A radar echo nowcasting method and system based on wavelet domain double attention

Abstract

The present application relates to the technical field of radar meteorological data processing, and particularly relates to a radar echo nowcasting method and system based on wavelet domain double attention. After obtaining a historical radar echo grid point sequence, one-dimensional two-dimensional discrete wavelet transform is performed on each frame to obtain a wavelet domain historical sequence containing a low-frequency approximation subband and multiple high-frequency detail subbands; a double attention encoder is used to extract high-resolution historical encoding features and low-resolution historical encoding features; for each future prediction time, based on the current decoding state and future time representation, multi-scale historical features are retrieved and fused through local spatial mixing and temporal cross-attention at each level to output a wavelet domain prediction result; and a future radar echo grid point sequence is obtained through inverse two-dimensional discrete wavelet transform. The present application can separate different scale echo structures, adaptively call multi-scale historical information, and improve the prediction performance of high-intensity echo events, and can be used for severe convective weather monitoring and short-time precipitation warning.

Description

Technical Field

[0001] This invention relates to the field of radar meteorological data processing technology, specifically to a radar echo nowcasting method and system based on wavelet domain dual attention. Background Technology

[0002] Radar echo nowcasting is a technique that infers changes in echo intensity and spatial distribution over a short future period based on gridded radar echo images from continuous historical timeframes. This technology can be applied to scenarios such as severe convective weather monitoring, short-term precipitation warnings, urban flooding risk alerts, and refined meteorological services.

[0003] Existing deep learning-based radar echo nowcasting methods typically perform convolutional coding, recurrent prediction, or spatiotemporal attention modeling directly on historical radar images in the pixel domain. For example, methods based on recurrent neural networks such as ConvLSTM and PredRNN capture the temporal evolution of echoes by constructing temporal memory units; methods based on attention mechanisms such as Earthformer and Rainformer model spatiotemporal dependencies through global or local attention mechanisms. These methods can learn the patterns of echo movement and intensity changes to some extent, but they are still prone to problems such as blurred prediction boundaries, attenuation of local extreme values, and missed detection of high-intensity events when dealing with complex evolution processes such as strong local echo cores, abrupt boundary changes, cell merging, and rapid enhancement or dissipation.

[0004] The main reasons for the above problems are as follows: First, radar echo fields contain both large-scale organizational structures and local detailed changes. When modeling uniformly in the pixel domain, information at different scales is coupled, and the model tends to prioritize fitting weak echoes and background areas, which account for a higher proportion, while neglecting the characterization of high-intensity local structures. Second, existing decoding methods typically rely on fixed historical frames, the last-moment encoding state, or fixed skip connections to generate future prediction results. This makes it difficult to adaptively call historical information with different temporal correlations for different future prediction times and spatial locations, leading to increased cumulative errors in multi-step predictions and making it difficult to accurately capture the evolution trend of high-intensity echoes. Third, strong echo events account for a limited proportion of the training samples, and uniform pixel errors (such as mean square error) tend to bias the training process towards low-intensity regions, affecting the model's ability to recognize medium- and high-intensity echo events. Summary of the Invention

[0005] The purpose of this invention is to propose a radar echo nowcasting method and system based on wavelet domain dual attention, which can separate echo structures at different scales, jointly model temporal evolution relationships and local spatial relationships, and adaptively retrieve multi-scale historical information according to the current prediction state.

[0006] According to a first aspect of the present disclosure, a radar echo proximity prediction method based on wavelet domain dual attention is provided, comprising the following steps: Obtain a historical radar echo grid sequence consisting of radar echo grid images from multiple historical moments; A first-order two-dimensional discrete wavelet transform is performed frame by frame on each radar echo grid image in the historical radar echo grid sequence to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. The wavelet domain history sequence is input into a dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features. For each future prediction time, low-resolution query features are generated based on the current decoding state and the corresponding future time representation. Low-resolution historical encoding features after local spatial mixing are used as keys and values, and low-resolution retrieval features are obtained through temporal cross attention. The low-resolution retrieval features are reorganized and spatially upsampled to obtain high-resolution query features. High-resolution historical encoding features after local spatial mixing are used as keys and values, and high-resolution retrieval features are obtained through temporal cross attention. By fusing the high-resolution query features and the high-resolution retrieval features, the wavelet domain prediction result for the current future prediction time is obtained; The current decoding state is updated based on the low-resolution retrieval features, and the prediction processing corresponding to each future prediction time is repeatedly executed; and Perform an inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results for each future prediction time to obtain the future radar echo grid sequence.

