A 3-hourly kilometer-level sea surface wind speed scatterometer SAR fusion deep learning downscaling method

Abstract

This invention discloses a SAR fusion deep learning downscaling method for scatterometers at the kilometer level over a 3-hour period, comprising the following steps: taking multi-source low-resolution satellite fused wind speeds as input, learning the large-scale background structure through a residual U-Net and introducing high-resolution SAR spatial features to generate an initial high-resolution background field; performing spatial consistency correction on the background field with SAR observations as the primary constraint; using the low-resolution interpolated field and high-resolution SAR wind speeds as conditions, modeling the corrected residual bias using a residual diffusion probability model to generate high-frequency details; and introducing an effective region mask throughout the training and evaluation process to shield interference from land and missing measurement areas, and achieving controllable fusion of large-scale and detailed data through adjustable weights during the inference stage. This invention effectively reduces the learning complexity of the diffusion model and improves the physical consistency and detail fidelity of the downscaling results.

Description

Technical Field

[0001] This invention relates to the field of meteorological and oceanographic remote sensing data processing technology, specifically to a 3-hour kilometer-level sea surface anemometer SAR fusion deep learning downscaling method. Background Technology

[0002] Sea surface wind speed is a key element in air-sea interaction, ocean dynamic processes, marine engineering safety, and disaster monitoring. While scatterometers can provide large-scale, stable, and continuous sea surface wind speed information, their spatial resolution is typically on the order of 10–25 km, making it difficult to characterize small- to medium-scale variations such as nearshore frontal structures, island wake effects, and strong wind bands. Synthetic Aperture Radar (SAR) can provide kilometer-level or even higher resolution sea surface wind speeds, effectively characterizing detailed sea surface structures; however, its observations suffer from sparse samples (less than six times per month) and limited spatial coverage, making it difficult to independently construct continuous wind speed products. Low-resolution fusion products lack small-scale weather system characteristics, while SAR observations are significantly insufficient in terms of coverage and frequency. The limitations of these two data sources overlap, severely restricting the in-depth development and refined application of sea surface wind field data, becoming a key bottleneck limiting research on small- to medium-scale air-sea processes and improving marine safety capabilities.

[0003] Existing methods mainly include interpolation, statistical correction, and deep learning downscaling methods based on U-Net, GAN, or diffusion models. However, existing technologies have at least the following shortcomings: First, ordinary interpolation methods only achieve spatial grid resampling and cannot recover true high-resolution wind speed details; second, single-stage CNNs or U-Nets are prone to outputting mean-based results driven by training loss, leading to smoothed wind speed gradients, local strong winds, and fine-scale boundary structures; third, when directly using DDPM to model complete high-resolution wind speeds, the model needs to learn both large-scale background and small-scale random structures simultaneously, resulting in high training difficulty and computational cost. In particular, wind speeds inferred from random noise may be distorted, thus reducing the fidelity of large-scale fields; fourth, satellite wind speed samples contain land areas, coastal boundaries, and invalid observation areas. If training loss and statistical normalization are calculated across the entire region, invalid areas will interfere with model parameter updates, reducing the quality of effective sea area reconstruction. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a deep learning downscaling method for SAR fusion of kilometer-level sea surface wind speed scatterometers at a 3-hour interval. This method constructs a three-stage modeling framework of "background field + SAR correction + bias generation" and introduces effective region masks in data construction, training loss, residual target, evaluation index and inference correction to solve the problems existing in the background technology.

[0005] Technical Solution: The present invention provides a deep learning-based downscaling method for 3-hour kilometer-level sea surface wind speed scatterometer SAR fusion, comprising the following steps:

[0006] Step 1: Acquire multi-source low-resolution satellite wind speed data and high-resolution SAR satellite wind speed data, perform spatiotemporal matching and systematic error correction with buoy observation data respectively, unify the time and spatial resolution of the multi-source low-resolution satellite wind speed data after correction to eliminate the systematic bias of each satellite, and fuse them into low-resolution fused wind speed. Obtain the low-resolution interpolation field through spatial resampling, and process the SAR satellite wind speed into a high-resolution target field.

