A three-dimensional time-varying field prediction method for radar detection

Abstract

The application discloses a three-dimensional time-varying field prediction method for radar detection, and the method comprises the following steps: obtaining original three-dimensional time-varying field data obtained through radar detection; generating corresponding position codes according to the spatial dimensions of the input original three-dimensional time-varying field data; performing feature fusion on the original three-dimensional time-varying field data and the position codes in the channel dimension to obtain fused data; performing an asymmetric spatial transformation operation on the fused data according to a preset asymmetric three-dimensional Patch size to obtain asymmetric spatial transformed data; inputting the asymmetric spatial transformed data into a pre-trained deep learning model to obtain prediction data; and wherein the pre-trained deep learning model is obtained by training based on historical three-dimensional time-varying field data. The asymmetric spatial transformation operation of the method can adapt to the anisotropic structure of the atmospheric field strength observed by the radar, and improves the physical rationality and interpretability of the subsequent model.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, and specifically relates to a three-dimensional time-varying field prediction method for radar detection. Background Technology

[0002] Weather radar is capable of high spatiotemporal resolution and continuous three-dimensional detection of the atmospheric environment, making it a crucial technology for acquiring three-dimensional time-varying atmospheric fields such as precipitation, wind fields, and cloud formations. Currently, deep learning models have shown great application potential in weather forecasting and atmospheric science research. However, significant technical challenges remain when processing high-resolution time-varying three-dimensional atmospheric field data. The memory usage and computational overhead of general-purpose spatiotemporal models (such as Video Swin Transformer (Liu et al., 2021)) vary with the number of voxels. It grows exponentially. Taking typical radar observations or multivariate three-dimensional atmospheric data as an example, when the data has a certain time step, vertical layer number and horizontal resolution (e.g., time step T=12, vertical layer number D=12, horizontal resolution H=W=400), directly using a general video model often exceeds the video memory limit of a single graphics processing unit (GPU), affecting practical applications.

[0003] Traditional patching methods, such as the Video Swin Transformer, typically assume spatial isotropy and compress data by uniformly dividing voxels into N equal parts. However, this method has significant limitations for continuous atmospheric fields with clear physical meaning. Due to the significant anisotropy of the three-dimensional atmospheric field (the number of vertical layers is much smaller than the horizontal grid resolution, and the horizontal correlation scale is significantly larger than the vertical direction), coupled with the nonlinear vertical distribution of its physical quantities with height, a simple uniform spatial partitioning method can easily disrupt the physical consistency of the atmospheric field boundary and even introduce physically unreasonable pseudo-structures.

[0004] Furthermore, existing positional encoding techniques, such as sinusoidal encoding in Transformers or encoding extended to spatially continuous fields in NeRF (Mildenhallet al., 2022), while providing positional information for the model, are not specifically optimized for the characteristics of discrete grids, especially the atmospheric field with significant anisotropy obtained from radar observations. These methods may fail to effectively capture the physical properties of the atmospheric field that decay exponentially with altitude in the vertical direction, and also fail to fully utilize the differences in atmospheric horizontal and vertical resolution for efficient compression.

[0005] Therefore, how to reduce memory and computational overhead in high-resolution atmospheric field data processing, make the patch partitioning method match the real physical characteristics of the atmospheric field in radar observation data, and make the position coding effectively capture its anisotropy and multi-scale structure has become a key problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a three-dimensional time-varying field prediction method for radar detection, so as to solve the problem that traditional uniform patch partitioning cannot adapt to the anisotropic structure of atmospheric field strength observed by radar.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a three-dimensional time-varying field prediction method for radar detection, comprising the following steps:

[0008] S1: Acquire the raw three-dimensional time-varying field data obtained from radar detection;

[0009] S2: Generate the corresponding position code based on the spatial dimension of the input original three-dimensional time-varying field data;

[0010] S3: The original three-dimensional time-varying field data and the position code are fused in the channel dimension to obtain the fused data;

[0011] S4: Perform an asymmetric spatial transformation operation on the fused data according to the preset asymmetric three-dimensional patch size to obtain asymmetric spatial transformed data; the asymmetric three-dimensional patch size is different from other dimensions in at least one dimension;

[0012] S5: Input the data after the asymmetric spatial transformation into a pre-trained deep learning model to obtain prediction data; wherein, the pre-trained deep learning model is trained based on historical three-dimensional time-varying field data;

[0013] The asymmetric spatial transformation operation includes asymmetric rearrangement, which specifically includes:

[0014] Based on the asymmetric 3D Patch size, the spatial dimension of the feature-stitched data is divided by integer division to obtain multiple 3D data blocks;

[0015] The three-dimensional data block is flattened to the channel dimension to obtain the data after asymmetric rearrangement.

[0016] By performing spatial transformation operations on asymmetric three-dimensional block sizes, adaptive processing is performed on the differences in feature distribution of radar detection data in different spatial dimensions. Compared with symmetric block transformation methods, it can more accurately capture the non-uniform spatial features of radar time-varying fields and reduce the loss of feature information.

