Precipitation nowcasting method and system based on macro-micro decoupling

Abstract

The present application relates to a kind of based on macro-micro decoupling precipitation nowcasting method and system.The method includes: obtaining historical radar echo sequence is spliced and flattened in time dimension, feature extraction is carried out by spatial encoder, deep layer hidden feature map is input into anisotropic gate state space module, macroscopic advection evolution characteristic is extracted and decoding is carried out, and initial prediction base map is generated;Initial prediction base map is spliced with historical radar echo sequence, as macroscopic constraint condition input into residual repair network of fusion boundary perception attention, carries out detail repair under the constraint of initial prediction base map, and generates precipitation nowcasting echo graph;Using composite loss function, the prediction model of anisotropic gate state space module and residual repair network is trained, and the trained prediction model is used for precipitation nowcasting;Macroscopic physical accuracy and micro image high fidelity can be considered in prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent weather forecasting technology, and in particular to a precipitation nowcasting method and system based on macro-micro decoupling. Background Technology

[0002] Precipitation nowcasting, based on continuous meteorological radar observation echoes, infers the evolution of short-term precipitation clouds. It is a core technological support for real-time early warning of severe convective meteorological disasters such as heavy rainfall, thunderstorms, and hail, and possesses irreplaceable practical value for urban flood control, geological disaster prevention, and transportation meteorological support. Current mainstream technologies primarily employ deep learning-based temporal modeling schemes, divided into two main technical routes: deterministic prediction models and probabilistic generative models. Each type of scheme has its own advantages and disadvantages in forecast positioning accuracy and image detail restoration.

[0003] Among them, deterministic models (such as recurrent neural networks, convolutional neural networks, and Transformers) can capture the overall motion trend, but as the prediction time increases, the predicted images will quickly become smooth and blurred, thus seriously underestimating the area of ​​extreme heavy precipitation. In addition, existing model architectures (including models that adopt spatiotemporal decoupling mechanisms) mostly use isotropic feature traversal methods when extracting spatial features. They treat precipitation images as static two-dimensional matrices and process them uniformly. This completely ignores the fact that the advection motion of real meteorological fluids is driven by large-scale continuous wind fields and has strong direction dependence (anisotropy). Due to the lack of inductive bias for such meteorological physical characteristics, existing models are often inaccurate and prone to positional shifts when capturing complex and intense strong convective advection trajectories. Probabilistic models (such as generative adversarial models and diffusion models) aim to synthesize high-frequency details, but due to the excessive degrees of freedom in their generation process, the predicted rainfall echoes are extremely inaccurate in terms of macroscopic location and are prone to positional distortion. At the same time, methods such as diffusion models rely on lengthy multi-step Markov denoising iterations, resulting in extremely high inference latency and huge computational load, making it difficult to meet the real-time early warning needs of resource-constrained scenarios in meteorological operations.

[0004] Current mainstream precipitation forecasting models often handle the seemingly contradictory tasks of "macroscopic trajectory deduction" and "microscopic detail generation" simultaneously within a single network branch. This structural lack of decoupling inevitably leads to a tug-of-war during optimization: either minimizing pixel-level errors results in mean regression, leading to smoothed and blurred images, or achieving realistic local textures sacrifices the accuracy of overall physical location. Even existing solutions employing spatiotemporal decoupling have failed to achieve the separation of constraints between macroscopic advection evolution and microscopic high-frequency details in terms of both physical logic and network structure.

[0005] Therefore, traditional precipitation nowcasting methods often suffer from problems such as smooth and ambiguous forecast results and difficulty in meeting the needs of real-time early warning response because they fail to effectively decouple the two conflicting tasks of "macro-trajectory extrapolation" and "micro-detail generation" and lack the ability to perceive the anisotropic motion characteristics of meteorological fluids. Summary of the Invention

[0006] Based on this, in order to solve the above-mentioned technical problems, a precipitation nowcasting method and system based on macro-micro decoupling is provided, which can take into account the macro-physical accuracy of the forecast and the high fidelity of the micro-image, reduce the inference delay of the high-fidelity model, and realize real-time and accurate early warning of extreme disasters in meteorological operations.

[0007] A precipitation nowcasting method based on macro-micro decoupling, the method comprising:

[0008] Historical radar echo sequences are acquired, and after the historical radar echo sequences are spliced ​​and flattened in the time dimension, downsampling features are extracted by a spatial encoder to generate a deep latent feature map.

[0009] The deep latent feature map is input into the anisotropic gated state space module, which includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macroscopic advection evolution features and decode them to generate an initial prediction base map.

[0010] The initial predicted base map and the historical radar echo sequence are concatenated and stitched together in the channel dimension, and the macroscopic constraint is input into the residual repair network of the fusion boundary perception attention. The residual repair network performs high-frequency detail repair under the constraint of the initial predicted base map to generate a precipitation nowcast echo map.

[0011] A composite loss function is used to perform end-to-end joint optimization training on the forecast model consisting of the anisotropic gated state space module and the residual repair network. The trained forecast model is then used to achieve nowcasting of precipitation.

[0012] In one embodiment, a historical radar echo sequence is acquired, and after the historical radar echo sequence is spliced ​​and flattened in the time dimension, downsampling feature extraction is performed through a spatial encoder to generate a deep latent feature map, including:

[0013] A historical radar echo sequence composed of two-dimensional radar echo images at consecutive time steps is acquired, and the historical radar echo sequence is preprocessed.

[0014] The preprocessed historical radar echo sequences are concatenated and stitched together in the time dimension to obtain a three-dimensional tensor, and the three-dimensional tensor is input into a spatial encoder composed of several concatenated two-dimensional convolutional blocks.

[0015] After being downsampled step by step by the spatial encoder, a deep latent feature map is output.

[0016] In one embodiment, the deep latent feature map is input into an anisotropic gated state space module, which includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract and decode macroscopic advection evolution features and generate an initial prediction base map, including:

[0017] For each spatial location of the deep latent feature map, a dynamic routing weight matrix corresponding to the four diagonal motion directions is adaptively generated through layer normalization and multilayer perceptron calculation.

[0018] Based on the dynamic routing weight matrix, the deep hidden feature map is scanned in parallel along four diagonal directions to obtain four sets of scanning sequences with different motion trends.

[0019] The four sets of scan sequences are respectively input into the state-space model with shared parameters for long-term time-series modeling, resulting in four sets of modeled sequences.

[0020] The four sets of modeled sequences are reversed to be restored into four sets of directional sensing feature maps. The dynamic routing weight matrix is ​​used to perform element-wise weighted fusion of the four sets of directional sensing feature maps to generate the fused macroscopic advection spatiotemporal features.

[0021] The macroscopic advection spatiotemporal features are input into the spatial decoder and subjected to step-by-step upsampling to generate an initial prediction base map.

[0022] In one embodiment, the initial predicted base map and the historical radar echo sequence are concatenated and stitched together at the channel dimension, and the result is input as a macroscopic constraint to a residual repair network that fuses boundary perception attention. The residual repair network then performs high-frequency detail repair under the constraint of the initial predicted base map to generate a precipitation nowcast echo map, including:

[0023] Maintaining the spatial height and width dimensions of the initial predicted base map and the historical radar echo sequence are consistent, they are concatenated and stitched together in the channel dimension to obtain a joint feature map;

[0024] The joint feature map is input into the residual repair network built on U-Net. After being downsampled step by step by the encoder, it reaches the bottleneck layer to obtain the bottleneck layer feature map.

[0025] In the bottleneck layer, the high-frequency channels in the feature map of the bottleneck layer are enhanced by a channel attention mechanism to obtain a channel-enhanced feature map.

