A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on a double-channel CNN

Abstract

The application discloses a distributed D-InSAR three-dimensional atmospheric wind field inversion method based on a double-channel CNN, solves the problem of inconsistent inversion results under different imaging conditions, enables the model to simultaneously capture the time evolution characteristics and spatial distribution characteristics of the atmospheric state, and learns the physical characteristic representation irrelevant to the imaging conditions, thereby improving the accuracy, stability and generalization ability of the wind field inversion, overcoming the problem of insufficient generalization ability of the traditional deep learning method, and establishing a multi-scale feature extraction system from local turbulence details to large-scale wind field trends through the progressive structure of the multi-level convolution layer and the pooling layer, thereby solving the problem of insufficient perception ability of the traditional single-scale feature extraction method to different spatial scale atmospheric phenomena, and improving the representation ability of the network to the complex atmospheric turbulence structure.

Description

Technical Field

[0001] This application relates to the field of data inversion technology, and in particular to a distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN. Background Technology

[0002] In the field of atmospheric measurement and meteorological observation, the acquisition of atmospheric wind fields mainly relies on a combination of various observation methods and numerical simulation methods. These traditional methods play an important role within their respective applicable ranges, but they also have limitations. Commonly used field observation methods, such as weather balloons, can directly acquire wind speed and direction along the vertical profile, providing high-precision upper-level wind speed data. However, their transmission frequency and spatial density are limited, making it difficult to meet the needs of high spatiotemporal resolution observation within the tropospheric boundary layer. Fixed or mobile lidar can continuously monitor wind fields within a certain altitude range and provide good temporal resolution, but their detection range, vertical coverage, and sensitivity are limited by instrument configuration and atmospheric conditions, and their equipment cost and maintenance complexity are high. Satellite remote sensing provides the possibility for inferring large-scale continuous horizontal wind fields, but it is usually affected by cloud cover, viewing angle, and sensor resolution. It lacks the ability to acquire wind speed information for lower-level and vertically layered structures, making it difficult to directly provide high-precision vertical profile data.

[0003] Spaceborne synthetic aperture radar (SAR) possesses all-weather, all-day, and high-resolution Earth observation capabilities. SAR signals are affected by refractive index and water vapor distribution as they travel through the atmosphere, resulting in significant atmospheric phase delay during interferometric processing. Using spaceborne SAR re-orbit observations, traditional differential interferometric SAR (D-InSAR) techniques can eliminate topographic and flat-ground phases, separating the atmospheric differential interferometric phase map between two observations. However, this phase only reflects the integral effect along the radar line of sight and cannot provide information on the vertical structure of atmospheric water vapor. The inversion results based on this only reflect the horizontally projected wind speed, failing to retrieve the vertical wind field components and making it difficult to further invert the three-dimensional atmospheric wind field.

[0004] Distributed synthetic aperture radar (SAR) employs multiple distributed SAR satellites for observation. Combined with differential interferometry (D-InSAR) technology, simultaneous or quasi-simultaneous multi-angle atmospheric differential interferometric phase maps can be obtained, leading to the three-dimensional refractive index distribution. Due to the high sensitivity of radar signal phase to water vapor, the three-dimensional refractive index distribution primarily characterizes the integrated water vapor distribution in each atmospheric layer at two time points during satellite reorbit observations. Obtaining a precise time-delayed three-dimensional atmospheric refractive index distribution allows for understanding the movement of three-dimensional water vapor masses during the time delay, thereby enabling accurate measurement of the three-dimensional distribution of atmospheric wind speed.

[0005] However, due to the drastic changes in the morphology of low-level water vapor masses and the weakening of high-level water vapor mass characteristics in tomographic inversion results, existing methods such as registration and thresholding methods have low accuracy in retrieving three-dimensional wind fields. Therefore, it is urgent to propose a new method to solve the problem of insufficient three-dimensional wind field inversion capability at present. Convolutional neural networks (CNNs) have shown advantages in image feature extraction, providing an effective approach to solving the aforementioned image feature extraction problem. Summary of the Invention

[0006] The main objective of this application is to provide a distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN to solve the problems mentioned in the background.

[0007] To achieve the above objectives, this application provides the following technical solution: A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN is described below: S1. Data acquisition: The target area is synchronously observed by multiple distributed synthetic aperture radar satellites at short time intervals to acquire radar observation data under different imaging conditions. The imaging conditions are determined by the radar operating frequency band, incident angle, orbital geometric parameters and interference time interval. At the same time, the three-dimensional refractive index field and tomographic wind field data of the target area are acquired. S2, Delay Inversion: Differential interferometry is performed on radar observation data to calculate the atmospheric tropospheric slant range delay and map it to the interferometric phase. A three-dimensional differential tropospheric refractive index tomographic image of the target area is obtained through three-dimensional tomographic inversion. S3. Working condition modeling: Extract the imaging working condition parameters corresponding to each group of three-dimensional differential refractive index data, normalize the parameters and integrate them into a standardized description vector; S4. Data Normalization: Based on the imaging condition parameter description vector, the amplitude normalization and spatial statistical characteristic standardization of the three-dimensional differential refractive index data under different conditions are performed. S5. Dual-channel construction: Select two sets of standardized three-dimensional differential refractive index data within a short time interval, convert them into grayscale images, and stitch them together along the channel dimension in chronological order to generate a dual-channel input tensor. S6. Feature Extraction: Construct a convolutional neural network model containing a dual-channel input layer, multi-level convolutional layers, pooling layers, flattening layers, and fully connected layers. Introduce imaging condition parameter description vectors into the convolutional layers for feature modulation. The formula for convolution operation is:

[0008] in, The input feature map; For the first k One convolutional kernel; bk This is the bias; the number of output channels is the same as the number of convolution kernels.

[0009] After the convolution operation, the ReLU function is applied to non-linearly transform the output, helping the network learn more complex features and preventing gradient vanishing. The formula for the ReLU function is:

[0010] S7. Joint Training: The tomographic wind field data is used as a supervision label and paired with the dual-channel input tensor to form a training sample set. A hybrid batch training strategy is adopted and feature consistency constraints are applied to samples of different working conditions with the same physical wind field. The network parameters are iteratively updated through backpropagation. S8. Model Validation: Input the validation sample set into the trained network model, calculate the error between the predicted wind field and the reference wind field, and evaluate the inversion accuracy of the model under unknown imaging conditions. S9, Wind Field Output: Acquire new distributed radar observation data, process it according to the above procedure, input it into the verified network model, and output the three-dimensional wind speed and wind direction of the target area at the corresponding observation time.

[0011] Preferably, in step S1, the specific method is as follows: S1.1 Deploy multiple distributed synthetic aperture radar satellites to conduct short-time synchronous observations of the target area. Each satellite images the same area under different orbital positions, incident angles, and operating frequency bands, forming a radar echo dataset covering multiple combinations of observation geometry and system parameters. Record the operating frequency band, incident angle, orbital geometry parameters, and interferometric time interval corresponding to each set of data, and establish a working condition identification code to map the observation data with the working condition parameters. S1.2 Obtain three-dimensional atmospheric refractive index field data of the target area during the observation period through meteorological detection equipment measurement or atmospheric physical model simulation, obtain three-dimensional wind field data of the corresponding spatiotemporal range using wind measuring radar or meteorological reanalysis data, perform spatiotemporal registration of refractive index field data with radar observation data, and associate the three-dimensional wind field data as true value labels with radar observation data to form training samples containing input features and output labels.

[0012] Preferably, in step S2, the specific method is as follows: S2.1. The multi-temporal radar observation data obtained in step S1 are registered and interferometrically processed. The three-dimensional refractive index field data obtained in step S1.2 is combined with satellite orbital parameters to calculate the atmospheric tropospheric slant range delay from the satellite to the ground pixel under each observation angle. The calculated slant range delay is mapped to the phase component of the complex interferogram. Interference terms such as thermal noise, residual ionospheric phase error and orbital error are superimposed to generate a simulated differential interferometric phase map. The differential tropospheric delay observation value is extracted from the interferometric phase map under different observation geometry. S2.2. Based on the multi-view observation geometry of the distributed satellite system, construct the three-dimensional spatial observation equation of the troposphere. Use the differential tropospheric delayed observation values ​​extracted in step S2.1 as the observation vector. Solve the observation equation using algebraic reconstruction or regularized iterative algorithm to reconstruct the three-dimensional differential tropospheric refractive index spatial distribution of the target area within a short time interval. Generate a differential refractive index tomographic image containing height layer information. Slice the tomographic image by height layer and convert it into a two-dimensional matrix format for storage.

