A multi-band radar data spatial downscaling method based on data fusion

Abstract

The present invention provides a multi-band radar data spatial downscaling method based on data fusion, including: Step 1, obtaining S-band and X-band observation data from weather radars; Step 2, performing data preprocessing on the obtained observation data; Step 3, constructing a spatial downscaling model; Step 4, training, validating, and testing the spatial downscaling model with the observation data after data preprocessing; Step 5, performing spatial downscaling on the S-band observation data according to the trained spatial downscaling model; Step 6, determining whether the generated data is close to the true value. The present invention can utilize a deep learning neural network to enable the model to extract and fuse the features of the S-band observation data and the X-band observation data, convert the originally low-resolution S-band data into high-resolution data, not only meeting the requirements for high-resolution radar data in practical applications but also reducing the dependence on hardware facilities.

Description

Technical Field

[0001] The present invention relates to the fields of weather radar and spatial downscaling, and particularly to a method for spatial downscaling of multi-band radar data based on data fusion. Background Art

[0002] Weather radar is a key tool for monitoring and analyzing atmospheric phenomena in meteorology. With the development of technology, spatial downscaling technology based on deep learning is gradually being applied to the reconstruction process of weather radar data. Currently, many research teams have successively proposed innovative methods for spatial downscaling based on super-resolution reconstruction. However, all these existing methods have an inherent defect, that is, the low-resolution data required by them are all obtained from the corresponding high-resolution data through a specific processing flow. This data acquisition method makes the low-resolution data depend on the high-resolution data at the source, and may thus cause problems such as doubts about data authenticity, poor data reliability, and potential data consistency issues in practical applications. Ultimately, it affects the universality and effectiveness of such methods in different application scenarios, and it is difficult to fully meet the requirements of actual meteorological monitoring and analysis for high-quality and highly reliable radar data.

[0003] Currently, no algorithm focuses on fusing the features of S-band observation data and X-band observation data to construct a spatial downscaling model. Summary of the Invention

[0004] Object of the Invention: The technical problem to be solved by the present invention is to provide a method for spatial downscaling of multi-band radar data based on data fusion in view of the deficiencies of the prior art. This method can convert the original low-resolution S-band data into high-resolution data, meet the requirements for high-resolution radar data in practical applications, and at the same time reduce the dependence on hardware facilities.

[0005] The method of the present invention includes the following steps:

[0006] Step 1, obtaining S-band and X-band observation data from a weather radar;

[0007] Step 2, performing data preprocessing on the obtained observation data;

[0008] Step 3, constructing a spatial downscaling model;

[0009] Step 4, training, validating, and testing the spatial downscaling model with the observation data after data preprocessing;

[0010] Step 5, performing spatial downscaling of the S-band observation data according to the trained spatial downscaling model;

[0011] Step 6, determining whether the generated data is close to the true value.

[0012] Step 2 includes:

[0013] Step 2.1, time-space matching of S-band and X-band observation data:

[0014] According to the time identification information contained in the original observation file name, the X-band observation data and the S-band radar are time-calibrated so that the X-band observation data and the S-band radar are aligned in the time dimension;

[0015] According to the latitude and longitude information corresponding to different stations, the S-band observation data is clipped according to the latitude and longitude values;

[0016] Step 2.2, rotate and align the X-band observation data:

[0017] By analyzing the historical observation data of the X-band, the current angle deviation value of the X-band is determined;

[0018] Based on the angle deviation value, the radar scanning center is used as the rotation reference point, and a rotation matrix algorithm based on the trigonometric function principle is used to transform the coordinate system of the X-band observation data;

[0019] Step 2.3, interpolate the X-band observation data to fill in the missing data:

[0020] Based on the distribution characteristics of the valid data area acquired by the X-band radar, the data range of invalid values ​​in the data is calculated to define the scanning blind area range and boundary conditions where data is missing;

[0021] Select an interpolation algorithm, use the valid data points around the scanning blind area as the reference data source, perform data estimation according to the established interpolation rules, and fill the estimated data into the corresponding missing positions;

[0022] Step 2.4, remove low-quality data: perform quality assessment on the X-band observation data and the S-band observation data, and delete the low-quality radar data from the X-band observation data and the S-band observation data; use the X-band observation data as high-resolution input data and the S-band observation data as low-resolution input data.