[0007] Step 2: Construct an effective region mask to extract regions that have effective observations in both the low-resolution interpolation field and the high-resolution target field, forming model training samples;

[0008] Step 3: Using the low-resolution interpolated field as input, learn the large-scale background structure through the residual U-Net network, output the bounded bias term and add it to the input to generate the initial high-resolution background field;

[0009] Step 4: Using SAR observations in the high-resolution target field as the dominant constraint, perform spatial consistency correction on the initial high-resolution background field to obtain the SAR-corrected background field;

[0010] Step 5: Construct a conditional residual diffusion probability model. Using the channel stitching of the low-resolution interpolation field and the high-resolution target field as conditional input, perform diffusion modeling on the residual deviation between the SAR correction background field and the high-resolution target field, learn the distribution of the residual deviation, and generate the residual deviation.

[0011] Step 6: Weighted fusion of the deviations between the SAR corrected background field and the small-scale weather system characteristics generated by the conditional residual diffusion probability model to obtain the final high-resolution wind speed.

[0012] Furthermore, in step 1, the wind speed observations from multiple satellites, after quality control and systematic error correction, are divided into consecutive 3-hour time windows according to the observation time. The observations are then mapped to a unified grid with a horizontal resolution of 12.5km × 12.5km using the nearest grid merging method. The effective wind speeds falling into the same grid within each time window are then weighted and arithmetically averaged to obtain the low-resolution fused wind speeds from multiple satellites. The low-resolution satellite wind speed data selected, which have been verified to have high data quality through actual measurements, include: the ASCAT-A / B / C scatterometers carried by the MetOp-A / B / C satellites, the scatterometers carried by the HY-2B / C / D satellites, the WindRAD scatterometer carried by the FY-3E satellite, the SSMIS radiometer carried by the DMSP F17 / F18 satellites, and the AMSR2 radiometer carried by the GCOM-W satellite.

[0013] Furthermore, in step 3, the residual U-Net network adopts an encoder-decoder structure and introduces a skip connection mechanism. The encoder gradually reduces the spatial resolution and increases the feature channels through downsampling; the decoder gradually restores the spatial resolution through upsampling, and the skip connection is used to preserve multi-scale spatial information.

[0014] Furthermore, in step 3, a bounded function is set at the output of the residual U-Net network to constrain the magnitude of the deviation term, so that the deviation term is kept within a physically reasonable range.

[0015] Furthermore, in step 4, the spatial consistency correction adopts the least squares regression method, which uses SAR observations as a reference to perform linear correction on the initial high-resolution background field, so that the corrected background field is optimally matched with the SAR observations in terms of spatial structure.

[0016] Furthermore, in step 5, the conditional residual diffusion probability model is guided by a multi-condition tensor formed by splicing the low-resolution interpolation field and the high-resolution target field by channel, and the residual bias is used as the diffusion object. This bias contains a large amount of small-scale weather system wind field characteristic information. The bias distribution under given conditions is learned through the forward noise addition and reverse noise reduction process, so that the generated residual bias is guided by both the large-scale background constraint and SAR observation information.

[0017] Furthermore, in step 5, the effective region mask is used throughout the model training and loss calculation process to shield against interference from land and missing measurement areas.

[0018] Furthermore, in step 6, weighted fusion is achieved through adjustable weighting coefficients. These coefficients are used to adjust the contribution ratio between the SAR corrected background field and the small-scale deviation generated by the conditional residual diffusion probability model, thereby achieving a controllable balance between large-scale wind speed and small-scale details.

[0019] The present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement any of the methods described herein.

[0020] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods described herein.

[0021] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0022] 1. By decomposing the downscaling task into three stages, namely “background field modeling, SAR correction, and bias diffusion generation”, the large-scale deterministic structure and the small-scale random structure are explicitly separated, thereby reducing the model learning complexity and improving the accuracy of the downscaling results.

[0023] 2. By using the SAR-led spatial consistency correction mechanism, the background field is transformed from a statistically significant reconstruction result to a highly consistent result with physical observation constraints, thereby improving the accuracy of the reconstruction results.