[0013] Preferably, in step S3, the specific method is as follows: S3.1 Extract the imaging condition parameters corresponding to each group of three-dimensional differential refractive index tomography images from the condition identification code established in step S1.1, including radar operating frequency band, incident angle, orbital altitude, baseline length and interferometric time interval. Classify and organize the extracted condition parameters according to radar system parameters and observation geometric parameters, establish an index mapping table between three-dimensional differential refractive index data and imaging condition parameters, and record the complete set of condition parameters corresponding to each group of tomography images to ensure the traceability of condition information in subsequent processing. S3.2. The extracted imaging parameters are normalized according to their types, the radar operating frequency band is mapped to the zero-to-one interval, the incident angle is normalized according to the maximum observation angle, the orbital height and baseline length are standardized according to the system design range, the interferometric time interval is normalized according to the maximum time window, and the normalized parameters are spliced ​​into a fixed-dimensional standardized description vector in a preset order, which is used as the input parameters for subsequent feature modulation.

[0014] Preferably, in step S4, the specific method is as follows; S4.1 Read the standardized description vector constructed in step S3.2, calculate the theoretical atmospheric refractive index dynamic range under the corresponding operating conditions based on the radar operating frequency band and incident angle parameters in the vector, perform linear transformation on the three-dimensional differential refractive index data generated in step S2.2 according to the calculated dynamic range, map the refractive index values ​​under different operating conditions to the zero-to-one interval, compensate for the systematic deviation of refractive index amplitude caused by the difference in radar system parameters, and store the transformed data in a unified numerical format. S4.2 Calculate the mean and standard deviation of each set of three-dimensional differential refractive index data in the spatial dimension. Determine the spatial correlation scale based on the orbital geometry parameters and interference time interval in the standardized description vector of step S3.2. Perform a standardization transformation on the refractive index data by subtracting the mean and dividing the standard deviation. Interpolate and resample the transformed data according to a uniform spatial sampling interval to compensate for the spatial distribution characteristic deviation caused by the difference in observation geometry. Organize the resampled data into a three-dimensional array according to the height layer.

[0015] Preferably, in step S5, the specific method is as follows; S5.1 Select two sets of differential refractive index data of the same target area within a short time interval from the three-dimensional array generated in step S4.2, extract two-dimensional slices of the same height layer from the two sets of data, quantize and map the refractive index values ​​in each two-dimensional slice according to the gray level from zero to 255, generate the corresponding eight-bit grayscale image, store each pixel value of the grayscale image as the corresponding element of the two-dimensional matrix, and mark the two sets of grayscale images as the first time phase and the second time phase respectively according to the order of observation time. S5.2. Take the first temporal grayscale image matrix generated in step S5.1 as the first channel input and the second temporal grayscale image matrix as the second channel input. Stack the two two-dimensional matrices into a three-dimensional tensor along the channel dimension. The first and second dimensions of the three-dimensional tensor correspond to the number of rows and columns of the image, respectively, and the third dimension corresponds to the number of channels and has a value of two. Use the stitched three-dimensional tensor as the input data of the convolutional neural network to retain temporal information and spatial distribution information.

[0016] Preferably, in step S6, the specific method is as follows; S6.1 Construct a dual-channel input layer to receive the three-dimensional tensor generated in step S5.2. Set the first-level convolutional layer to contain thirty-two 3x3 convolutional kernels to perform convolution operations on the input tensor. The output of the convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Set the second-level convolutional layer to contain sixty-four 3x3 convolutional kernels to perform convolution operations on the pooled feature map. The output of the second-level convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Construct a flattening layer to convert the pooled two-dimensional feature map into a one-dimensional feature vector. S6.2. Input the standardized description vector generated in step S3.2 into the feature modulation module. The feature modulation module maps the description vector into a modulation weight vector with the same number of channels as the output feature map of the first-level convolutional layer through a fully connected layer. The modulation weight vector is multiplied element-wise by channel with the output feature map of the first-level convolutional layer to construct a fully connected layer containing 128 neurons to receive the flattened feature vector. The output layer contains the same number of neurons as the wind speed and wind direction prediction dimension. The output layer directly outputs the three-dimensional wind field prediction value.

[0017] Preferably, in step S7, the specific method is as follows; S7.1. The tomographic wind field data obtained in step S1.2 is used as a supervision label and paired with the dual-channel input tensor generated in step S5.2 according to the spatiotemporal correspondence to form training samples. Independent samples from the training samples with different data sources from the training set are selected to construct a validation sample set. The training samples are grouped and labeled according to the imaging condition type. Samples from different imaging conditions are randomly selected in each training batch and mixed. Each batch contains samples covering at least three different imaging condition types. The condition identifier and physical wind field identifier corresponding to each sample are recorded. S7.2 Input the mixed batch samples into the convolutional neural network model constructed in step S6, calculate the mean square error between the wind field prediction value output by the network and the supervision label as the prediction loss, extract the feature maps output by the first convolutional layer for samples with the same physical wind field label but different operating condition labels, calculate the cosine distance between these feature maps as the feature consistency loss, and weight the prediction loss and the feature consistency loss to form the total loss function. Calculate the gradient of the total loss function with respect to the network parameters through the backpropagation algorithm, and update the network parameters using an adaptive learning rate optimization algorithm.

[0018] Preferably, in step S8, the specific method is as follows; S8.1 Input the dual-channel input tensor of the verification sample set constructed in step S7.1 into the convolutional neural network model trained in step S7.2. The network model outputs the corresponding three-dimensional wind field prediction value. Extract the reference wind field data paired with the input tensor from the verification sample set. Calculate the root mean square error and mean absolute error of the predicted wind field and the reference wind field in the wind speed dimension. Calculate the mean angle deviation and standard deviation of the angle deviation of the predicted wind field and the reference wind field in the wind direction dimension. Group and statistically analyze the error index under each working condition according to the imaging working condition type corresponding to the verification sample. S8.2 Count the number of imaging conditions in the verification sample set that do not appear in the training sample set, calculate the root mean square error of wind speed and the wind direction angle deviation of these unknown conditions, compare the error index of the unknown conditions with the error index of the known conditions, calculate the percentage increase of the unknown condition error relative to the known condition error, when the percentage increase is less than the preset threshold, the model is determined to pass the verification, record the model parameters that have passed the verification and save them as a model file, and transfer the model file to step S9 for wind field inversion.

[0019] Preferably, in step S9, the specific method is as follows; S9.1. New radar echo data is obtained by synchronously observing the target area in a short time interval using multiple distributed synthetic aperture radar satellites. Differential interferometry processing and three-dimensional tomographic inversion are performed on the radar echo data to generate a three-dimensional differential refractive index tomographic image. Imaging parameters corresponding to the new observation data are extracted and normalized to generate a standardized description vector. Amplitude normalization and spatial statistical characteristic standardization transformation are performed on the three-dimensional differential refractive index data. Two sets of standardized data within a short time interval are selected and converted into grayscale images. The two grayscale images are stitched together along the channel dimension to generate a dual-channel input tensor. S9.2 Read the model file saved in step S8.2 and load the convolutional neural network model parameters into memory. Input the dual-channel input tensor and standardized description vector generated in step S9.1 into the dual-channel input layer and feature modulation module of the network model, respectively. The network model outputs a prediction vector containing wind speed values ​​and wind direction angles. Reorganize the prediction vector into a three-dimensional array according to the three dimensions of height, longitude and latitude. Each element of the three-dimensional array stores the wind speed and wind direction data of the corresponding spatial location. Write the three-dimensional array into a network general data format file.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention solves the problem of inconsistent inversion results under different imaging conditions, enabling the model to simultaneously capture the temporal evolution and spatial distribution characteristics of atmospheric conditions, and learn physical feature representations independent of imaging conditions, thereby improving the accuracy, stability and generalization ability of wind field inversion and overcoming the problem of insufficient generalization ability of traditional deep learning methods.

[0021] 2. The progressive structure of multi-level convolutional layers and pooling layers in this invention establishes a multi-scale feature extraction system from local turbulence details to large-scale wind field trends, which solves the problem that traditional single-scale feature extraction methods are insufficient in their ability to perceive atmospheric phenomena at different spatial scales, and improves the network's ability to represent complex atmospheric turbulence structures.

[0022] 3. The introduction of the working condition adaptive feature modulation mechanism in this invention breaks through the limitations of traditional fixed-weight convolutional networks. Through dynamic weight adjustment driven by working condition parameters, the network can adaptively adjust the feature response according to different imaging conditions, realizing working condition-independent physical feature learning. Compared with traditional methods, it improves the model's cross-working condition generalization ability and adaptability to unknown working conditions.