[0007] In one embodiment, the first-level two-dimensional discrete wavelet transform is a first-level Haar discrete wavelet transform, the low-frequency approximation subband is an LL subband, and the plurality of high-frequency detail subbands include LH subband, HL subband and HH subband; The wavelet domain prediction result for the current future prediction time is generated by independent linear output heads, which include LL subband output head, LH subband output head, HL subband output head and HH subband output head.

[0008] In one embodiment, the dual attention encoder includes a high-resolution feature encoding branch and a low-resolution feature encoding branch; the high-resolution feature encoding branch performs temporal self-attention calculation along the time dimension at each fixed spatial location, and performs spatial self-attention calculation within a local spatial window at each historical moment to obtain the high-resolution historical encoded features; the low-resolution feature encoding branch obtains the high-resolution historical encoded features by downsampling the high-resolution historical encoded features only along the spatial dimension, and performs temporal self-attention calculation and local window spatial self-attention calculation on the downsampled features to obtain the low-resolution historical encoded features.

[0009] In one embodiment, the local spatial mixing includes: performing deep convolution on the historical encoded features at each historical moment at the corresponding resolution to fuse the historical information of the current spatial location and its adjacent spatial locations; the temporal cross-attention implementation includes: at each fixed spatial location, arranging the historical encoded features after local spatial mixing into key sequences and value sequences along the historical time dimension, using the corresponding query features as query sequences, and performing multi-head cross-attention calculation to obtain intermediate attention features: in, express Depthwise convolution, and These represent query mapping, key mapping, and value mapping, respectively. This indicates cross-attention calculation; After performing a residual connection between the intermediate attention features and the query features, the corresponding retrieval features are obtained through a feedforward network.

[0010] In one embodiment, the low-resolution retrieval features are reorganized by dual attention blocks and then spatially upsampled using nearest neighbor spatial interpolation and two-dimensional convolutional mapping; the high-resolution query features and the high-resolution retrieval features are concatenated in the channel dimension and then fused by dual attention blocks.

[0011] In one embodiment, when training a prediction model for performing the method, a total training loss is calculated based on the predicted radar echo grid sequence and the target radar echo grid sequence of the training samples. The total training loss includes a basic reconstruction loss and a multi-threshold event loss. The basic reconstruction loss includes an intensity-adaptive pixel loss and a neighborhood consistency loss. The multi-threshold event loss is obtained by weighting event losses corresponding to multiple echo intensity thresholds. The parameters of the prediction model are updated according to the total training loss.

[0012] In one embodiment, the intensity-adaptive pixel loss is defined as: in, This represents the total number of spatiotemporal pixels involved in the loss calculation. Indicates the first Predicted values ​​at each location Indicates the first The target value at each location. This is the intensity weighting coefficient; The neighborhood consistency loss is defined as: in, This represents the predicted radar echo grid sequence. Represents the target radar echo grid sequence. Indicates the pooling window is Local average pooling operation; The multi-threshold event loss is composed of multiple echo intensity thresholds. The corresponding Focal event loss is obtained by weighting the losses. For the m-th threshold, the event loss is: in, Represents the class balance parameter. Indicates the focus parameter; From the target value Relative to threshold Generated binary event labels: and , The temperature parameter is the value at the corresponding threshold. The multi-threshold event loss is: in, This indicates the number of echo intensity thresholds. Indicates the first The event loss weight corresponding to each echo intensity threshold Indicates the first Each echo intensity threshold corresponds to Losses from the incident.