[0026] The gradient prior features of the channel enhancement feature map in the horizontal and vertical directions are extracted. The gradient prior features are fused with the channel enhancement feature map after pooling. A spatial attention mask is generated by convolution operation. The spatial attention mask is multiplied element-wise with the channel enhancement feature map to obtain the bottleneck feature map after boundary awareness enhancement.

[0027] The bottleneck feature map enhanced by the boundary perception is input into the residual repair network. A spatial mask branch and a high-frequency residual branch are separated through a dual-branch spatial gating residual structure. Based on the initial prediction base map, the spatial mask and the high-frequency intensity residual are multiplied element by element and then superimposed onto the initial prediction base map to generate a precipitation nowcast echo map.

[0028] In one embodiment, the bottleneck feature map enhanced by boundary awareness is input into the residual repair network. A spatial mask branch and a high-frequency residual branch are separated through a dual-branch spatially gated residual structure. Using the initial prediction base map as a reference, the spatial mask and the high-frequency intensity residual are multiplied element-wise and then superimposed onto the initial prediction base map to generate a precipitation nowcast echo map, including:

[0029] The bottleneck feature map enhanced by the boundary awareness is input into the decoder of the residual repair network for progressive upsampling, and the bottleneck layer feature map output by the corresponding layer of the encoder is received through skip connections and concatenated, and then the repair feature map is output after convolution.

[0030] The repaired feature map is equally divided into a first branch and a second branch in the channel dimension. The first branch is mapped to a spatial mask through a first activation function, and the second branch is mapped to a high-frequency intensity residual through a second activation function.

[0031] After multiplying the spatial mask element-wise with the high-frequency intensity residual, the result is added element-wise to the initial prediction base map, and the addition result is cropped to a preset effective value range of radar echo intensity to output a precipitation nowcast echo map.

[0032] In one embodiment, the method further includes:

[0033] Obtain a pre-constructed radar echo training set, which contains several sample pairs, each of which contains a historical radar echo input sequence and a corresponding future radar echo target sequence.

[0034] Construct a prediction model consisting of the anisotropic gated state space module and the residual repair network;

[0035] The historical radar echo input sequence is input into the prediction model, and after forward propagation, it passes through the anisotropic gated state space module and the residual repair network in sequence to output the predicted echo map.

[0036] In one embodiment, a composite loss function is used to perform end-to-end joint optimization training on the forecast model consisting of the anisotropic gated state space module and the residual repair network. The trained forecast model is then used to achieve nowcasting of precipitation, including:

[0037] Construct a composite loss function that includes a weighted sum of weighted mean squared error loss, generative adversarial loss, perceptual loss, and auxiliary loss;

[0038] The total loss value between the predicted echo map and the future radar echo target sequence is calculated based on the composite loss function.

[0039] The gradients of each learnable parameter in the prediction model are calculated using the backpropagation algorithm based on the total loss value, and the optimizer is used to update each learnable parameter based on the gradients.

[0040] The forward propagation, total loss calculation, backpropagation and parameter update are performed iteratively until the preset training termination condition is met, and the trained forecast model is obtained for nowcasting precipitation.

[0041] In one embodiment, the weighted mean square error loss is used to apply different weights to pixels with different precipitation intensities to constrain the regression accuracy of extreme precipitation areas; the generative adversarial loss is used to constrain the visual realism of the predicted echo map; the perceptual loss is used to constrain the semantic consistency between the predicted echo map and the target echo map in the high-level feature space; and the auxiliary loss is specifically used to supervise the initial predicted base map to constrain the task decoupling between macroscopic advection evolution extraction and microscopic high-frequency detail restoration.

[0042] A precipitation nowcasting system based on macro-micro decoupling, the system comprising:

[0043] The data acquisition and encoding module is used to acquire historical radar echo sequences, splice and flatten the historical radar echo sequences in the time dimension, and then perform downsampling feature extraction through a spatial encoder to generate a deep latent feature map.

[0044] The macroscopic advection inference module is used to input the deep latent feature map into the anisotropic gated state space module. The anisotropic gated state space module includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macroscopic advection evolution features and decode them to generate an initial prediction base map.

[0045] The micro-detail repair module is used to concatenate and stitch the initial prediction base map and the historical radar echo sequence in the channel dimension, and input them as macro-constraints to the residual repair network of the fusion boundary perception attention. The residual repair network performs high-frequency detail repair under the constraint of the initial prediction base map to generate a precipitation nowcast echo map.

[0046] The joint training module is used to perform end-to-end joint optimization training of the forecast model consisting of the anisotropic gated state space module and the residual repair network using a composite loss function, and the trained forecast model is used to realize precipitation nowcasting.

[0047] The aforementioned precipitation nowcasting method and system based on macro-micro decoupling achieves deep latent feature maps by temporally flattening historical radar echo sequences and performing layered downsampling via a spatial encoder, thus eliminating redundant pixel noise in the original radar images. An anisotropic gated state space module with dynamic routing, multi-directional scanning, and anisotropic weighted fusion mechanisms adaptively learns multi-directional motion weights for the anisotropic advection characteristics driven by meteorological fluid wind fields, extracting the macroscopic movement and evolution patterns of large-scale precipitation systems and outputting an initial prediction base map, which can suppress the overall positional shift of strong convective clouds. The initial prediction base map, as a macroscopic constraint, is stitched together with the original radar sequence and fed into a residual repair network with fusion boundary perception attention. This residual repair network performs directional high-frequency detail repair, reducing false echo artifacts in rainless background areas. A composite loss function with sub-objective constraints is used for end-to-end joint training of the overall model. The sub-loss terms supervise the accuracy of the macroscopic base map, the fidelity of the microscopic image, and the fitting effect of extreme precipitation intensity, enabling real-time and accurate early warning of extreme disasters in meteorological operations. Attached Figure Description

[0048] Figure 1 This is an application environment diagram of a precipitation nowcasting method based on macro-micro decoupling in one embodiment;

[0049] Figure 2 This is a flowchart illustrating a precipitation nowcasting method based on macro-micro decoupling in one embodiment;

[0050] Figure 3 This is a block diagram of a precipitation nowcasting system based on macro-micro decoupling in one embodiment;

[0051] Figure 4 This is a schematic diagram of the architecture of a precipitation nowcasting method and system based on macro-micro decoupling in one embodiment;

[0052] Figure 5 This is a schematic diagram showing a comprehensive comparison of the computational overhead and inference efficiency of the model under single-sequence inference in one embodiment.

[0053] Figure 6 This is a comparison of experimental results of this application on the MeteoNet dataset in one embodiment;

[0054] Figure 7 This is an example of the ablation results of each core module of this application on MeteoNet;

[0055] Figure 8 This is a visualization of the present application on MeteoNet in one embodiment;

[0056] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] The precipitation nowcasting method based on macro-micro decoupling provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer equipment 110. Computer device 110 can acquire historical radar echo sequences, stitch and flatten them in the time dimension, and then extract downsampled features through a spatial encoder to generate a deep latent feature map. Computer device 110 can input the deep latent feature map into an anisotropic gated state space module, which includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion, to extract and decode macroscopic advection evolution features, generating an initial prediction base map. Computer device 110 can concatenate and stitch the initial prediction base map with the historical radar echo sequences in the channel dimension, using this as a macroscopic constraint condition input to a residual repair network for fusion boundary perception attention. Under the constraint of the initial prediction base map, the residual repair network performs high-frequency detail repair, generating a precipitation nowcast echo map. Computer device 110 can use a composite loss function to perform end-to-end joint optimization training on the forecast model composed of the anisotropic gated state space module and the residual repair network, obtaining a trained forecast model for precipitation nowcasting. Among them, computer equipment 110 may include, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, and other devices.