[0023] 4. The end-to-end network structure design of this invention integrates feature extraction, operating condition modulation, and wind field prediction into a unified deep learning framework, avoiding the error accumulation problem in traditional multi-step processing methods, improving the automation level and overall accuracy of the inversion process, and providing an efficient and reliable technical means for the operational application of atmospheric wind fields in distributed radar systems. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the steps of the method described in this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] Example 1: Please refer to Figure 1 A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN is presented below: S1. Data acquisition: The target area is synchronously observed by multiple distributed synthetic aperture radar satellites at short time intervals to acquire radar observation data under different imaging conditions. The imaging conditions are determined by the radar operating frequency band, incident angle, orbital geometry parameters and interference time interval. At the same time, the three-dimensional refractive index field and tomographic wind field data of the target area are acquired. S2, Delay Inversion: Differential interferometry is performed on radar observation data to calculate the atmospheric tropospheric slant range delay and map it to the interferometric phase. A three-dimensional differential tropospheric refractive index tomographic image of the target area is obtained through three-dimensional tomographic inversion. S3. Working condition modeling: Extract the imaging working condition parameters corresponding to each set of three-dimensional differential refractive index data, normalize the parameters and integrate them into a standardized description vector; S4. Data Normalization: Based on the imaging condition parameter description vector, the amplitude normalization and spatial statistical characteristic standardization of the three-dimensional differential refractive index data under different conditions are performed. S5. Dual-channel construction: Select two sets of standardized three-dimensional differential refractive index data within a short time interval, convert them into grayscale images, and stitch them together along the channel dimension in chronological order to generate a dual-channel input tensor. S6. Feature Extraction: Construct a convolutional neural network model containing a dual-channel input layer, multi-level convolutional layers, pooling layers, flattening layers, and fully connected layers. Introduce imaging condition parameter description vectors into the convolutional layers for feature modulation. The formula for convolution operation is:

[0029] in, The input feature map; For the first k One convolutional kernel; b k This is the bias; the number of output channels is the same as the number of convolution kernels.

[0030] After the convolution operation, the ReLU function is applied to non-linearly transform the output, helping the network learn more complex features and preventing gradient vanishing. The formula for the ReLU function is:

[0031] S7. Joint Training: The tomographic wind field data is used as a supervision label and paired with the dual-channel input tensor to form a training sample set. A hybrid batch training strategy is adopted and feature consistency constraints are applied to samples of different working conditions with the same physical wind field. The network parameters are iteratively updated through backpropagation. S8. Model Validation: Input the validation sample set into the trained network model, calculate the error between the predicted wind field and the reference wind field, and evaluate the inversion accuracy of the model under unknown imaging conditions. S9, Wind Field Output: Acquire new distributed radar observation data, process it according to the above procedure, input it into the verified network model, and output the three-dimensional wind speed and wind direction of the target area at the corresponding observation time.

[0032] In this embodiment: Step S1 utilizes short-interval synchronous observations from multiple distributed synthetic aperture radar satellites and a condition identification technique to achieve the beneficial effects of constructing radar observation datasets covering various imaging conditions and obtaining supervised labels, thereby improving the diversity of training samples and the reliability of ground truth. Condition identification establishes a clear correspondence between each set of observation data and its imaging parameters, providing a traceable parameter source for subsequent condition parameter normalization and feature modulation. This avoids the problem of insufficient model generalization ability caused by missing condition information in traditional methods, and improves the standardization of data management and the accuracy of subsequent processing.

[0033] Step S2, through differential interferometry and three-dimensional tomographic inversion techniques, achieves the beneficial effect of reconstructing the three-dimensional spatial distribution of refractive index from two-dimensional phase observation, improving the vertical resolution and spatial coverage of the atmospheric refractive index field. The calculation of slant range delay and phase mapping process simulates the error characteristics under real observation conditions, improving the realism of the tomographic inversion results. The generated three-dimensional differential refractive index tomographic image contains spatial variation information of atmospheric refractive index in the vertical and horizontal directions, providing high-resolution input data for subsequent neural network extraction of atmospheric turbulence features and wind field inversion, overcoming the problem of insufficient vertical resolution caused by traditional methods that only utilize two-dimensional phase information.

[0034] Step S3, through the extraction and normalization of imaging condition parameters and the construction of standardized description vectors, achieves a unified representation of condition parameters with different dimensions and numerical ranges, improving the comparability and usability of condition information. The standardized description vectors enable the quantification of differences between different imaging conditions using a unified numerical form, providing clear condition information input for subsequent data standardization and feature modulation. The establishment of the index mapping table ensures the traceability of the correspondence between condition parameters and refractive index data, avoiding data processing errors caused by chaotic condition information in traditional methods, and improving the reliability and repeatability of the entire inversion process.

[0035] Step S4 eliminates the systematic differences in refractive index data under different imaging conditions through adaptive amplitude normalization and spatial statistical characteristic standardization techniques, improving data consistency and comparability. The normalization strategy based on the dynamic range of the operating condition parameter calculation theory maps data from different radar systems and observation geometries to a unified numerical range, avoiding the poor adaptability of traditional fixed threshold methods to different operating conditions. Spatial statistical characteristic standardization makes the data under different operating conditions more consistent in spatial correlation and distribution, creating conditions for neural networks to extract operating condition-independent physical features and improving the model's cross-operating condition generalization ability.

[0036] Step S5, by selecting two sets of standardized refractive index data within a short time interval and converting them into a dual-channel input tensor, effectively encodes the temporal variation information of atmospheric state, improving the spatiotemporal feature representation capability of the input data. The quantization mapping of grayscale images converts continuous refractive index values ​​into discrete grayscale levels suitable for convolutional neural network processing, reducing data dimensionality and computational complexity. The dual-channel stitching structure enables the network to simultaneously perceive the spatial distribution and temporal evolution characteristics of atmospheric refractive index at two different times. Compared to traditional single-phase input methods, this captures the dynamic changes in atmospheric turbulence, improving the accuracy of wind speed and direction prediction.

[0037] Step S6 constructs a neural network model containing a dual-channel input layer and multi-level convolutional layers, and introduces an adaptive feature modulation mechanism based on operating conditions. This enables multi-scale feature extraction, from local turbulence details to large-scale wind field trends, improving the network's ability to perceive atmospheric phenomena at different spatial scales. The feature modulation mechanism driven by the operating condition parameter description vector adaptively adjusts the feature responses under different operating conditions in the convolutional layers, allowing the network to learn atmospheric physical feature representations independent of imaging conditions. Compared to traditional fixed-weight convolutional networks, this enhances the model's adaptability to different radar systems and observation geometries, improving the model's cross-operating-condition generalization performance and inversion accuracy.

[0038] Step S7 establishes a stable mapping relationship across different operating conditions by employing a hybrid batch training strategy and a feature consistency constraint mechanism, thereby improving the model's generalization ability and inversion stability. The hybrid batch training strategy exposes the network to samples from multiple imaging conditions during the same training process, avoiding the overfitting problem caused by traditional single-condition training. The feature consistency constraint mechanism guides the network to learn condition-invariant physical feature representations by calculating the cosine distance between feature maps of samples from the same physical wind field under different operating conditions. The weighted combination of prediction loss and consistency loss ensures wind field prediction accuracy while constraining the consistency of the feature space, improving the model's adaptability and practicality to unknown operating conditions.

[0039] Step S8, through the calculation of multi-dimensional error indices and the comparison of errors under known and unknown operating conditions, achieves a comprehensive evaluation of the model's wind field inversion accuracy and cross-condition generalization ability, improving the scientific rigor and reliability of model validation. The root mean square error and mean absolute error of wind speed, as well as the angle deviation indices of wind direction, accurately quantify the difference between the predicted results and the true values. The comparative analysis of errors under unknown and known operating conditions objectively evaluates the model's cross-condition generalization ability by calculating the percentage increase in error. The threshold-based validation mechanism ensures that only models with satisfactory generalization performance can be used in practical applications, improving the safety and reliability of model application.

[0040] Step S9, by reusing the data processing methods and model parameter loading mechanisms from previous steps, realizes a complete application process from new observation data to wind field inversion results, improving the automation and timeliness of atmospheric wind field inversion. The reuse of data processing methods ensures consistency between new and training data in format and feature space, while the model parameter loading and dual-input prediction mechanism enable the network to respond quickly to new observation data. The three-dimensional array organizes wind field information according to height, longitude, and latitude, and the output of the network's general data format file ensures compatibility between the inversion results and existing meteorological data systems, improving the practicality and scalability of the wind field products.

[0041] This invention achieves high-precision cross-condition inversion of three-dimensional atmospheric wind fields from distributed differential interferometric synthetic aperture radar (DISAR) through a combination of multiple steps, improving the model's adaptability to different radar systems and observation geometries. Compared to traditional methods, this invention solves the problem of inconsistent inversion results under different imaging conditions, enabling the model to simultaneously capture the temporal evolution and spatial distribution characteristics of atmospheric states, and learn physical feature representations independent of imaging conditions. This improves the accuracy, stability, and generalization ability of wind field inversion, overcoming the insufficient generalization ability of traditional deep learning methods.