[0024] 3. By using an effective region masking mechanism, interference from land, coastlines, and missing measurement areas on the model training process is avoided, thereby improving the accuracy of effective sea area reconstruction;

[0025] 4. By introducing the deviation weighting coefficient γ, the controllable adjustment between large-scale wind speed and small-scale details is achieved, thereby generating a sea surface wind speed product that combines large-scale consistency with high-resolution detail expression. Attached Figure Description

[0026] Figure 1 The on-orbit time of each satellite in this invention;

[0027] Figure 2 This is a schematic diagram of the present invention;

[0028] Figure 3 This is the Residual UNet network structure of the present invention;

[0029] Figure 4 This is the Residual DDPM network structure of the present invention;

[0030] Figure 5 The figures are the 12.5km multi-source satellite fusion wind speed, the 12.5km downscaled to 1km satellite fusion wind speed, and the SAR wind speed results (1km resolution), respectively. The color scale unit is m / s. Detailed Implementation

[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0032] like Figure 2 As shown, this embodiment of the invention provides a deep learning-based downscaling method for fusion of SAR (Shockwave scattering) at the kilometer level over a 3-hour period, comprising the following steps:

[0033] Step 1: Data processing of multi-source satellite, SAR observations, and buoy corrections. For example... Figure 1As shown, based on the observation range and on-orbit time of multiple satellites, the time resolution of the unified multi-source satellite fusion data is set to 3 hours per interval, and the horizontal resolution is 12.5km × 12.5km. The multi-source low-resolution satellite wind speed data selected, which have been verified to have high data quality through actual measurements, include: the ASCAT-A / B / C scatterometers carried by MetOp-A / B / C satellites, the scatterometers carried by HY-2B / C / D satellites, the WindRAD scatterometer carried by FY-3E satellite, the SSMIS radiometer carried by DMSP F17 / F18 satellites, and the AMSR2 radiometer carried by GCOM-W satellite. The specific process is as follows:

[0034] First, raw sea surface wind speed data were acquired from 10 low-resolution satellites, including the ASCAT-A / B / C scatterometer aboard the MetOp-A / B / C satellites, the scatterometer aboard the HY-2B / C / D satellites, the WindRAD scatterometer aboard the FY-3E satellite, the SSMIS radiometer aboard the DMSP F17 / F18 satellites, and the AMSR2 radiometer aboard the GCOM-W satellite. This data was denoted as [missing data]. For the raw data, quality control and outlier removal are performed first; simultaneously, buoy-observed wind speed data are acquired, and the buoy observations are spatiotemporally matched with the data from each satellite, using the least squares method for linear regression. For the s-th satellite participating in the fusion, a univariate linear regression model is constructed:

[0035] ;

[0036] Where y is the measured wind speed of the buoy that matches the observation of the s-th satellite at the same spatiotemporal location; Let be the slope of the linear regression model for the s-th satellite; This represents the original wind speed observation value from the s-th satellite; Let be the intercept of the linear regression model for the s-th satellite; The random error term is represented by s, where s is the satellite number participating in the fusion. This is based on the spatiotemporal matching of buoy observations and satellite data. Using a set of valid samples, the parameters are estimated using the least squares method. and That is, minimizing the sum of squared residuals:

[0037] ;

[0038] in, Let be the regression parameters of the s-th satellite to be estimated; arg min represents the parameter combination that minimizes the objective function; k and b are candidate slopes and intercepts in the optimization process; denoted as the total number of valid samples after the s-th satellite and buoy complete spatiotemporal matching; i is the sample number. Let be the buoy wind speed of the i-th matched sample; Let be the original wind speed of the s-th satellite at the i-th matching sample; This represents the summation of the squared residuals over all valid samples.

[0039] By setting the partial derivatives of the objective function with respect to k and b to zero, we can obtain the solution. and The original wind speeds of the satellites were systematically corrected using the k and b parameters of each satellite. The corrected wind speeds are as follows:

[0040] ;

[0041] in, The wind speed of the s-th satellite after systematic error correction; The original satellite wind speed before correction; and , respectively, are the slope and intercept estimated from the matching sample of the s-th satellite and the buoy; s is the satellite number.

[0042] After completing quality control and systematic error correction for the wind speed data from each satellite, each day is divided into eight consecutive 3-hour time windows. The m-th fusion time window is represented as:

[0043] ;

[0044] in, This represents the m-th 3-hour fusion time window; m is the time window index. When m=0, it corresponds to 00:00 to 03:00; when m=7, it corresponds to 21:00 to 24:00. Only satellite wind speed data whose observation time falls within the current time window are used for fusion during that period.