[0059] In one embodiment, such as Figure 2 As shown, a precipitation nowcasting method based on macro-micro decoupling is provided, including the following steps:

[0060] Step 202: Obtain historical radar echo sequences. After splicing and flattening the historical radar echo sequences in the time dimension, extract downsampled features using a spatial encoder to generate a deep latent feature map.

[0061] Computer equipment can acquire historical radar observation data and perform initial feature encoding, specifically by acquiring continuous radar echo sequences over a past period. The sequence contains two-dimensional radar echo images at multiple consecutive time steps, reflecting the intensity and location distribution of recent precipitation. Computer equipment can process the input radar echo sequence. The image is stitched and flattened in the time dimension and then input into a preset spatial encoder. The spatial encoder performs downsampling feature extraction through a series of convolution operations, mapping the original image dimension sequence into a deep latent feature map with smaller spatial resolution and higher channel dimension.

[0062] Specifically, in one embodiment, the provided method for precipitation nowcasting based on macro-micro decoupling may further include a data acquisition and encoding process. The specific process includes: acquiring a historical radar echo sequence composed of two-dimensional radar echo images at continuous time steps, and preprocessing the historical radar echo sequence; concatenating and stitching the preprocessed historical radar echo sequence in the time dimension to obtain a three-dimensional tensor, and inputting the three-dimensional tensor into a spatial encoder composed of several concatenated two-dimensional convolutional blocks; and outputting a deep latent feature map after downsampling processing by the spatial encoder.

[0063] Computer equipment can acquire two-dimensional radar echo images with continuous fixed time intervals and uniform size. Multiple frames of two-dimensional echo images are combined to form a historical radar echo sequence. Standardized preprocessing operations are then performed on this historical radar echo sequence. Preprocessing may include removing invalid and abnormal pixels, aligning geographic grid ranges, and normalizing radar reflectivity values ​​to eliminate data bias caused by observation noise, resulting in a well-organized and usable standardized radar echo time series sample. Next, the computer equipment can perform a time-series cascade stitching operation on the preprocessed historical radar echo sequence. Specifically, while maintaining the height and width spatial dimensions of each individual radar image, the two-dimensional echo images corresponding to all time steps in the sequence are stacked and merged sequentially along the channel dimension, converting the original four-dimensional time series data into a three-dimensional feature tensor that integrates all historical precipitation information.

[0064] Next, the computer equipment can input the stitched 3D feature tensor into a pre-built spatial encoder. The spatial encoder consists of multiple cascaded 2D convolutional blocks, each containing a 2D convolutional layer, a layer normalization layer, a nonlinear activation layer, and a downsampling layer with stride. The 3D feature tensor is processed layer by layer through each convolutional block, with spatial dimension downsampling performed synchronously at each block, progressively compressing the height and width of the feature map while increasing the number of feature channels. The convolutional operations layer by layer strip away shallow, pixel-level irrelevant background information from the original image, abstracting and extracting deep spatial semantic features such as precipitation cloud contours, convection distribution, and gradient strength. After the spatial encoder's downsampling process across all convolutional blocks, a deep latent feature map with lower spatial resolution, richer channel dimensions, and a condensed, complete temporal precipitation spatial information can be output.

[0065] Step 204: Input the deep latent feature map into the anisotropic gated state space module. The anisotropic gated state space module includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macroscopic advection evolution features and decode them to generate an initial prediction base map.

[0066] To capture the directional dependence of real meteorological fluids driven by large-scale continuous wind fields, this embodiment abandons the isotropic static traversal method in existing visual state space models and innovatively introduces an anisotropic gated state space module (AG-Mamba). That is, after extracting the deep latent feature map, the computer can send the deep latent feature map into the designed anisotropic gated state space module to perform macroscopic advection evolution modeling. This module integrates two core mechanisms: dynamic routing multi-directional state space scanning and anisotropic weighted fusion. The overall computational goal is to extract macroscopic advection evolution features from the latent features that characterize the large-scale movement of precipitation clouds, and after decoding, output an initial prediction base map that retains only the overall precipitation distribution structure.

[0067] In one embodiment, the provided precipitation nowcasting method based on macro-micro decoupling may further include a macro-advection extrapolation process, specifically including: for each spatial location of the deep latent feature map, adaptively generating dynamic routing weight matrices corresponding to the four diagonal motion directions through layer normalization and multilayer perceptron calculation; based on the dynamic routing weight matrices, scanning the deep latent feature map in parallel along the four diagonal directions to obtain four sets of scanning sequences with different motion trends; inputting the four sets of scanning sequences into a state-space model with shared parameters for long-term time-series modeling to obtain four sets of modeled sequences; inversely recovering the four sets of modeled sequences into four sets of directional perceptron feature maps, and using the dynamic routing weight matrix to perform element-wise weighted fusion of the four sets of directional perceptron feature maps to generate fused macro-advection spatiotemporal features; inputting the macro-advection spatiotemporal features into a spatial decoder for step-by-step upsampling processing to generate an initial prediction base map.

[0068] The computer device can extract macroscopic advection evolution features based on anisotropic feature routing and generate a coherent base map. The computer device can input the obtained deep hidden feature map into the anisotropic gated state space module (AG-Mamba) to capture the macroscopic advection motion features of the precipitation system driven by large-scale wind fields.

[0069] For each independent spatial coordinate location of the feature vector in the deep latent feature map, the computer device can unify the numerical distribution of the features at that location through layer normalization, eliminating computational biases caused by differences in feature scale across different regions. The normalized features are then input into a lightweight multilayer perceptron (MLP) to complete feature mapping transformation. After normalization activation processing, weight values ​​for the four diagonal motion directions are adaptively generated for the current spatial location. The four-way weights corresponding to all spatial locations together constitute the dynamic routing weight matrix of the entire feature map. Different regions have different cloud distributions and precipitation gradients, resulting in different generated four-way weight values, achieving pixel-by-pixel differentiated directional weight allocation that conforms to local wind field motion patterns. In other words, the computer device can perform calculations using layer normalization and multilayer perceptrons (MLPs) on the input feature map to adaptively learn and generate a directional weight matrix for each spatial location. Specifically, the anisotropic gated state space module utilizes layer normalization (LN) and multilayer perceptron (MLP) to design a lightweight routing network that adaptively learns for each spatial location and generates dynamic routing weight matrices for the four directions. : ;

[0070] Next, the computer equipment can simultaneously launch four independent sequence scanning operations on the normalized deep latent feature map according to the four types of diagonal motion directions defined by the output dynamic routing weight matrix. The four scanning processes are executed in parallel without waiting for each other. Each scan completely traverses the entire latent feature map along the corresponding diagonal path, and finally outputs four sets of one-dimensional scan sequences that can represent the trend of precipitation movement in a single direction. Then, the scan sequences obtained from the four scans are respectively input into the same set of parameter-shared state space models to carry out time-series evolution modeling. Thanks to the computational characteristics of the state space model that expands linearly with the sequence length, the long-term precipitation evolution law contained in each set of scan sequences is mined, and four sets of modeled sequences after time-series information mining are obtained. The parameter-shared design can significantly reduce the overall number of parameters of the module and reduce the consumption of computing power.