[0017] Furthermore, in S2, generating the corresponding position code includes:

[0018] S2.1: Normalize the original spatial coordinates of the input raw three-dimensional time-varying field data to obtain normalized spatial coordinates;

[0019] S2.2: Generate multiple wavelength parameters with logarithmic interval distribution based on the preset position encoding dimension;

[0020] S2.3: Calculate the vertical basis based on the vertical dimension coordinates in the normalized spatial coordinates;

[0021] S2.4: Based on the normalized spatial coordinates and multiple wavelength parameters, calculate the horizontal sine basis and the horizontal cosine basis;

[0022] S2.5: The vertical basis, the horizontal sine basis, and the horizontal cosine basis are concatenated along the channel dimension to obtain the corresponding position code.

[0023] By using Earth-specific multi-scale location coding to construct independent location bases for the vertical and horizontal dimensions, and employing logarithmically spaced wavelength parameters, location coding can effectively capture the anisotropy and multi-scale structure of the atmospheric field, thereby enhancing the ability of subsequent models to perceive and express meteorological characteristics in different spatial directions and at different scales.

[0024] Furthermore, in S5:

[0025] The channel dimension of the data after the asymmetric spatial transformation is reduced to obtain compressed data;

[0026] The compressed data is input into a pre-trained deep learning model to obtain predicted data. Dimensionality reduction effectively reduces the computational cost of subsequent time-series core modules, improves the processing efficiency of high-resolution three-dimensional atmospheric field data, and facilitates data processing for subsequent deep learning models.

[0027] Furthermore, the dimensionality reduction process employs either reversible or irreversible dimensionality transformation operations. Using reversible dimensionality transformation operations for dimensionality reduction can further compress the number of data channels without losing key feature information and ensuring that the information can be completely recovered. Using irreversible dimensionality transformation to reduce the channel dimension can effectively compress the number of data channels while retaining key feature information of the atmospheric field.

[0028] Furthermore, the original three-dimensional time-varying field data is a multidimensional tensor with dimension . Where B is the batch size, T is the time step, C is the number of channels, D is the number of vertical layers, H is the horizontal height, W is the horizontal width, and the size of the spatial dimension is... ;

[0029] The size of the asymmetric patch is: ,in, p z The size of the patch in the vertical dimension. p x The dimensions of the patch in the horizontal and vertical dimensions. p y Let Patch be the size in the horizontal width dimension, and or .

[0030] This invention addresses the anisotropic characteristics of atmospheric fields where the number of vertical layers is much smaller than the horizontal grid resolution. By differentiating the asymmetric patch size in the vertical and horizontal dimensions, this invention can accurately adapt to the strong anisotropic characteristics of atmospheric fields obtained from radar observations. This ensures that the patch division matches the true spatial distribution structure of the atmospheric field, avoiding physical boundary destruction and unreasonable pseudo-structures caused by uniform segmentation. This provides a more reasonable data foundation for subsequent feature extraction and spatiotemporal modeling, and enhances the model's ability to capture the vertical structure and horizontal evolution of the atmosphere.

[0031] Furthermore, the feature fusion is performed as feature concatenation, and the dimension of the concatenated data is [missing information]. In S4, the asymmetric spatial transformation operation is an asymmetric rearrangement, which specifically includes the following steps:

[0032] Based on the asymmetric patch size Regarding spatial dimensions conduct Integer division; resulting in multiple 3D data blocks;

[0033] The three-dimensional data block is flattened to the channel dimension to obtain data after asymmetric spatial transformation. The dimension of the data after asymmetric spatial transformation is... ;

[0034] Among them, the number of channels after rearrangement Vertical height after rearrangement The horizontal height after rearrangement The horizontal width after rearrangement .

[0035] By losslessly rearranging three-dimensional spatial voxels to the channel dimension, the number of channels is significantly increased while the spatial dimension is reduced, thus effectively reducing the memory usage and computational overhead of high-resolution time-varying three-dimensional atmospheric field data.

[0036] Furthermore, when the spatial dimension Cannot be During integer division, zero-value padding is performed on the spatial dimension data, specifically including the following steps:

[0037] Calculate the fill amount for the spatial dimension, including the horizontal height fill amount. Horizontal width fill amount and vertical fill volume ;in, Indicates rounding up;

[0038] The corresponding boundary of the spatial dimension is supplemented with zero values ​​corresponding to the filling amount;

[0039] Update the spatial dimension to , as well as .

[0040] By using zero-value padding, the spatial dimension of the three-dimensional time-varying field data can meet the divisibility requirement of the asymmetric block size, ensuring that the asymmetric spatial transformation operation can be executed stably and completely, and avoiding incomplete data segmentation, feature loss or calculation abnormalities caused by dimension mismatch.

[0041] Furthermore, in S4, the asymmetric spatial transformation operation is an asymmetric downsampling operation, which effectively reduces data resolution and computational load while adapting to the anisotropic spatial structure of the atmospheric field, further reducing memory usage and model computation overhead, and improving the processing efficiency of high-resolution three-dimensional atmospheric field data.

[0042] Furthermore, after obtaining the predicted data, an inverse symmetric spatial transformation operation is performed on the predicted data to restore the spatial dimension of the predicted data to the spatial dimension of the original three-dimensional time-varying field data, thus obtaining the data after the inverse asymmetric spatial transformation.

[0043] By performing an inverse symmetric spatial transformation operation, the spatial dimension is restored to be consistent with the original three-dimensional time-varying field data. This can restore the complete spatial structure and physical distribution of the high-resolution atmospheric field, ensuring that the final output data is consistent with the original observation data in terms of dimension, spatial topology, and physical meaning. This makes it convenient for practical applications such as weather forecasting, precipitation inversion, and radar data product generation.