[0013] According to a second aspect of the present disclosure, a radar echo proximity prediction system based on wavelet domain dual attention is provided, comprising: The data acquisition module acquires a sequence of historical radar echo grid points, which consists of radar echo grid point images from multiple historical moments. The wavelet transform module performs a first-level two-dimensional discrete wavelet transform on each radar echo grid image in the historical radar echo grid sequence frame by frame to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. The dual attention coding module inputs the wavelet domain history sequence into the dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features; The low-resolution dynamic historical retrieval module generates low-resolution query features based on the current decoding state and the corresponding future time representation for each future prediction time. It uses the low-resolution historical encoding features after local spatial mixing as keys and values, and obtains low-resolution retrieval features through temporal cross attention. The high-resolution dynamic historical retrieval module performs feature recombination and spatial upsampling on the low-resolution retrieval features to obtain high-resolution query features. The high-resolution historical encoding features after local spatial mixing are used as keys and values, and high-resolution retrieval features are obtained through temporal cross attention. The feature fusion and prediction module fuses the high-resolution query features and the high-resolution retrieval features to obtain the wavelet domain prediction result for the current future prediction time. The decoding state update module updates the current decoding state based on the low-resolution retrieval features and repeatedly executes the prediction processing corresponding to each future prediction time. The inverse wavelet reconstruction module performs an inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results at each future prediction time to obtain the future radar echo grid sequence.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned radar echo proximity prediction method based on wavelet domain dual attention.

[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned radar echo proximity prediction method based on wavelet domain dual attention.

[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. By separating low-frequency organizational structures and high-frequency local details through frame-by-frame two-dimensional discrete wavelet transform, the model can perform spatiotemporal predictions in a feature space with explicit scale separation. This transform decomposes the radar echo field into LL subbands and LH, HL, and HH subbands, where the LL subband retains the main outline of the large-scale echo, and the high-frequency subbands represent local gradients and boundary changes in different directions. Compared to directly modeling uniformly in the pixel domain, this invention explicitly decouples physical features at different scales, enabling the model to learn large-scale evolution trends and abrupt changes in local structures separately. This enhances the model's ability to represent strong echo boundaries and local structural changes, and alleviates the problems of boundary ambiguity and extreme value attenuation in the prediction results.

[0017] 2. This design models the temporal evolution dependencies and local spatial structure relationships at fixed spatial locations using both temporal self-attention and local window spatial self-attention. Temporal self-attention extracts the evolutionary relationships between different historical moments along the time dimension, while local window spatial self-attention independently extracts the spatial structure relationships within a local area at each historical moment. This design reduces the computational complexity of performing global attention calculations on all spatiotemporal locations while specifically enhancing the modeling capabilities for key structures such as local echo cores and abrupt boundary changes.

[0018] 3. By employing a temporal cross-attention retrieval mechanism that uses the current decoding state as the query and historical features after local spatial mixing as the key and value, historical information at different time steps and scales can be adaptively invoked at different future prediction moments. This mechanism dynamically retrieves the most relevant spatiotemporal patterns from historical encoded features at each prediction step based on the current decoding state, rather than relying on fixed historical frames or static skip connections. This improves the utilization of historical information in multi-step predictions and reduces accumulated errors.

[0019] 4. By combining intensity-adaptive pixel loss, neighborhood consistency loss, and multi-threshold event loss, the training process assigns greater regression weight to high-intensity echo regions. At the same time, the model's ability to identify medium-to-high intensity events is enhanced by Focal event loss corresponding to multiple echo intensity thresholds, effectively improving the prediction performance of the prediction model in strong echo events. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0021] Figure 1 This is a schematic diagram of the overall architecture of the radar echo nowcasting model based on wavelet domain dual attention in this invention.

[0022] Figure 2This is a schematic diagram of the structure of the dual attention encoder and dynamic history retrieval decoder of the present invention.

[0023] Figure 3 This is a schematic diagram of the dual attention module of the present invention.

[0024] Figure 4 This is a schematic diagram of the spatial location-time self-attention module of the present invention.

[0025] Figure 5 This is a schematic diagram of the structure of the frame-by-frame local window space self-attention module of the present invention.

[0026] Figure 6 This is a schematic diagram of the structure of the dynamic history retrieval module of the present invention. Detailed Implementation

[0027] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0030] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0031] Example 1: like Figure 1 As shown, this embodiment provides a radar echo nowcasting method based on wavelet domain dual attention, including the following steps: S1. Obtain a historical radar echo grid sequence consisting of radar echo grid images from multiple historical moments; In this embodiment, the historical input sequence is assumed to be... The future target sequence is ,in, include Radar echo grid images of historical moments, include Radar echo grid images for each future predicted time. The spatial size of each radar echo grid image is [missing information]. The number of channels is In one example, It can be set to 7. It can be set to 6. and It can be set to 128. It can be set to 1. It should be noted that the above frame count, spatial size, and number of channels are merely examples and do not constitute a limitation on the present invention.