[0071] The computer equipment can reverse-engineer four sets of modeled one-dimensional sequences into four sets of directional sensing feature maps with the same spatial size as the original latent feature maps. It then retrieves the generated dynamic routing weight matrix and, for each identical spatial coordinate, performs element-wise weighted fusion on the four sets of directional sensing feature maps using the corresponding four-directional weight values. A higher weight percentage indicates a greater contribution of the advection motion in that diagonal direction to the current grid precipitation evolution. After fusion, a macroscopic advection spatiotemporal feature is obtained, integrating all directional motion information and conforming to the anisotropic advection physical characteristics of meteorological fluids. The weighted fusion macroscopic advection spatiotemporal feature is then input into the built-in spatial decoder of the module. This decoder consists of cascaded multi-level upsampling convolutional units. The feature maps undergo progressive upsampling operations by the decoder, continuously improving spatial resolution and gradually restoring them to the original spatial scale of the radar observation image. During decoding, only macroscopic topological information such as the outline of large-scale precipitation clouds, the center position of heavy precipitation, and the overall movement trend of the convective system is retained, while fragmented, non-dynamically significant local pixel textures are filtered out. Finally, an initial prediction base map with a coherent topological structure and accurate overall cloud position is output.

[0072] That is, after the four parallel-generated scanning sequences are input into a shared Mamba model for long-range spatiotemporal modeling, the computer equipment can use a dynamic weight matrix to perform anisotropic element-wise weighted fusion. The fused macroscopic spatiotemporal features are then fed into a spatial decoder for upsampling, ultimately generating an initial prediction base map. Through adaptive interpolation of dynamic weights, the network can implicitly and continuously approximate the macroscopic true advection trajectory in any direction within a discrete grid; the initial prediction base map generated in this stage is decoded. It is not a disordered or ambiguous transitional product, but a predictive base map that already has a high degree of consistency and physical coherence in macroscopic topology, which constitutes the absolute physical constraint premise for subsequent microscopic repair.

[0073] Step 206: The initial predicted base map and the historical radar echo sequence are concatenated and stitched together in the channel dimension. This is used as a macroscopic constraint input to the residual repair network of the fusion boundary perception attention. The residual repair network performs high-frequency detail repair under the constraint of the initial predicted base map to generate a precipitation nowcast echo map.

[0074] Computer equipment can output a coherent initial prediction basemap. With the original input The samples are cascaded and stitched along the channel dimension to form a base sample with macroscopic constraints. The whole sample is then input into a microscopic residual repair network (including BDAU-Net and SGRM) for strong convection high-frequency detail injection.

[0075] In one embodiment, a precipitation nowcasting method based on macro-micro decoupling may further include a micro-detail restoration process. This process specifically includes: maintaining the spatial height and width dimensions of the initial prediction base map consistent with historical radar echo sequences, and cascading and stitching them along the channel dimension to obtain a joint feature map; inputting the joint feature map into a residual restoration network built on U-Net, which is downsampled step-by-step by the encoder to reach the bottleneck layer, obtaining a bottleneck layer feature map; in the bottleneck layer, strengthening the high-frequency channels in the bottleneck layer feature map through a channel attention mechanism to obtain a channel-enhanced feature map; and extracting the channel-enhanced feature map... The gradient prior features in the horizontal and vertical directions are fused with the channel enhancement feature map after pooling. A spatial attention mask is generated through convolution, and the spatial attention mask is multiplied element-wise with the channel enhancement feature map to obtain the bottleneck feature map after boundary awareness enhancement. The bottleneck feature map after boundary awareness enhancement is input into the residual repair network. The spatial mask branch and the high-frequency residual branch are separated through a two-branch spatial gating residual structure. Based on the initial prediction base map, the spatial mask and the high-frequency intensity residual are multiplied element-wise and superimposed on the initial prediction base map to generate the precipitation nowcast echo map.

[0076] The computer device aligns the spatial height and width dimensions of the initial predicted base map and the historical radar echo sequence, ensuring a one-to-one correspondence between the spatial pixel positions of the two feature maps. Then, it concatenates and stacks the two maps along the feature channel dimension to generate a joint feature map. Next, the computer device inputs the stitched joint feature map into a residual inpainting network built on a U-Net architecture. The joint feature map first flows through the network's encoding branch, which progressively compresses the spatial dimensions and extracts multi-scale deep features through multi-layer convolution and stride downsampling operations. These features are passed layer by layer to the bottleneck layer at the center of the network, outputting the bottleneck layer feature map. The computer device can adaptively evaluate the importance of each channel feature in the bottleneck layer feature map using a channel attention mechanism, automatically increasing the weights of high-frequency channels carrying fine texture information of strong convection and extreme precipitation, while suppressing low-frequency ineffective channels corresponding to rainless backgrounds. After processing, a channel-enhanced feature map highlighting precipitation details is obtained. The computer equipment can extract gradient prior features in the horizontal and vertical dimensions of the channel enhancement feature map. Regions with higher gradient amplitudes correspond to areas with drastic dynamic changes, such as precipitation fronts and convective cells. The gradient prior features are fused with the channel enhancement feature map after pooling compression, and then a spatial attention mask that distinguishes active areas from stable backgrounds is generated through convolution. The spatial attention mask is multiplied element-wise with the channel enhancement feature map to filter out stable background features without precipitation changes, retaining only the effective features of convective fronts, thus obtaining the bottleneck feature map after boundary awareness enhancement.

[0077] The computer equipment can feed the bottleneck feature map enhanced with boundary awareness into the decoding branch of the residual repair network for upsampling and restoration. A dual-branch spatially gated residual structure is then connected to the network output, splitting the network output features into two independent branches: one branch generates a spatial mask branch through activation operations, used to mark whether micro-details need to be added to each spatial region; the other branch generates a high-frequency intensity residual branch through activation operations, storing fine texture information of extreme precipitation and convective fronts. During the calculation, the initial prediction base map is used as the precipitation distribution benchmark. The spatial mask and the high-frequency intensity residual are multiplied element-wise, and the multiplication result is superimposed on the initial prediction base map. In convectively active regions with high activation values ​​in the spatial mask, the high-frequency residual is fully superimposed to restore the true micro-texture of precipitation; finally, a nowcast precipitation echo map is output.

[0078] In one embodiment, a precipitation nowcasting method based on macro-micro decoupling may further include the process of generating a precipitation nowcasting echo map. The specific process includes: inputting the bottleneck feature map enhanced by boundary awareness into the decoder of the residual repair network for step-by-step upsampling, and then stitching together the bottleneck layer feature maps output by the corresponding layer of the encoder via skip connections, followed by convolution to output a repaired feature map; equally splitting the repaired feature map into a first branch and a second branch in the channel dimension, mapping the first branch to a spatial mask using a first activation function, and mapping the second branch to a high-frequency intensity residual using a second activation function; multiplying the spatial mask and the high-frequency intensity residual element-wise, then adding them element-wise to the initial prediction base map, and cropping the sum to a preset effective range of radar echo intensity to output the precipitation nowcasting echo map.

[0079] Specifically, the computer device can input the bottleneck feature map enhanced with boundary awareness into the decoder built into the residual inpainting network. The decoder is composed of multiple layers of upsampling convolutional units cascaded sequentially, performing upsampling operations on the feature map layer by layer to restore spatial resolution. Simultaneously, based on the skip connection structure inherent in the U-Net architecture, it receives the shallow feature maps output from each corresponding layer of the network encoder and concatenates and fuses them with the currently decoded features along the channel dimension, thereby preserving precipitation texture information at different scales. The concatenated and fused features are then processed by convolutional layers to complete feature integration, outputting a repaired feature map integrating multi-scale information. The computer device can equally divide the repaired feature map output by the decoder along the channel dimension into two independent branches, serving as a spatial mask generation branch and a high-frequency residual generation branch, respectively. The first branch inputs a first activation function for mapping processing and outputs a spatial mask with values ​​limited to the range of 0 to 1. The mask is used to mark whether each spatial grid allows the superposition of microscopic high-frequency details. The second branch inputs a second activation function for mapping processing and outputs a high-frequency intensity residual carrying convective frontal and extreme precipitation fine texture information.