[0045] The effective observations from each satellite are mapped to a regular latitude and longitude grid with a spatial resolution of 12.5 km. The grid index corresponding to the nth effective observation from the s-th satellite data source is:

[0046] ;

[0047] ;

[0048] in, and These are the latitudinal and meridional grid indices corresponding to the observation, respectively. and These are the latitude and longitude of the nth observation point on the s-th satellite, respectively; , ; ; This indicates rounding to the nearest integer. The target grid range is... , .

[0049] Define the valid observation indicator function for date d and the m-th 3-hour time window:

[0050] ;

[0051] in, This is the valid observation indicator function; d represents the date; m represents the 3-hour time window index; s represents the satellite data source number; n represents the satellite observation point number; i and j are the latitudinal and longitudinal indices of the spatial grid, respectively.

[0052] Within the m-th 3-hour time window of date d, the cumulative wind speed of all valid satellite observations at grid point (i,j) is:

[0053] ;

[0054] The corresponding total number of valid observations is:

[0055] ;

[0056] in, This is the sum of all valid wind speeds falling on grid point (i,j) within the current 3-hour time window; The total number of valid observations within the grid; S is the total number of satellite or sensor data sources participating in the fusion. This represents the number of valid observation points for the s-th satellite within the m-th time window on date d; Sea surface wind speed after quality control and systematic error correction.

[0057] The 3-hourly merged wind speed for grid point (i,j) is calculated as follows:

[0058] ;

[0059] in, For the m-th 3-hour time window of date d, the multi-source satellite fused sea surface wind speed at grid point (i,j); when there are valid observations for this grid, the arithmetic mean of all valid observations is taken; when there are no valid observations for this grid, it is set as a missing value.

[0060] This multi-source satellite fusion method adopts an equal-weighted average at the observation point level, with each effective observation point having the same weight in the fusion calculation. From the perspective of satellite data sources, the equivalent weight of each satellite is adaptively determined by the number of effective observations within the corresponding 3-hour time period and spatial grid. The more effective observations there are, the greater their contribution to the fusion result, thus making full use of the complementary information of different satellites in terms of observation time, orbital position, and spatial coverage.

[0061] After obtaining the satellite-fused wind speed at 12.5 km every 3 hours, bilinear interpolation was used to resample the low-resolution wind speed to the target high-resolution grid, resulting in the low-resolution interpolated field. This field serves as the input background field for subsequent models.

[0062] Simultaneously, high-resolution SAR satellite raw wind speed data matching the study area were acquired. Similar to the multi-satellite fusion method, the raw SAR satellite wind speed data was spatiotemporally gridded at 1km intervals over 3 hours. Then, buoy data and the same method were used for systematic error correction. Finally, the corrected 1km-3-hour high-resolution SAR satellite wind speed was obtained, denoted as the high-resolution SAR target field. It and They have the same spatial range and grid size.

[0063] Step 2: Constructing effective region masks and pairing samples.

[0064] Since satellite wind speed data may contain land areas, coastal boundaries, invalid observations, or missing data, this invention constructs an effective region mask M to limit the effective range for model training, normalized statistics, and residual calculation. For each grid point (i,j), if both the low-resolution interpolation field and the high-resolution target field are valid at that location, the mask value is 1; otherwise, it is 0.

[0065] ;

[0066] in, represents the effective region mask value at grid point (i,j); i and j are the row and column indices of the spatial grid, respectively. This is the interpolated value of the low-resolution wind speed at this grid point; This represents the value of the high-resolution SAR target field at this grid point; a value of 1 indicates that both wind speed fields are valid at this location, while a value of 0 indicates that the location is land, a missing area, or at least one invalid input.

[0067] This forms the training sample triples:

[0068] ;

[0069] in, For low-resolution interpolation fields; M represents the high-resolution SAR target field; M is the effective region mask with the same spatial size as both; the three together constitute a model training sample.

