A method and system for fine classification and prediction of severe convective weather by fusing geostationary satellite data

Abstract

The application relates to the technical field of weather prediction, in particular to a strong convective weather refined classification and prediction method and system fusing stationary satellite data. The method comprises the following steps: performing data preprocessing based on obtained multi-source cross-modal meteorological data; constructing a joint space-time classification and prediction network model fusing stationary satellite data, including constructing a cross-modal feature fusion model, SimVP-based hidden space nonlinear space-time deduction, ConvLSTM-based disaster sequence decoding and multi-task classification; performing end-to-end joint loss and model optimization of the constructed network model driven by heterogeneous true values; performing post-processing consistency constraint on the optimization result based on single forward reasoning and a morphological filtering mechanism; and outputting a prediction result. The application breaks through the single physical observation bottleneck and significantly improves the refined decoupling accuracy of complex convective multi-disasters.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, and in particular to a method and system for refined classification and forecasting of severe convective weather by fusing geostationary satellite data. Background Technology

[0002] Severe convective weather (such as short-duration heavy rainfall, thunderstorms, strong winds, hail, and lightning) is characterized by its sudden onset, short lifespan, and high destructive power, making it one of the main meteorological factors causing severe natural disasters. Different types of severe convective disasters pose significantly different threats to agriculture, transportation, aviation, and the safety of people's lives and property. Therefore, disaster prevention and mitigation efforts require drastically different response measures and early warning needs for different types of severe convective weather. This necessitates that meteorological forecasts not only accurately predict the occurrence of severe convective weather but also provide refined classification and forecasting of its specific disaster types.

[0003] However, existing severe convection forecasting methods generally suffer from the following technical shortcomings:

[0004] First, existing methods do not provide detailed classifications of severe convective weather, making it difficult to meet the needs of multi-hazard forecasting.

[0005] Traditional numerical weather prediction (NWP) or radar extrapolation methods are limited by computational resources and physical parameterization schemes, making it difficult to provide high spatiotemporal resolution multi-hazard forecasts in a very short time. In recent years, although deep learning technology has been widely introduced into the field of meteorological forecasting, existing related models often focus only on the prediction of a single hazard type (such as only for short-term heavy precipitation), lacking the ability to simultaneously perform fine-grained classification of multiple associated natural hazards such as heavy precipitation, hail, thunderstorms, and lightning within a unified network framework. This makes it difficult for forecast results to guide differentiated emergency response efforts.

[0006] Second, relying solely on single meteorological data for deep learning classification and recognition yields poor results.

[0007] The atmospheric physical processes within severe convective systems are extremely complex and highly nonlinear. Currently, most deep learning-based severe convective classification methods typically rely solely on single-dimensional radar echo data for hazard type identification. Since a single data source cannot comprehensively characterize key physical features such as cloud top temperature and water vapor distribution during different hazards (especially hail and strong winds), this "direct identification" approach, lacking multi-dimensional feature assistance, results in insufficient feature extraction by the model in complex weather conditions. Therefore, existing models often suffer from low accuracy, high false alarm rates, and high false negative rates when performing fine-grained classification of multiple hazards.

[0008] Furthermore, the atmospheric thermodynamic processes within severe convective weather systems are extremely complex and highly nonlinear, making accurate and precise classification and early warning of associated disasters an urgent need for disaster prevention and mitigation. In existing forecasting operations, while ground-based weather radar can capture precipitation structure and echo intensity within clouds, its single radar reflectivity characteristic often presents a bottleneck in information representation when dealing with convective cells such as hail and thunderstorms that have similar morphologies but vastly different disaster-causing mechanisms, making it difficult to directly and accurately identify the specific disaster type. In contrast, while geostationary meteorological satellites can provide rich macroscopic thermodynamic prior information such as cloud top temperature, phase state, and water vapor distribution, they exhibit significant cross-modal differences compared to ground-based radar in terms of physical dimensions, observation altitude, and information sensitivity. Meanwhile, existing meteorological prediction models based on conventional convolutional or recurrent neural networks often focus only on coarse-grained regression extrapolation of single elements (such as radar echoes), lacking the ability to jointly model the long-term spatiotemporal dependence of multi-source data. They are difficult to effectively decouple the complex evolution characteristics of multiple disasters occurring concurrently in a single network, and are prone to problems such as disaster location offset, underreporting of extreme disasters (such as hail), and blurred boundaries of multiple classifications.

[0009] In summary, the field of meteorological forecasting urgently needs a forecasting method that can overcome the limitations of a single data source, effectively mine complex spatiotemporal evolution characteristics, and perform refined classification of various specific disaster types of severe convective weather. Summary of the Invention

[0010] To address the aforementioned problems, this invention provides a method and system for refined classification and forecasting of severe convective weather by fusing geostationary satellite data.

[0011] Firstly, the present invention provides a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data, employing the following technical solution:

[0012] A method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data includes:

[0013] Acquire multi-source, multimodal meteorological data;

[0014] Data preprocessing is performed based on the acquired multi-source, cross-modal meteorological data;

[0015] Construct a joint spatiotemporal classification and forecasting network model that integrates geostationary satellite data, including building a cross-modal feature fusion model, latent space nonlinear spatiotemporal extrapolation based on SimVP, and disaster sequence decoding and multi-task classification based on ConvLSTM;

[0016] Heterogeneous truth-driven end-to-end joint loss and model optimization are applied to the constructed network model.

[0017] Post-processing consistency constraints are applied to the optimization results based on a single forward inference and morphological filtering mechanism.

[0018] Output the prediction results.

[0019] Secondly, a refined classification and forecasting system for severe convective weather that integrates geostationary satellite data includes:

[0020] The data acquisition module is configured to acquire multi-source, cross-modal meteorological data.

[0021] The preprocessing module is configured to perform data preprocessing based on the acquired multi-source cross-modal meteorological data;

[0022] The model building module is configured to build a joint spatiotemporal classification and forecasting network model that integrates geostationary satellite data, including building a cross-modal feature fusion model, latent space nonlinear spatiotemporal extrapolation based on SimVP, and disaster sequence decoding and multi-task classification based on ConvLSTM.

[0023] The model optimization module is configured to perform heterogeneous truth-driven end-to-end joint loss and model optimization on the constructed network model;

[0024] The constraint module is configured to perform post-processing consistency constraints on the optimization results based on a single forward inference and morphological filtering mechanism.

[0025] The prediction module is configured to output prediction results.

[0026] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data.

[0027] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, which are adapted to be loaded and executed by the processor to provide a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data.

[0028] In summary, the present invention has the following beneficial technical effects:

[0029] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in the following four aspects:

[0030] (1) Break through the bottleneck of single physical observation and significantly improve the precision of fine decoupling of complex convection multi-hazard.

[0031] To address the shortcomings of existing methods that rely solely on a single radar, which can easily lead to misjudgments of convective cells with similar disaster-causing mechanisms, this invention introduces multi-channel observations from geostationary meteorological satellites and constructs a cross-modal feature fusion model based on a channel attention mechanism. This network structure explicitly and nonlinearly fuses the microscopic three-dimensional precipitation structure from radar with the macroscopic cloud top thermodynamic priors from satellites at the feature level. Compared to single-modal input, this mechanism significantly enriches the physical information representation dimensions of the convection core, successfully decoupling four major disaster types—short-duration heavy precipitation, hail, thunderstorms, and lightning—within a unified network, effectively reducing the false alarm and missed alarm rates for extreme disasters in complex environments.

[0032] (2) Completely solve the end-to-end training barrier that makes it difficult to coordinate and optimize heterogeneous cross-frequency meteorological data.

[0033] To address the significant temporal asynchrony between radar observations (6 minutes), satellite multi-channel data (10 minutes), and real precipitation labels (GPM, 30 minutes), this invention designs a rigorous mathematical alignment scheme. By creatively introducing a "time-dimensional accumulation operator" at the loss calculation stage, high-frequency precipitation prediction slices are temporally aggregated to strictly align with low-frequency precipitation ground truth values. This theoretical design directly penetrates the time pooling layer for MSE gradient backpropagation, breaking down the computational silos between high-frequency microscopic deduction and low-frequency macroscopic verification, and ensuring the stability of gradient updates under cross-frequency heterogeneous multimodal data.