[0044] Furthermore, after obtaining the data after the inverse asymmetric spatial transformation, channel processing is performed on the data to remove the position encoding channels, thus obtaining the final prediction result.

[0045] By pruning the front of the channel dimension from the data recovered by inverse spatial transformation and removing the extra channels introduced by position encoding, a final prediction result containing only physical variables is obtained.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] The asymmetric rearrangement operation proposed in this invention rearranges the spatial dimension information of the input data to the channel dimension without loss through a preset asymmetric patch size, while reducing spatial resolution and significantly reducing memory and computational overhead. Addressing the anisotropic characteristics of the atmospheric field, where the number of vertical layers is much smaller than the horizontal grid resolution, the invention differentiates the asymmetric patch size in the vertical and horizontal dimensions, ensuring that the patch partitioning method matches the true physical characteristics of the atmospheric field in radar observation data, avoiding the physical inconsistencies and irrationalities that may result from traditional uniform patches. By constructing independent position bases for the vertical and horizontal dimensions respectively and employing logarithmically spaced wavelength parameters, the position encoding can effectively capture its anisotropy and multi-scale structure. Attached Figure Description

[0048] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a three-dimensional time-varying field prediction method for radar detection in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the basis functions according to an embodiment of the present invention. Figure 2 (a) in the figure is a schematic diagram of the linear basis functions in an embodiment of the present invention; Figure 2 (b) in the figure is a schematic diagram of the exponentially decaying basis function in an embodiment of the present invention; Figure 2 (c) in the figure is a schematic diagram of the sinusoidal basis functions in an embodiment of the present invention; Figure 2 (d) in the figure is a schematic diagram of the cosine basis functions in an embodiment of the present invention;

[0051] Figure 3 The structural block diagram of the module for implementing the three-dimensional time-varying field prediction method for radar detection according to the embodiments of the present invention is shown below;

[0052] Figure 4 This is a performance evaluation index diagram of the precipitation forecasting method for three-dimensional time-varying field prediction based on radar detection according to an embodiment of the present invention. Figure 4 (a) in the figure is a curve showing the change of the CSI index with the forecast lead time; Figure 4 (b) in the figure is a curve showing the change of the POD index with the forecast lead time; Figure 4 (c) in the figure is a curve showing the change of the FAR index with the forecast lead time; Figure 4 (d) in the figure is a curve showing the change of the HSS index with the forecast lead time; Figure 4(e) in the figure is a curve showing the change of the BIAS index with the forecast lead time; Figure 4 (f) in the figure is a radar chart of CSI, POD, FAR, HSS and BIAS indicators for each forecast lead time. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Example 1

[0056] The three-dimensional time-varying field prediction method for radar detection proposed in this invention introduces Atmosphere-Aware Patch Encoding (AAPE) to specifically optimize the spatial continuity and nonlinear vertical structure characteristics of the atmospheric field with altitude. Figure 1 As shown, the method includes the following steps:

[0057] S1: Acquire the raw three-dimensional time-varying field data obtained from radar detection.

[0058] First, the raw input atmospheric data The input is fed into the processing system of this embodiment of the invention, wherein, Let X be the dimension of the atmospheric data, B be the batch size, T be the time step, C be the number of channels, D be the number of vertical layers, H be the horizontal height, and W be the horizontal width.

[0059] S2: Generate the corresponding position code based on the spatial dimension of the input original three-dimensional time-varying field data.

[0060] Independent of the main data stream, based on the spatial dimension of the input data The corresponding location code γ is generated through the Earth Specific Multiscale Location Coding (ESPE) module. The ESPE module constructs 2D horizontal and 3D codes based on normalized coordinates and specifically introduces vertical basis functions (such as linear or exponential decay functions) to capture the multiscale characteristics of the atmospheric field and the physical properties in the vertical direction.

[0061] Specifically, the process of generating location codes includes the following steps:

[0062] S2.1: The original spatial coordinates of the input raw three-dimensional time-varying field data Normalization is performed to obtain the normalized spatial coordinates. ;

[0063] S2.2: Based on the preset position encoding dimension pe_dim Dynamically generate multiple wavelength parameters with logarithmically spaced distribution. ;

[0064] S2.3: Calculate the vertical basis based on the vertical dimension coordinates in the normalized spatial coordinates. The vertical basis can be configured as a linear function. or exponential decay function linear functions Specific images such as Figure 2 As shown in (a), the value of this function varies with the normalization height. Linear growth, exponential decay function Specific images such as Figure 2 As shown in (b), where a>0, the value of the function varies with the normalization height. Both basis functions exhibit exponential decay to simulate the decay of atmospheric density with altitude; both basis functions are normalized at altitude. Visualization is performed within the range; optionally, the vertical basis can also be configured as a trigonometric function basis, which includes a vertical sine basis and a vertical cosine basis;

[0065] In addition, the following alternatives can be used for vertical bases to accommodate different atmospheric physical properties or computational needs:

[0066] In one embodiment, a polynomial function basis is employed: higher-order polynomial functions can be used to fit the nonlinear variations of atmospheric physical quantities with altitude, for example... Wherein, coefficient a i The parameters can be preset or learnable. This approach provides more flexible capabilities for modeling vertical structures.