[0042] Example 2: Please refer to Figure 1 In step S1, the specific method is as follows: S1.1 Deploy multiple distributed synthetic aperture radar satellites to conduct short-time synchronous observations of the target area. Each satellite images the same area under different orbital positions, incident angles, and operating frequency bands, forming a radar echo dataset covering multiple combinations of observation geometry and system parameters. Record the operating frequency band, incident angle, orbital geometry parameters, and interferometric time interval corresponding to each set of data, and establish a working condition identification code to map the observation data with the working condition parameters. S1.2 Obtain three-dimensional atmospheric refractive index field data of the target area during the observation period through meteorological detection equipment measurement or atmospheric physical model simulation, obtain three-dimensional wind field data of the corresponding spatiotemporal range using wind measuring radar or meteorological reanalysis data, perform spatiotemporal registration of refractive index field data with radar observation data, and associate the three-dimensional wind field data as true value labels with radar observation data to form training samples containing input features and output labels.

[0043] In this embodiment: Step S1.1 involves deploying multiple distributed synthetic aperture radar satellites to conduct short-interval synchronous observations of the same area at different orbital positions, incident angles, and operating frequency bands. This enables the construction of radar echo datasets covering various combinations of observation geometry and system parameters, improving the diversity and coverage of training data. The establishment of operational condition identification codes establishes a clear mapping relationship between each set of radar observation data and its corresponding operating frequency band, incident angle, orbital geometry parameters, and interferometric time interval. This provides a traceable parameter source for subsequent operational condition parameter normalization and feature modulation, avoiding the data management chaos caused by missing or disordered operational condition information in traditional methods, and improving the standardization of data organization and the accuracy of subsequent processing.

[0044] Step S1.2 involves acquiring three-dimensional atmospheric refractive index field data through meteorological detection equipment or atmospheric physics model simulation, and obtaining three-dimensional wind field data using wind-measuring radar or meteorological reanalysis data. This achieves the fusion of multi-source reference data, improving the reliability of the ground truth of the supervision labels and the diversity of data sources. Spatiotemporal registration technology precisely aligns the refractive index field data with radar observation data in both time and space, ensuring consistency between input features and reference data. Associating the three-dimensional wind field data as ground truth labels with radar observation data to form complete training samples establishes a clear correspondence between input features and output labels, providing high-quality training data for subsequent supervised learning of the neural network and improving the effectiveness of model training.

[0045] Step S1, through the combination of the two sub-steps mentioned above, realizes a complete process from acquiring radar observation data under multiple operating conditions to constructing supervised training samples, bringing three important improvements compared to traditional methods. First, simultaneous observations from multiple satellites under different imaging conditions overcome the limitations of traditional single-condition data acquisition, providing a data foundation for cross-condition feature learning and improving the model's adaptability to different radar systems and observation geometries. Second, the establishment of operating condition identification codes enables standardized management of observation data and imaging parameters, solving the data processing difficulties caused by missing operating condition information in traditional methods, and improving the traceability of data management and the reliability of the processing flow. Third, multi-source reference data fusion and spatiotemporal registration techniques ensure the accuracy of the ground truth and the consistency of input and output of the training samples. Compared to traditional single-source data source methods, this improves the reliability of supervised labels, laying a solid foundation for establishing an accurate refractive index-to-wind field mapping relationship in the subsequent neural network, and improving the quality of model training and inversion accuracy.

[0046] Example 3: Please refer to Figure 1 In step S2, the specific method is as follows: S2.1. The multi-temporal radar observation data obtained in step S1 are registered and interferometrically processed. The three-dimensional refractive index field data obtained in step S1.2 is combined with satellite orbital parameters to calculate the atmospheric tropospheric slant range delay from the satellite to the ground pixel under each observation angle. The calculated slant range delay is mapped to the phase component of the complex interferogram. Interference terms such as thermal noise, residual ionospheric phase error and orbital error are superimposed to generate a simulated differential interferometric phase map. The differential tropospheric delay observation value is extracted from the interferometric phase map under different observation geometry. S2.2. Based on the multi-view observation geometry of the distributed satellite system, construct the three-dimensional spatial observation equation of the troposphere. Use the differential tropospheric delayed observation values ​​extracted in step S2.1 as the observation vector. Solve the observation equation using algebraic reconstruction or regularized iterative algorithm to reconstruct the three-dimensional differential tropospheric refractive index spatial distribution of the target area within a short time interval. Generate a differential refractive index tomographic image containing height layer information. Slice the tomographic image by height layer and convert it into a two-dimensional matrix format for storage.

[0047] In this embodiment: Step S2.1 involves registering and interferometrically processing multi-temporal radar observation data, and calculating the atmospheric tropospheric slant range delay using three-dimensional refractive index field data combined with satellite orbital parameters. This achieves the conversion from radar echo data to differential interferometric phase data, improving the extraction accuracy of atmospheric delay information. The process of mapping the slant range delay to the phase components of the complex interferogram establishes a quantitative relationship between atmospheric refractive index and phase observation. The method of generating a simulated differential interferometric phase map by superimposing interference terms such as thermal noise, residual ionospheric phase error, and orbital error realistically reproduces the error characteristics under actual observation conditions, improving the authenticity of the simulated data. Extracting differential tropospheric delay observation values ​​from interferometric phase maps under different observation geometries provides multi-view observation input for subsequent three-dimensional tomographic inversion, improving the reliability of the inversion results.

[0048] Step S2.1 involves registering and interferometrically processing multi-temporal radar observation data, and calculating the atmospheric tropospheric slant range delay using three-dimensional refractive index field data combined with satellite orbital parameters. This achieves the conversion from radar echo data to differential interferometric phase data, improving the accuracy of atmospheric delay information extraction. The process of mapping the slant range delay to the phase components of the complex interferogram establishes a quantitative relationship between atmospheric refractive index and phase observations. The method of generating a simulated differential interferometric phase map by superimposing interference terms such as thermal noise, residual ionospheric phase error, and orbital error realistically reproduces the error characteristics under actual observation conditions. The differential tropospheric delay observation values ​​extracted from different observation geometries provide multi-view observation inputs for subsequent three-dimensional tomographic inversion, improving the reliability of the inversion results.

[0049] Example 4: Please refer to Figure 1 In step S3, the specific method is as follows: S3.1 Extract the imaging condition parameters corresponding to each group of three-dimensional differential refractive index tomography images from the condition identification code established in step S1.1, including radar operating frequency band, incident angle, orbital altitude, baseline length and interferometric time interval. Classify and organize the extracted condition parameters according to radar system parameters and observation geometric parameters, establish an index mapping table between three-dimensional differential refractive index data and imaging condition parameters, and record the complete set of condition parameters corresponding to each group of tomography images to ensure the traceability of condition information in subsequent processing. S3.2. The extracted imaging parameters are normalized according to their types, the radar operating frequency band is mapped to the zero-to-one interval, the incident angle is normalized according to the maximum observation angle, the orbital height and baseline length are standardized according to the system design range, the interferometric time interval is normalized according to the maximum time window, and the normalized parameters are spliced ​​into a fixed-dimensional standardized description vector in a preset order, which is used as the input parameters for subsequent feature modulation.

[0050] In this embodiment: Step S3.1 extracts the imaging condition parameters corresponding to each group of three-dimensional differential refractive index tomographic images from the condition identification code, and organizes them according to radar system parameters and observation geometric parameters. This achieves structured management of condition information, improving the accuracy of parameter extraction and the standardization of organization. The establishment of the index mapping table creates a clear correspondence between the three-dimensional differential refractive index data and the imaging condition parameters, ensuring the traceability of the complete set of condition parameters corresponding to each group of tomographic images and avoiding data processing errors caused by missing or chaotic condition information in traditional methods. The classification and organization of condition parameters provides a basis for subsequently adopting different normalization strategies for different parameter types, improving the flexibility and accuracy of parameter processing.

[0051] Step S3.2 normalizes the numerical range of different types of imaging parameters, achieving a unified representation of parameters with different dimensions and numerical ranges, thus improving the comparability and usability of the imaging parameters. For different parameter types such as radar operating frequency band, incident angle, orbital altitude, baseline length, and interferometric time interval, appropriate normalization methods are used to map each parameter to a unified numerical range, eliminating the impact of parameter dimension differences on subsequent processing. The normalized parameters are then concatenated in a preset order into a standardized description vector of fixed dimensions, establishing a unified numerical representation of the imaging parameters. This provides a standardized input interface for the subsequent feature modulation module of the neural network, improving the compatibility between the imaging parameters and the network model.