[0080] The computer equipment can perform element-wise multiplication of the generated spatial mask with the high-frequency intensity residual. The result of the multiplication represents the effective high-frequency details that can be superimposed in each region. Then, the multiplication result is added element-wise to the initial prediction base map, superimposing the microscopic details onto the base map that locks the macroscopic precipitation structure. Finally, the overall feature values ​​after addition are truncated and cropped, and all pixel values ​​are constrained to the preset effective value range corresponding to the radar observation reflectivity. Abnormal extreme values ​​that do not conform to meteorological and physical laws are eliminated. Finally, a precipitation nowcast echo map with accurate macroscopic location, realistic microscopic details and no background false artifacts is output.

[0081] In other words, the computer device can concatenate the structurally coherent initial prediction base map with the original input along the channel dimension, and use this as a conditional guide to feed into the microscopic residual repair module. At this stage, the features first pass through Boundary Aware Directional Attention (BDA) at the bottleneck layer of U-Net. This mechanism not only applies Channel Attention (LCA) to enhance high-frequency channels, but also extracts absolute gradient prior features in the horizontal and vertical directions through spatial prior attention, generating a spatial attention mask that forces the network to precisely focus on the dynamically evolving strong convective front region. At the very end of the network output, to prevent high-frequency features from flooding the global space, a dual-branch Spatial Gated Residual Structure (SGRM) is employed. This structure decouples the network output into a spatial activation mask branch. With high-frequency intensity residual branch : ,in, This represents the final predicted echo plot of the difference branch output; M represents the initial prediction base map generated by the macro-prediction branch; M represents the spatial mask used to define the residual repair region; δ represents the high-frequency residuals predicted by the residual repair network; through the spatial mask... Strong constraints, forcing high-frequency residuals It only works in the active convection region where the mask is activated; once it leaves the active region indicated by the macroscopic base map, the mask value approaches 0, thereby mitigating the propagation path of high-frequency noise in the rainless background region at the physical structure level.

[0082] Specifically, the computer device can input the concatenated joint features into a U-Net architecture with skip connections. At the feature bottleneck layer of the network, channels containing high-frequency precipitation information are first enhanced through Channel Attention (LCA). Subsequently, a boundary-aware operator is introduced to extract the gradient prior features of latent features in the horizontal and vertical directions. These gradient prior features are then fused with pooling features and subjected to convolution to generate a spatial attention mask. This guides the residual repair network to accurately focus on and locate the region of strong convective fronts with dramatic dynamic evolution.

[0083] At the very end of the micro-repair network output, to prevent high-frequency features from overflowing, the computer device can introduce a Spatial Gated Residual Structure (SGRM) for forced structural decoupling; the output of the last convolutional layer is equally split into two branches: a spatial mask branch activated by Sigmoid. and the high-frequency intensity residual branch activated by Tanh .

[0084] The final high-frequency restoration process is strictly controlled by a dual-branch system. By introducing a gating mechanism, the computer equipment can use the generated coherent base map as the reference for mask filtering. In areas with active precipitation, the residual branch is responsible for injecting realistic high-frequency physical details; while in rainless areas or smooth background areas, the mask... The activation value approaches 0, which cuts off the superposition and spread of high-frequency artifacts and background noise in a way that attenuates by a factor of two, and finally outputs a high-quality, high-fidelity precipitation nowcast echo map.

[0085] In one embodiment, a precipitation nowcasting method based on macro-micro decoupling may further include a process of constructing a training set. The specific process includes: acquiring a pre-constructed radar echo training set, which contains several sample pairs, each containing a historical radar echo input sequence and a corresponding future radar echo target sequence; constructing a forecasting model composed of an anisotropic gated state space module and a residual repair network; inputting the historical radar echo input sequence into the forecasting model, and passing it forward through the anisotropic gated state space module and the residual repair network in sequence to output a predicted echo map.

[0086] Before training the model, the computer equipment can first build a radar echo training dataset for model parameter learning and build a complete precipitation nowcasting model to complete the preparation of training samples and network model.

[0087] Specifically, the computer equipment can acquire a pre-constructed radar echo training set. This training set contains multiple sample pairs, each consisting of a historical radar echo input sequence and its corresponding future radar echo target sequence. The historical radar echo input sequence comprises historical radar echoes observed over multiple consecutive time steps, serving as the model's input data. The future radar echo target sequence consists of radar echoes observed in the same geographical area during the corresponding time period after the input sequence, serving as ground truth labels for comparison during training. The computer equipment can build a precipitation nowcasting model, constructed from the aforementioned anisotropic gated state space module and a residual repair network incorporating boundary-aware attention. The modules are connected according to a predetermined data flow order. The historical radar echo input sequence sequentially passes through the spatial encoder and the anisotropic gated state space module to generate an initial prediction base map. This initial prediction base map is then concatenated with the historical radar echo input sequence and fed into the residual repair network, completing the overall model's network topology construction.

[0088] During the forward propagation of the model, the computer can send the historical radar echo input sequence from the training set into the already built prediction model. The data is passed forward sequentially according to the preset data flow. First, the spatial encoder completes temporal stitching and downsampling feature extraction to obtain a deep latent feature map. Then, it is input into the anisotropic gated state space module to complete multi-directional dynamic routing scanning, anisotropic weighted fusion, and decoding operations, and outputs an initial prediction base map. After that, the initial prediction base map and the historical radar echo input sequence are stitched together along the channel dimension to obtain a joint feature map, which is input into the residual repair network. After bottleneck layer boundary perception attention enhancement, decoder upsampling, and detailed repair processing of the dual-branch spatial gated residual structure, the prediction model finally outputs a prediction echo map, which is used to calculate the loss of subsequent and future radar echo target sequences.

[0089] Step 208: The forecast model consisting of anisotropic gated state space modules and residual repair network is jointly optimized and trained end-to-end using a composite loss function to obtain the trained forecast model for nowcasting precipitation.

[0090] Computer equipment can use pre-built radar echo training sets to perform end-to-end collaborative training of forecast models.

[0091] In one embodiment, a precipitation nowcasting method based on macro-micro decoupling may further include a model training process, specifically including: constructing a composite loss function containing a weighted sum of weighted mean square error loss, generative adversarial loss, perception loss, and auxiliary loss; calculating the total loss value between the predicted echo map and the future radar echo target sequence based on the composite loss function; calculating the gradient of each learnable parameter in the forecast model based on the total loss value using a backpropagation algorithm, and updating each learnable parameter using an optimizer based on the gradient; iteratively executing forward propagation, total loss value calculation, backpropagation, and parameter update until a preset training termination condition is met, resulting in a trained forecast model used for precipitation nowcasting.

[0092] Computer equipment can construct a composite loss function, which is a weighted sum of four loss terms: weighted mean squared error loss, generative adversarial loss, perceptual loss, and auxiliary loss. Each loss term is assigned a corresponding weight coefficient to coordinate the balance between different optimization objectives. The computer equipment can input the predicted echo map output by the model and the future radar echo target sequence in the sample pair into the composite loss function to calculate the total loss value corresponding to the current batch of samples. The total loss value comprehensively reflects the differences between the model output and the ground truth in multiple dimensions, such as macroscopic base map prediction bias, predicted image pixel error, and texture detail fidelity.

[0093] The computer equipment can use the backpropagation algorithm to solve the gradient of all learnable parameters in the prediction model layer by layer based on the calculated total loss value; then the preset optimizer reads the gradient values ​​corresponding to each parameter and updates all learnable parameters of each module in the prediction model according to the optimization update rules.