[0070] To avoid invalid regions affecting sample statistics, this invention calculates the normalized mean and standard deviation within the valid regions:

[0071] ;

[0072] in, and These are the mask mean and mask standard deviation of variable X within the effective region, respectively. Let X be the value of the wind speed field to be normalized at grid point (i, j), where X is all the valid data of the low-resolution interpolation field and the target field in the training samples; For effective area masking; Σ i,j This indicates summing over all spatial grids.

[0073] The normalized form is:

[0074] ;

[0075] in, The normalized wind speed value; The wind speed value before normalization; and It is divided into the mean and standard deviation within the effective region; ε is a small positive constant set to prevent the denominator from being zero; i and j are spatial grid indices.

[0076] Step 3: Residual UNet background field reconstruction.

[0077] like Figure 3 As shown, the Residual UNet network uses low-resolution interpolated fields. The input is used to extract multi-scale spatial features through an encoder, bottleneck layer, decoder, and skip connection structure. The overall network is a four-layer encoder-decoder symmetric structure: the encoder consists of four downsampling modules stacked sequentially, each containing two consecutive convolutional layers (3×3 kernel size, padding 1), and finally a 2×2 max pooling layer with a stride of 2 to halve the spatial resolution, while the number of feature channels increases progressively: input channel 1→32→64→128→256. The bottleneck layer contains a 3×3 convolutional block with 512 feature channels. The decoder upsamples through transposed convolutions (2×2 kernel size, stride 2) to gradually restore the spatial resolution, and concatenates the channel-dimensional features with the corresponding encoder layer through skip connections to preserve multi-scale spatial structure information. The number of feature channels in each layer of the decoder is: 512→256→128→64→32. The network output layer uses a 1×1 convolution to map the feature channels to single-channel bias terms. Except for the output layer, all intermediate convolutional layers of the network employ group normalization and the SiLU activation function.

[0078] The Residual UNet network does not directly output the complete high-resolution wind speed; instead, it outputs a bounded bias with amplitude constraints. First, the low-resolution interpolated field is standardized and input into the Residual U-Net network to obtain the unconstrained raw bias response. Then, a hyperbolic tangent function is used to constrain the raw bias response grid-by-grid to obtain the bounded bias in the physical wind speed space. Let the raw output of the Residual UNet be... The deviation is constrained using the tanh function:

[0079] ;

[0080] in, The bounded bias field output by the Residual UNet; To control the scaling factor of the upper limit of the deviation amplitude, it is set to 3.0 in actual calculations; tanh is the hyperbolic tangent function performed grid-by-grid; The parameter is The ResidualUNet nonlinear mapping is the network forward computation process consisting of an encoder, bottleneck layer, decoder, skip connections, and output convolutional layer. For all trainable parameters of Residual UNet, This is a low-resolution interpolation field. Since the hyperbolic tangent function ranges from -1 to 1, there is a bounded bias. The physical amplitude is limited to the range of (-3.0, 3.0).

[0081] The background field is obtained by adding the low-resolution interpolated field to the bounded bias:

[0082] ;

[0083] in, The initial high-resolution background field generated for Residual UNet; For low-resolution interpolation fields; This is the bounded bias field output by the Residual UNet.

[0084] This stage is used to learn stable large-scale background structures and introduce SAR details. Its training loss is calculated within the effective region mask M, using the mask mean absolute error:

[0085] ;

[0086] in, The mask mean absolute error loss of the Residual UNet; For effective area masking; The predicted wind speed at grid point (i, j) is the initial background field. Σ represents the SAR high-resolution target wind speed at the corresponding location; i,j This indicates summation over all spatial grids; i and j are spatial grid indices.

[0087] Step 4: SAR correction module.

[0088] Get the background field Then, the k and b parameters are solved according to the linear correction formula:

[0089] ;

[0090] Where arg min represents the parameter combination that minimizes the objective function; k and b are the linear correction slope and intercept that minimize the sum of squared errors; For SAR high-resolution target fields; This is the initial high-resolution background field.

[0091] Then, the background field is corrected according to the formula:

[0092] ;

[0093] in, The background field is linearly corrected by SAR observations; k and b are the correction slope and intercept obtained by least squares regression. This is the initial high-resolution background field.

[0094] This correction module uses SAR as a physical constraint to correct the initial background field, aligning it in terms of spatial structure primarily based on SAR observations. However, a post-correction bias still exists after correction. :

[0095] ;

[0096] in, For SAR high-resolution target fields; This is the background field after linear correction by SAR observation.