[0023] Step 3 includes:

[0024] Step 3.1, construct two independent feature extraction modules to process low-resolution input data and downsampled high-resolution input data respectively;

[0025] Multi-scale feature extraction for low-resolution input data: The S-band observation data is input into the dilated multi-scale convolutional network, and flows through the small receptive field convolution layer, the medium receptive field convolution layer and the large receptive field convolution layer in sequence to extract multi-scale features and generate a low-resolution LR feature map;

[0026] Downsample the high-resolution input data and extract multi-scale features: First, perform downsampling on the X-band observation data to reduce the spatial resolution to the same as that of the S-band observation data, generating the downsampled high-resolution input data HR Down Input, and simulating the data distribution under low-resolution input conditions. Second, input the high-resolution input data HR Down Input into the same dilated multi-scale convolutional network structure as the low-resolution branch, and successively pass through the convolutional layer with a small receptive field, the convolutional layer with a medium receptive field, and the convolutional layer with a large receptive field to extract multi-scale features, generating the downsampled high-resolution HR Down feature map. Through the symmetric feature extraction structure, the model learns the common features of low-resolution and high-resolution data.

[0027] Step 3.2, construct a data fusion module:

[0028] The data fusion module deeply fuses the features extracted from the two branches to generate a comprehensive feature map: First, perform a channel concatenation operation, concatenate the low-resolution LR feature map after multi-scale feature extraction and the downsampled high-resolution HR Down feature map along the channel dimension. The concatenated feature map successively passes through a convolutional layer, a batch normalization layer, and a rectified linear unit to achieve deep fusion of the features from the low-resolution LR and downsampled high-resolution HR Down data.

[0029] The convolutional layer in the data fusion module is a two-dimensional convolutional layer. The input is the concatenated 128-channel feature map. It slides a 3×3 convolutional kernel on the feature map for convolution operations and finally outputs a 64-channel feature map. The batch normalization layer is a two-dimensional batch normalization layer used to normalize the data to a distribution with a mean of 0 and a variance of 1. The rectified linear unit is the ReLU activation function used to perform a non-linear transformation on the feature map after batch normalization. The ReLU activation function is:

[0030] ,

[0031] where is the input value after being processed by the batch normalization layer, is the output value after passing through the ReLU activation function. If is greater than 0, the output is equal to ; if is less than or equal to 0, the output is equal to 0;

[0032] Step 3.3, construct an upsampling module: The upsampling module includes a channel attention module, two residual dense blocks, a convolutional layer, and bicubic interpolation upsampling.

[0033] The channel attention module, two residual dense blocks, the convolutional layer for expanding the number of channels, bicubic interpolation upsampling, and the convolutional layer for adjusting the number of channels are connected in sequence: First is the channel attention module, followed by two residual dense blocks, then through the convolutional layer for expanding the number of channels, and then upsampling using the bicubic interpolation function F.interpolate to complete the conversion from low-resolution input to high-resolution output; finally, using the convolutional layer for adjusting the number of channels, the final upsampling result with the same number of channels as the input image is output;

[0034] Among them, in the channel attention module, first receive the concatenated feature map output by the data fusion module; then perform adaptive average pooling and adaptive max pooling operations on the concatenated feature map respectively; subsequently add the results of average pooling and max pooling and input them into the fully connected network composed of two linear layers, ReLU activation function, and Sigmoid function; then adjust and expand the shape of the attention weights to the same size as the concatenated feature map; finally, multiply the concatenated feature map and the attention weights element-wise to obtain the feature map adjusted by the channel attention mechanism;

[0035] The following operations are performed on the residual dense block:

[0036] Step a1, when forward propagating, receive the feature map f adjusted by the channel attention mechanism; first perform a convolutional operation on the feature map f, and then obtain the feature map f1 through the LeakyReLU activation function; then concatenate the feature map f and the feature map f1 in the channel dimension, perform convolution and activate through the LeakyReLU function to obtain the feature map f2;

[0037] Step a2, according to step a1, continuously concatenate the obtained feature maps f1 and f2 in the channel dimension and input them into the convolutional layer in the residual dense block to obtain the feature maps f3 and f4 in sequence; finally, concatenate all the feature maps f1, f2, f3, and f4 in the channel dimension and perform convolution to obtain the feature map f5, multiply the feature map f5 by 0.2 and add it to the feature map f to obtain the output of the residual dense block;

[0038] The formula for upsampling using the bicubic interpolation function F.interpolate is:

[0039] ,

[0040] where The function represents using the bicubic interpolation method to upsample the fused feature map processed by the convolutional layer for expanding the number of channels to the specified size , obtaining the high-resolution feature map represents the height of the high-resolution image, represents the width of the high-resolution image;

[0041] Step 3.4, modify the loss function: introduce the height-weight adjustment loss, perform a linear mapping according to the input height information to associate the height with the weight, and at the same time introduce the Sobel gradient loss to enhance the edge information, specifically including:

[0042] The formula for the weight loss function is:

[0043] ,

[0044] ,

[0045] where, represents the total loss, represents the L1 loss, represents the Sobel gradient loss, represents the weight of the Sobel gradient loss, n represents the number of samples, represents the total loss of the i-th sample, represents the height loss corresponding to the i-th sample, where i = 1, 2,..., n, represents the weighted loss;

[0046] By successively constructing the feature extraction module, data fusion module, and upsampling module, a spatial downscaling model is finally obtained.

[0047] Step 4 includes:

[0048] Step 4.1, divide the dataset:

[0049] Take the S-band and X-band observation data after data preprocessing as data pairs, and divide them into training set, validation set, and test set according to a certain proportion; take the S-band observation data as the low-resolution data and the X-band observation data as the high-resolution radar data, and the low-resolution data and high-resolution data correspond one by one;

[0050] Step 4.2, train the spatial downscaling model according to the divided dataset:

[0051] In the training stage, input m S-band observation data in the training set and m X-band observation data , and the output prediction result is the image after the spatial downscaling of the model, represents the m-th S-band observation data, represents the m-th X-band observation data, where m is the number of data in the training set.

[0052] Step 4.2 further includes: according to the weighted loss function of the spatial downscaling model, through the backpropagation algorithm, updating the iterative weight parameters, and performing iterative training in a loop until the model converges.

[0053] In step 5, perform spatial downscaling of the S-band observation data according to the trained spatial downscaling model; perform visualization processing on the data generated by the spatial downscaling model, and compare it with the real X-band observation data.

[0054] In step 6, judge whether the generated data is close to the real value according to the evaluation index structural similarity SSIM. The formula is:

[0055] ,

[0056] where the luminance comparison function , the contrast comparison function and the structure comparison function are calculated according to the following formulas respectively:

[0057] ,

[0058] ,

[0059] ,

[0060] where represents the X-band observation data, represents the S-band observation data, and are respectively the mean value of and the and are respectively the standard deviation of and the is the and covariance of , and are constants, is the weight coefficient. When , it means that when calculating SSIM, the luminance comparison function , the contrast comparison function and the structure comparison function have the same influence weight on the final result. When the two images are the same, .

[0061] The present invention also provides an electronic device, including a processor and a memory. The memory stores program codes. When the program codes are executed by the processor, the processor is caused to execute the steps of the method described above.

[0062] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction runs on a computer, the steps of the method described above are executed.

[0063] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention accurately calibrates S-band and X-band observation data, avoiding data problems caused by spatio-temporal differences; meanwhile, it comprehensively improves the quality of X-band observation data, which is superior to conventional simple processing methods. (2) The present invention adopts hole multi-scale convolution, breaking through the limitation of a single scale, comprehensively excavating data features from multiple dimensions; innovatively fusing multi-scale features to achieve deep fusion, providing rich information for subsequent upsampling. (3) The present invention can utilize a deep learning neural network to enable the model to extract and fuse the features of S-band and X-band observation data, converting the originally low-resolution S-band data into high-resolution data, not only meeting the demand for high-resolution radar data in practical applications but also reducing the dependence on hardware facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flowchart of the present invention.

[0065] Figure 2 is a flowchart of data preprocessing of the present invention.

[0066] Figure 3 is a schematic structural diagram of the spatial downscaling model of the present invention.

[0067] Figure 4 is a schematic structural diagram of the feature extraction module of the present invention.

[0068] Figure 5 is a schematic structural diagram of the data fusion module of the present invention.

[0069] Figure 6 is a schematic diagram of the training, verification and testing idea of the spatial downscaling model of the present invention.

[0070] Figure 7 is an effect diagram of the multi-band radar data spatial downscaling method based on data fusion of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0071] The following further detailed description of the present invention is made in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become clearer.