[0052] Step S3, through the combination of the two sub-steps mentioned above, achieves a complete conversion process from raw operating condition parameters to standardized descriptive vectors, bringing three significant improvements compared to traditional methods. First, the structured extraction and classification of operating condition parameters solves the problem of chaotic operating condition information management in traditional methods. The establishment of the index mapping table ensures the traceability of the correspondence between operating condition parameters and refractive index data, improving the standardization of data management and the reliability of the processing flow. Second, adopting appropriate normalization methods for different parameter types is more refined than the traditional unified normalization strategy, fully considering the physical meaning and numerical characteristics of different parameters, improving the accuracy and rationality of normalization processing. Third, the construction of fixed-dimensional standardized descriptive vectors establishes a standardized interface between operating condition information and the neural network, enabling the differences between different imaging conditions to be quantified and transmitted through a unified numerical form. This provides key parameter inputs for the subsequent implementation of the operating condition adaptive feature modulation mechanism, enhancing the model's perception and adaptability to different imaging conditions, and laying the foundation for high-precision wind field inversion across operating conditions.

[0053] Example 5: Please refer to Figure 1 In step S4, the specific method is as follows; S4.1 Read the standardized description vector constructed in step S3.2, calculate the theoretical atmospheric refractive index dynamic range under the corresponding operating conditions based on the radar operating frequency band and incident angle parameters in the vector, perform linear transformation on the three-dimensional differential refractive index data generated in step S2.2 according to the calculated dynamic range, map the refractive index values ​​under different operating conditions to the zero-to-one interval, compensate for the systematic deviation of refractive index amplitude caused by the difference in radar system parameters, and store the transformed data in a unified numerical format. S4.2 Calculate the mean and standard deviation of each set of three-dimensional differential refractive index data in the spatial dimension. Determine the spatial correlation scale based on the orbital geometry parameters and interference time interval in the standardized description vector of step S3.2. Perform a standardization transformation on the refractive index data by subtracting the mean and dividing the standard deviation. Interpolate and resample the transformed data according to a uniform spatial sampling interval to compensate for the spatial distribution characteristic deviation caused by the difference in observation geometry. Organize the resampled data into a three-dimensional array according to the height layer.

[0054] In this embodiment: Step S4.1 calculates the theoretical atmospheric refractive index dynamic range under the corresponding operating condition based on the radar operating frequency band and incident angle parameters in the standardized description vector, and performs an operating condition-adaptive linear transformation on the three-dimensional differential refractive index data. This achieves amplitude uniformity of refractive index data under different radar system parameters, improving data comparability and consistency. The normalization strategy based on operating condition parameters to calculate the theoretical dynamic range is more accurate than the traditional fixed threshold method. It can fully consider the influence of different radar operating frequencies and incident angles on the sensitivity of atmospheric refractive index observation, effectively compensate for the systematic deviation of refractive index amplitude caused by differences in radar system parameters, provide standardized input data for subsequent neural network processing, and improve the model's adaptability to data under different operating conditions.

[0055] Step S4.2 calculates the mean and standard deviation of each set of three-dimensional differential refractive index data in the spatial dimension, and performs a standardization transformation based on the spatial correlation scale determined by the orbital geometry parameters and the interferometric time interval. This unifies the spatial statistical characteristics of the refractive index data under different observation geometry conditions, improving the spatial consistency of the data. The standardization transformation of subtracting the mean and dividing by the standard deviation eliminates the mean offset and variance differences of the data under different operating conditions, making the data distribution tend to a standard normal distribution, which is convenient for the neural network to extract stable feature representations. Interpolation and resampling according to a unified spatial sampling interval compensates for the spatial distribution characteristic deviation caused by differences in observation geometry. The resampled data is organized into a three-dimensional array according to the height layer, providing a unified input format for subsequent grayscale image conversion and improving the standardization of the data processing flow.

[0056] Step S4, through the combination of the two sub-steps mentioned above, achieves a complete conversion process from the original three-dimensional differential refractive index data to a standardized three-dimensional array, bringing three significant improvements compared to traditional methods. First, the condition-adaptive amplitude normalization strategy calculates the theoretical dynamic range based on different radar system parameters, which is more accurate and flexible than the traditional fixed threshold normalization method. It effectively compensates for systematic deviations under different operating conditions, solving the problem of poor adaptability of traditional methods to different radar systems and observation conditions, and improving the accuracy of data standardization. Second, the statistical characteristic standardization method based on operating condition parameters to determine spatial correlation scales fully considers the influence of different observation geometry conditions on the spatial distribution characteristics of atmospheric refractive index. This is more reasonable than the traditional unified standardization method, improving the consistency of spatial statistical characteristics of data under different operating conditions. Third, the dual standardization processing of amplitude and spatial dimensions establishes a complete data standardization system, making the refractive index data under different imaging conditions tend to be consistent in both numerical magnitude and spatial distribution characteristics. This creates conditions for neural networks to learn condition-independent physical feature representations, improving the model's cross-condition generalization ability and inversion accuracy.

[0057] Example 6: Please refer to Figure 1 In step S5, the specific method is as follows; S5.1 Select two sets of differential refractive index data of the same target area within a short time interval from the three-dimensional array generated in step S4.2, extract two-dimensional slices of the same height layer from the two sets of data, quantize and map the refractive index values ​​in each two-dimensional slice according to the gray level from zero to 255, generate the corresponding eight-bit grayscale image, store each pixel value of the grayscale image as the corresponding element of the two-dimensional matrix, and mark the two sets of grayscale images as the first time phase and the second time phase respectively according to the order of observation time. S5.2. Take the first temporal grayscale image matrix generated in step S5.1 as the first channel input and the second temporal grayscale image matrix as the second channel input. Stack the two two-dimensional matrices into a three-dimensional tensor along the channel dimension. The first and second dimensions of the three-dimensional tensor correspond to the number of rows and columns of the image, respectively, and the third dimension corresponds to the number of channels and has a value of two. Use the stitched three-dimensional tensor as the input data of the convolutional neural network to retain temporal information and spatial distribution information.

[0058] In this embodiment: Step S5.1 selects two sets of differential refractive index data of the same target area within a short time interval from a three-dimensional array, and extracts two-dimensional slices of the same height layer for gray-scale quantization mapping. This realizes the conversion from continuous refractive index values ​​to discrete gray-scale images, improving the data visualization capability and computational efficiency. The refractive index values ​​are quantized and mapped according to gray levels from zero to 255 to generate an eight-bit gray-scale image, reducing data storage space and computational complexity while preserving the main characteristic information of the spatial distribution of refractive index. The two sets of gray-scale images are labeled as the first and second time phases according to the order of observation time, providing a basis for preserving the temporal evolution characteristics of atmospheric state in the subsequent dual-channel input construction, and improving the temporal expression capability of the input data.

[0059] Step S5.2 stacks the grayscale image matrices of the first and second time phases into a three-dimensional tensor along the channel dimension, achieving unified encoding of temporal and spatial distribution information, thus improving the information density and expressive power of the input data. The first and second dimensions of the three-dimensional tensor, corresponding to the number of rows and columns of the image, preserve the spatial distribution characteristics of the refractive index. The third dimension, corresponding to the number of channels and having a value of two, preserves the atmospheric state information at both time points, enabling the convolutional neural network to simultaneously perceive the spatial distribution and temporal evolution characteristics of atmospheric refractive index. Compared to traditional single-phase input methods, the dual-channel input structure can capture the dynamic changes of atmospheric turbulence, providing richer input information for the neural network to extract spatiotemporal coupling features and improving the accuracy of wind field inversion.

[0060] Step S5, through the combination of the two sub-steps mentioned above, achieves a complete conversion process from a standardized three-dimensional array to a dual-channel input tensor, bringing three significant improvements compared to traditional methods. First, the gray-scale standardization mapping converts continuous refractive index values ​​into discrete gray levels suitable for convolutional neural network processing. This reduces data dimensionality and computational complexity while preserving key feature information, solving the problems of low computational efficiency and large storage space caused by continuous numerical input in traditional methods, thus improving data processing efficiency and practicality. Second, the selection and temporal labeling of two sets of refractive index data within a short time interval establishes a clear expression of atmospheric state temporal evolution information. Compared to traditional single-phase input methods, this captures the dynamic changes in atmospheric turbulence, providing richer physical information for wind field inversion. Third, the dual-channel stacked structure uniformly encodes temporal and spatial distribution information into a three-dimensional tensor, establishing a standardized input format adapted to convolutional neural networks. This enables the network to simultaneously extract spatial and temporal features, improving the model's ability to perceive atmospheric dynamic processes and the accuracy of wind field inversion, laying a data foundation for subsequent cross-condition feature learning and high-precision inversion.

[0061] Example 7: Please refer to Figure 1 In step S6, the specific method is as follows; S6.1 Construct a dual-channel input layer to receive the three-dimensional tensor generated in step S5.2. Set the first-level convolutional layer to contain thirty-two 3x3 convolutional kernels to perform convolution operations on the input tensor. The output of the convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Set the second-level convolutional layer to contain sixty-four 3x3 convolutional kernels to perform convolution operations on the pooled feature map. The output of the second-level convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Construct a flattening layer to convert the pooled two-dimensional feature map into a one-dimensional feature vector. S6.2. Input the standardized description vector generated in step S3.2 into the feature modulation module. The feature modulation module maps the description vector into a modulation weight vector with the same number of channels as the output feature map of the first-level convolutional layer through a fully connected layer. The modulation weight vector is multiplied element-wise by channel with the output feature map of the first-level convolutional layer to construct a fully connected layer containing 128 neurons to receive the flattened feature vector. The output layer contains the same number of neurons as the wind speed and wind direction prediction dimension. The output layer directly outputs the three-dimensional wind field prediction value.