[0097] The corrected bias will be learned and generated using a subsequent Residual DDPM model.

[0098] Step 5: Conditional Residual DDPM Bias Generation Model.

[0099] like Figure 4As shown, the conditional Residual DDPM network does not directly generate complete high-resolution wind speeds, but rather models the diffusion of biases that still exist after SAR correction. Its conditional input is formed by concatenating a low-resolution interpolated field with the actual high-resolution SAR wind speeds channel by channel.

[0100] ;

[0101] Where C is the condition tensor of the conditional Residual DDPM; square brackets indicate concatenation along the channel dimension; For low-resolution interpolation fields; This represents a high-resolution SAR target field.

[0102] The input to the diffusion network is determined by the corrected bias. Together with the conditional input C, the concatenated network input is: ;

[0103] in, The input conditional Residual DDPM is the corrected bias sample; square brackets indicate stitching along the channel dimension; C is the conditional tensor consisting of the low-resolution interpolation field and the SAR high-resolution target field.

[0104] The encoding layer sequentially extracts features of 48×256×256, 96×128×128, 192×64×64, and 192×32×32, with a bottleneck layer dimension of 192×32×32. The decoding layer recovers the spatial scale step-by-step through upsampling and skip connections, and finally predicts a noise field of the same size as the input residual through the output layer. The network's direct output is the noise prediction value. ;

[0105] in, For a noise prediction network with parameter θ, under input correction bias conditions , θ is the predicted noise output under condition C and time step t; θ is the set of trainable parameters of the network; t is the current diffusion time step; ε is the standard Gaussian noise actually added during the forward diffusion process.

[0106] During the training phase, the corrected deviation was analyzed. Forward diffusion noise addition is performed at time step t:

[0107] ;

[0108] ;

[0109] in, Indicates the deviation after correction given no noise. Generate the noisy bias state at time step t under the condition of The conditional probability distribution; The cumulative signal retention coefficient; The remaining bias after correction is obtained from the difference between the high-resolution target wind speed field and the background field after SAR spatial consistency correction. To limit the impact of anomalous bias on the diffusion model training, in actual calculations... The amplitude in the physical quantity space is truncated to [-2.0, 2.0] m / s. for The deviation state after adding noise at the t-th diffusion time step; To and Random noise of the same dimension follows a multivariate standard Gaussian distribution with a mean of 0 and a covariance matrix of identity matrix I. ; I is the identity matrix.

[0110] In the actual calculation, the total diffusion time steps were set to T=200. Noise variance scheduling coefficient. Using a linear scheduling method, the value is 0.0001 at the first diffusion time step and 0.02 at the 200th diffusion time step, with an increment of 0.0001 between adjacent time steps. Therefore:

[0111] ;

[0112] in, The variance scheduling coefficient for adding noise at the t-th diffusion time step, with a value ranging from 0.0001 to 0.02; t is the diffusion time step index.

[0113] The single-step signal retention coefficient at time step t Defined as:

[0114] ;

[0115] Cumulative signal retention coefficient Defined as:

[0116] ;

[0117] in, This represents the cumulative product of the retained coefficients of all single-step signals from time step 1 to time step t; k is the cumulative product index; ∏ represents the chain multiplication operation. During model training, for each training sample, a time step t is randomly and uniformly selected from 200 diffusion time steps, and the corresponding noise bias state is generated according to the above parameters. .

[0118] The training objective of conditional residual DDPM is to predict the added noise. The loss function is the mean square error of mask noise prediction:

[0119] ;

[0120] in, The mean square error loss for mask noise prediction in conditional Residual DDPM; For effective area masking; This represents the noise predicted by the network at grid point (i, j); This represents the actual noise added at this grid point during forward diffusion; Σ i,j This indicates summation over all spatial grids; i and j are spatial grid indices.

[0121] During model training, for each sample, a time step t is randomly and uniformly selected from T=200 diffusion time steps:

[0122] ;

[0123] The time step t is sequentially processed through a 192-dimensional sinusoidal position code and a multilayer perceptron mapping to obtain the time step embedding vector:

[0124] ;

[0125] in, This is the 192-dimensional embedding vector corresponding to time step t; Represents the sine-cosine position code; This represents a multilayer perceptron consisting of two fully connected layers and a SiLU activation function.