[0072] An embodiment of the present invention proposes a multi - band radar data spatial downscaling method based on data fusion, as follows Figure 1 shown, including the following steps:

[0073] Step 1, obtain S - band and X - band observation data from weather radar:

[0074] Obtain S - band and X - band observation data from weather radar. The S - band observation data is a networked CAPPI (Composite Azimuthal Plane Position Indicator) product of reflectivity factor, and the X - band observation data is a CAPPI product of reflectivity factor of a phased - array radar. The spatial resolution of the X - band observation data is 0.0025, the spatial resolution of the S - band observation data is 0.01, and the time resolution of both is 6 minutes. The height range of the data covers from 2000 meters to 6000 meters and is divided into height layers every 500 meters.

[0075] Step 2, perform data pre - processing on the obtained observation data. The data processing flow chart is as Figure 2 shown, specifically including the following steps:

[0076] Step 2.1, perform spatio - temporal matching on the S - band and X - band observation data:

[0077] First, according to the time identification information contained in the original observation file name, perform time calibration operations on the X - band observation data and the S - band radar, so that the X - band observation data and the S - band radar are aligned in the time dimension;

[0078] Secondly, since the S - band radar is a networked CAPPI product of reflectivity factor, it is necessary to perform precise cropping processing on the S - band observation data according to the longitude and latitude information corresponding to different stations to ensure its spatial consistency with the X - band observation data;

[0079] Step 2.2, perform rotation alignment on the X - band observation data:

[0080] Since there is a certain degree of scanning angle deviation in the X - band radar, this deviation will have an adverse impact on the precise alignment between different radar data. It is necessary to perform precise rotation adjustment operations on the X - band observation data to make its scanning angle consistent with the angle reference of the S - band observation data;

[0081] First, determine its current angle deviation value through the analysis of the historical scanning data of the X - band radar; secondly, based on the determined deviation value, with the radar scanning center as the rotation reference point, use the rotation matrix algorithm based on the principle of trigonometric functions to perform transformation operations on the coordinate system of the X - band observation data;

[0082] Step 2.3, perform interpolation on the X - band radar to fill in missing data:

[0083] In view of the problem of missing data in some areas caused by the scanning characteristics in the X-band observation data, it is necessary to interpolate and fill the missing data of the X-band radar to ensure the integrity and continuity of the data;

[0084] First, based on the distribution characteristics of the effective data area obtained by the X-band radar, calculate the data range of invalid values in the data to accurately define the range of the scanning blind area with missing data and its boundary conditions;

[0085] Secondly, select an appropriate interpolation algorithm, such as an interpolation algorithm based on weighted average of adjacent points, bilinear interpolation algorithm, spline interpolation algorithm or Kriging interpolation algorithm, etc. Use the effective data points around the scanning blind area as the reference data source, and estimate the data according to the established interpolation rules, and fill the estimated data into the corresponding missing positions;

[0086] Step 2.4, eliminate low-quality data:

[0087] Conduct quality assessment on the X-band observation data and S-band observation data. Through the preset determination standard of the effective data volume threshold, identify the radar data with a relatively small proportion of effective data. For example, if the proportion of zero values in a certain type of data exceeds the threshold of 30%, then this type of data will be determined as the data with a relatively small proportion of effective data, that is, low-quality data; delete the low-quality radar data from the X-band observation data and S-band observation data; use the X-band observation data as the high-resolution input data and the S-band observation data as the low-resolution input data.

[0088] Step 3, construct a spatial downscaling model. The overall structure diagram of the spatial downscaling model is as Figure 3 shown, including a feature extraction module, a data fusion module, and an upsampling module:

[0089] Step 3.1, construct two independent feature extraction modules to process the low-resolution input data LowResolution Input (LR Input) and the downsampled high-resolution input data High Resolution DownInput (HR Down Input) respectively; the specific structure of the module is as Figure 4 shown:

[0090] Perform multi-scale feature extraction on the low-resolution input data (LR Input): Input the S-band observation data into the dilated multi-scale convolutional network, and flow through the small receptive field convolutional layer (dilation rate 1, convolutional kernel 3×3), medium receptive field convolutional layer (dilation rate 2, convolutional kernel 3×3), and large receptive field convolutional layer (dilation rate 4, convolutional kernel 3×3) in sequence to extract multi-scale features and generate a low-resolution LR feature map;

[0091] Downsample the high-resolution input data (HR Input) and extract multi-scale features: First, perform downsampling on the X-band observation data to reduce the spatial resolution to the same as that of the S-band observation data, generating the downsampled high-resolution input data HR Down Input to simulate the data distribution under low-resolution input conditions. Second, input the high-resolution input data HR Down Input into the same dilated multi-scale convolutional network structure as the low-resolution branch, and successively pass through the convolutional layer with a small receptive field, the convolutional layer with a medium receptive field, and the convolutional layer with a large receptive field to extract multi-scale features, generating the downsampled high-resolution HR Down feature map. Through the symmetric feature extraction structure, the model learns the common features of low-resolution and high-resolution data.