[0067] In one embodiment, a piecewise function basis is employed: based on the different physical properties that the atmosphere may exhibit at different altitudes (such as the boundary layer, troposphere, and stratosphere), a piecewise function is designed as a vertical basis, and different function forms are used in different altitude ranges (e.g., linear growth at the bottom, gradual growth in the middle, and exponential decay at the top) to more finely characterize the vertical structure.

[0068] In one embodiment, wavelet basis functions are employed: wavelet basis functions such as Haar wavelets and Daubechies wavelets are introduced to encode vertical location information. Wavelets have good localization properties and multi-resolution analysis capabilities, making them more suitable for capturing abrupt changes or local features in the vertical structure of the atmosphere.

[0069] In one embodiment, a combination of radial basis functions (RBFs) is employed: a vertical basis is constructed using a linear combination of a set of radial basis functions (e.g., Gaussian functions) with adjustable center points and widths. .in, These parameters can be either fixed or learnable. This approach provides more powerful nonlinear fitting capabilities.

[0070] In one embodiment, a learned vertical embedding is employed: vertical base Instead of predefined mathematical functions, it directly extracts the normalized height from a small neural network (e.g., an MLP). z The embedding vector is learned from the data. This approach offers maximum flexibility, enabling fully adaptive capture of the atmospheric vertical structure.

[0071] S2.4: Based on the normalized spatial coordinates and multiple wavelength parameters Calculate the horizontal sine basis , and horizontal cosine basis , The specific images of the horizontal sine basis and the horizontal cosine basis are as follows: Figure 2 (c) and Figure 2 As shown in (d);

[0072] S2.5: The vertical basis, the horizontal sine basis, the horizontal cosine basis, and the optional vertical sine basis and vertical cosine basis are concatenated in the channel dimension to obtain the corresponding position code, forming the final ESPE vector.

[0073] Direct location encoding avoids interference with the numerical semantics of the original physical variables, ensuring physical consistency. A logarithmically spaced multi-scale sine / cosine basis is introduced on the normalized spatial coordinates, and especially in the vertical direction, a linearly or exponentially decaying vertical basis capable of capturing atmospheric physical laws is introduced. This encoding can accurately capture the multi-scale characteristics and altitude-varying vertical structure of the atmospheric field. Through multi-scale frequency injection and specially designed vertical bases (linear or exponential bases), it can more accurately capture the complex anisotropy and multi-scale physical structure of the atmospheric field, thus achieving an advantage in accuracy.

[0074] In typical radar and multivariate 3D atmospheric data prediction tasks (such as radar reflectivity prediction and wind field prediction), the method of this invention saves video memory and computing power, and uses indicators such as mean square error (MSE) and structural similarity (SSIM) for evaluation. The prediction accuracy can be maintained at the same level as or even better than that of the prior art.

[0075] S3: The original three-dimensional time-varying field data and the position code are fused in the channel dimension to obtain the fused data.

[0076] The original physical variable X is directly concatenated with the ESPE γ generated in S2 along the channel dimension. This direct concatenation method avoids the interference of traditional 1×1×1 convolutional mixing on the numerical semantics of the original physical variable, thus better maintaining physical consistency. The concatenated result is... Its dimensions become , where PE is the number of channels for position encoding.

[0077] Furthermore, feature fusion can also be performed in the following ways:

[0078] In one embodiment, a 1 × 1 × 1 convolution is used to fuse the original 3D time-varying field data and the position encoding vector as an alternative to the direct concatenation method. The original 3D time-varying field data and the position encoding vector are concatenated along the channel dimension and then input into a lightweight 1 × 1 × 1 3D convolutional layer. This convolutional layer learns the kernel parameters to achieve adaptive fusion of the two information, and outputs the fused feature vector for subsequent model calculations.

[0079] In one embodiment, Multilayer Perceptron (MLP) fusion is employed: ESPE can fuse the original 3D time-varying field data at the feature level through one or more MLP layers before inputting to the core temporal module, rather than simply concatenating channels. Introducing MLP layers provides stronger nonlinear expressive power for the feature fusion process.

[0080] S4: Perform an asymmetric spatial transformation operation on the fused data according to the preset asymmetric three-dimensional patch size to obtain the asymmetric spatial transformed data;

[0081] The asymmetric three-dimensional patch size differs from the other dimensions in at least one dimension.

[0082] The size of the asymmetric patch is: ,in, p z The size of the patch in the vertical dimension. p x The dimensions of the patch in the horizontal and vertical dimensions. py Let Patch be the size in the horizontal width dimension, and or .

[0083] spliced Performing an asymmetric spatial transformation operation, in one embodiment, is an asymmetric reshuffle. The operation addresses the anisotropic characteristics of the atmospheric field (the number of vertical layers D is much smaller than the number of horizontal layers H and W) through an asymmetric design. This asymmetric rearrangement involves adapting the fused data to an asymmetric three-dimensional patch size, adjusting the spatial dimensions... conduct Divide the data into multiple 3D data blocks by integer division, and then flatten these 3D data blocks to the channel dimension. Then, convert the input tensor into... The voxel blocks of size are rearranged losslessly to the channel dimension, thus significantly increasing the number of channels while simultaneously reducing the spatial dimension. The rearranged result is Z, whose number of channels becomes... The spatial dimension is reduced to , , In one embodiment, the size of the asymmetric patch is... The horizontal height is the same as the horizontal width.