[0126] In each temporal conditional convolutional module of the conditional Residual DDPM network, the temporal step embedding vector is injected into the corresponding feature layer after linear mapping:

[0127] ;

[0128] in, This is the temporal condition vector of the l-th temporal conditional convolutional module; and These represent the weights and biases of the temporal mapping layer of this module. The temporal conditional vector is added channel-by-channel to the group-normalized feature map of the first convolutional layer, enabling the network to perceive the current diffusion stage and perform corresponding noise prediction.

[0129] Through the training method described above, the model learns the corrected bias noise distribution under given low-resolution interpolation fields and high-resolution real wind speed conditions of SAR, thereby generating small-scale residual biases with random details that still exist after SAR correction during the inference stage. .

[0130] Step 6: The background field and deviation are fused to output high-resolution wind speed results.

[0131] The SAR-corrected background field is fused with the residual bias obtained from conditional Residual DDPM inference to obtain the final high-resolution wind speed:

[0132] ;

[0133] in, This is the final output of a high-resolution sea surface wind speed field; The background field after SAR observation correction; The residual bias generated by conditional Residual DDPM; This is the bias fusion weighting coefficient, with a value range of 0 ≤ ≤1, when When, the output only retains the SAR-corrected background field; when At that time, small-scale bias details generated by DDPM were introduced, and The larger the value, the greater the contribution of small-scale details; this can be adjusted... It can strike a balance between large-scale wind speed and small-scale details.

[0134] In actual reasoning and calculation, take:

[0135] ;

[0136] The residual bias generated by conditional Residual DDPM is superimposed on the SAR-corrected background field, and the final output is:

[0137] ;

[0138] The above settings enhance small-scale wind speed gradients, local extrema, and boundary details while maintaining the stability of large-scale wind speed structures.

[0139] Through the above steps, this invention achieves downscaling reconstruction guided by SAR satellite wind speed, from multi-source low-resolution satellite wind speeds to high-resolution wind speeds at 1km intervals over 3 hours. The Residual UNet background field reconstruction ensures stable and reliable large-scale background while incorporating high-resolution SAR details. The SAR correction module guarantees the spatial physical consistency between the reconstructed background field and the SAR data. Residual DDPM generates small-scale random biases under the constraints of high-resolution SAR wind speeds and low-resolution interpolated fields. An effective region mask is used throughout the entire process of data processing, model training, bias construction, and result output, thereby improving the accuracy and controllability of high-resolution wind speed reconstruction in complex nearshore areas. This invention effectively addresses the problems of missing small-scale weather system features in existing low-resolution fused data and insufficient SAR observation range and frequency.

[0140] like Figure 5 The diagram shows a comparison of the wind speed obtained from multi-source satellite fusion at 12.5 km, the 1 km wind speed obtained by downscaling using the method of this invention, and the SAR wind speed corrected from the original measured wind speed from the buoy. The left side shows the 12.5 km fused wind speed, which has coarse spatial resolution and makes it difficult to distinguish small- and medium-scale features. The right side shows the original SAR wind speed corrected from the measured wind speed from the buoy, which is rich in detail but has limited temporal and spatial coverage. The middle side shows the 1 km downscaled wind speed output by this invention, which is superior to the extremely limited sporadic coverage of SAR in terms of temporal and spatial continuity and observation area. It achieves a temporal resolution of 3 hours per hour, a significant improvement over the less than 6 observations per month for SAR. It is significantly superior to the 12.5 km fused wind speed in terms of detail, and can more clearly present more features of small-scale weather systems (such as wind speed gradients and local extreme value areas), balancing the integrity of the large-scale background field with the fidelity of small-scale details, effectively compensating for the shortcomings of a single data source.