[0094] The computer equipment can iteratively execute the complete process of forward propagation of samples to output predicted echo maps, calculate the total loss value, backpropagation to obtain gradients, and optimizer to update network parameters. The training process continuously repeats this iterative process, constantly correcting the network parameters until the preset training termination condition is met. The training termination condition can be set to the total loss no longer decreasing significantly, or the number of iterations reaching the preset maximum number of iterations. When the training termination condition is met, training stops, the network parameters are saved, and the trained prediction model is obtained.

[0095] During model optimization, computer equipment can employ a composite loss function that precisely balances pixel-level accuracy with generated visual quality. : ;in, , , All are pre-defined non-negative weight coefficients, used to adjust the contributions of generative adversarial loss, perceptual loss, and auxiliary loss to the total loss function, respectively; the weighted mean square error loss serves as the basic reconstruction loss, with its weight set to 1. These weights are used to balance the pixel-level accuracy and visual realism of echo prediction; the weighted mean square error (... The focus is on the regression of extreme precipitation areas to ensure macroscopic conservation; generating counter-loss ( ) and perceived loss ( ) is used in conjunction with auxiliary losses to sculpt the mesoscale structure and microscale texture of strong convective fronts without disrupting macroscopically coherent topology. This is specifically used for intermediate basemaps generated by AG-Mamba. Supervise the process to prevent macro-level advection extraction and micro-level repair from overstepping their boundaries, ensuring the effectiveness of the decoupled architecture.

[0096] In one embodiment, the weighted mean square error loss is used to apply different weights to pixels with different precipitation intensities to constrain the regression accuracy of extreme precipitation areas; the generative adversarial loss is used to constrain the visual realism of the predicted echo map; the perceptual loss is used to constrain the semantic consistency between the predicted echo map and the target echo map in the high-level feature space; and the auxiliary loss is specifically used to supervise the initial predicted base map to constrain the task decoupling between macroscopic advection evolution extraction and microscopic high-frequency detail restoration.

[0097] In one embodiment, a precipitation nowcasting method and system based on macro-micro decoupling is provided. By introducing a dynamic directional routing mechanism based on AG-Mamba, the directional induction bias of meteorological fluid evolution is explicitly injected into the network, thereby solving the advection evolution prediction error caused by the lack of physical direction awareness in existing models. By proposing a macro-micro decoupling architecture, the prediction process is strictly divided into two physical stages: generating a coherent macro base map and high-frequency detail repair based on mask constraints. The coherent macro base map is used as a pre-constraint condition. In the micro repair stage, a dual-branch gating module is used to accurately reduce the propagation path of high-frequency noise in rainless areas, thereby solving the dual problems of many artifacts in generative models and easy smoothing in deterministic models. By deeply integrating the state-space model (SSM) with linearly scalable computational efficiency O(L) with the decoupling mask architecture, high-fidelity repair can be achieved in one step without multi-step sampling, thereby solving the high latency problem in resource-constrained scenarios.

[0098] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by 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 flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0099] In one embodiment, such as Figure 3 As shown, a precipitation nowcasting system based on macro-micro decoupling is provided, including: a data acquisition and encoding module 310, a macro-advection extrapolation module 320, a micro-detail repair module 330, and a joint training module 340, wherein:

[0100] The data acquisition and encoding module 310 is used to acquire historical radar echo sequences, splice and flatten the historical radar echo sequences in the time dimension, and then perform downsampling feature extraction through a spatial encoder to generate a deep latent feature map.

[0101] The macro-advection inference module 320 is used to input the deep latent feature map into the anisotropic gated state space module. The anisotropic gated state space module includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macro-advection evolution features and decode them to generate an initial prediction base map.

[0102] The micro-detail repair module 330 is used to concatenate and stitch the initial prediction base map and historical radar echo sequences in the channel dimension, and input them as macro-constraints to the residual repair network of the fusion boundary perception attention. Under the constraint of the initial prediction base map, the residual repair network performs high-frequency detail repair and generates a precipitation nowcast echo map.

[0103] The joint training module 340 is used to perform end-to-end joint optimization training of the forecast model consisting of an anisotropic gated state space module and a residual repair network using a composite loss function. The trained forecast model is then used to achieve nowcasting of precipitation.

[0104] In one embodiment, the data acquisition and encoding module 310 is further configured to acquire a historical radar echo sequence composed of two-dimensional radar echo images at continuous time steps, and preprocess the historical radar echo sequence; concatenate and stitch the preprocessed historical radar echo sequence in the time dimension to obtain a three-dimensional tensor, and input the three-dimensional tensor into a spatial encoder composed of several concatenated two-dimensional convolutional blocks; after being downsampled step by step by the spatial encoder, a deep latent feature map is output.

[0105] In one embodiment, the macroscopic advection inference module 320 is further used to adaptively generate dynamic routing weight matrices corresponding to the four diagonal motion directions for each spatial location of the deep latent feature map through layer normalization and multilayer perceptron calculation; based on the dynamic routing weight matrices, the deep latent feature map is scanned in parallel along the four diagonal directions to obtain four sets of scanning sequences with different motion trends; the four sets of scanning sequences are input into a state space model with shared parameters for long-term time-series modeling to obtain four sets of modeled sequences; the four sets of modeled sequences are reverse-converted to four sets of directional perceptron feature maps, and the four sets of directional perceptron feature maps are fused element-wise using the dynamic routing weight matrix to generate fused macroscopic advection spatiotemporal features; the macroscopic advection spatiotemporal features are input into a spatial decoder for step-by-step upsampling processing to generate an initial prediction base map.

[0106] In one embodiment, the micro-detail restoration module 330 is further configured to maintain the spatial height and width dimensions of the initial prediction base map and the historical radar echo sequence consistent, and to perform cascaded stitching in the channel dimension to obtain a joint feature map; the joint feature map is input into a residual restoration network based on U-Net, and after being downsampled step by step by the encoder, it reaches the bottleneck layer to obtain a bottleneck layer feature map; in the bottleneck layer, the high-frequency channels in the bottleneck layer feature map are enhanced through a channel attention mechanism to obtain a channel enhanced feature map; the gradient prior features in the horizontal and vertical directions of the channel enhanced feature map are extracted, and the gradient prior features are fused with the pooled channel enhanced feature map, and a spatial attention mask is generated through a convolution operation; the spatial attention mask is multiplied element-wise with the channel enhanced feature map to obtain a bottleneck feature map enhanced by boundary awareness; the bottleneck feature map enhanced by boundary awareness is input into the residual restoration network, and a spatial mask branch and a high-frequency residual branch are separated through a dual-branch spatial gating residual structure; with the initial prediction base map as a reference, the spatial mask and the high-frequency intensity residual are multiplied element-wise and superimposed on the initial prediction base map to generate a precipitation nowcast echo map.

[0107] In one embodiment, the micro-detail restoration module 330 is further configured to input the bottleneck feature map enhanced by boundary awareness into the decoder of the residual restoration network for step-by-step upsampling, and to receive the bottleneck layer feature map output by the corresponding layer of the encoder through skip connections for splicing, and output the restored feature map after convolution; to split the restored feature map equally into a first branch and a second branch in the channel dimension, to map the first branch to a spatial mask through a first activation function, and to map the second branch to a high-frequency intensity residual through a second activation function; to multiply the spatial mask and the high-frequency intensity residual element-wise, and then add them element-wise to the initial prediction base map, and to crop the sum to a preset effective value range of radar echo intensity, and output the precipitation nowcast echo map.

[0108] In one embodiment, the joint training module 340 is further configured to acquire a pre-constructed radar echo training set, which contains several sample pairs, each containing a historical radar echo input sequence and a corresponding future radar echo target sequence; construct a prediction model consisting of an anisotropic gated state space module and a residual repair network; input the historical radar echo input sequence into the prediction model, and pass it through the anisotropic gated state space module and the residual repair network in sequence via forward propagation to output a predicted echo map.