[0032] S2. Perform a first-level two-dimensional discrete wavelet transform on each radar echo grid image in the historical radar echo grid sequence frame by frame to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. In this embodiment, the historical input sequence Each radar echo grid image in the dataset undergoes a first-order two-dimensional discrete wavelet transform frame by frame. For any historical moment... Radar echo grid image After transformation, LL subband, LH subband, HL subband and HH subband are obtained, and the four subbands are spliced ​​along the channel dimension to obtain wavelet domain features. The wavelet domain features corresponding to all historical moments constitute the wavelet domain history sequence. This embodiment employs a first-order Haar discrete wavelet transform. After the first-order Haar discrete wavelet transform, the spatial size is reduced from... Become The number of channels is determined by Expand to The LL subband preserves the main outline and large-scale organization information of the echo field, while the LH, HL, and HH subbands are used to characterize local gradients, boundary variations, and texture undulations in different directions. The high-frequency subband does not directly correspond to high-intensity echo event labels, but is used to enhance the expression of local structural changes.

[0033] S3. Input the wavelet domain history sequence into a dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features; like Figure 2 As shown, the dual attention encoder and dynamic history retrieval decoder include a high-resolution coding stage, a low-resolution coding stage, a low-resolution dynamic history retrieval stage, and a high-resolution dynamic history retrieval stage. The wavelet domain history sequence W is input into the dual attention encoder to extract historical coding features at different spatial scales.

[0034] like Figure 3 As shown, the dual attention module includes a temporal self-attention unit for each spatial location and a frame-by-frame local window spatial self-attention unit connected in sequence. The input features are processed sequentially by the temporal self-attention unit for each spatial location and the frame-by-frame local window spatial self-attention unit to obtain spatiotemporally enhanced features. In the high-resolution coding stage, the wavelet domain input is first mapped to the basic dimensional features through a linear mapping, and spatial location representation is added; then, temporal self-attention calculation, local window spatial self-attention calculation, and feedforward network calculation are performed sequentially to obtain high-resolution historical encoded features. .

[0035] like Figure 4 As shown, temporal self-attention calculation is performed independently at each fixed spatial location. Specifically, for the same spatial location... These historical features, forming a length of [length] along the time dimension, are [characteristics]. The feature sequence is used, and after adding the temporal location representation, multi-head self-attention calculation is performed to extract the evolutionary correlation between the spatial location at different historical moments.

[0036] like Figure 5 As shown, the local window spatial self-attention computation is performed independently at each historical time step. The spatial features of that historical time step are divided into multiple local windows, and multi-head self-attention computation is performed within each local window to extract the spatial structural relationships within the local area. The size of the local window can be set to... , Or other dimensions that are divisible by the feature map space size; in one example, the local window size is set to... In obtaining high-resolution historical coding features Then, downsampling is performed only along the spatial dimension without changing the temporal dimension to obtain low-resolution features. After adding spatial location representations to the low-resolution features, temporal self-attention calculation and local window spatial self-attention calculation are repeated to obtain low-resolution historical encoded features. .

[0037] S4. For each future prediction time, generate low-resolution query features based on the current decoding state and the corresponding future time representation. Use the low-resolution historical coding features after local spatial mixing as keys and values, and obtain low-resolution retrieval features through temporal cross attention. like Figure 6 As shown, the dynamic history retrieval module uses the current decoding state and future time representation to form query features, and uses the historical encoding features processed by deep convolution to form keys and values, and performs cross-attention calculation along the time dimension at a fixed spatial location. The initial decoding state consists of low-resolution historical encoding features. The characteristics of the last historical moment are determined, and represented as in, This represents the last moment of the historical input sequence. For the first... The first future prediction moment will be compared with the current decoding state. Adding the future time representations together yields the low-resolution query features: in, Indicates the first The low-resolution decoding state corresponding to each future predicted time. Indicates the first The future time representation corresponding to each future prediction moment. Encoding low-resolution historical features. Input the characteristics of each historical moment respectively Deep convolution is used to obtain low-resolution historical features containing local information about neighboring spatial locations. Subsequently, at each fixed spatial location, the low-resolution historical features are arranged along the time dimension into a key sequence and a value sequence, and the low-resolution query features are then... As a query sequence, perform multi-head cross-attention computation to obtain intermediate attention features: in, express Depthwise convolution, and These represent query mapping, key mapping, and value mapping, respectively. This represents cross-attention computation. The intermediate attention features are residually concatenated with the low-resolution query features, and then processed through layer normalization and a feedforward network to obtain the first... Low-resolution retrieval features corresponding to each future prediction time: in, Indicates the normalization layer. This represents a feedforward network.