Claims

1. A deep learning-based downscaling method for fusion of SAR data from sea surface anemometers at the kilometer level, characterized in that, Includes the following steps: Step 1: Acquire multi-source low-resolution satellite wind speed data and high-resolution SAR satellite wind speed data, perform spatiotemporal matching and systematic error correction with buoy observation data respectively, unify the time and spatial resolution of the corrected multi-source low-resolution satellite wind speed data, fuse them into low-resolution fused wind speed data, obtain a low-resolution interpolated field through spatial resampling, and process the SAR satellite wind speed data into a high-resolution target field. Step 2: Construct an effective region mask to extract regions that have effective observations in both the low-resolution interpolation field and the high-resolution target field, forming model training samples; Step 3: Using the low-resolution interpolated field as input, learn the large-scale background structure through the residual U-Net network, output the bounded bias term and add it to the input to generate the initial high-resolution background field; Step 4: Using SAR observations in the high-resolution target field as the dominant constraint, perform spatial consistency correction on the initial high-resolution background field to obtain the SAR-corrected background field; Step 5: Construct a conditional residual diffusion probability model. Using the channel stitching of the low-resolution interpolation field and the high-resolution target field as conditional input, perform diffusion modeling on the residual deviation between the SAR correction background field and the high-resolution target field, learn the distribution of the residual deviation, and generate the residual deviation. Step 6: Weighted fusion of the deviations between the SAR corrected background field and the small-scale weather system characteristics generated by the conditional residual diffusion probability model to obtain the final high-resolution wind speed.

2. The method for downscaling SAR fusion at the kilometer-level sea surface anemometer according to claim 1, characterized in that, In step 1, the wind speed observations from multiple satellites, after quality control and systematic error correction, are divided into consecutive 3-hour time windows according to the observation time. The observations are then mapped to a unified grid with a horizontal resolution of 12.5km×12.5km using the nearest grid merging method. The effective wind speeds falling into the same grid within each time window are then averaged using equal weights to obtain the low-resolution fused wind speeds from the multi-source satellites. Low-resolution satellite wind speed data were selected from the following: ASCAT-A / B / C scatterometers carried by MetOp-A / B / C satellites, scatterometers carried by HY-2B / C / D satellites, WindRAD scatterometers carried by FY-3E satellites, SSMIS radiometers carried by DMSP F17 / F18 satellites, and AMSR2 radiometers carried by GCOM-W satellites.

3. The method for downscaling SAR fusion at the kilometer-level sea surface anemometer according to claim 1, characterized in that, In step 3, the residual U-Net network adopts an encoder-decoder structure and introduces a skip connection mechanism. The encoder gradually reduces the spatial resolution and increases the feature channels through downsampling. The decoder gradually restores spatial resolution through upsampling, and skip connections are used to preserve multi-scale spatial information.

4. The deep learning downscaling method for fusion of SAR (Self-Range Anemometer) at the kilometer level with a 3-hour interval as described in claim 1, characterized in that... In step 3, a bounded function is set at the output of the residual U-Net network to constrain the magnitude of the deviation term.

5. The method for downscaling SAR fusion at the kilometer-level sea surface anemometer according to claim 1, characterized in that, In step 4, the spatial consistency correction uses the least squares regression method to perform linear correction on the initial high-resolution background field with SAR observations as a reference, so that the corrected background field is optimally matched with the SAR observations in terms of spatial structure.

6. The method for downscaling SAR fusion at the kilometer-level sea surface anemometer according to claim 1, characterized in that, In step 5, the conditional residual diffusion probability model is guided by a multi-conditional tensor formed by splicing the low-resolution interpolation field and the high-resolution target field by channel, and the residual bias is used as the diffusion object. The deviation contains a large amount of wind field characteristic information of small-scale weather systems. By learning the deviation distribution under given conditions through the forward noise addition and reverse noise reduction process, the generated residual deviation is guided by both the large-scale background constraint and SAR observation information.

7. The deep learning downscaling method for fusion of SAR (Self-Range Anemometer) at the kilometer level with a 3-hour interval as described in claim 1, characterized in that... In step 5, the effective region mask is used throughout the model training and loss calculation process to shield the interference from land and missing measurement areas.

8. The method for downscaling SAR fusion at the kilometer level for sea surface anemometers according to claim 1, characterized in that, In step 6, weighted fusion is achieved through adjustable weighting coefficients, which are used to adjust the contribution ratio between the SAR corrected background field and the small-scale bias generated by the conditional residual diffusion probability model.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method of any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-8.

Images