[0062] In this embodiment: Step S6.1 constructs a network structure comprising a dual-channel input layer, multiple levels of convolutional layers, pooling layers, and flattening layers, achieving multi-scale feature extraction from a dual-channel input tensor to a one-dimensional feature vector, thus improving the network's ability to perceive atmospheric phenomena at different spatial scales. The first-level convolutional layer contains thirty-two 3x3 convolutional kernels to extract local turbulence details, while the second-level convolutional layer contains sixty-four 3x3 convolutional kernels to extract larger-scale wind field trend features. This progressive feature extraction through two levels of convolutional layers establishes a multi-scale feature representation from local to global. The configuration of activation functions and pooling layers reduces feature dimensionality and computational complexity while ensuring feature extraction capabilities. The flattening layer provides a standardized input format for subsequent fully connected layer processing, improving the rationality of the network structure.

[0063] Step S6.2 involves inputting the standardized description vector into the feature modulation module and mapping it to a modulation weight vector through a fully connected layer. This weight vector is then multiplied element-wise by channel with the output feature map of the first-level convolutional layer. This achieves adaptive modulation of the feature extraction process by the operating conditions, improving the network's adaptability to different imaging conditions. The feature modulation module dynamically adjusts the response intensity of different channel features according to the operating conditions, enabling the network to learn atmospheric physical feature representations independent of the imaging conditions. Compared to traditional fixed-weight convolutional networks, this enhances the model's cross-condition generalization ability. The fully connected layer, containing 128 neurons, maps the extracted feature vector to a high-level semantic representation. The output layer directly outputs the 3D wind field prediction value, establishing an end-to-end mapping relationship from the refractive index image to the wind field parameters, thus improving the automation of the inversion process.

[0064] Step S6, through the combination of the two sub-steps mentioned above, realizes the construction of a dual-channel convolutional neural network model incorporating an adaptive feature modulation mechanism for operating conditions, bringing three significant improvements compared to traditional methods. First, the progressive structure of multi-level convolutional and pooling layers establishes a multi-scale feature extraction system, from local turbulence details to large-scale wind field trends, solving the problem of insufficient perception of atmospheric phenomena at different spatial scales by traditional single-scale feature extraction methods, and improving the network's ability to represent complex atmospheric turbulence structures. Second, the introduction of the adaptive feature modulation mechanism for operating conditions breaks through the limitations of traditional fixed-weight convolutional networks. Through dynamic weight adjustment driven by operating condition parameters, the network can adaptively adjust its feature response according to different imaging conditions, achieving operating condition-independent physical feature learning. Compared to traditional methods, this improves the model's cross-operating condition generalization ability and adaptability to unknown operating conditions. Third, the end-to-end network structure design integrates feature extraction, operating condition modulation, and wind field prediction into a unified deep learning framework, avoiding the error accumulation problem in traditional multi-step processing methods, improving the automation level and overall accuracy of the inversion process, and providing an efficient and reliable technical means for the operational application of atmospheric wind fields in distributed radar systems.

[0065] Example 8: Please refer to Figure 1 In step S7, the specific method is as follows; S7.1. The tomographic wind field data obtained in step S1.2 is used as a supervision label and paired with the dual-channel input tensor generated in step S5.2 according to the spatiotemporal correspondence to form training samples. Independent samples from the training samples with different data sources from the training set are selected to construct a validation sample set. The training samples are grouped and labeled according to the imaging condition type. Samples from different imaging conditions are randomly selected in each training batch and mixed. Each batch contains samples covering at least three different imaging condition types. The condition identifier and physical wind field identifier corresponding to each sample are recorded. S7.2 Input the mixed batch samples into the convolutional neural network model constructed in step S6, calculate the mean square error between the wind field prediction value output by the network and the supervision label as the prediction loss, extract the feature maps output by the first convolutional layer for samples with the same physical wind field label but different operating condition labels, calculate the cosine distance between these feature maps as the feature consistency loss, and weight the prediction loss and the feature consistency loss to form the total loss function. Calculate the gradient of the total loss function with respect to the network parameters through the backpropagation algorithm, and update the network parameters using an adaptive learning rate optimization algorithm.

[0066] In this embodiment: Step S7.1 uses tomographic wind field data as supervision labels paired with dual-channel input tensors to form training samples, and selects independent samples to construct a validation set, achieving high-quality sample organization required for supervised learning, thus improving the reliability of training data and the objectivity of evaluation. Samples are grouped and labeled according to imaging condition type, and samples from different imaging conditions are randomly selected and mixed in each training batch to ensure that the batch covers at least three imaging condition types. This allows the network to simultaneously encounter data distributions under multiple observation conditions during training, avoiding the model overfitting problem caused by traditional single-condition training. Recording condition identifiers and physical wind field identifiers provides accurate sample pairing criteria for subsequent feature consistency constraints, improving the targeting of the training strategy and its cross-condition learning capability.

[0067] Step S7.2 involves inputting mixed batch samples into the network model, calculating the mean squared error between the predicted value and the supervision label as the prediction loss, and extracting feature maps of samples with the same physical wind field but different operating conditions to calculate the cosine distance as the feature consistency loss. This achieves dual optimization of prediction accuracy and feature robustness, improving the comprehensiveness of the model's learning objectives. The two losses are weighted and summed to form the total loss function. The network parameters are updated through backpropagation and adaptive learning rate optimization algorithms. This forces the network to learn physical feature representations independent of operating conditions while pursuing prediction accuracy. Compared to traditional training strategies based solely on prediction loss, this enhances the model's adaptability to changes in imaging conditions and establishes a stable mapping relationship between differential refractive index and three-dimensional wind field.

[0068] Step S7, through the combination of the two sub-steps mentioned above, realizes a complete training process with cross-condition joint training and feature consistency constraints, bringing three important improvements compared to traditional methods. First, the mixed organization of multi-condition samples within a batch breaks through the limitations of traditional single-condition or simple mixed training, enabling the network to simultaneously learn the data distribution under different observation conditions in each parameter update, improving the model's generalization ability and adaptability to unknown conditions. Second, the introduction of feature consistency loss constraints ensures that the feature representations of samples from different conditions under the same physical wind field tend to be consistent, solving the problem that traditional deep learning methods are easily affected by imaging parameters and learn condition-related pseudo-features. This guides the network to focus on the essence of atmospheric physics, improving the stability and physical interpretability of the inversion results. Third, the synergistic design of the dual-objective loss function and adaptive optimization strategy enhances feature robustness while ensuring prediction accuracy during training. Compared to traditional single-loss training, it improves the accuracy and reliability of wind field inversion under the variable observation conditions of actual distributed radar systems, providing a more robust technical guarantee for operational applications.

[0069] Example 9: Please refer to Figure 1 In step S8, the specific method is as follows; S8.1 Input the dual-channel input tensor of the verification sample set constructed in step S7.1 into the convolutional neural network model trained in step S7.2. The network model outputs the corresponding three-dimensional wind field prediction value. Extract the reference wind field data paired with the input tensor from the verification sample set. Calculate the root mean square error and mean absolute error of the predicted wind field and the reference wind field in the wind speed dimension. Calculate the mean angle deviation and standard deviation of the angle deviation of the predicted wind field and the reference wind field in the wind direction dimension. Group and statistically analyze the error index under each working condition according to the imaging working condition type corresponding to the verification sample. S8.2 Count the number of imaging conditions in the verification sample set that do not appear in the training sample set, calculate the root mean square error of wind speed and the wind direction angle deviation of these unknown conditions, compare the error index of the unknown conditions with the error index of the known conditions, calculate the percentage increase of the unknown condition error relative to the known condition error, when the percentage increase is less than the preset threshold, the model is determined to pass the verification, record the model parameters that have passed the verification and save them as a model file, and transfer the model file to step S9 for wind field inversion.

[0070] In this embodiment: Step S8.1 obtains the three-dimensional wind field prediction value by feeding the dual-channel input tensor of the validation sample set into the trained network model, and compares it with the reference wind field data to calculate the root mean square error of wind speed, mean absolute error, and mean and standard deviation of wind direction angle deviation, thereby achieving a quantitative evaluation of the model's prediction accuracy and improving the objectivity and comprehensiveness of the validation process. The method of grouping and statistically analyzing error indicators under each imaging condition according to the imaging condition type ensures that the validation not only focuses on overall performance but also reveals the differences in model performance under different observation conditions, promptly identifying performance shortcomings under specific conditions, and providing targeted guidance for subsequent model optimization. This avoids the problem of traditional overall average evaluation masking differences under different conditions.