[0038] S5. Perform feature recombination and spatial upsampling on the low-resolution retrieval features to obtain high-resolution query features. Use the high-resolution historical coding features after local spatial mixing as keys and values, and obtain high-resolution retrieval features through temporal cross attention. Specifically, low-resolution retrieval features After feature reorganization using dual attention blocks, spatial upsampling is performed through nearest neighbor spatial interpolation and two-dimensional convolutional mapping to obtain high-resolution query features: in, This indicates a dual attention module. This represents nearest neighbor spatial interpolation. This represents a two-dimensional convolutional mapping. It encodes high-resolution historical features. Input the characteristics of each historical moment respectively Depth convolution yields high-resolution historical features that include local neighborhood information; Using high-resolution historical features as keys and values, multi-head cross-attention computation is performed at each fixed spatial location to obtain high-resolution retrieval features: S6. By fusing the high-resolution query features and the high-resolution retrieval features, the wavelet domain prediction result for the current future prediction time is obtained; High-resolution query features High-resolution retrieval features The features are concatenated along the channel dimension and then fused using dual attention blocks to obtain the fused features: Fusion features Input the LL subband output head, LH subband output head, HL subband output head, and HH subband output head respectively to obtain the first... The four wavelet sub-band prediction results corresponding to each future prediction time are combined along the channel dimension to obtain the wavelet domain prediction result for the current future prediction time.

[0039] S7. Update the current decoding state based on the low-resolution retrieval features, and repeat the prediction processing corresponding to each future prediction time. After completing the prediction of the current future time, the low-resolution retrieved features will be used. The decoded state as the next future prediction time, i.e. Then, the above prediction process is repeated until the result is obtained. Wavelet domain prediction results for the predicted future time.

[0040] S8. Perform inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results for each future prediction time to obtain the future radar echo grid sequence.

[0041] In this embodiment, the prediction model is trained using a joint loss consisting of the basic reconstruction loss and the multi-threshold event loss. The total training loss is expressed as: in, Indicates the loss from basic reconstruction. Indicates the loss for multi-threshold events. and This represents the corresponding loss weight. In one example... Set to 0.60, Set to 0.40. The base reconstruction loss includes intensity-adaptive pixel loss. and neighborhood consistency loss , can be represented as: in, and These represent the weights of the intensity-adaptive pixel loss and the neighborhood consistency loss, respectively. In one example, and Both can be set to 0.5. For any pixel position... The intensity-adaptive pixel loss is expressed as: in, This represents the total number of spatiotemporal pixels involved in the loss calculation. Indicates the first Predicted values ​​at each location Indicates the first The target value at each position. For the intensity weighting coefficient, in one example, It can be set to 10. By dynamically weighting the pixel error according to the target echo intensity, high-intensity echo regions receive greater regression constraints during training. The neighborhood consistency loss is expressed as: in, This represents the predicted radar echo grid sequence. Represents the target radar echo grid sequence. Indicates the pooling window is Local average pooling operation. Neighborhood consistency loss is used to reduce excessive error penalties caused by small-scale spatial displacements and enhance local spatial continuity constraints. Multi-threshold event loss includes multiple echo intensity thresholds corresponding to different events. Event loss. Before training, radar echo data is normalized to a preset numerical range, and the original echo intensity threshold is converted into a normalized threshold according to the same normalization rule. For the first Each echo intensity threshold is used to generate a corresponding binary event label based on the target value. Predict radar echo grid values With normalized threshold The difference divided by the corresponding temperature parameter , and enter The function yields the probability of continuous events: in, Indicates the first The first echo intensity threshold The probability of a continuous event corresponding to each position. The probability of the m-th echo intensity threshold. The loss from the event is represented as follows: in, Represents the class balance parameter. This represents the focusing parameter. The multi-threshold event loss is expressed as... in, This indicates the number of echo intensity thresholds. Indicates the first The event loss weight corresponding to each echo intensity threshold Indicates the first Each echo intensity threshold corresponds to Event Loss. By weighted averaging of event losses at different echo intensity thresholds, the prediction model's ability to identify echo events of varying intensity levels is considered during training. In one example, the echo intensity thresholds can be set to 133, 160, 181, and 219, with corresponding weights of 0.1, 0.2, 0.3, and 0.4, and corresponding temperature parameters of 0.1, 0.08, 0.05, and 0.02. The focusing parameter for loss can be set to 2.0, and the class balance parameter can be set to 0.55. These values ​​are preferred examples and do not constitute a limitation on the scope of protection of this invention.