[0034] (3) Take into account both the global dynamics simulation and the reconstruction fidelity of the local extreme disaster area.

[0035] To address the issues of strong echo position drift and morphological dissipation that are common in conventional deep learning models, this invention constructs a deep architecture that combines SimVP latent space inference with ConvLSTM temporal decoding. SimVP avoids the error accumulation of autoregressive extrapolation and maintains the long-range nonlinear dynamic evolution of the system. Simultaneously, by introducing an improved pixel-level focal loss into the loss function, the gradient of simple background pixels is attenuated, forcing the network's learning focus to shift towards the core of disasters with extremely low occurrence frequencies (such as hail and strong winds). The combination of these two approaches significantly improves the sharpness and positioning accuracy of the output multi-hazard probability distribution field at convective boundaries.

[0036] (4) Achieve efficient single-time inference and map generation, and provide quantitative forecast confidence support.

[0037] At the operational deployment level, for a fixed 301×301 observation grid, this invention abandons the computationally redundant sliding window stitching scheme and innovatively adopts a single-pass inference strategy of "edge filling + network forward + reverse center pruning". Combined with morphological opening operations, it greatly reduces computational latency while eliminating stitching artifacts, achieving near real-time and efficient forecasting. Furthermore, this invention introduces a lightweight time-of-test (TTA) mechanism at the output end, directly outputting a quantitative distribution map of forecast uncertainty by calculating the pixel-level variance of multi-branch perturbation predictions. This structural feature effectively avoids the risk of "overconfidence" in the model under rare weather conditions, providing a reliable basis with significant engineering guidance for disaster prevention decision-making. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall process for implementing a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data according to the present invention.

[0039] Figure 2 This is a schematic diagram of the data preprocessing process for a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data, as described in this invention.

[0040] Figure 3 This is a schematic diagram of a model for implementing a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data, as per the present invention.

[0041] Figure 4 This is a schematic diagram of the short-term heavy precipitation GPM forecast results for a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data, as implemented in this invention.

[0042] Figure 5 This diagram illustrates the thunderstorm and gale forecast results of a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data, as implemented in this invention. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] Example 1

[0045] Reference Figure 1 This embodiment of a method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data includes:

[0046] S1 Multi-source cross-modal meteorological data acquisition

[0047] The severe convective weather forecast observation inputs and related truth labels required for this invention are mainly divided into the following three categories, covering the coastal areas of China, and are obtained through different channels:

[0048] 1. First type of data: Multi-channel radar observation data and high-frequency disaster record tags

[0049] This type of data is used to capture the three-dimensional structural features of severe convective systems and the actual situation of ground disasters. The data comes from the historical research dataset provided by the National Meteorological Information Center. Applications are submitted to the National Meteorological Information Center to obtain the data. Specifically, it includes:

[0050] Reflectivity categories: Combined reflectivity (CR), 0.5km height reflectivity (R05), 1.5km height reflectivity (R15).

[0051] Gradient and intensity classes: horizontal reflectance gradient (RG1, RG2), vertical integrated liquid water content (VIL).

[0052] Velocity field types: radial velocity at 0.5km altitude (V05), radial velocity at 1.5km altitude (V15).

[0053] These channels, with a sampling frequency of once every 6 minutes, constitute the core sensing input for the model to perceive the intensity of convection in the lower atmosphere.

[0054] Disaster record tags: Simultaneously acquire observation records of severe convective disasters aligned with radar space, covering detailed occurrence information of thunderstorms, strong winds (maximum wind speed), hail (ground observation), and lightning.

[0055] 2. Second type of data: Geostationary meteorological satellite observation data

[0056] This type of data is used to provide prior information on the thermodynamics of cloud tops. Full data from the Himawari-8 (H8) satellite was downloaded from the official P-Tree distribution system of the Japan Aerospace Exploration Agency (JAXA) and cropped for the coastal region of China. This invention selects 12 spectral channels to construct a high-dimensional satellite feature vector:

[0057] Visible / Near Infrared Channels (VIS Channels): Extract albedo data from channels 1 to 6. to These channels can clearly reflect the microphysical structure, optical thickness, and cloud top roughness of daytime clouds, which helps to identify the top texture of strong convective cloud clusters.

[0058] Infrared Channels: Extract blackbody brightness temperature (TBB) data, specifically including... , , , , and These channels cover the water vapor absorption zone and infrared window region, and can effectively invert cloud top height, thickness, and overshooting top characteristics during convective bursts.

[0059] 3. Third type of data: True GPM (Global Precipitation Measurement) precipitation data

[0060] This data was downloaded from NASA's GES DISC platform, showing the full range of GPM IMERG precipitation data.

[0061] Spatial clipping was performed on the coastal areas of China to extract the precipitation rate as the regression true value for the short-duration heavy precipitation (ERL) forecasting mission.

[0062] The data has a spatial resolution of 0.1° and a temporal resolution of 30 minutes. Through the time window dynamic convergence mechanism in the subsequent S2 step, it is strictly registered with radar / satellite data at a 6-minute frequency.

[0063] S2 Data Preprocessing: Strict Registration and Alignment of Spatiotemporal Sequences

[0064] After acquiring multi-source heterogeneous meteorological data, this invention completely solves the serious asynchronous heterogeneity problem in physical dimensions, spatial resolution, and temporal sampling frequency of ground-based observations (radar), space-based observations (satellites), and external verification ground truth (GPM and disaster records) by constructing a sophisticated spatial resampling algorithm and dynamic time window convergence mechanism. This stage is a key foundation for ensuring that subsequent deep learning networks can extract robust cross-modal spatiotemporal features and avoid gradient divergence during model training.

[0065] S2.1 Spatial Mesh Unified Resampling and Edge Fill Enhancement

[0066] To address the complex underlying surface features of China's coastal areas, this invention performs refined spatial matching of acquired data from 8 radar channels and 12 satellite channels. Ground-based radar data is typically based on polar coordinate scanning, while H8 satellite and GPM data are based on isotropic projections. Ground-based radar base data (reflectivity factor) and multi-channel observation data (visible light VIS and infrared IR) from the geostationary meteorological satellite H8 are acquired within the target area. Let the original radar image be... Satellite imagery is Using bilinear interpolation operators Both are resampled and aligned to the same target 2D mesh. .

[0067] Specifically, let the four adjacent known grid points on the original image to be processed be... Its corresponding data value is For any floating-point coordinate on the target grid The bilinear interpolation is calculated as follows:

[0068] ,

[0069] ,

[0070] ,

[0071] Based on the above calculations, for the center of each target pixel : After unified projection, the spatial dimension is extracted according to the actual coverage area. The local sub-block tensor of a pixel. To accommodate the integer division characteristics of multi-level downsampling operations in subsequent spatiotemporal depth networks, an edge padding operator is defined before the data is input into the network. (such as zero padding or reflection padding), will The feature map is expanded to a size that is an integer multiple of 32 in the nearest neighborhood, i.e. :

[0072] ,

[0073] Among them, the number of radar channels Number of satellite channels .

[0074] Simultaneously, a standardization operator is defined to normalize the maximum and minimum values ​​of the data from each channel, mapping them to... The interval. And a reverse center clipping operator is predefined. Size restoration for network decoding output:

[0075] ,

[0076] S2.2 Timing registration and physical stitching of multi-source observation data at the input end

[0077] For radar (assuming the time series is...) (with a 6-minute sampling interval) and satellite (assuming the time series is...) To address the asynchronous sampling issue (with a 10-minute sampling interval), this step performs strict time-dimensional alignment and data-level stitching. For any radar time... Find the satellite frame with the closest time distance for timestamp registration:

[0078] ,

[0079] Considering the satellite's 10-minute sampling frequency, a time error tolerance is set to ensure the continuity of the radar timing and prevent the loss of any sampling steps. It is 5 minutes (half of the satellite sampling period). When the conditions are met... At that time, the radar tensor and the registered satellite tensor are directly concatenated in the input channel dimension to form the synchronous multimodal physical input frame for the current moment:

[0080] ,

[0081] Extracting history Each time step constructs a strictly continuous historical spatiotemporal tensor sequence as input to the subsequent network: ,

[0082] S2.3 Dynamic time window aggregation of heterogeneous data at the tag end (core alignment mechanism)

[0083] This paper addresses the severe heterogeneity between model-predicted high-frequency sequences and true low-frequency precipitation data (GPM data, 30-minute intervals). It assumes the model will continuously output data over the next 30 minutes. (Example) ) 6-minute interval precipitation prediction slice sequence ,in To achieve the target time True value of 30-minute GPM precipitation Strict alignment; define the time-dimensional accumulation operator before loss calculation. : For disaster tags such as hail and strong winds, which have high-frequency characteristics recorded every 6 minutes, a point-to-point frame-by-frame mapping is maintained. This mechanism, through mathematical-level time pooling, completely establishes an end-to-end computation graph for high- and low-frequency cross-modal learning.

