data from the optimized multi-channel time-series data; A multi-channel data fidelity term based on constant brightness and constant gradient is constructed, and combined with a smoothing constraint term, a variational optical flow energy functional is formed. The variational optical flow energy functional is minimized using a numerical optimization algorithm to calculate the motion displacement vector of each pixel between two consecutive frames. The motion displacement vectors are integrated in space to generate the two-dimensional motion field, which describes the velocity and direction of the observation field in two-dimensional space.

4. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 1, characterized in that, The preliminary spatiotemporal extrapolation of the two-dimensional motion field, combining the semi-Lagrange algorithm with mass conservation constraints, generates preliminary optical flow extrapolation results, including: Using the current observation field as the initial field, and based on the two-dimensional motion field, a semi-Lagrange algorithm is used to track the position of each grid point in the future. During the tracking process, mass conservation constraints are applied to the physical quantities of each grid point to ensure that the total mass remains constant during the motion. Based on the tracked future time grid point positions, the physical quantities of the initial field are mapped onto the prediction grid of the future time rules through interpolation methods to generate the preliminary extrapolation field for the future time. The process of tracking, applying constraints, mapping, and generating a preliminary extrapolation field is repeated for multiple consecutive future moments to generate the preliminary extrapolation result of the optical flow in time series form.

5. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 1, characterized in that, The design and application of the adaptive weight fusion mechanism include: Calculate the motion stability index of the preliminary extrapolation result of the optical flow during the target prediction period; Based on the pre-set forecast task requirements, determine the intensity of the requirement for physical consistency of forecast results and the intensity of the requirement for capturing complex nonlinear features. Using the motion stability index, the physical consistency requirement intensity, and the complex nonlinear feature capture requirement intensity as inputs, a learnable weight allocation network dynamically generates weight values ​​assigned to the preliminary extrapolation result of the optical flow and weight values ​​assigned to the deep learning prediction result. The generated weight values ​​are used to perform a weighted average fusion of the preliminary extrapolation result of the optical flow and the prediction result of the deep learning to generate the fused extrapolation result.

6. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 5, characterized in that, The calculation of the motion stability index of the preliminary extrapolation result of the optical flow during the target prediction period includes: Within the forecast period corresponding to the preliminary extrapolation result of the optical flow, the variation amplitude of the two-dimensional motion field between adjacent extrapolated fields is calculated; The mean and variance of the aforementioned variation amplitude over the entire spatial region are statistically analyzed and used as a measure of motion consistency. Analyze the temporal continuity of the direction of the two-dimensional motion field and calculate the standard deviation of the direction angle; The motion stability index is obtained by combining the mean and variance of the variation amplitude and the standard deviation of the direction angle through linear combination.

7. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 1, characterized in that, The non-physical bias correction process for the fused extrapolation results, based on atmospheric dynamics equations and radiative transfer physical constraints, includes: Multi-channel physical quantities, including temperature, humidity, water vapor, and cloud water, are extracted from the fusion extrapolation results. Substitute the extracted multi-channel physical quantities into a simplified set of atmospheric dynamics equations and check whether the spatiotemporal evolution of the multi-channel physical quantities satisfies the basic dynamic and thermodynamic relationships described by the set of equations. Substitute the aforementioned multi-channel physical quantities into the radiative transfer equation to calculate the simulated satellite channel radiation values; By comparing the simulated satellite channel radiation values ​​with the physically reasonable theoretical radiation value range, grid points and channels with non-physical deviations are identified. The physical quantities of the grid points and channels with non-physical deviations are adjusted to satisfy the constraints of the atmospheric dynamics equation and the radiative transfer equation, while minimizing the modification of the original fused extrapolation results, thus completing the non-physical deviation correction process and generating the corrected multi-channel extrapolation field.

8. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 7, characterized in that, The adjustment of the physical quantity values ​​of the grid points and channels identified as having non-physical deviations includes: For each grid point with non-physical deviation, a local optimization problem is constructed with the satisfaction of physical constraints as the optimization objective and the adjustment of physical quantities as the optimization variables. The physical constraints include local equilibrium constraints derived from the atmospheric dynamics equations and reasonable range constraints for channel radiation values ​​defined by the radiative transfer equations. The local optimization problem is solved by an iterative optimization algorithm to obtain the physical quantity correction value that maximizes the satisfaction of physical constraints and minimizes the adjustment amount. Replace the corresponding grid point and channel values ​​in the original fusion extrapolation result with the physical quantity correction value to complete the correction of a single grid point; Traverse all grid points with non-physical deviations, perform corrections sequentially, and generate the corrected multi-channel extrapolation field that satisfies global physical consistency.

9. The optical flow-deep learning satellite multi-channel data extrapolation method according to claim 2, characterized in that, The step of performing adaptive spatial filtering on each single-channel time-series data sequence using the selected filtering kernel function to generate denoised single-channel data includes: For each single-channel time-series data sequence, the size parameters of the filtering kernel function are dynamically determined based on the noise variance in the noise statistics. The selected filtering kernel function is convolved with the neighborhood of each pixel in the current single-channel time-series data sequence to calculate the filtered pixel value. The weight coefficients of the filtering kernel function are adaptively adjusted based on the local noise variance and image gradient information to preserve image edge features while denoising. The convolution operation results are normalized to generate denoised single-channel data; The denoised single-channel data is recombined in time series order to form a denoised single-channel time series data sequence.

Images