[0042] To verify the technical effectiveness of the method of this invention, a prediction experiment was conducted using the publicly available SEVIR-LR radar echo dataset. In the experiment, seven consecutive frames of historical radar echo grid images were input, and the prediction of the subsequent six frames of radar echo grid images was performed. The Critical Success Index (CSI), Probability of Detection (POD), Fair Threat Score (ETS), and Frequency Bias were used for evaluation. Higher values ​​for CSI, POD, and ETS indicate better event prediction performance. In Tables 2 and 3, CSI-H, POD-H, ETS-H, and Bias-H represent the average values ​​of CSI, POD, ETS, and Bias at echo intensity thresholds of 160, 181, and 219, respectively. A closer Bias-H to 1 indicates a smaller predicted echo area bias. As shown in Table 1, the method of this invention achieves higher CSI, POD, and ETS values ​​than the listed comparative methods at echo intensity thresholds of 160, 181, and 219, indicating that the method of this invention helps improve the identification and detection capabilities of high-intensity radar echo events. As shown in Table 2, with both temporal and spatial attention enabled and a local window size of 4×4, after converting from the pixel domain to the wavelet domain, CSI-H and ETS-H improved from 0.235 and 0.231 to 0.250 and 0.246, respectively. Adding depthwise convolution to the wavelet domain further improved CSI-H, POD-H, and ETS-H to 0.261, 0.382, and 0.257. In the listed structural configurations, the complete structural model achieved CSI-H, POD-H, and ETS-H of 0.268, 0.425, and 0.265, respectively, with a Bias-H of 0.982. Removing temporal attention or local window spatial attention resulted in a decrease in CSI-H, POD-H, and ETS-H, indicating that joint modeling of temporal evolution and local spatial structural relationships helps improve the prediction performance of high-intensity echo events. As shown in Table 3, compared with using only the mean squared error loss, the addition of multi-threshold event loss improved CSI-H, POD-H, and ETS-H. After applying the complete joint loss, CSI-H, POD-H, and ETS-H reached 0.268, 0.425, and 0.265, respectively, while Bias-H reached 0.982. This indicates that the combined use of intensity-adaptive pixel loss, neighborhood consistency loss, and multi-threshold event loss helps improve the prediction capability of high-intensity echo events and reduces the predicted echo area bias.

[0043] Table 1. Comparison of event prediction performance of different methods at high echo intensity thresholds. Table 2 Comparison of high-threshold event prediction performance under different structural configurations Table 3. Comparison of high-threshold event prediction performance under different training loss configurations Example 2: This embodiment provides a radar echo nowcasting system based on wavelet domain dual attention, including: The data acquisition module acquires a sequence of historical radar echo grid points, which consists of radar echo grid point images from multiple historical moments. The wavelet transform module performs a first-level two-dimensional discrete wavelet transform on each radar echo grid image in the historical radar echo grid sequence frame by frame to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. The dual attention coding module inputs the wavelet domain history sequence into the dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features; The low-resolution dynamic historical retrieval module generates low-resolution query features based on the current decoding state and the corresponding future time representation for each future prediction time. It uses the low-resolution historical encoding features after local spatial mixing as keys and values, and obtains low-resolution retrieval features through temporal cross attention. The high-resolution dynamic historical retrieval module performs feature recombination and spatial upsampling on the low-resolution retrieval features to obtain high-resolution query features. The high-resolution historical encoding features after local spatial mixing are used as keys and values, and high-resolution retrieval features are obtained through temporal cross attention. The feature fusion and prediction module fuses the high-resolution query features and the high-resolution retrieval features to obtain the wavelet domain prediction result for the current future prediction time. The decoding state update module updates the current decoding state based on the low-resolution retrieval features and repeatedly executes the prediction processing corresponding to each future prediction time. The inverse wavelet reconstruction module performs an inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results at each future prediction time to obtain the future radar echo grid sequence.