[0092] Step 3.2, construct a data fusion module, and the specific structure of the module is as Figure 5 shown:

[0093] The data fusion module deeply fuses the features extracted from the two branches to generate a comprehensive feature map: First, perform a channel concatenation operation, concatenate the low-resolution LR feature map after multi-scale feature extraction and the downsampled high-resolution HR Down feature map along the channel dimension. The concatenated feature map successively passes through a convolutional layer, a batch normalization layer, and a rectified linear unit to achieve deep fusion of the features from the low-resolution LR and downsampled high-resolution HR Down data.

[0094] The convolutional layer in the data fusion module is a two-dimensional convolutional layer. The input is the concatenated 128-channel feature map, and a 3×3 convolutional kernel slides on the feature map for convolution operations, finally outputting a 64-channel feature map. The batch normalization layer is a two-dimensional batch normalization layer used to normalize the data to a distribution with a mean of 0 and a variance of 1. The rectified linear unit is the ReLU activation function used to perform a non-linear transformation on the feature map after batch normalization. The ReLU activation function is:

[0095] ,

[0096] where is the input value after being processed by the batch normalization layer, is the output value after passing through the ReLU activation function. If is greater than 0, the output is equal to ; if is less than or equal to 0, the output is equal to 0;

[0097] Step 3.3, construct an upsampling module: The upsampling module includes a channel attention module, two residual dense blocks, a convolutional layer, and bicubic interpolation upsampling.

[0098] The channel attention module, two residual dense blocks, the convolutional layer for expanding the number of channels, bicubic interpolation upsampling, and the convolutional layer for adjusting the number of channels are connected in sequence: First is the channel attention module, which can adaptively adjust the feature responses of each channel and enhance the feature expression ability; then are two residual dense blocks, which combine the advantages of residual and dense connections, can alleviate the problem of gradient disappearance and promote feature propagation and reuse; then through the convolutional layer for expanding the number of channels, the number of channels of the output feature map is expanded to provide sufficient information for the subsequent upsampling operation; then use the bicubic interpolation function F.interpolate for upsampling to complete the conversion from low-resolution input to high-resolution output; finally use the convolutional layer for adjusting the number of channels to output the final upsampling result with the same number of channels as the input image;

[0099] Among them, in the channel attention module, first receive the concatenated feature map output by the data fusion module; then perform adaptive average pooling and adaptive max pooling operations on the concatenated feature map respectively; then add the results of average pooling and max pooling and input them into the fully connected network composed of two linear layers, ReLU activation function, and Sigmoid function; then adjust and expand the shape of the attention weights to the same size as the concatenated feature map; finally multiply the concatenated feature map element-wise by the attention weights to obtain the feature map adjusted by the channel attention mechanism;

[0100] The following operations are performed on the residual dense block:

[0101] Step a1, when forward propagating, receive the feature map f adjusted by the channel attention mechanism; first perform a convolutional operation on the feature map f, and then obtain the feature map f1 through the LeakyReLU activation function; then concatenate the feature map f and the feature map f1 in the channel dimension, perform convolution and activate through the LeakyReLU function to obtain the feature map f2;

[0102] Step a2, according to step a1, continuously concatenate the obtained feature maps f1 and f2 in the channel dimension and input them into the convolutional layer in the residual dense block to obtain the feature maps f3 and f4 in sequence; finally, concatenate all the feature maps f1, f2, f3, and f4 in the channel dimension and perform convolution to obtain the feature map f5, multiply the feature map f5 by 0.2 and add it to the feature map f to obtain the output of the residual dense block;

[0103] The formula for upsampling using the bicubic interpolation function F.interpolate is:

[0104] ,

[0105] where The function represents using the bicubic interpolation method to upsample the fused feature map processed by the convolutional layer with the expanded number of channels to the specified size to obtain a high-resolution feature map , completing the conversion from low-resolution input to high-resolution output; where represents the height of the high-resolution image, represents the width of the high-resolution image;