[0084] Furthermore, asymmetric spatial transformations can also be implemented through the following alternative schemes:

[0085] In one embodiment, asymmetric pooling and upsampling are employed. This operation is irreversible, but in scenarios where computational cost is extremely critical and some information loss is acceptable, asymmetric max pooling, average pooling, or sparse convolution can be used to reduce spatial resolution. Correspondingly, decoding requires asymmetric upsampling operations (e.g., deconvolution or interpolation). This approach can simplify the model in certain tasks.

[0086] In one embodiment, multiple patch shapes are recombined using channel dimensions: Besides You can choose other asymmetric patch shapes, such as the patch size in the horizontal direction. They can be different to accommodate certain horizontal grids with stronger anisotropy.

[0087] In one embodiment, a compression method based on Fourier / wavelet transform is employed: the three-dimensional voxel block undergoes Fourier or wavelet transform in the spatial domain, and coefficients are truncated in the transform domain, retaining only low-frequency components to achieve data compression. This method can naturally capture the multi-scale structural information of the signal.

[0088] The asymmetric patch design fully considers the significant differences between the horizontal and vertical dimensions of the atmospheric field, avoiding the physical inconsistencies and irrationalities that may occur with traditional uniform patches. The exponential basis introduced in the vertical direction directly draws upon atmospheric physical laws, enhancing the model's physical rationality and interpretability to a certain extent, making the modeling of the atmospheric vertical profile more accurate. This enhances the physical friendliness and interpretability of the embodiments of this invention.

[0089] S5: Input the data after the asymmetric spatial transformation into a pre-trained deep learning model for feature extraction and prediction to obtain predicted data. The deep learning model takes 10 frames of past three-dimensional time-varying field data as input, performs time-series feature extraction and prediction tasks through the deep learning model of the core time-series processing module (e.g., Transformer model or 3D recurrent neural network), and outputs 10 frames of future three-dimensional time-varying field data.

[0090] The asymmetric spatial transformation operation includes asymmetric rearrangement, which specifically includes:

[0091] Based on the asymmetric 3D patch size, the spatial dimension of the feature-stitched data is divided by integer division to obtain multiple 3D data blocks; wherein, the feature-stitched data is obtained through S3, specifically, the original 3D time-varying field data and the position code are feature-stitched in the channel dimension to obtain the feature-stitched data;

[0092] The three-dimensional data block is flattened to the channel dimension to obtain the data after asymmetric rearrangement.

[0093] The aforementioned asymmetric reversible rearrangement, Earth-specific multi-scale location encoding, and direct connection splicing strategy are organically combined and applied to a high-resolution time-varying three-dimensional atmospheric field deep learning model to maintain or even improve prediction accuracy while significantly reducing memory and computational overhead.

[0094] Furthermore, in one embodiment, in S5, the channel dimension of the data after the asymmetric spatial transformation is reduced to obtain compressed data. The compressed data The data is input into a deep learning model for feature extraction and prediction to obtain the predicted data. .

[0095] In one embodiment, the dimensionality reduction process includes bottleneck compression of the channel dimension of the data after the asymmetric spatial transformation, in order to compress the number of channels and obtain compressed data. .

[0096] In one embodiment, a reversible dimensionality transformation operation is used to reduce the channel dimensions of the data after the asymmetric spatial transformation. To further reduce the computational cost of the subsequent time-series core module, a 1 × 1 × 1 three-dimensional convolutional layer can be used after the asymmetric spatial transformation to reduce the high-dimensional channels. Compressed to lower dimensions The above steps are optional and can be omitted in scenarios where high accuracy is required and increased computational overhead is permissible.

[0097] Furthermore, bottleneck compression can also be performed in the following ways:

[0098] In one embodiment, linear projection is employed: linear dimensionality reduction is performed using a simple fully connected layer (or an equivalent 1 × 1 × 1 convolution), transforming... Mapped to .

[0099] In one embodiment, an attention mechanism is used for compression: a channel attention mechanism (e.g., a Squeeze-and-Excitation module) is introduced to learn the importance weights of different channels, and the channels are weighted or selectively compressed based on the importance weights to retain more important information.

[0100] In one embodiment, a sparsity compression operation is employed: channel dimensionality reduction is achieved by learning sparse matrix multiplication, thereby theoretically saving computational resources.

[0101] Through the above operations, the method of this embodiment of the invention can achieve significant savings in memory and computational overhead, and shrink the space lattice through asymmetric reversible rearrangement. Times. For example, in In a typical configuration, the spatial lattice is reduced by a factor of 32. Combined with channel bottleneck compression, this achieves approximately 5-15 times savings in GPU memory and computing power. This significantly lowers the hardware barrier for single-card training and inference of high-resolution atmospheric models, making it possible to process more complex models and higher-resolution data with limited computing resources. Furthermore, the position encoding and compression methods are universal and can be seamlessly embedded as general modules into various deep learning-based high-resolution time-varying 3D atmospheric field modeling tasks, such as short-term weather forecasting, climate simulation, and radar data analysis. It has strong versatility and broad application prospects.