[0084] S2.4 Lifetime Threshold Filtering and Effective Sample Set Construction for Severe Convective Events

[0085] In order to eliminate transient noise interference caused by short lifespans and lack of complete evolution cycles, and to ensure that subsequent spatiotemporal deep networks (SimVP and ConvLSTM) can learn physically meaningful long-term evolution laws, this invention introduces a strict lifespan threshold filtering mechanism based on spatiotemporal pairing.

[0086] For each strong convective event after initial alignment, its temporal continuity is traced. Let the duration of a certain strong convective process on the time axis be denoted as . Calculate the total duration of its life history. Set a minimum lifespan threshold. Hour (i.e. 90 minutes, strictly corresponding to 15 consecutive 6-minute radar / tag observation frames).

[0087] Define the filtering indicator function for sample set construction. :

[0088] ,

[0089] like If the event has a complete and continuous dynamic development process, its corresponding multi-source synchronous data sequence and label pair will be formally included in the final training / validation sample set; if If any such occurrence is considered an isolated, sporadic convection, a system that dissipates too quickly, or observational noise, it will not be included in the dataset and will be removed. This screening process greatly purifies the spatiotemporal sequences, providing a high-quality data foundation for the model to capture long-range dynamic characteristics.

[0090] S2.5 Stepping sliding window along the time axis and construction of large-scale sample sets

[0091] After completing the registration and event filtering of multi-source data, in order to fully expand the amount of data required for deep network training and enable the model to learn the continuous dynamic characteristics of strong convective systems at different evolution stages, this invention employs a step-sliding window mechanism along the time axis to construct large-scale sample pairs. Specifically, for any continuous time series of a strong convective event that satisfies the S2.4 life history threshold condition, a time sliding window is set. This window contains a length of... Historical input sequences (such as) Frame (corresponding to the past 30 minutes) and length is Future predicted label sequences (such as A frame (corresponding to the next 60 minutes). This sliding window is shifted along the time axis, with a sliding step size of 6 minutes (i.e., strictly aligning with the radar's highest temporal resolution). Each time the window slides backward by one step, a new pair of spatiotemporal training samples that highly overlaps in time with the previous sample is extracted and generated. This mechanism segments a single, long-lived strong convection event into a large number of independent samples with continuous time shifts, which not only achieves an exponential expansion of the training dataset, but also effectively enhances the temporal translation invariance of the spatiotemporal network to advection and morphological evolution.

[0092] To ensure the spatial generalization ability and class balance of the sample set, this invention further implements a geographic stratified sampling and multimodal synchronous data augmentation mechanism after generating independent sample pairs:

[0093] 1. Core Splitting Logic: Stratified sampling by province. To ensure spatial independence, this invention extracts province information from event names and isolates them during partitioning. This ensures that both the training and validation sets cover samples from all target provinces, thereby improving the model's cross-regional generalization ability. When extracting the validation set (e.g., at a 20% ratio), a weighted allocation strategy based on events is implemented. Within each province, validation samples are independently extracted from each event proportionally based on the duration of the strong convective event. This avoids extreme imbalances where long-lived events dominate the training set or short events are entirely allocated to the validation set, ensuring the spatiotemporal representativeness of the validation set. A fallback mechanism is also included: if stratified sampling fails due to extreme distributions, it automatically reverts to global random partitioning to ensure system robustness.

[0094] 2. Training-Specific Multimodal Data Augmentation. Addressing the extreme scarcity of positive samples such as those from thunderstorms and strong winds, this invention employs spatial augmentation operators in the training set, theoretically expanding the training sample size by four times through random horizontal and vertical flipping. A key requirement in this step is strict multimodal synchronization: radar sequences, satellite multi-channel sequences, and corresponding high- and low-frequency label sequences must undergo fully synchronized flipping transformations in space to prevent the cross-modal Transformer encoder from learning incorrect spatial correspondences. Furthermore, when constructing the validation and test sets, this invention strictly disables augmentation operations, ensuring that model evaluation is entirely based on real observations.

[0095] S3 Joint Spatiotemporal Classification and Forecasting Network Integrating Geostationary Satellite Data

[0096] A spatiotemporal neural network is constructed, encompassing cross-modal feature fusion, latent space inference, and decoupling of sequence-based multi-task operations. The network operates on a concatenated multimodal history tensor sequence. The input is the future time step, and the final output is the future time step. The probability distribution of various associated disasters. The overall architecture of this network consists of three parts connected in series: a cross-modal space encoder, a SimVP latent space translator, and a ConvLSTM multi-task decoder.

[0097] S3.1 Constructing a cross-modal feature fusion model (Spatial Encoder & Fusion Module)

[0098] In order to eliminate the differences in physical dimensions and extract deep joint representations, this step constructs a cross-modal feature fusion model for heterogeneous inputs that are physically spliced ​​only at the data level in S2.2.

[0099] The radar channel and satellite channel in the input sequence are decoupled and fed into a two-dimensional convolutional spatial encoder with independent weights. and Each element extracts its high-dimensional spatial representation. To avoid gradient diffusion of meteorological microscale features (such as local echo cores) as the network deepens, the encoder internally constructs... Residual blocks are stacked layers. Specifically, for the 1st... Layer residual module ( Let its input feature matrix be... Feature mapping operator in residual branch It includes 2D convolution (Conv2D), batch normalization (BN), and the LeakyReLU activation function. The residual mapping forward propagation process is strictly defined as:

[0100] ,

[0101] Through this skip connection, the network is able to fully preserve the sharpness of the edges of strongly convective cells when performing downsampling with a step size of 2. The multimodal high-dimensional representation finally extracted by the encoder is as follows:

[0102] ,

[0103] ,

[0104] in, The number of latent space feature channels (e.g., 64 dimensions). and Let be the size of the downsampled and compressed feature map. Define a fusion operator based on channel attention. Since the feature dependence weights of thunderstorms, strong winds, or hail on radar water condensate distribution and satellite cloud top brightness temperature change dynamically at different stages of their life cycle, this invention first stitches together two high-dimensional features in the channel dimension. .

[0105] Subsequently, global average pooling was used. By compressing the spatial dimension, channel statistics of the global receptive field are obtained. To capture the nonlinear interactions between channels, these statistics are fed into a multilayer perceptron (MLP) network. Mathematically, this MLP comprises two fully connected layers: dimension reduction and dimension upscaling (with weight matrices respectively...). and ,in (This is a dimensionality reduction scaling factor), supplemented with nonlinear activation, thereby generating the final channel weighted vector. :

[0106] ,

[0107] Finally, the adaptive weighted vector is multiplied by the concatenated features using the Hadamard product. ), to achieve substantial cross-modal feature-level fusion:

[0108] ,

[0109] Features after fusion It fully combines the radar's microstructure sensitivity with the satellite's thermodynamic priors.