[0071] Step S8.2 involves statistically validating imaging conditions not present in the training set, comparing the error indices of these unknown condition samples with those of known condition samples, and calculating the percentage increase in error for unknown conditions. This allows for an independent assessment of the model's cross-condition generalization ability, improving the rigor and reliability of the validation criteria. When the percentage increase is less than a preset threshold, the model is deemed to have passed validation, and the model parameters are recorded and saved. This ensures that only models with good adaptability to different location conditions can be used in practical applications. Compared to traditional methods that only validate on known data, this reduces the risk of model failure under real, variable observation conditions, improving the reliability and practical value of model deployment.

[0072] Step S8, through the combination of the two sub-steps mentioned above, realizes a complete model validation process including condition grouping evaluation and generalization capability verification for unknown conditions, bringing three important improvements compared to traditional methods. First, grouping statistical error indices by imaging condition type overcomes the limitations of traditional overall average evaluation, clearly revealing performance differences of the model under different radar systems and observation geometries, providing precise basis for targeted optimization, and improving the granularity and guiding value of validation diagnosis. Second, the specific introduction of error comparison and amplification threshold determination for unknown condition samples addresses the shortcomings of traditional validation limited to training data, allowing for independent and rigorous assessment of the model's generalization capability, and enhancing the representativeness of validation for the variable observation conditions of actual distributed radar systems. Third, the validation standard combining quantified error indices and threshold determination makes the model's pass conditions more objective and quantifiable, avoiding the uncertainty brought about by traditional subjective evaluation, ultimately improving the stability and reliability of the model's wind field inversion under cross-condition conditions, providing a more reliable technical guarantee for subsequent operational applications.

[0073] Example 10: Please refer to Figure 1 In step S9, the specific method is as follows; S9.1. New radar echo data is obtained by synchronously observing the target area in a short time interval using multiple distributed synthetic aperture radar satellites. Differential interferometry processing and three-dimensional tomographic inversion are performed on the radar echo data to generate a three-dimensional differential refractive index tomographic image. Imaging parameters corresponding to the new observation data are extracted and normalized to generate a standardized description vector. Amplitude normalization and spatial statistical characteristic standardization transformation are performed on the three-dimensional differential refractive index data. Two sets of standardized data within a short time interval are selected and converted into grayscale images. The two grayscale images are stitched together along the channel dimension to generate a dual-channel input tensor. S9.2 Read the model file saved in step S8.2 and load the convolutional neural network model parameters into memory. Input the dual-channel input tensor and standardized description vector generated in step S9.1 into the dual-channel input layer and feature modulation module of the network model, respectively. The network model outputs a prediction vector containing wind speed values ​​and wind direction angles. Reorganize the prediction vector into a three-dimensional array according to the three dimensions of height, longitude and latitude. Each element of the three-dimensional array stores the wind speed and wind direction data of the corresponding spatial location. Write the three-dimensional array into a network general data format file.

[0074] In this embodiment: Step S9.1 generates a refractive index tomographic image by performing differential interferometry processing and three-dimensional tomographic inversion on the new radar echo data, and sequentially completes the normalization of imaging parameters, standardization of data amplitude and spatial statistical characteristics, and stitching of dual-channel grayscale images. This achieves a complete preprocessing flow from raw observation data to standardized dual-channel input tensors, improving the consistency between the distribution of new data and data from the training phase. Strictly replicating the normalization and input construction methods used during training effectively compensates for data deviations caused by differences in radar systems and observation geometry in actual observations. This allows the input tensor to retain true atmospheric physical characteristics while reducing operational interference, providing high-quality standardized input for subsequent model predictions and avoiding the prediction accuracy degradation problem caused by inconsistent distribution of old and new data in traditional methods.

[0075] Step S9.2 loads the validated model parameters, feeding the dual-channel input tensor and standardized description vector into the network model to complete forward inference. The predicted vectors are then reorganized into a three-dimensional wind field array organized by altitude, longitude, and latitude, and finally written into a common data format file. This completes the post-processing from model inference to structured wind field output, improving the visualization of the inversion results and their compatibility with subsequent applications. The end-to-end prediction and three-dimensional reconstruction method directly generates spatially gridded wind field data that meets meteorological operational needs, avoiding the error accumulation and format inconsistencies caused by multi-step manual conversion in traditional methods. This improves the automation level of wind field inversion and the standardization of result delivery.

[0076] Step S9, through the combination of the two sub-steps mentioned above, realizes a complete operational inversion process from actual distributed radar observation data to standardized 3D wind field results, bringing three important improvements compared to traditional methods. First, it rigorously reproduces the normalization, standardization, and dual-channel input construction process of the training phase, solving the problem that traditional methods are easily affected by differences in operating conditions when processing new observation data. This ensures the model maintains high predictive consistency under real and varied observation conditions, improving the robustness and reliability of the inversion results. Second, the automated design of end-to-end model inference and 3D result reconstruction avoids error propagation and manual intervention in traditional multi-module serial processing, improving the efficiency and repeatability of the entire inversion chain. Third, it outputs 3D gridded wind field data in a universal format, directly meeting the needs of downstream applications such as weather forecasting and numerical model assimilation. Compared to traditional methods that only provide 2D or single-point wind field results, this significantly enhances the practical value and operational application potential of the results, providing an efficient and robust technical path for the large-scale application of distributed synthetic aperture radar systems in atmospheric wind field monitoring.

[0077] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0078] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN, characterized in that: The specific method is as follows: S1. Data acquisition: The target area is synchronously observed by multiple distributed synthetic aperture radar satellites at short time intervals to acquire radar observation data under different imaging conditions. The imaging conditions are determined by the radar operating frequency band, incident angle, orbital geometric parameters and interference time interval. At the same time, the three-dimensional refractive index field and tomographic wind field data of the target area are acquired. S2, Delay Inversion: Differential interferometry is performed on radar observation data to calculate the atmospheric tropospheric slant range delay and map it to the interferometric phase. A three-dimensional differential tropospheric refractive index tomographic image of the target area is obtained through three-dimensional tomographic inversion. S3. Working condition modeling: Extract the imaging working condition parameters corresponding to each group of three-dimensional differential refractive index data, normalize the parameters and integrate them into a standardized description vector; S4. Data Normalization: Based on the imaging condition parameter description vector, the amplitude normalization and spatial statistical characteristic standardization of the three-dimensional differential refractive index data under different conditions are performed. S5. Dual-channel construction: Select two sets of standardized three-dimensional differential refractive index data within a short time interval, convert them into grayscale images, and stitch them together along the channel dimension in chronological order to generate a dual-channel input tensor. S6. Feature Extraction: Construct a convolutional neural network model containing a dual-channel input layer, multi-level convolutional layers, pooling layers, flattening layers, and fully connected layers. Introduce imaging condition parameter description vectors into the convolutional layers for feature modulation. The formula for convolution operation is: in, The input feature map; For the first k One convolutional kernel; b k This is for bias; the number of output channels is the same as the number of convolution kernels. After the convolution operation, the ReLU function is applied to non-linearly transform the output, helping the network learn more complex features and preventing gradient vanishing. The formula for the ReLU function is: S7. Joint Training: The tomographic wind field data is used as a supervision label and paired with the dual-channel input tensor to form a training sample set. A hybrid batch training strategy is adopted and feature consistency constraints are applied to samples of different working conditions with the same physical wind field. The network parameters are iteratively updated through backpropagation. S8. Model Validation: Input the validation sample set into the trained network model, calculate the error between the predicted wind field and the reference wind field, and evaluate the inversion accuracy of the model under unknown imaging conditions. S9, Wind Field Output: Acquire new distributed radar observation data, process it according to the above procedure, input it into the verified network model, and output the three-dimensional wind speed and wind direction of the target area at the corresponding observation time.

2. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 1, characterized in that, In step S1, the specific method is as follows: S1.1 Deploy multiple distributed synthetic aperture radar satellites to conduct short-time synchronous observations of the target area. Each satellite images the same area under different orbital positions, incident angles, and operating frequency bands, forming a radar echo dataset covering multiple combinations of observation geometry and system parameters. Record the operating frequency band, incident angle, orbital geometry parameters, and interferometric time interval corresponding to each set of data, and establish a working condition identification code to map the observation data with the working condition parameters. S1.2 Obtain three-dimensional atmospheric refractive index field data of the target area during the observation period through meteorological detection equipment measurement or atmospheric physical model simulation, obtain three-dimensional wind field data of the corresponding spatiotemporal range using wind measuring radar or meteorological reanalysis data, perform spatiotemporal registration of refractive index field data with radar observation data, and associate the three-dimensional wind field data as true value labels with radar observation data to form training samples containing input features and output labels.

3. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 2, characterized in that, In step S2, the specific method is as follows: S2.