[0106] Step 3.4, modify the loss function: Usually, as the height increases, the quality and accuracy of the data received by the weather radar will be affected to a certain extent, and the data reliability will decrease accordingly. Therefore, a height weight adjustment loss is introduced to linearly map according to the input height information to associate height with weight, with higher weights for lower heights; at the same time, a Sobel gradient loss is introduced to enhance edge information, specifically including:

[0107] In the loss calculation process, first calculate the original L1 loss, which reflects the difference between the model output and the true label; then calculate the Sobel gradient loss, which measures the deviation between the Sobel gradient of the model output image and the Sobel gradient of the true label image; then combine these two losses to obtain the total loss; subsequently, multiply the total loss by the height weight and take the average of the results to finally obtain the weighted loss, so that samples with lower heights correspond to higher weights and have a greater impact during the loss calculation process. The weight loss function formula is:

[0108] ,

[0109] ,

[0110] where, represents the total loss, represents the L1 loss, represents the Sobel gradient loss, represents the weight of the Sobel gradient loss, n represents the number of samples, represents the total loss of the i-th sample, represents the height loss corresponding to the i-th sample, where i = 1, 2,..., n, represents the weighted loss.

[0111] By successively constructing a feature extraction module, a data fusion module, and an upsampling module, a spatial downscaling model is finally obtained.

[0112] Step 4, train, validate, and test the spatial downscaling model according to the data after data preprocessing. The flowchart is as Figure 6 shown, specifically including:

[0113] Step 4.1, Divide the dataset:

[0114] Take the S-band and X-band observation data after data preprocessing as data pairs, and divide them into a training set, a validation set, and a test set according to a ratio (such as 7:2:1); take the S-band observation data as the low-resolution data and the X-band observation data as the high-resolution radar data, with the low-resolution data and the high-resolution data corresponding one by one.

[0115] Step 4.2, Train the spatial downscaling model according to the divided dataset:

[0116] In the training stage, input m S-band observation data in the training set and m X-band observation data , and the output prediction result is the image after spatial downscaling by the model, denotes the m-th S-band observation data, denotes the m-th X-band observation data, where m is the number of data in the training set, and the value range is from 7000 to 10000, covering data in different time periods.

[0117] According to the weighted loss function of the spatial downscaling model, update the iterative weight parameters through the backpropagation algorithm, and perform cyclic iterative training for this step until the model converges.

[0118] Step 5, Perform spatial downscaling on the S-band observation data according to the trained spatial downscaling model;

[0119] Perform spatial downscaling on the S-band observation data according to the trained spatial downscaling model; perform visualization processing on the data generated by the model and compare it with the real X-band observation data.

[0120] Step 6, Judge whether the generated data is close enough to the real value;

[0121] Compare according to the evaluation index structural similarity (SSIM). SSIM is based on the brightness of the image , contrast and structure to calculate. For two images and , the SSIM formula is:

[0122] ,

[0123] where the brightness comparison function , the contrast comparison function and the structure comparison function are calculated as follows:

[0124] ,

[0125] ,

[0126] ,

[0127] wherein represents X-band observation data, represents S-band observation data, and are respectively the mean value of and and are respectively the standard deviation of and is and the covariance of , and are constants, is the weight coefficient. When , it means that when calculating SSIM, the luminance comparison function , the contrast comparison function and the structure comparison function have the same influence weight on the final result. When the two images are the same, .

[0128] After the model is trained, it is used to perform spatial downscaling on the S-band radar data, and then the data generated by the model is compared with the X-band radar data. If the generated data is close enough, the model can be put into use. Figure 7 This is the effect diagram of the multi-band radar data spatial downscaling method based on data fusion of the present invention. As shown in Figure 7 , when the S-band radar data is input and processed by the model, high-resolution S-band radar data after spatial downscaling is obtained.

[0129] The present invention provides a multi-band radar data spatial downscaling method based on data fusion. There are many methods and ways to specifically implement this technical solution. The above description is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by the existing technology.