[0110] S3.2 Latent Space Nonlinear Spatiotemporal Predictor Based on SimVP

[0111] Cross-modal fusion feature sequences The input is fed into the SimVP core inference module. This step utilizes a three-tiered collaborative architecture of encoder, translator, and decoder to achieve a nonlinear mapping from the physical observation field to the future dynamic evolution trend. The specific logic is as follows:

[0112] 1. Encoder Stage: Spatial Collapse and Physical Consistency Mapping of Heterogeneous Features

[0113] The encoder consists of stacked cross-modal residual downsampling units, serving as the core front-end module of SimVP. Its function is to compress features from spatial features to a high-dimensional latent space. Specifically, the encoder uses the cross-modal fused feature sequence output in step S3.1. As input, it is downsampled by a 2D convolution with a stride of 2. While reducing spatial redundancy, a deep "physical feature distillation" is performed. This distillation calculation process is defined as follows: In this process, the encoder can compress the fused high-dimensional tensor to an extreme degree, mapping it into a sequence of latent variables with spatiotemporal continuity. In a meteorological and physical sense, this step corresponds to the “three-dimensional physical field inversion” of the current initial state of the atmosphere. That is, by learning the potential correlation between the distribution of water condensate in the cloud and the temperature gradient at the cloud top through a convolutional neural network, the “initial seed nucleus of convection” features that are sufficient to support the subsequent long-range evolution projection.

[0114] 2. Translator Stage: Non-autoregressive spatiotemporal manifold decoupling from multi-scale scenes

[0115] To deeply adapt to the multi-scale and highly nonlinear dynamic evolution characteristics of severe convective weather, this invention performs physical scene decoupling and reconstruction of the latent space extrapolation process. Let the spatiotemporal translator within this module be... It employs a fully convolutional non-autoregressive architecture, unlike traditional recurrent networks (RNNs), and performs "end-to-end timeline translation" directly on the latent space manifold. The translator consists of stacked latent space Inception convolutional modules, which, for the input latent space features... The local forward inference process is defined in the following specific physical scenario:

[0116] (1) Scenario 1: Cloud bottom-cloud top thermodynamic cross-channel interaction (deep convection core locking scenario)

[0117] First use Convolution is used for cross-channel information interaction and dimensionality reduction. Physically, this operation corresponds to the nonlinear coupling between radar hydrophobic distribution (such as VIL) and satellite blackbody brightness temperature (TBB), and is used to accurately extract feature vectors characterizing the intensity of vertical upward motion in deep convection systems. This invention defines it as the deep convection core feature map. :

[0118] ,

[0119] (2) Scenario 2: Local microscale convection generation and dissipation simulation (hail / isolated thunderstorm scenario)

[0120] use Grouped convolution focuses on small local receptive fields. This branch is specifically designed to learn the rapid morphological distortions and dramatic increases and decreases in internal echo intensity of isolated supercells or hail cloud nuclei with short lifespans, ranging from a few kilometers to tens of kilometers. This invention defines it as microscale evolutionary inference features. :

[0121] ,

[0122] (3) Scenario 3: Advection extrapolation of mesoscale convective systems (squall line / large-scale short-term heavy precipitation scenario)

[0123] use Grouped convolutions are used to obtain the global spatial receptive field. This branch is responsible for capturing the overall movement vectors and steering airflow characteristics of mesoscale convective complexes (MCS) or linear squall lines spanning hundreds of kilometers. This invention defines these as mesoscale advection extrapolation features. :

[0124] ,

[0125] (4) Scenario 4: Nested coupling of multi-scale dynamic features

[0126] This invention concatenates and reconstructs the different physical scale features derived above, adds them to the original input residual, and outputs the complete multi-scale dynamic features of this layer. :

[0127] ,

[0128] (5) Scenario 5: Echo shape preservation under spatiotemporal smoothing constraints (anti-dissipation scenario)

[0129] This invention achieves global perception of evolutionary dynamics through the Inception group module within the translator. Mathematically, this is equivalent to solving the discrete form of atmospheric dynamics partial differential equations in latent space. It captures the translation vectors guiding airflow through large-scale grouped convolution and captures the nonlinear bursts of local convection through small-scale convolution, thereby ensuring the system's ability to perceive future changes. Within a given time frame, it can still maintain a clear single-unit structure and accurate intensity distribution. This non-autoregressive mapping mechanism effectively maintains the "morphological fidelity" of the strong convective system at the physical level, avoiding the "echo dissipation" problem caused by the rapid attenuation of echo intensity in the later stages of the forecast due to error accumulation, and preventing the location of the disaster area from shifting.

[0130] After the above deduction, the translator Will The frame history state is non-autoregressively mapped to one time. The future hidden state sequence of the frame:

[0131] ,

[0132] This pure spatial convolution operation directly translates states across the time axis, completely avoiding the error accumulation during autoregressive extrapolation in traditional recurrent neural networks, and effectively maintaining the continuity of cloud evolution and the integrity of the landing area structure.

[0133] 3. Decoder Stage: Multi-scale Feature Reconstruction and Pixel-Level Disaster Localization

[0134] The decoder is responsible for deducing the future hidden state sequence. It restores the full-resolution pixel-level forecast product. The decoder internally employs a nested architecture of transposed convolution and skip connections.

[0135] Scenario 6: Multi-scale detail compensation (refined area reconstruction scenario)

[0136] To achieve high-resolution 301×301 pixel service output, the decoder incorporates shallow spatial details from the encoder's initial stages through cross-layer connections during the progressive upsampling process. In meteorological terms, this process achieves a dynamic nesting of "macroscopic dynamic trends" and "local terrain-sensitive features." This multi-scale reconstruction mechanism enables the model's output forecast field to not only possess accurate mesoscale movement trends but also to delineate the "sharp edges" of hail nuclei or thunderstorm winds with pixel-level precision. Through this high-fidelity reconstruction, the model can accurately transform the dynamic predictions of latent space into binary warning areas in a geographic coordinate system, thereby resolving the common boundary ambiguity and location offset problems in multi-class forecasts.

[0137] The technical advantage of this step lies in the decoupling of meteorological physical logic and deep learning operators through the three-level collaboration of the SimVP architecture: the encoder is responsible for initial state field inversion, the translator is responsible for dynamic nonlinear deduction, and the decoder is responsible for high-fidelity reconstruction of the landing area. This link completely establishes an end-to-end computational graph for severe convective weather, from "microscopic structure perception" to "long-range evolution simulation".

[0138] S3.3 Disaster Sequence Decoding and Multi-task Classification Based on ConvLSTM (Decoder & Multi-taskHead)

[0139] The future spacetime characteristics that will be deduced Frame by frame, the data is fed into a ConvLSTM network to extract the long-term temporal dependencies of specific disaster types during their occurrence and dissipation. At any prediction step... The update operator inside ConvLSTM uses convolution ( ) and element-wise multiplication ( This enables dynamic evolution simulation of the complex life cycle of strong convective systems. The local forward extrapolation process is defined by the following physical scenario:

[0140] (1) Scenario 7: Dissipation Convection Single Component Weight Decay (Forget Gate Control Scenario) The system uses the forget gate Automatically identify and attenuate the characteristics of convective systems about to enter the dissipation phase. Physically, when the echo intensity at a historical moment shows a decreasing trend and the satellite cloud top brightness temperature rises, the forgetting gate will reduce the cell state from the previous moment. The weights are used to simulate the energy dissipation process of a convective system:

[0141] ,

[0142] (2) Scenario 8: Energy replenishment of newly formed strong convection cells (input gate control scenario) The system uses the input gate The newly generated convective kernel, deduced from the latent space, is integrated into long-term memory. This operation corresponds to the enhanced updraft characteristics captured by ground observations or satellite data in the early stages of a strong convective eruption, and uses a gating mechanism to superimpose these instantaneous new energies onto the cellular state. middle:

[0143] ,

[0144] ,

[0145] (3) Scenario 9: Spatiotemporal saliency extraction of disaster characteristics (output gate control scenario) The system utilizes the output gate From the current state that incorporates temporal memory, the feature components that are most representative of specific disasters (such as hail nuclei and strong wind gusts) are selected to generate the final hidden state. :

[0146] ,

[0147] ,

[0148] (4) Scenario 10: Multi-scale resolution restoration and physical quantity decoupling (decoding and multi-task scenario) Subsequently, the high-dimensional hidden state that incorporates temporal memory will be integrated. Feed space decoder The decoder, through stacked transposed convolutions and bilinear upsampling layers, is responsible for progressively restoring the feature map size from the latent space scale to the physical observation size, including padding boundaries. :

[0149] ,

[0150] Next, the upsampled full-resolution features The feed consists of 4 groups in parallel A multi-task decoding head composed of convolutions. This parallel design achieves substantial decoupling of the precipitation regression task from the hail, thunderstorm wind, and lightning classification tasks in the feature space, effectively blocking gradient interference between tasks with different physical dimensions. Finally, utilizing... The operator restores the original resolution and outputs the probability matrices for four independent disasters:

[0151] ,

[0152] Channel identification These four tensors, each with a clear physical probability meaning, constitute the final refined multi-class forecast results for severe convection.