[0109] In one embodiment, the joint training module 340 is further configured to construct a composite loss function that includes a weighted sum of weighted mean square error loss, generative adversarial loss, perceptual loss, and auxiliary loss; calculate the total loss value between the predicted echo map and the future radar echo target sequence based on the composite loss function; calculate the gradient of each learnable parameter in the forecast model based on the total loss value using the backpropagation algorithm, and update each learnable parameter based on the gradient using an optimizer; iteratively execute forward propagation, total loss value calculation, backpropagation, and parameter update until the preset training termination condition is met, and obtain the trained forecast model for nowcasting precipitation.

[0110] In one embodiment, the weighted mean square error loss is used to apply different weights to pixels with different precipitation intensities to constrain the regression accuracy of extreme precipitation areas; the generative adversarial loss is used to constrain the visual realism of the predicted echo map; the perceptual loss is used to constrain the semantic consistency between the predicted echo map and the target echo map in the high-level feature space; and the auxiliary loss is specifically used to supervise the initial predicted base map to constrain the task decoupling between macroscopic advection evolution extraction and microscopic high-frequency detail restoration.

[0111] In one embodiment, a precipitation nowcasting method and system based on macro-micro decoupling can be applied to, for example... Figure 4 The architecture shown mainly includes a macroscopic advection evolution prediction module, a residual repair module, and a spatial gating / prior attention mechanism.

[0112] Specifically, the macroscopic advection evolution prediction module is mainly responsible for capturing large-scale motion trends. The input is a sequence of historical radar echo images. First, a reshape operation is performed to adjust the spatiotemporal data into a tensor format suitable for network processing. The encoder contains multiple processing units. The model uses a variant based on a state-space model (such as Mamba) and incorporates anisotropic gating mechanisms to efficiently extract long-term temporal dependencies and macroscopic motion features. The decoder restores the extracted features to spatial dimensions and outputs an initial prediction base map that contains the main motion trends but may lack high-frequency details.

[0113] The residual repair module is mainly responsible for detail enhancement and boundary sharpening, stitching the coarse prediction base map generated in the first stage with the original input historical sequence in the channel dimension; the spatial gating module is the core component of the repair network, which adaptively determines which areas need to be repaired and which areas remain unchanged based on the coarse prediction map; after repair, the final, more detailed, high-resolution radar echo prediction map is generated.

[0114] In this embodiment, the coherent positional anchoring provided by the front-end AG-Mamba completely relieves the burden of global inference for the back-end micro-repair network; while the back-end SGRM gating module relies entirely on the coherent base map of the front end to implement precise local high-frequency injection. The deep coupling of the two not only solves the effect of the smooth decay of deterministic models over time, but also effectively suppresses the background artifacts and positional distortions common in generative models. In addition, thanks to the O(L) linear scaling computational efficiency of the state-space model, this architecture can complete high-fidelity repair in one step without the need for cumbersome multi-step Markov denoising sampling (such as diffusion models), and the single-sequence inference time is extremely low (tested to be extremely low), greatly reducing computational overhead and meeting the real-time requirements of extreme weather warnings.

[0115] In one embodiment, taking the MeteoNet dataset as an example, a comprehensive comparison of the computational overhead and inference efficiency of various models on the MeteoNet dataset under single-sequence inference is shown. Figure 5 As shown, the model parameters, computational complexity (FLOPs), peak memory usage, and single-sequence inference time are illustrated. This application significantly improves quantitative prediction indicators such as the Critical Success Index (CSI) and the visual evaluation effect based on LPIPS while accurately capturing the evolution trend of macroscopic meteorological advection. This improvement overcomes the severe smoothing fuzziness and attenuation effects that occur in traditional deterministic forecast models as the prediction lead time increases, effectively avoiding underestimation of extreme heavy precipitation areas. Furthermore, relying on the excellent computational efficiency of the state-space model, which expands linearly with sequence length (O(L)), this application can complete high-fidelity repair of microscopic details in one step without relying on cumbersome multi-step iterative sampling (such as diffusion models); it achieves extremely low single-sequence inference time, completely breaking through the bottleneck of traditional probabilistic generative models being difficult to implement and computationally expensive in resource-constrained meteorological early warning operations.

[0116] In one embodiment, the overall quantitative comparison results of the model in this application with each baseline model on the MeteoNet dataset are as follows: Figure 6As shown in the figure, we selected the MeteoNet public radar dataset for evaluation. We used the Critical Success Index (CSI) and the Heidke Skill Score (HSS) to evaluate the precipitation event recognition capability under different reflectivity thresholds, and denoted the average values ​​of the results for each threshold as CSI-M and HSS-M. We used the CSI corresponding to higher reflectivity thresholds to evaluate the predictive capability of stronger precipitation echoes. We used CSI-pool4 with a local pooling window size of 4 to evaluate the spatial localization capability when allowing for finite neighborhood offsets. We used the Structural Similarity Index (SSIM) and the Learned Perceptual Image Patch Similarity (LPIPS) to evaluate structural consistency and perceptual differences. MeteoNet covers two regions in mainland France, with a radar frame interval of 5 minutes; each sample contains 25 radar images. This model was compared with seven representative precipitation nowcasting models. The deterministic baselines include ConvLSTM, EarthFormer, SimVP, PredRNN, and AlphaPre: ConvLSTM performs spatiotemporal recursive modeling through convolutionally gated recurrent units; EarthFormer models a wide range of dependencies through a spatiotemporal attention structure; SimVP directly generates future sequences through a parallel non-recurrent convolutional architecture; PredRNN propagates historical states step by step through cascaded spatiotemporal memory units; and AlphaPre models the amplitude and phase information of the precipitation field in the frequency domain. The probabilistic generation baselines include PreDiff and DiffCast: PreDiff generates future radar sequences using a conditional latent spatial diffusion process, while DiffCast compensates for local stochastic variations in deterministic predictions through diffuse detail modeling. The model in this application exhibits a relatively balanced performance across precipitation event recognition, strong echo retention, spatial positioning, and perception quality. Visual comparisons of predictions from different models are shown below. Figure 8 As shown.

[0117] Ablation experiments were conducted on the model prediction method and system of this application. The experimental results are as follows: Figure 7 As shown, the results indicate that adaptively assigning directional weights based on input features helps extract direction-related information about the evolution of horizontal flow. Jointly utilizing boundary information in both horizontal and vertical directions provides more complete structural constraints for micro-detail restoration. By controlling the compensation amplitude through residual branching and limiting the compensation location through spatial gating branching, the spread of unconstrained residuals in non-target areas can be reduced, achieving a better overall balance between classification prediction, spatial localization, and perception quality.

[0118] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a macro-micro decoupling-based precipitation nowcasting method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0119] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0120] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a precipitation nowcasting method based on macro-micro decoupling.

[0121] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a precipitation nowcasting method based on macro-micro decoupling.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0123] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0124] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A precipitation nowcasting method based on macro-micro decoupling, characterized in that, The method includes: Historical radar echo sequences are acquired, and after the historical radar echo sequences are spliced ​​and flattened in the time dimension, downsampling features are extracted by a spatial encoder to generate a deep latent feature map. The deep latent feature map is input into the anisotropic gated state space module, which includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macroscopic advection evolution features and decode them to generate an initial prediction base map. The initial predicted base map and the historical radar echo sequence are concatenated and stitched together in the channel dimension, and the macroscopic constraint is input into the residual repair network of the fusion boundary perception attention. The residual repair network performs high-frequency detail repair under the constraint of the initial predicted base map to generate a precipitation nowcast echo map. A composite loss function is used to perform end-to-end joint optimization training on the forecast model consisting of the anisotropic gated state space module and the residual repair network. The trained forecast model is then used to achieve nowcasting of precipitation.