[0044] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0045] Example 3: An electronic device is provided for running the aforementioned "Radar Echo Near-Time Prediction Method Based on Wavelet Domain Dual Attention". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S8 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain a historical radar echo grid sequence consisting of radar echo grid images from multiple historical moments; S2. Perform a first-level two-dimensional discrete wavelet transform on each radar echo grid image in the historical radar echo grid sequence frame by frame to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. S3. Input the wavelet domain history sequence into a dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features; S4. For each future prediction time, generate low-resolution query features based on the current decoding state and the corresponding future time representation. Use the low-resolution historical coding features after local spatial mixing as keys and values, and obtain low-resolution retrieval features through temporal cross attention. S5. Perform feature recombination and spatial upsampling on the low-resolution retrieval features to obtain high-resolution query features. Use the high-resolution historical coding features after local spatial mixing as keys and values, and obtain high-resolution retrieval features through temporal cross attention. S6. By fusing the high-resolution query features and the high-resolution retrieval features, the wavelet domain prediction result for the current future prediction time is obtained; S7. Update the current decoding state based on the low-resolution retrieval features, and repeatedly execute the prediction processing corresponding to each future prediction time; and S8. Perform inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results for each future prediction time to obtain the future radar echo grid sequence.

[0046] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0047] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S8 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0048] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0049] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0050] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A radar echo nowcasting method based on wavelet domain dual attention, characterized in that, Includes the following steps: Obtain a historical radar echo grid sequence consisting of radar echo grid images from multiple historical moments; A first-order two-dimensional discrete wavelet transform is performed frame by frame on each radar echo grid image in the historical radar echo grid sequence to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. The wavelet domain history sequence is input into a dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features. For each future prediction time, low-resolution query features are generated based on the current decoding state and the corresponding future time representation. Low-resolution historical encoding features after local spatial mixing are used as keys and values, and low-resolution retrieval features are obtained through temporal cross attention. The low-resolution retrieval features are reorganized and spatially upsampled to obtain high-resolution query features. High-resolution historical encoding features after local spatial mixing are used as keys and values, and high-resolution retrieval features are obtained through temporal cross attention. By fusing the high-resolution query features and the high-resolution retrieval features, the wavelet domain prediction result for the current future prediction time is obtained; The current decoding state is updated based on the low-resolution retrieval features, and the prediction processing corresponding to each future prediction time is repeated. as well as Perform an inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results for each future prediction time to obtain the future radar echo grid sequence.

2. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 1, characterized in that, The first-level two-dimensional discrete wavelet transform is a first-level Haar discrete wavelet transform, the low-frequency approximation subband is the LL subband, and the multiple high-frequency detail subbands include the LH subband, the HL subband, and the HH subband. The wavelet domain prediction result for the current future prediction time is generated by independent linear output heads, which include LL subband output head, LH subband output head, HL subband output head and HH subband output head.

3. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 1, characterized in that, The dual attention encoder includes a high-resolution feature encoding branch and a low-resolution feature encoding branch. The high-resolution feature encoding branch performs temporal self-attention calculation along the time dimension at each fixed spatial location and performs spatial self-attention calculation within the local spatial window at each historical moment to obtain the high-resolution historical encoded features. The low-resolution feature encoding branch obtains the high-resolution historical encoded features by downsampling them only along the spatial dimension and performing temporal self-attention calculation and local window spatial self-attention calculation on the downsampled features to obtain the low-resolution historical encoded features.

4. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 1, characterized in that, The local spatial mixing includes: performing deep convolution on the historical encoded features at each historical moment under the corresponding resolution to fuse the historical information of the current spatial location and its adjacent spatial locations; the temporal cross-attention implementation includes: at each fixed spatial location, arranging the historical encoded features after local spatial mixing into key sequences and value sequences along the historical time dimension, using the corresponding query features as query sequences, and performing multi-head cross-attention calculation to obtain intermediate attention features: in, express Depthwise convolution, and These represent query mapping, key mapping, and value mapping, respectively. This indicates cross-attention calculation; After performing a residual connection between the intermediate attention features and the query features, the corresponding retrieval features are obtained through a feedforward network.

5. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 4, characterized in that, The low-resolution retrieval features are reorganized by dual attention blocks and then spatially upsampled using nearest neighbor spatial interpolation and two-dimensional convolutional mapping. The high-resolution query features and the high-resolution retrieval features are concatenated in the channel dimension and then fused by dual attention blocks.

6. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 1, characterized in that, When training the prediction model used to perform the method, a total training loss is calculated based on the predicted radar echo grid sequence and the target radar echo grid sequence of the training samples. The total training loss includes a basic reconstruction loss and a multi-threshold event loss. The basic reconstruction loss includes an intensity-adaptive pixel loss and a neighborhood consistency loss. The multi-threshold event loss is obtained by weighting the event losses corresponding to multiple echo intensity thresholds. The parameters of the prediction model are updated according to the total training loss.

7. The radar echo proximity prediction method based on wavelet domain dual attention as described in claim 6, characterized in that, The intensity-adaptive pixel loss is defined as: in, This represents the total number of spatiotemporal pixels involved in the loss calculation. Indicates the first Predicted values ​​at each location Indicates the first The target value at each location. This is the intensity weighting coefficient; The neighborhood consistency loss is defined as: in, This represents the predicted radar echo grid sequence. Represents the target radar echo grid sequence. Indicates the pooling window is Local average pooling operation; The multi-threshold event loss is composed of multiple echo intensity thresholds. The corresponding Focal event loss is obtained by weighting the losses. For the m-th threshold, the event loss is: in, Represents the class balance parameter. Indicates the focus parameter; From the target value Relative to threshold Generated binary event labels: and , The temperature parameter is the corresponding threshold value. The multi-threshold event loss is: in, This indicates the number of echo intensity thresholds. Indicates the first The event loss weight corresponding to each echo intensity threshold Indicates the first Each echo intensity threshold corresponds to Losses from the incident.

8. A radar echo nowcasting system based on wavelet domain dual attention, characterized in that, include: The data acquisition module acquires a sequence of historical radar echo grid points, which consists of radar echo grid point images from multiple historical moments. The wavelet transform module performs a first-level two-dimensional discrete wavelet transform on each radar echo grid image in the historical radar echo grid sequence frame by frame to obtain a wavelet domain historical sequence containing low-frequency approximate sub-bands and multiple high-frequency detail sub-bands. The dual attention coding module inputs the wavelet domain history sequence into the dual attention encoder to obtain high-resolution history coding features and low-resolution history coding features; The low-resolution dynamic historical retrieval module generates low-resolution query features based on the current decoding state and the corresponding future time representation for each future prediction time. It uses the low-resolution historical encoding features after local spatial mixing as keys and values, and obtains low-resolution retrieval features through temporal cross attention. The high-resolution dynamic historical retrieval module performs feature recombination and spatial upsampling on the low-resolution retrieval features to obtain high-resolution query features. The high-resolution historical encoding features after local spatial mixing are used as keys and values, and high-resolution retrieval features are obtained through temporal cross attention. The feature fusion and prediction module fuses the high-resolution query features and the high-resolution retrieval features to obtain the wavelet domain prediction result for the current future prediction time. The decoding state update module updates the current decoding state based on the low-resolution retrieval features and repeatedly executes the prediction processing corresponding to each future prediction time. as well as The inverse wavelet reconstruction module performs an inverse two-dimensional discrete wavelet transform on the wavelet domain prediction results at each future prediction time to obtain the future radar echo grid sequence.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the radar echo proximity prediction method based on wavelet domain dual attention as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a radar echo proximity prediction method based on wavelet domain dual attention as described in any one of claims 1-7.

Images