Claims

1. A multi-band radar data spatial downscaling method based on data fusion, characterized in that: The following steps are involved: Step 1, obtaining S-band and X-band observation data from weather radar; Step 2, preprocessing the acquired observation data; Step 3, construct a spatial downscaling model; Step 4, training, verifying and testing the spatial downscaling model through the observation data after data preprocessing; Step 5, perform spatial downscaling of the S-band observation data according to the trained spatial downscaling model; Step 6, determine whether the generated data is close to the true value; Step 2 includes: Step 2.1, time-space matching of S-band and X-band observation data: According to the time identification information contained in the original observation file name, the X-band observation data and the S-band radar are time-calibrated so that the X-band observation data and the S-band radar are aligned in the time dimension; According to the latitude and longitude information corresponding to different stations, the S-band observation data is clipped according to the latitude and longitude values; Step 2.2, rotation and alignment of X-band observation data: By analyzing the historical observation data of the X-band, the current angle deviation value of the X-band is determined; Based on the angle deviation value, the radar scanning center is used as the rotation reference point, and a rotation matrix algorithm based on the principle of trigonometric function is used to transform the coordinate system of the X-band observation data; Step 2.3, interpolate the X-band observation data to fill in the missing data: Based on the distribution characteristics of the valid data area acquired by the X-band radar, the data range of invalid values ​​in the data is calculated to define the scanning blind area range and boundary conditions where data is missing; Select an interpolation algorithm, use the valid data points around the scanning blind area as the reference data source, perform data estimation according to the established interpolation rules, and fill the estimated data into the corresponding missing positions; Step 2.4, remove low-quality data: perform quality assessment on the X-band observation data and the S-band observation data, and remove low-quality radar data from the X-band observation data and the S-band observation data; use the X-band observation data as high-resolution input data and the S-band observation data as low-resolution input data; Step 3 includes: Step 3.1, construct two independent feature extraction modules to process low-resolution input data and downsampled high-resolution input data respectively; Multi-scale feature extraction for low-resolution input data: The S-band observation data is input into the dilated multi-scale convolutional network, and flows through the small receptive field convolution layer, the medium receptive field convolution layer and the large receptive field convolution layer in sequence to extract multi-scale features and generate a low-resolution LR feature map; Down-sampling and multi-scale feature extraction of high-resolution input data: First, down-sampling operation is performed on X-band observation data to reduce the spatial resolution to the same level as S-band observation data, and down-sampled high-resolution input data HR Down Input is generated to simulate the data distribution under low-resolution input conditions; secondly, the high-resolution input data HRDown Input is input into the same hole multi-scale convolutional network structure as the low-resolution branch, and flows through the small receptive field convolution layer, the medium receptive field convolution layer and the large receptive field convolution layer in sequence to extract multi-scale features and generate a down-sampled high-resolution HRDown feature map. Through the symmetrical feature extraction structure, the model learns the common features of low-resolution and high-resolution data; Step 3.2, build data fusion module: The data fusion module deeply fuses the features extracted from the two branches to generate a comprehensive feature map: first, a channel splicing operation is performed to splice the low-resolution LR feature map after multi-scale feature extraction with the downsampled high-resolution HRDown feature map according to the channel dimension. The spliced ​​feature map passes through the convolution layer, batch normalization layer and rectified linear unit in turn to achieve deep fusion of the features from the low-resolution LR and downsampled high-resolution HR Down data; The convolution layer in the data fusion module is a two-dimensional convolution layer, whose input is the spliced ​​128-channel feature map, and the convolution operation is performed by sliding a 3×3 convolution kernel on the feature map, and finally outputs a 64-channel feature map; the batch normalization layer is a two-dimensional batch normalization layer, which is used to normalize the data to a distribution with a mean of 0 and a variance of 1; the rectified linear unit is a ReLU activation function, which is used to perform nonlinear transformation on the batch-normalized feature map; the ReLU activation function is: and relu =max(0,x norm ), where x norm is the input value after batch normalization layer processing, y relu is the output value after the ReLU activation function. If x norm Greater than 0, output y relu Equal to x norm ; if x norm Less than or equal to 0, output y relu is equal to 0; Step 3.3, construct an upsampling module: the upsampling module includes a channel attention module, two residual dense blocks, a convolutional layer and bicubic interpolation upsampling; The channel attention module, two residual dense blocks, a convolutional layer for expanding the number of channels, bicubic interpolation upsampling, and a convolutional layer for adjusting the number of channels are connected in sequence: first, the