[0153] S4 Heterogeneous Truth-Driven End-to-End Joint Loss and Model Optimization

[0154] In multi-hazard severe convective weather forecasting tasks, two major technical barriers are encountered: first, the severe class imbalance of extreme hazards such as hail, thunderstorms, and strong winds in the spatiotemporal grid (i.e., the proportion of positive sample pixels is extremely small); second, there is a severe frequency domain asynchrony between the high-frequency predicted slices output by the model and the true low-frequency GPM precipitation values. To address these challenges, this invention designs a four-dimensional coupled heterogeneous multi-task loss function system. This drives the network to perform end-to-end robust parameter optimization.

[0155] S4.1 Pixel-level Focal Loss Constraints for Extremely Small Sample Disasters

[0156] For hail, thunderstorms, and lightning—three types of disasters that cause severe damage but occur very infrequently—if conventional cross-entropy loss is used, the update of the backward gradient is easily overwhelmed by a massive number of simple, negative background pixels that are "disaster-free." Let's consider a disaster channel. In the predicted time step The actual binary tag matrix is The model outputs the prediction probability matrix as follows: Define the spatial pixel domain as An improved pixel-level binary classification Focal Loss is used to suppress the weights of simple background pixels, forcing the model to focus on extreme, strong convection cores:

[0157] ,

[0158] The parameters are defined as follows:

[0159] : Indicates a specific disaster channel (Such as hail, thunderstorms, strong winds, or lightning) in the predicted time step The pixel-level classification loss value.

[0160] : Represents the set of spatial pixel domains, that is, the two-dimensional geographic grid space output by the model (such as 301×301).

[0161] : Represents the total number of pixels in the spatial pixel domain, used for average normalization of global pixel-level loss.

[0162] : Indicates traversing the grid space Each pixel coordinate set .

[0163] : Represents the class balance prior factor, used to adjust the proportion of positive and negative samples contributing to the total loss. For extreme disasters with severe imbalance, this coefficient is used to compensate for the weight of positive samples.

[0164] : Represents the focusing parameter for mining difficult samples. It is used to suppress the weight contribution of simple background pixels (easy-to-classify samples), forcing the model to focus on the extremely strong convection cores that are difficult to predict and have high error during backpropagation.

[0165] : Indicates at time step Pixel position Disaster type The actual binary tag value (0 indicates no disaster, 1 indicates disaster).

[0166] : Indicates the model at time step Output pixel position Belongs to the category of disaster The predicted probability estimate. The total time-series loss of this channel is... .

[0167] S4.2 Joint Optimization of Asynchronous Heavy Precipitation Across Time Window Regression

[0168] For short-duration heavy precipitation channels verified based on GPM, the mid-time convergence operator in S2.3 is used. Cumulative prediction results Calculate its true GPM value corresponding to the 30-minute spatial resolution. Mean square error (MSE):

[0169] ,

[0170] The mean squared error here directly penetrates the time accumulation operator for gradient backpropagation, realizing implicit supervision of macroscopic low-frequency labels for microscopic high-frequency extrapolation. The parameters are defined as follows:

[0171] : Represents the mean squared error loss (MSE) value for short-duration heavy precipitation missions, used to measure the deviation between the model's predicted cumulative precipitation and the true value observed by satellite.

[0172] : Represents the value extracted from GPM satellite data at the target time. The corresponding 30-minute interval precipitation rate true tensor.

[0173] :express The norm square operation, by calculating the sum of squares of the pixel-by-pixel differences between the predicted tensor and the true tensor, enables sensitive capture of errors in the core area of ​​severe convective precipitation and reverse gradient propagation.

[0174] S4.3 Global Multi-Task Joint Loss and Backpropagation Optimization

[0175] Construct a global joint optimization objective and use dynamic weight hyperparameters. Balancing the convergence rates of the four tasks:

[0176] ,

[0177] The parameters are defined as follows:

[0178] : Represents the total joint loss function when the model is trained globally end-to-end, which serves as the objective function for backpropagation optimization of global parameters.

[0179] : Represents four predefined dynamic weight hyperparameters, corresponding to precipitation regression, hail classification, thunderstorm wind classification and lightning classification tasks, respectively. They are used to balance the convergence speed between different tasks and prevent model training bias due to differences in sample size.

[0180] : Represents the pixel-level focal loss for hail classification tasks, used to improve classification performance when hail samples are extremely scarce.

[0181] : Indicates the pixel-level focus loss in the thunderstorm and gale classification task. By suppressing the weight of background pixels, the detection rate of the core area of ​​the gale is improved.

[0182] : Represents the pixel-level focus loss for lightning classification tasks, enabling fine-grained spatial constraints on the probability of lightning occurrence.

[0183] During network training, the AdamW optimizer (incorporating L2 regularization to prevent overfitting) is employed, combined with a cosine annealing learning rate scheduling strategy. Historical multi-source tensor sequences are used as input, with... End-to-end back-propagation of global parameters is driven by the objective function. The optimal weight model is automatically saved based on the critical success index (CSI) and probe rate (POD) on the validation set.

[0184] S5 Business-Oriented Near Real-Time Inference and Post-Processing Consistency Constraints

[0185] In actual meteorological forecasting operations, for fixed target observation areas, in order to ensure the efficiency and continuity of outputting disaster coverage areas, this invention designs a single forward inference and morphological filtering mechanism.

[0186] S5.1 Single-pass Inference Based on Boundary Padding and Precise Clipping

[0187] Given that the target business area is fixed To avoid the additional memory overhead and computational latency caused by complex sliding window stitching, this invention employs a simplified, single-inference-efficient strategy for spatial resolution. After the service flow acquires the latest radar and satellite data, it directly calls the edge-filling operator in S2.1. Pad the input tensor to the specified size. .

[0188] After the network performs a single forward pass, the reverse center pruning operator is directly called at the output. Remove redundant boundary fill regions:

[0189] ,

[0190] The parameters are defined as follows:

[0191] : Represents the global continuous probability field prediction tensor of the final output of the network, with a spatial resolution of 301×301 pixels, representing the probability distribution of each type of disaster within the target area.

[0192] : indicates a single forward propagation process of the joint spatiotemporal classification prediction model constructed in this invention.

[0193] : indicates the edge padding operator, which pads the original data at the input from 301×301 to the nearest integer multiple of 32 (i.e., 320×320).

[0194] : Represents the input multi-source synchronous meteorological data tensor sequence to be used for near real-time inference.

[0195] : Represents the real tensor space to which the output probability field belongs.

[0196] This mechanism ensures the entire frame The absolute coherence of the probability field in its spatial structure avoids any form of gaps and maximizes the timeliness of nowcasting.

[0197] S5.2 Morphological noise filtering and binarized landing area generation

[0198] The global continuous probability field directly output by the network Perform morphological opening (i.e., erosion followed by dilation) to set the structuring element. :

[0199] ,

[0200] The symbols are defined as follows:

[0201] : Represents the global continuous probability field prediction value output by the model in a single forward inference.

[0202] : Represents the morphological erosion operator, used to reduce the edges of high-value regions in a probability field to peel away discrete noise patches with extremely small surface areas and very short lifespans.

[0203] : Represents the morphological dilation operator, used to restore and smooth the boundary topology of the main convection system after an erosion operation.

[0204] : Represents a predefined morphological structuring element, which determines the sensitivity of the filtering operation to noise size.

[0205] This operation, in a physics sense, filters out discrete noise patches with extremely small areas and very short lifespans without disrupting the boundary topology of the main convection system. Finally, the optimal probability cutoff threshold is determined based on statistical calibration of the validation set. Generate binarized early warning maps of various disaster types. :

[0206] ,

[0207] in This is an indicator function.