2. The precipitation nowcasting method based on macro-micro decoupling according to claim 1, characterized in that, Historical radar echo sequences are acquired, and after being spliced ​​and flattened along the time dimension, downsampling features are extracted using a spatial encoder to generate a deep latent feature map, including: A historical radar echo sequence composed of two-dimensional radar echo images at consecutive time steps is acquired, and the historical radar echo sequence is preprocessed. The preprocessed historical radar echo sequences are concatenated and stitched together in the time dimension to obtain a three-dimensional tensor, and the three-dimensional tensor is input into a spatial encoder composed of several concatenated two-dimensional convolutional blocks. After being downsampled step by step by the spatial encoder, a deep latent feature map is output.

3. The precipitation nowcasting method based on macro-micro decoupling according to claim 1, characterized in that, The deep latent feature map is input into the anisotropic gated state space module, which includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract and decode macroscopic advection evolution features and generate an initial prediction base map, including: For each spatial location of the deep latent feature map, a dynamic routing weight matrix corresponding to the four diagonal motion directions is adaptively generated through layer normalization and multilayer perceptron calculation. Based on the dynamic routing weight matrix, the deep hidden feature map is scanned in parallel along four diagonal directions to obtain four sets of scanning sequences with different motion trends. The four sets of scan sequences are respectively input into the state-space model with shared parameters for long-term time-series modeling, resulting in four sets of modeled sequences. The four sets of modeled sequences are reversed to be restored into four sets of directional sensing feature maps. The dynamic routing weight matrix is ​​used to perform element-wise weighted fusion of the four sets of directional sensing feature maps to generate the fused macroscopic advection spatiotemporal features. The macroscopic advection spatiotemporal features are input into the spatial decoder and subjected to step-by-step upsampling to generate an initial prediction base map.

4. The precipitation nowcasting method based on macro-micro decoupling according to claim 1, characterized in that, The initial predicted base map and the historical radar echo sequence are concatenated and stitched together in the channel dimension, and this is used as a macroscopic constraint input to the residual repair network that fuses boundary perception attention. The residual repair network then performs high-frequency detail repair under the constraint of the initial predicted base map to generate a precipitation nowcast echo map, including: Maintaining the spatial height and width dimensions of the initial predicted base map and the historical radar echo sequence are consistent, they are concatenated and stitched together in the channel dimension to obtain a joint feature map; The joint feature map is input into the residual repair network built on U-Net. After being downsampled step by step by the encoder, it reaches the bottleneck layer to obtain the bottleneck layer feature map. In the bottleneck layer, the high-frequency channels in the feature map of the bottleneck layer are enhanced by a channel attention mechanism to obtain a channel-enhanced feature map. The gradient prior features of the channel enhancement feature map in the horizontal and vertical directions are extracted. The gradient prior features are fused with the channel enhancement feature map after pooling. A spatial attention mask is generated by convolution operation. The spatial attention mask is multiplied element-wise with the channel enhancement feature map to obtain the bottleneck feature map after boundary awareness enhancement. The bottleneck feature map enhanced by the boundary perception is input into the residual repair network. A spatial mask branch and a high-frequency residual branch are separated through a dual-branch spatial gating residual structure. Based on the initial prediction base map, the spatial mask and the high-frequency intensity residual are multiplied element by element and then superimposed onto the initial prediction base map to generate a precipitation nowcast echo map.

5. The precipitation nowcasting method based on macro-micro decoupling according to claim 4, characterized in that, The bottleneck feature map enhanced by boundary awareness is input into the residual repair network. A spatial mask branch and a high-frequency residual branch are separated through a dual-branch spatially gated residual structure. Using the initial prediction base map as a reference, the spatial mask and the high-frequency intensity residual are multiplied element-wise and then superimposed onto the initial prediction base map to generate a precipitation nowcast echo map, including: The bottleneck feature map enhanced by the boundary awareness is input into the decoder of the residual repair network for progressive upsampling, and the bottleneck layer feature map output by the corresponding layer of the encoder is received through skip connections and concatenated, and then the repair feature map is output after convolution. The repaired feature map is equally divided into a first branch and a second branch in the channel dimension. The first branch is mapped to a spatial mask through a first activation function, and the second branch is mapped to a high-frequency intensity residual through a second activation function. After multiplying the spatial mask element-wise with the high-frequency intensity residual, the result is added element-wise to the initial prediction base map, and the addition result is cropped to a preset effective value range of radar echo intensity to output a precipitation nowcast echo map.

6. The precipitation nowcasting method based on macro-micro decoupling according to claim 1, characterized in that, The method further includes: Obtain a pre-constructed radar echo training set, which contains several sample pairs, each of which contains a historical radar echo input sequence and a corresponding future radar echo target sequence. Construct a prediction model consisting of the anisotropic gated state space module and the residual repair network; The historical radar echo input sequence is input into the prediction model, and after forward propagation, it passes through the anisotropic gated state space module and the residual repair network in sequence to output the predicted echo map.

7. The precipitation nowcasting method based on macro-micro decoupling according to claim 6, characterized in that, An end-to-end joint optimization training of the forecast model consisting of the anisotropic gated state space module and the residual repair network is performed using a composite loss function. The resulting trained forecast model is used to achieve nowcasting of precipitation, including: Construct a composite loss function that includes a weighted sum of weighted mean squared error loss, generative adversarial loss, perceptual loss, and auxiliary loss; The total loss value between the predicted echo map and the future radar echo target sequence is calculated based on the composite loss function. The gradients of each learnable parameter in the prediction model are calculated using the backpropagation algorithm based on the total loss value, and the optimizer is used to update each learnable parameter based on the gradients. The forward propagation, total loss calculation, backpropagation and parameter update are performed iteratively until the preset training termination condition is met, and the trained forecast model is obtained for nowcasting precipitation.

8. The precipitation nowcasting method based on macro-micro decoupling according to claim 7, characterized in that, The weighted mean square error loss is used to apply different weights to pixels with different precipitation intensities to constrain the regression accuracy of extreme precipitation areas; the generative adversarial loss is used to constrain the visual realism of the predicted echo map; the perceptual loss is used to constrain the semantic consistency between the predicted echo map and the target echo map in the high-level feature space; the auxiliary loss is specifically used to supervise the initial predicted base map to constrain the task decoupling between macroscopic advection evolution extraction and microscopic high-frequency detail restoration.

9. A precipitation nowcasting system based on macro-micro decoupling, characterized in that, The system includes: The data acquisition and encoding module is used to acquire historical radar echo sequences, splice and flatten the historical radar echo sequences in the time dimension, and then perform downsampling feature extraction through a spatial encoder to generate a deep latent feature map. The macroscopic advection inference module is used to input the deep latent feature map into the anisotropic gated state space module. The anisotropic gated state space module includes multi-directional state space scanning based on dynamic routing and anisotropic weighted fusion to extract macroscopic advection evolution features and decode them to generate an initial prediction base map. The micro-detail repair module is used to concatenate and stitch the initial prediction base map and the historical radar echo sequence in the channel dimension, and input them as macro-constraints to the residual repair network of the fusion boundary perception attention. The residual repair network performs high-frequency detail repair under the constraint of the initial prediction base map to generate a precipitation nowcast echo map. The joint training module is used to perform end-to-end joint optimization training of the forecast model consisting of the anisotropic gated state space module and the residual repair network using a composite loss function, and the trained forecast model is used to realize precipitation nowcasting.

Images