[0102] Following the core time-series processing steps, an inverse spatial transformation operation is performed on the predicted data. In one embodiment, the inverse spatial transformation operation is an inverse reshuffle. This restores the spatial dimension of the predicted data to the spatial dimension of the original three-dimensional time-varying field data, resulting in data after the inverse asymmetric spatial transformation. The core timing processing module outputs... Then, perform the inverse transformation of the asymmetric rearrangement operation. This reorganizes the rearranged information in the channel dimension back into the spatial dimension, restoring the original spatial resolution of the data.

[0103] The spatial-to-channel rearrangement is a pure element-wise rearrangement. By losslessly rearranging three-dimensional spatial voxels to the channel dimension without any information loss, it achieves significant savings in video memory and computational overhead for high-resolution time-varying three-dimensional atmospheric field data. The existence of the inverse transform ensures that the original spatial structure can be completely recovered when the model is output, ensuring the complete recovery of information during compression and decompression, and guaranteeing the reliability of the entire process and the accuracy of the prediction results.

[0104] After the inverse spatial transformation operation, a strip position encoding (Strip PE) operation is performed. This involves channel processing of the data after the inverse asymmetric spatial transformation, removing the position encoding channels to obtain the final prediction result. The output after the inverse spatial transformation. The channel dimension may include additional dimensions introduced by positional encoding. This embodiment of the invention precisely removes the positionally encoded channels by pruning the front part of the channel dimension, thereby restoring the original target output channel number C and obtaining the final prediction result. .

[0105] like Figure 3 As shown, the method of this embodiment of the invention mainly consists of the following functional modules or components.

[0106] Input module: Receives raw, high-resolution, time-varying three-dimensional atmospheric field data. .

[0107] Earth-Specific Multiscale Location Coding Module: This module is used to generate location codes γ for atmospheric field data that capture its multiscale characteristics and anisotropy (especially vertical physical properties). It consists of the following units:

[0108] Coordinate normalization unit: used to normalize the original spatial coordinates Convert to normalized coordinates .

[0109] Wavelength generation unit: used to generate wavelengths based on wavelengths of ... pe_dim Dynamically generate multiple wavelength parameters with logarithmic intervals .

[0110] Vertical basis generation unit: used to calculate vertical basis It can be configured as a linear function. and exponential decay function .

[0111] Horizontal and vertical sine / cosine basis generation units: calculated based on normalized coordinates and wavelength. , and , .

[0112] Encoding and splicing unit: The vertical basis, horizontal sine / cosine basis, and optional vertical sine / cosine basis are spliced ​​together in the channel dimension to form the final ESPE vector.

[0113] The ESPE module runs independently of the main data processing path, and its output is fused with the original data in subsequent steps, while the main path leads directly to the next module, retaining the original input. .

[0114] Direct fusion module: Located between the input module and the asymmetric spatial transformation module, it directly fuses the channels of the original input data X with the channels of ESPE γ to form augmented data containing location information. This module includes a channel dimension splicing operator.

[0115] The asymmetric spatial transformation module is located between the direct-connect fusion module and the bottleneck compression module (or the core timing processing module), and is used to adjust the size of the asymmetric patch according to a preset value. Input data The spatial dimensional information is losslessly rearranged to the channel dimension while reducing spatial resolution. Its core objective is to achieve efficient asymmetric data compression. This module is implemented using pixel unshuffle or similar data recombination logic, requiring that D, H, and W can be... Divisibility: When a dimension is not divisible, zero-padding is used to make the padded dimension divisible. The specific calculation steps are as follows:

[0116] The data is zero-padded, and the padding amount for each dimension is calculated. Indicates rounding up:

[0117] ,

[0118] ,

[0119] ,

[0120] And fill in the corresponding number of zero-value elements at the dimension boundaries of the data, and update the dimension to [value]. , as well as This allows the updated D, H, W to be... Divisible by.

[0121] Bottleneck compression module (optional) Located between the asymmetric spatial transformation module and the core temporal processing module, when this module is activated, it uses a 1 × 1 × 1 three-dimensional convolutional layer to transform the high-dimensional channels after spatial transformation. Compressed to lower dimensions This saves subsequent computing resources.

[0122] The core temporal processing module, TemporalCore, is located between the bottleneck compression module (or asymmetric spatial transformation module) and the inverse spatial transformation module. It performs the core computational logic of the deep learning model, such as feature extraction, temporal modeling, and prediction. It can be implemented using any deep learning architecture suitable for 3D temporal data, such as 3D Transformer, 3DResNet, or a structure combined with recurrent neural networks.

[0123] Inverse space transformation module Located between the core timing processing module and the stripping location encoding module, this function acts as the inverse operation of the asymmetric spatial transformation module. It reorganizes the channel dimension information output by the core timing processing module back into the spatial dimension, restoring the original spatial resolution of the data. This is achieved through pixel shuffle or similar data recombination logic.

[0124] The location encoding stripping module, located after the inverse spatial transformation module, is the output of the entire process. It removes the extra channels introduced by location encoding from the data recovered from the inverse spatial transformation by pruning the beginning of the channel dimensions, thus obtaining the final prediction result containing only physical variables. This is implemented by the channel pruning operator.