[0208] S6 Forecast Uncertainty Estimation and Refined Multi-Hazard Operational Mapping

[0209] To reduce the risk of "overconfidence" in models under complex terrain obstruction or extremely rare weather conditions, and to enhance the engineering interpretability of forecast products, this invention introduces implicit uncertainty quantification assessment in the final mapping stage.

[0210] S6.1 Cell-level confidence estimation based on test-time augmentation (TTA)

[0211] During business inference, a lightweight independent and identically distributed Gaussian random perturbation is applied to the input multimodal historical spatiotemporal tensor sequence. Data generated based on the same moment. One (example) The perturbation input branch with minute changes is executed in parallel with a single inference in S5.1 to obtain the corresponding set of predicted probabilities. .

[0212] Calculate each pixel in Mean under sub-inference with pixel-level standard deviation :

[0213] , ,

[0214] The parameters are defined as follows:

[0215] : Indicates the pixel position Disaster type go through The average predicted probability value is obtained after multiple perturbation inferences. This average value replaces the original single forward propagation result as the final synthetic probability output, significantly enhancing the stability of the forecast.

[0216] : Indicates the pixel position Disaster type The pixel-level standard deviation of the predicted probability. This value directly quantifies the model's prediction uncertainty at this pixel. The larger the standard deviation, the lower the model's prediction confidence at that point.

[0217] : Represents the total number of small perturbation input branches generated based on Test-Time Augmentation (TTA) (i.e., the total number of parallel inference executions, as shown in the example). ).

[0218] : Indicates the perturbation branch index of the currently executing inference, with a value ranging from 1 to 1. .

[0219] : indicates the first In the inference branch with independent and identically distributed Gaussian random perturbations, the model output is located at... Disaster types The predicted probability.

[0220] : Indicates a specific category of severe convective disasters (such as short-duration heavy rainfall, hail, thunderstorms, strong winds, and lightning).

[0221] S6.2 Multi-dimensional refined product synthesis output

[0222] The system ultimately outputs a synchronized spatiotemporally aligned multi-hazard product package, including:

[0223] (1) Probability distribution field: Provides an infinitely continuous probability map of the occurrence of short-term heavy precipitation, hail, thunderstorms and strong winds and lightning within the future forecast period (updated every 6 minutes). .

[0224] (2) Detailed landing area map: Provides the spatial landing area boundary of disaster confirmation after morphological filtering and threshold truncation. .

[0225] (3) Uncertainty confidence map: Provides an uncertainty distribution map generated based on TTA. This provides reliable quantitative references for disaster prevention decision-makers when faced with discrepancies in model forecasts, effectively avoiding excessive emergency response caused by false alarms.

[0226] Reference Figure 2 This embodiment details the execution path of the data preprocessing stage. The system first receives GPM precipitation data, satellite observation data, and ground-based radar data. It aligns data from different projected coordinate systems using a spatial resampling operator and addresses the size adaptation issue during network computation using edge padding and center clipping operators. In the time dimension, the system executes timestamp alignment logic to find the optimal matching frames between radar and satellite data, and uses a time-dimensional accumulation operator to resolve the frequency domain heterogeneity issue between low-frequency GPM labels and high-frequency model predictions. For the sample set, the system filters out complete strong convection lifecycle events based on duration (ΔT) and performs sliding window segmentation and stratified sampling, ultimately constructing a high-quality training and validation base.

[0227] Reference Figure 3 This embodiment demonstrates the core components of a joint spatiotemporal classification and forecasting network. The left side of the model is a cross-modal feature fusion module. Radar and satellite data are processed by independent residual encoders to extract spatial features, which are then adaptively fused using a channel attention mechanism. The middle part of the model is a SimVP-based latent space translator, internally composed of multiple layers of Inception convolutional modules. Grouped convolutions reduce the number of parameters while capturing multi-dimensional evolutionary dynamics. The right side of the model is the decoder. The derived latent state sequence is processed by ConvLSTM to extract long-term temporal dependencies. Finally, a multi-task decoder head decouples the output for short-term heavy precipitation regression, hail, thunderstorm wind, and lightning classification tasks.

[0228] Reference Figure 4This embodiment demonstrates the model's synthetic forecast performance for short-duration heavy precipitation. The figure compares the actual GPM observation field with the model's predicted field. Through the time-dimensional accumulation operator and MSE loss constraint described in this invention, the predicted field maintains a high degree of consistency with the satellite ground truth in terms of precipitation intensity distribution, location of heavy precipitation centers, and morphological evolution of precipitation bands, effectively capturing the macroscopic dynamic trends of strong convective systems.

[0229] Reference Figure 5 This embodiment demonstrates the refined classification and forecasting results for thunderstorm and strong wind disasters. The figure compares the actual disaster labels with the binarized landing area map generated by the model. Thanks to the improved pixel-level focal loss that suppresses simple background pixels, the model can still accurately identify the core triggering area of ​​strong wind disasters even when positive samples are extremely scarce. The landing area edges are sharp and there are no obvious false alarm patches, proving the robustness of this method in handling extremely small sample disaster types.

[0230] Experimental verification

[0231] This embodiment aims to verify the effectiveness of the proposed method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data through specific comparative and ablation experiments.

[0232] 1. Experimental Dataset Construction and Parameter Setting

[0233] Based on historical severe convective events in the target sea area and coastal region, those consistent with life history were selected. The event sequence is based on a threshold condition. To ensure spatial independence and generalization ability verification, stratified sampling is strictly performed according to province. The AdamW optimizer is used during training, with an initial learning rate set to... It also incorporates a cosine annealing strategy for dynamic decay. The batch size is set to 16.

[0234] 2. Evaluation Metrics

[0235] For precipitation regression tasks, root mean square error (RMSE) and mean absolute error (MAE) are used for evaluation. For refined classification tasks of hail, thunderstorm winds, and lightning, due to the extreme imbalance between positive and negative samples, the meteorological standard Critical Success Index (CSI), Probability of Detection (POD), and False Alarm Rate (FAR) are used for comprehensive evaluation. The number of positive samples (Hits) of correctly predicted disasters. These are false alarms. The specific formula for the number of missed reports (Misses) is defined as follows:

[0236] ,

[0237] ,

[0238] ,

[0239] 3. Comparison and Ablation Experimental Design (Ablation Studies)

[0240] To fully verify the multi-source data fusion strategy and the independent contribution of each component in the network architecture, this invention designed the following five sets of progressive comparative experiments:

[0241] Experiment 1 (Single Radar Baseline Model): Only multi-channel radar sequence data (combined reflectivity, radial velocity, etc.) are input; satellite data is not input. Spatiotemporal extrapolation is performed using a conventional ConvLSTM architecture. This experiment aims to establish a basic performance baseline for deep learning classification and recognition relying solely on single-dimensional radar echo data.

[0242] Experiment 2 (Early Fusion Data-Level Alignment Model): Radar and satellite data were used, but only physical space grid resampling and coarse temporal window matching were performed before inputting into the network (Early Fusion). Features were directly concatenated along the channel dimension and input into a regular temporal network without going through a dedicated cross-modal encoder and latent space SimVP inference.

[0243] Experiment 3 (Feature-Level Synchronous Fusion Network - Invention Architecture but Without Focal Loss): This experiment employs the cross-modal feature fusion model designed in this invention (including a fusion operator based on channel attention) and a joint architecture of SimVP and ConvLSTM. However, in the classification multi-task decoding stage, only the conventional cross-entropy loss function is used to evaluate the feature extraction capability of the network architecture itself.

[0244] Experiment 4 (Pure Satellite End-to-End Baseline Validation): To rigorously explore the independent representation potential of geostationary meteorological satellites under complex disaster-causing mechanisms, this experiment used only 12 spectral channels of the Himawari-8 (H8) satellite as a single input source. In terms of training strategy, no external source data was used for model pre-training; end-to-end training from scratch was performed directly on the target severe convective event dataset. This experiment stripped away radar structural features, focusing on validating the model's ability to extract convective evolution trends from pure infrared and visible light sequences.