1. The multi-temporal radar observation data obtained in step S1 are registered and interferometrically processed. The three-dimensional refractive index field data obtained in step S1.2 is combined with satellite orbital parameters to calculate the atmospheric tropospheric slant range delay from the satellite to the ground pixel under each observation angle. The calculated slant range delay is mapped to the phase component of the complex interferogram. Interference terms such as thermal noise, residual ionospheric phase error and orbital error are superimposed to generate a simulated differential interferometric phase map. The differential tropospheric delay observation value is extracted from the interferometric phase map under different observation geometry. S2.

2. Based on the multi-view observation geometry of the distributed satellite system, construct the three-dimensional spatial observation equation of the troposphere. Use the differential tropospheric delayed observation values ​​extracted in step S2.1 as the observation vector. Solve the observation equation using algebraic reconstruction or regularized iterative algorithm to reconstruct the three-dimensional differential tropospheric refractive index spatial distribution of the target area within a short time interval. Generate a differential refractive index tomographic image containing height layer information. Slice the tomographic image by height layer and convert it into a two-dimensional matrix format for storage.

4. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 3, characterized in that, In step S3, the specific method is as follows: S3.1 Extract the imaging condition parameters corresponding to each group of three-dimensional differential refractive index tomography images from the condition identification code established in step S1.1, including radar operating frequency band, incident angle, orbital altitude, baseline length and interferometric time interval. Classify and organize the extracted condition parameters according to radar system parameters and observation geometric parameters, establish an index mapping table between three-dimensional differential refractive index data and imaging condition parameters, and record the complete set of condition parameters corresponding to each group of tomography images to ensure the traceability of condition information in subsequent processing. S3.

2. The extracted imaging parameters are normalized according to their types, the radar operating frequency band is mapped to the zero-to-one interval, the incident angle is normalized according to the maximum observation angle, the orbital height and baseline length are standardized according to the system design range, the interferometric time interval is normalized according to the maximum time window, and the normalized parameters are spliced ​​into a fixed-dimensional standardized description vector in a preset order, which is used as the input parameters for subsequent feature modulation.

5. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 4, characterized in that, In step S4, the specific method is as follows; S4.1 Read the standardized description vector constructed in step S3.2, calculate the theoretical atmospheric refractive index dynamic range under the corresponding operating conditions based on the radar operating frequency band and incident angle parameters in the vector, perform linear transformation on the three-dimensional differential refractive index data generated in step S2.2 according to the calculated dynamic range, map the refractive index values ​​under different operating conditions to the zero-to-one interval, compensate for the systematic deviation of refractive index amplitude caused by the difference in radar system parameters, and store the transformed data in a unified numerical format. S4.2 Calculate the mean and standard deviation of each set of three-dimensional differential refractive index data in the spatial dimension. Determine the spatial correlation scale based on the orbital geometry parameters and interference time interval in the standardized description vector of step S3.

2. Perform a standardization transformation on the refractive index data by subtracting the mean and dividing the standard deviation. Interpolate and resample the transformed data according to a uniform spatial sampling interval to compensate for the spatial distribution characteristic deviation caused by the difference in observation geometry. Organize the resampled data into a three-dimensional array according to the height layer.

6. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 5, characterized in that, In step S5, the specific method is as follows; S5.1 Select two sets of differential refractive index data of the same target area within a short time interval from the three-dimensional array generated in step S4.2, extract two-dimensional slices of the same height layer from the two sets of data, quantize and map the refractive index values ​​in each two-dimensional slice according to the gray level from zero to 255, generate the corresponding eight-bit grayscale image, store each pixel value of the grayscale image as the corresponding element of the two-dimensional matrix, and mark the two sets of grayscale images as the first time phase and the second time phase respectively according to the order of observation time. S5.

2. Take the first temporal grayscale image matrix generated in step S5.1 as the first channel input and the second temporal grayscale image matrix as the second channel input. Stack the two two-dimensional matrices into a three-dimensional tensor along the channel dimension. The first and second dimensions of the three-dimensional tensor correspond to the number of rows and columns of the image, respectively, and the third dimension corresponds to the number of channels and has a value of two. Use the stitched three-dimensional tensor as the input data of the convolutional neural network to retain temporal information and spatial distribution information.

7. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 6, characterized in that, In step S6, the specific method is as follows; S6.1 Construct a dual-channel input layer to receive the three-dimensional tensor generated in step S5.

2. Set the first-level convolutional layer to contain thirty-two 3x3 convolutional kernels to perform convolution operations on the input tensor. The output of the convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Set the second-level convolutional layer to contain sixty-four 3x3 convolutional kernels to perform convolution operations on the pooled feature map. The output of the second-level convolutional layer is connected to a 2x2 max pooling layer after passing through a modified linear unit activation function. Construct a flattening layer to convert the pooled two-dimensional feature map into a one-dimensional feature vector. S6.

2. Input the standardized description vector generated in step S3.2 into the feature modulation module. The feature modulation module maps the description vector into a modulation weight vector with the same number of channels as the output feature map of the first-level convolutional layer through a fully connected layer. The modulation weight vector is multiplied element-wise by channel with the output feature map of the first-level convolutional layer to construct a fully connected layer containing 128 neurons to receive the flattened feature vector. The output layer contains the same number of neurons as the wind speed and wind direction prediction dimension. The output layer directly outputs the three-dimensional wind field prediction value.

8. The distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 7, characterized in that, In step S7, the specific method is as follows; S7.

1. The tomographic wind field data obtained in step S1.2 is used as a supervision label and paired with the dual-channel input tensor generated in step S5.2 according to the spatiotemporal correspondence to form training samples. Independent samples from the training samples with different data sources from the training set are selected to construct a validation sample set. The training samples are grouped and labeled according to the imaging condition type. Samples from different imaging conditions are randomly selected in each training batch and mixed. Each batch contains samples covering at least three different imaging condition types. The condition identifier and physical wind field identifier corresponding to each sample are recorded. S7.2 Input the mixed batch samples into the convolutional neural network model constructed in step S6, calculate the mean square error between the wind field prediction value output by the network and the supervision label as the prediction loss, extract the feature maps output by the first convolutional layer for samples with the same physical wind field label but different operating condition labels, calculate the cosine distance between these feature maps as the feature consistency loss, and weight the prediction loss and the feature consistency loss to form the total loss function. Calculate the gradient of the total loss function with respect to the network parameters through the backpropagation algorithm, and update the network parameters using an adaptive learning rate optimization algorithm.

9. A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 8, characterized in that, In step S8, the specific method is as follows; S8.1 Input the dual-channel input tensor of the verification sample set constructed in step S7.1 into the convolutional neural network model trained in step S7.

2. The network model outputs the corresponding three-dimensional wind field prediction value. Extract the reference wind field data paired with the input tensor from the verification sample set. Calculate the root mean square error and mean absolute error of the predicted wind field and the reference wind field in the wind speed dimension. Calculate the mean angle deviation and standard deviation of the angle deviation of the predicted wind field and the reference wind field in the wind direction dimension. Group and statistically analyze the error index under each working condition according to the imaging working condition type corresponding to the verification sample. S8.2 Count the number of imaging conditions in the verification sample set that do not appear in the training sample set, calculate the root mean square error of wind speed and the wind direction angle deviation of these unknown conditions, compare the error index of the unknown conditions with the error index of the known conditions, calculate the percentage increase of the unknown condition error relative to the known condition error, when the percentage increase is less than the preset threshold, the model is determined to pass the verification, record the model parameters that have passed the verification and save them as a model file, and transfer the model file to step S9 for wind field inversion.

10. A distributed D-InSAR three-dimensional atmospheric wind field inversion method based on dual-channel CNN according to claim 9, characterized in that, In step S9, the specific method is as follows; S9.

1. New radar echo data is obtained by synchronously observing the target area in a short time interval using multiple distributed synthetic aperture radar satellites. Differential interferometry processing and three-dimensional tomographic inversion are performed on the radar echo data to generate a three-dimensional differential refractive index tomographic image. Imaging parameters corresponding to the new observation data are extracted and normalized to generate a standardized description vector. Amplitude normalization and spatial statistical characteristic standardization transformation are performed on the three-dimensional differential refractive index data. Two sets of standardized data within a short time interval are selected and converted into grayscale images. The two grayscale images are stitched together along the channel dimension to generate a dual-channel input tensor. S9.2 Read the model file saved in step S8.2 and load the convolutional neural network model parameters into memory. Input the dual-channel input tensor and standardized description vector generated in step S9.1 into the dual-channel input layer and feature modulation module of the network model, respectively. The network model outputs a prediction vector containing wind speed values ​​and wind direction angles. Reorganize the prediction vector into a three-dimensional array according to the three dimensions of height, longitude and latitude. Each element of the three-dimensional array stores the wind speed and wind direction data of the corresponding spatial location. Write the three-dimensional array into a network general data format file.

Images