[0125] The forward propagation process of the AAPE method is as follows: First, atmospheric data is input. Then, through the Earth-specific location encoding module, 2D horizontal and 3D codes are generated. The ESPE is combined with the input data X through a direct-connect fusion module to form... Dimension becomes . Then it enters the asymmetric space transformation module, and the number of channels becomes... The spatial dimension is reduced to Then, bottleneck compression can be selectively performed using 1 × 1 × 1 convolutions to reduce the number of channels from... Down to The compressed data is input into the core time-series processing module (e.g., Transformer). The processed output undergoes inverse spatial transformation to restore the spatial dimension, and finally, the positional encoding is removed to obtain the final predicted atmospheric data Y, with the dimension restored to [value missing]. .

[0126] Figure 4This is a precipitation forecast performance evaluation index diagram obtained by the three-dimensional time-varying field prediction method for radar detection according to the present invention. The horizontal axis represents the number of forecast frames generated at 6-minute intervals within one hour (a total of 10 frames), and the vertical axis represents the value of the corresponding scoring index. Figure 4 (a) in the figure is a graph showing the change of the CSI (Critical Success Index) with forecast lead time; Figure 4 (b) in the figure is a curve showing the change of the POD (Probability of Detection) index with the forecast lead time; Figure 4 (c) in the figure is a curve of the FAR (False Alarm Ratio) index changing with the forecast lead time; Figure 4 (d) in the figure is a graph showing the HSS (Heidke Skill Score) index as a function of the number of forecast frames; Figure 4 (e) in the figure is a curve showing the change of the BIAS (Bias Score) index with the forecast lead time; Figure 4 In the diagram, (f) represents the radar chart of the time-mean CSI, POD, FAR, HSS, and BIAS indicators. Using five meteorological forecasting evaluation indicators—CSI, POD, FAR, HSS, and BIAS—combined with time-series changes and radar mean radar charts, the predictive performance of the model under different thresholds (thr=20 / 35) was quantified. The method of this embodiment demonstrates excellent overall prediction accuracy and target detection capability, with a low false alarm rate and a high degree of matching between predicted and actual echo quantities. Table 1 shows the quantitative results of existing mainstream models on radar echo datasets; relevant content can be found in the published literature "Wang Y, Wu H, Zhang J, Gao Z, Wang J, Yu PS, Long M, 2022. PredRNN: A RecurrentNeuralNetwork for Spatiotemporal Predictive Learning".

[0127] Table 1 Quantitative results of radar echo dataset

[0128]

[0129] Existing methods suffer from problems such as blurred echo edges and loss of small-scale edge echoes. However, the extrapolation results obtained by the method of this invention show that the main body of the echo is distributed in a regional cluster, with strong core echoes concentrated in the upper and middle parts of the detection area, and the edges are transitioned with medium and low intensity echoes. The method of this invention can accurately capture the edge contour features of low-threshold echoes, completely preserve the small echo clusters in the edge area, and retain more echo details. It has a better capture and preservation effect on small-scale echoes, providing information support for echo generation and dissipation inference.

[0130] In terms of objective testing indicators, the CSI score of the method in this embodiment is higher than that of the prior art. The echo prediction CSI of the method in this embodiment is about 0.5 (threshold is 35dBZ). Even compared with the CSI value of the prior art at the threshold of 30dBZ (about 0.3), the performance of the method in this embodiment still has a significant advantage, providing quantitative indicator support for the prediction effectiveness of the method.

[0131] The method of this invention provides an asymmetric rearrangement with the following advantages in terms of physical rationality and reversibility: Traditional patching methods typically involve uniform segmentation, failing to consider the strong anisotropy of the atmospheric field (the significant difference between vertical and horizontal dimensions). This invention proposes an asymmetric patch and designs a lossless, reversible rearrangement operation from space to channel. It can achieve asymmetric dimensional transformation while ensuring information integrity, and guarantees that the inverse operation can accurately restore the original spatial structure. Simultaneously, this design must physically conform to the anisotropic characteristics of the atmospheric field to avoid introducing unreasonable physical artifacts. This requires a deep understanding of the mathematical principles of data rearrangement and the physical properties of the atmosphere.

[0132] The refined design of Earth-specific multi-scale location coding provided by the method of this invention has the following advantages: Existing location coding (such as sinusoidal coding and NeRF coding) lacks consideration of the specific physical laws of the atmospheric field. This invention not only introduces logarithmically spaced wavelengths covering the multi-scale spectrum, but more importantly, it introduces a linear or exponential attenuation basis in the vertical direction. The technical challenge of this invention lies in how to select and combine these basis functions so that they can effectively encode the positional information of discrete grids while accurately capturing the physical characteristics of atmospheric density, temperature, and other physical quantities that decay exponentially with altitude, thereby improving the model's ability to understand vertical structures. Simultaneously, the design of direct-connection splicing positional encoding avoids numerical semantic interference that may result from convolutional mixing, representing an improvement and breakthrough over existing positional encoding mechanisms.

[0133] The method in this invention also achieves synergistic optimization and balance of the overall process: integrating ESPE, asymmetric reversible rearrangement, and optional bottleneck compression into a unified framework, achieving an optimal balance between saving memory and computational overhead, prediction accuracy, and physical rationality. For example, while bottleneck compression can save resources, it is prone to information loss; how to weigh its applicable scenarios and design a strategy to remove positional encoding to ensure the purity of the final output are all technical challenges that require careful design and verification in this invention.