channel attention module, followed by two residual dense blocks, and then the convolutional layer for expanding the number of channels, and then the bicubic interpolation function F.interpolate is used for upsampling to complete the conversion from low-resolution input to high-resolution output; finally, the convolutional layer for adjusting the number of channels is used to output the final upsampling result with the same number of channels as the input image; Among them, in the channel attention module, the spliced ​​feature map output by the data fusion module is first received; then the spliced ​​feature map is adaptively averaged and adaptively max-pooled respectively; then the results of average pooling and max-pooling are added together and input into a fully connected network consisting of two linear layers, ReLU activation function and Sigmoid function; then the shape of the attention weight is adjusted to the same size as the spliced ​​feature map; finally, the spliced ​​feature map is element-wise multiplied with the attention weight to obtain the feature map adjusted by the channel attention mechanism; The residual dense block performs the following operations: Step a1, during forward propagation, receive the feature map f adjusted by the channel attention mechanism; first, perform a convolution operation on the feature map f, and then activate it through the LeakyReLU function to obtain the feature map f1; then, concatenate the feature map f and the feature map f1 in the channel dimension, perform convolution and activate it through the LeakyReLU function to obtain the feature map f2; Step a2: According to step a1, the obtained feature maps f1 and f2 are continuously concatenated in the channel dimension and input into the convolution layer in the residual dense block to obtain feature maps f3 and f4 in turn; finally, all feature maps f1, f2, f3 and f4 are concatenated in the channel dimension and convolved to obtain feature map f5, and feature map f5 is multiplied by 0.2 and added to feature map f to obtain the output of the residual dense block; The formula for upsampling the bicubic interpolation function F.interpolate is: and high_res =BicubicInterpolate(y up_conv ,H high ,W high ), The BicubicInterpolate function represents the use of the bicubic interpolation method to fused the feature map y after the convolution layer with the expanded number of channels up_conv Upsample to the specified size H high ×W high , get the high-resolution feature map y high_res , complete the conversion from low-resolution input to high-resolution output; where H high Represents the height of the high-resolution image, W high Indicates the width of the high-resolution image; Step 3.4, modify the loss function: introduce height weight adjustment loss, perform linear mapping based on the input height information, associate height with weight, and introduce Sobel gradient loss to enhance edge information, including: The weight loss function formula is: L total =L1+α×L sobel , Among them, L total represents the total loss, L1 represents the L1 loss, L sobel represents the Sobel gradient loss, α represents the weight of the Sobel gradient loss, n represents the number of samples, represents the total loss of the i-th sample, w (i) represents the height loss corresponding to the i-th sample, where i = 1, 2, ..., n, L weighted represents the weighted loss; By sequentially constructing the feature extraction module, data fusion module, and upsampling module, we finally obtained the spatial downscaling model; Step 4 includes: Step 4.1, divide the data set: The S-band and X-band observation data after data preprocessing are used as data pairs and divided into training set, validation set and test set according to proportion; the S-band observation data are used as low-resolution data, and the X-band observation data are used as high-resolution radar data, and the low-resolution data and high-resolution data correspond one to one; Step 4.2: Train the spatial downscaling model based on the divided data set: In the training phase, the m S-band observation data {x1,…,x m } and m X-band observation data {y1,…,y m }, the output prediction result is the image after model space downscaling, x m represents the mth S-band observation data, y m represents the mth X-band observation data, where m is the number of data in the training set; Step 4.2 also includes: updating the iterative weight parameters according to the weighted loss function of the spatial downscaling model through a back propagation algorithm, and performing cyclic iterative training until the model converges; In step 5, the S-band observation data is spatially downscaled according to the trained spatial downscaling model; the data generated by the spatial downscaling model is visualized and compared with the real X-band observation data; In step 6, the evaluation index structural similarity SSIM is used to determine whether the generated data is close to the true value. The formula is: SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ , Among them, the calculation formulas of brightness comparison function l(x,y), contrast comparison function c(x,y) and structure comparison function s(x,y) are: Where y represents the X-band observation data, x represents the S-band observation data, μ x and μ y are the mean of x and the mean of y, σ x and σ y are the standard deviation of x and the standard deviation of y, σ xy is the covariance of x and y, C1, C2 and C3 are constants, α, β, γ are weight coefficients; In step 6, when α=β=γ=1, it means that when calculating SSIM, the brightness comparison function l(x, y), the contrast comparison function c(x, y) and the structure comparison function s(x, y) have the same influence weight on the final result. When the two images are the same, SSIM(x, y)=1.

2. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to claim 1.

3. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to claim 1 are executed.

Images