[0134] References:

[0135] Liu Z, Ning J, Cao Y, Wei Y, Zhang Z, Lin S,&Hu H (2021). Video Swin Transformer (No. arXiv:2106.13230).

[0136] Mildenhall B, Srinivasan PP, Tancik M, Barron JT, Ramamoorthi R,&Ng R(2022). NeRF: representing scenes as neural radiance fields for viewsynthesis. Communications of the ACM , 65 (1), 99–106.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] The above embodiments should be understood as being used only to illustrate the present invention more clearly, and not to limit the scope of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. A three-dimensional time-varying field prediction method for radar detection, characterized in that, Includes the following steps: S1: Acquire the raw three-dimensional time-varying field data obtained from radar detection; S2: Generate the corresponding position code based on the spatial dimension of the input original three-dimensional time-varying field data; S3: The original three-dimensional time-varying field data and the position code are fused in the channel dimension to obtain the fused data; S4: Perform an asymmetric spatial transformation operation on the fused data according to the preset asymmetric three-dimensional patch size to obtain asymmetric spatial transformed data; the asymmetric three-dimensional patch size is different from other dimensions in at least one dimension; S5: Input the data after the asymmetric spatial transformation into a pre-trained deep learning model to obtain prediction data; wherein, the pre-trained deep learning model is trained based on historical three-dimensional time-varying field data; The asymmetric spatial transformation operation includes asymmetric rearrangement, which specifically includes: Based on the asymmetric 3D Patch size, the spatial dimension of the feature-stitched data is divided by integer division to obtain multiple 3D data blocks; The three-dimensional data block is flattened to the channel dimension to obtain the data after asymmetric rearrangement.

2. The three-dimensional time-varying field prediction method for radar detection according to claim 1, characterized in that, In S2, generating the corresponding position code includes: S2.1: Normalize the original spatial coordinates of the input raw three-dimensional time-varying field data to obtain normalized spatial coordinates; S2.2: Generate multiple wavelength parameters with logarithmic interval distribution based on the preset position encoding dimension; S2.3: Calculate the vertical basis based on the vertical dimension coordinates in the normalized spatial coordinates; S2.4: Based on the normalized spatial coordinates and multiple wavelength parameters, calculate the horizontal sine basis and the horizontal cosine basis; S2.5: The vertical basis, the horizontal sine basis, and the horizontal cosine basis are concatenated along the channel dimension to obtain the corresponding position code.

3. The three-dimensional time-varying field prediction method for radar detection according to claim 1, characterized in that, In S5: The channel dimension of the data after the asymmetric spatial transformation is reduced to obtain compressed data; The compressed data is input into a pre-trained deep learning model to obtain prediction data.

4. The three-dimensional time-varying field prediction method for radar detection according to claim 3, characterized in that, The dimensionality reduction process employs either reversible or irreversible dimensionality transformation operations.

5. The three-dimensional time-varying field prediction method for radar detection according to claim 1, characterized in that, The original three-dimensional time-varying field data is a multidimensional tensor with dimensions of . Where B is the batch size, T is the time step, C is the number of channels, D is the number of vertical layers, H is the horizontal height, W is the horizontal width, and the size of the spatial dimension is... ; The size of the asymmetric patch is: ,in, p z The size of the patch in the vertical dimension. p x The dimensions of the patch in the horizontal and vertical dimensions. p y Let Patch be the size in the horizontal width dimension, and or .

6. The three-dimensional time-varying field prediction method for radar detection according to claim 5, characterized in that, The feature fusion is performed as feature concatenation, and the dimension of the concatenated data is [dimension value missing]. In S4, the asymmetric spatial transformation operation is an asymmetric rearrangement, which specifically includes the following steps: Based on the asymmetric patch size Regarding spatial dimensions conduct Integer division; resulting in multiple 3D data blocks; The three-dimensional data block is flattened to the channel dimension to obtain data after asymmetric spatial transformation. The dimension of the data after asymmetric spatial transformation is... ; Among them, the number of channels after rearrangement Vertical height after rearrangement The horizontal height after rearrangement The horizontal width after rearrangement .

7. The three-dimensional time-varying field prediction method for radar detection according to claim 6, characterized in that, When the spatial dimension Cannot be During integer division, zero-value padding is performed on the spatial dimension data, specifically including the following steps: Calculate the fill amount for the spatial dimension, including the horizontal height fill amount. Horizontal width fill amount and vertical fill volume ;in, Indicates rounding up; The corresponding boundary of the spatial dimension is supplemented with zero values ​​corresponding to the filling amount; Update the spatial dimension to , as well as .

8. The three-dimensional time-varying field prediction method for radar detection according to claim 1, characterized in that, In S4, the asymmetric spatial transformation operation is an asymmetric downsampling operation.

9. The three-dimensional time-varying field prediction method for radar detection according to claim 1 or 3, characterized in that, After obtaining the prediction data, an inverse asymmetric spatial transformation operation is performed on the prediction data to restore the spatial dimension of the prediction data to the spatial dimension of the original three-dimensional time-varying field data, thus obtaining the data after the inverse asymmetric spatial transformation.

10. The three-dimensional time-varying field prediction method for radar detection according to claim 9, characterized in that, After obtaining the data after the inverse asymmetric spatial transformation, channel processing is performed on the transformed data to remove the position encoding channels, thus obtaining the final prediction result.

Images