[0245] Experiment 5 (Complete Method of the Invention): Integrating all the technical features of the present invention, including rigorous temporal registration, cross-modal feature fusion, SimVP latent space inference, and pixel-level focal loss (FocalLoss) introduced for extreme small sample disasters.

[0246] 4. Analysis of Experimental Results and Technical Effects

[0247] Experimental results show that Experiment 1 performed reasonably well in short-term heavy precipitation prediction, but its CSI score was extremely low in classifying associated disasters such as hail and thunderstorms, and its FAR false alarm rate was high, confirming the information representation bottleneck of a single radar reflectivity feature. Experiment 2 introduced satellite data, but due to the lack of deep feature-level decoupling and alignment, cross-modal differences led to unstable network training gradients, resulting in limited improvement in classification accuracy and even offsets in the localization of some high-frequency landing areas. Experiment 4, without relying on external pre-training or radar assistance, successfully captured the long-range advection evolution characteristics of cloud tops solely based on the thermodynamic prior information from H8 satellite data. In forecasts with a lead time of more than 60 minutes, its large-scale system movement trend score was better than Experiment 1, demonstrating the independent value of end-to-end training with satellite data. Experiment 5 (this invention) achieved the best results in all indicators. Figure 3 The joint spatiotemporal classification forecasting model architecture shown in this invention fully leverages the advantages of cross-modal feature fusion and nonlinear spatiotemporal extrapolation. Compared to Experiment 3, by introducing an improved pixel-level binary classification Focal Loss to suppress the weights of simple background pixels, the model significantly improves the detection rate (POD) of small-sample disaster cores with extremely low occurrence frequencies, such as hail and strong winds, by a significant percentage point, completely solving the problem of high underreporting rates for extreme disasters. Figure 5 The comparison chart of thunderstorm and gale forecast results is shown below. Figure 5 In the text, (a) represents the truth value of thunderstorm winds. Figure 5 In the diagram, (b) represents the model's predicted value. The predicted landing area output by this invention not only accurately pinpoints the disaster core but also has sharp edges, closely matching the true labels. Simultaneously, the joint framework of SimVP and ConvLSTM effectively maintains the fidelity of the landing area morphology, such as... Figure 4 The short-duration heavy precipitation (GPM) forecast results are shown in the figure. Figure 4 In this context, (a) represents the truth value of GPM. Figure 4 In the diagram, (b) represents the model's predicted value. The model's predicted field maintains a high degree of consistency with the satellite's true value in terms of precipitation intensity distribution and morphological evolution, demonstrating the breakthrough of this invention in the collaborative optimization of multi-source cross-frequency meteorological data.

[0248] Example 2

[0249] This embodiment provides a refined classification and forecasting system for severe convective weather that integrates geostationary satellite data.

[0250] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data.

[0251] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a refined classification and forecasting method for severe convective weather that integrates geostationary satellite data.

[0252] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data, characterized in that, include: Acquire multi-source, multimodal meteorological data; Data preprocessing is performed based on the acquired multi-source, cross-modal meteorological data; Construct a joint spatiotemporal classification and forecasting network model that integrates geostationary satellite data, including building a cross-modal feature fusion model, latent space nonlinear spatiotemporal extrapolation based on SimVP, and disaster sequence decoding and multi-task classification based on ConvLSTM; Heterogeneous truth-driven end-to-end joint loss and model optimization are applied to the constructed network model. Post-processing consistency constraints are applied to the optimization results based on a single forward inference and morphological filtering mechanism. Output the prediction results; The construction of the cross-modal feature fusion model includes, for heterogeneous inputs that are only physically concatenated at the data level, to eliminate differences in physical dimensions and extract deep joint representations, constructing a cross-modal feature fusion model, decoupling the radar channel and satellite channel in the input sequence, and feeding them separately into a two-dimensional convolutional spatial encoder with independent weights. and Extract their respective high-dimensional spatial representations; to avoid gradient vanishing of meteorological microscale features when the network is deepened, the encoder internally constructs... The residual convolutional modules are stacked in layers, specifically, for the ... Layer residual module Let its input feature matrix be Feature mapping operator in residual branch The residual mapping forward propagation process, which includes 2D convolution (Conv2D), batch normalization (BN), and LeakyReLU activation functions, is defined as follows: ,in Indicates the first The feature matrix output by the layer residual module; This represents a nonlinear activation function for a modified linear unit with leakage. Represents the feature mapping operator in the residual branch; Indicates the first The input feature matrix of the layer residual module; Indicates the first Layer feature mapping operator The learnable weight parameter matrix corresponding to the internal convolutional layers; : Represents the element-wise addition operation of tensors. Skip connections enable the network to preserve the sharpness of strong convection unit edges during downsampling. The encoder ultimately extracts the multimodal high-dimensional representation as follows: , ,in, The number of latent space feature channels. and Let the size of the downsampled and compressed feature map be defined; then, a fusion operator based on the channel attention mechanism is defined. Two high-dimensional features are spliced ​​together at the channel dimension. Then, global average pooling is used. By compressing the spatial dimension, channel statistics of the global receptive field are obtained; to capture the nonlinear interactions between channels, these statistics are fed into a multilayer perceptron (MLP) network, thereby generating the final channel weighted vector. : Finally, the adaptive weighted vector and the concatenated features are multiplied by the Hadamard product. To achieve substantial cross-modal feature-level fusion: Features after fusion It fully combines the radar's microstructure sensitivity with the satellite's thermodynamic priors.

2. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 1, characterized in that, The data preprocessing based on the acquired multi-source cross-modal meteorological data includes, after acquiring the multi-source heterogeneous meteorological data, employing a spatial resampling algorithm and a dynamic time window convergence mechanism to address the severe asynchronous heterogeneity issues in physical dimensions, spatial resolution, and temporal sampling frequency between ground-based observations, space-based observations, and external verification true values. Firstly, refined spatial matching is performed on the acquired 8 types of radar channel data and 12 types of satellite channel data. Let the original radar image be... Satellite imagery is Using bilinear interpolation operators Both are resampled and aligned to the same target 2D mesh. Specifically, let the four adjacent known grid points on the original image to be processed be... Its corresponding data value is For any floating-point coordinate on the target grid Bilinear interpolation is calculated as follows: , , For the center of each target pixel : After unified projection, the spatial dimension is extracted according to the actual coverage area. The local sub-block tensor of a pixel; to accommodate the integer division characteristics of multi-level downsampling operations in subsequent spatiotemporal depth networks, an edge padding operator is defined before the data is input into the network. ,Will The feature map is expanded to a size that is an integer multiple of 32 in the nearest neighborhood, i.e. : The number of radar channels Number of satellite channels Simultaneously, a standardization operator is defined to normalize the maximum and minimum values ​​of the data in each channel, mapping them to... The interval is defined, and the reverse center clipping operator is predefined. Size restoration for network decoding output: ,in, This indicates the reverse center clipping operator; : Represents the predicted feature map; This represents the feature map or result tensor of the final output; Represents the target tensor space of the mapping.

3. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 2, characterized in that, The data preprocessing based on the acquired multi-source cross-modal meteorological data also includes temporal registration and physical stitching of the input multi-source observation data. It addresses the asynchronous sampling problem between radar and satellite by performing strict time-dimension alignment and data-level stitching for any radar time point. Find the satellite frame with the closest time distance for timestamp registration. : ,in, Represents the optimization operator; This represents a set of timestamp sequences from geostationary meteorological satellite observations. Represents the set of time series data from geostationary meteorological satellites Any candidate sampling timestamp in the dataset; Indicates the timestamp of ground-based radar observation; This represents the absolute time difference between the radar target timestamp and the satellite candidate timestamp, setting a time error tolerance to ensure the continuity of radar timing. When satisfied At that time, the radar tensor and the registered satellite tensor are concatenated along the input channel dimension to form the synchronous multimodal physical input frame for the current moment: and extract history Each time step constructs a strictly continuous historical spatiotemporal tensor sequence as input to the subsequent network: Then, to address the severe heterogeneity between the model's predicted high-frequency sequences and the true labels of low-frequency precipitation, the model is assumed to continuously output... Precipitation forecast slice sequence at 6-minute intervals ,in To be in line with the target time True value of 30-minute GPM precipitation Strict alignment; define the time-dimensional accumulation operator before loss calculation. : Finally, to eliminate transient noise interference from processes with short lifespans and lacking complete evolutionary cycles, let the duration of a strong convective process on the time axis be... Calculate the total duration of its life history. Set a minimum lifespan time threshold Define the filtering indicator function for sample set construction. : ,like If the event has a complete and continuous dynamic development process, its corresponding multi-source synchronous data sequence and label pair will be formally included in the final training and validation sample set; if If the data is not included in the dataset, it will be considered as an occasional isolated convection, a system that dissipates too quickly, or observational noise, and will not be included in the dataset.

4. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 3, characterized in that, The SimVP-based latent space nonlinear spatiotemporal extrapolation includes fusing cross-modal feature sequences. The input is fed into the SimVP core inference module, which constructs a three-level collaborative architecture of encoder, spatiotemporal translator, and decoder to achieve a nonlinear mapping from the physical observation field to the future dynamic evolution trend. In the encoder stage, spatial collapse and physical consistency mapping of heterogeneous features are performed, and the encoder fuses feature sequences across modalities. As input, physical feature distillation is performed through a convolution operation with a stride of 2. The calculation process is as follows: This maps features to a sequence of latent variables. In the translator stage, non-autoregressive spatiotemporal manifold extrapolation and multi-scale scene decoupling are performed. To deeply adapt to the multi-scale, highly nonlinear dynamic evolution characteristics of severe convective weather, the latent space extrapolation process undergoes physical scene decoupling and reconstruction. Let the spatiotemporal translator within this module be... Unlike traditional recurrent neural networks (RNNs), it employs a fully convolutional non-autoregressive architecture, performing end-to-end time-axis flipping on the latent space manifold. The translator consists of stacked latent space Inception convolutional modules, which, for the input latent space features... The local forward inference process is defined as several physical scenarios: Scenario 1: Cloud base-cloud top thermodynamic cross-channel interaction. First, it uses... Convolution performs cross-channel information interaction and dimensionality reduction, which physically corresponds to the nonlinear coupling between radar hydrophobic distribution and satellite blackbody brightness temperature, and is defined as deep convection core feature mapping. : Scenario 2: Local microscale convection generation and dissipation simulation, utilizing... Grouped convolution focuses on small local receptive fields, defined as microscale evolutionary features. : Scenario 3: Advection extrapolation of mesoscale convective systems, utilizing... Grouped convolutions acquire the global spatial receptive field, defined as mesoscale advection inference features. : Scenario 4: Nested coupling of multi-scale dynamic features. The derived features at different physical scales are concatenated and reconstructed, and then added to the original input residual to output the complete multi-scale dynamic features of this layer. : Scenario 5: Echo morphology preservation under spatiotemporal smoothing constraints, achieving global perception of evolutionary dynamics through the Inception module within the translator; After the above deduction, the translator... Will The frame history state is non-autoregressively mapped to one time. The future hidden state sequence of the frame: Finally, in the decoder stage, multi-scale feature reconstruction and pixel-level disaster localization are performed, and the future hidden state sequence is deduced based on the decoder. Restored to full-resolution pixel-level forecast products.

5. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 4, characterized in that, The ConvLSTM-based disaster sequence decoding and multi-task classification includes the deduced future spatiotemporal features. The data is fed frame by frame into a ConvLSTM network to extract the long-term temporal dependencies of specific disaster types during their occurrence and dissipation. This data is then applied at any prediction step. The update operator inside ConvLSTM uses convolution. element-wise multiplication To achieve dynamic evolution simulation of the complex life history of strong convective systems, the local forward extrapolation process is defined as several physical scenarios: the weight decay of dissipating convective cells, through a forgetting gate. The system automatically identifies and attenuates the characteristics of convective systems that are about to enter the dissipation phase. When the echo intensity at a historical moment shows a decreasing trend and the satellite cloud top brightness temperature rises, the forgetting gate will reduce the cell state at the previous moment. The weights are used to simulate the energy dissipation process of a convective system: ; Energy replenishment for newly formed strong convective cells, via the input gate Integrating the newly generated convective core deduced from the latent space into long-term memory corresponds to the enhanced updraft characteristics captured by ground observations or satellite data in the early stages of a strong convective eruption. Through a gating mechanism, these instantaneous new energies are superimposed onto the cellular state. middle: , Disaster type characteristics spatiotemporal saliency extraction, using output gate From the current state of fused temporal memory, the feature components that are most representative of a specific disaster are selected to generate the final hidden state. : , Multi-scale resolution recovery and physical quantity decoupling will integrate the high-dimensional hidden states of temporal memory. Feed space decoder The decoder, through stacked transposed convolutions and bilinear upsampling layers, is responsible for progressively restoring the feature map size from the latent space scale to the physical observation size, including the padding boundaries. : Finally, the upsampled full-resolution features are... The feed consists of 4 groups in parallel A multi-task decoding head composed of convolutions, utilizing The operator restores the original resolution and outputs the probability matrices for four independent disasters: Channel identifier The tensor of the probability matrix of these four independent disaster types is the final refined multi-class forecast result of severe convection.

6. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 5, characterized in that, The process of applying heterogeneous truth-based end-to-end joint loss and model optimization to the constructed network model includes building a four-dimensional coupled heterogeneous multi-task loss function system. To drive the network to perform end-to-end robust parameter optimization, a disaster channel is first established. In the predicted time step The actual binary tag matrix is The model outputs the prediction probability matrix as follows: Define the spatial pixel domain as An improved pixel-level binary classification Focal Loss is used to suppress the weights of simple background pixels, forcing the model to focus on extreme, strong convection cores: ,in, Indicates a specific disaster channel In the predicted time step Pixel-level classification loss value; Represents the set of spatial pixel fields; : Represents the total number of pixels within the spatial pixel domain; Indicates traversing the grid space Each pixel coordinate set ; Represents the class-balanced prior factors; The modulation coefficients represent those used in hard sample mining. : Indicates at time step Pixel position Disaster type The actual binary tag value; Indicates the model at time step The output is the predicted probability estimate; then, for short-duration heavy precipitation channels validated by GPM, the time-converged operator is used. Cumulative prediction results Calculate its true GPM value corresponding to the 30-minute spatial resolution. Mean square error (MSE): ,in, : Represents the value extracted from GPM satellite data at the target time. The corresponding 30-minute interval precipitation rate true tensor; express The norm square operation is performed, and finally a global joint optimization objective is constructed using dynamic weight hyperparameters. Balancing the convergence rates of the four tasks: ,in, This represents the total joint loss function when the model is trained globally end-to-end. This represents four predefined dynamic weight hyperparameters, corresponding to precipitation regression, hail classification, thunderstorm wind classification, and lightning classification tasks, respectively.

7. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 6, characterized in that, The post-processing consistency constraint based on the single-pass forward inference and morphological filtering mechanism includes a simplified, efficient single-pass inference strategy to avoid the additional memory overhead and computational latency caused by complex sliding window stitching. After the business flow acquires the latest radar and satellite data, the edge-filling operator is used. Pad the input tensor to the specified size. After the network performs a single forward pass, the reverse center pruning operator is directly called at the output. Remove redundant boundary fill regions: ,in, This represents the tensor of the global continuous probability field predictions that the network will ultimately output. This represents a single forward propagation process of the joint spatiotemporal classification prediction model; Indicates the edge fill operator; Represents a tensor sequence of multi-source synchronous meteorological data; ensures the entire image... The absolute coherence of the probability field in its spatial structure avoids any form of gaps, maximizing the timeliness of nowcasting; then, the global continuous probability field directly output by the network... Perform morphological opening operation and set the structuring element. : ,in, This represents the global continuous probability field prediction value output by the model in a single forward inference. Represents the morphological erosion operator. Represents the morphological dilation operator. Represents predefined morphological structural elements; finally, the optimal probability truncation threshold is determined based on statistical calibration of the validation set. Generate binarized early warning maps of various disaster types. : ,in This is an indicator function.

8. The method for refined classification and forecasting of severe convective weather by fusing geostationary satellite data according to claim 7, characterized in that, The output prediction results include an implicit uncertainty quantification assessment introduced in the final mapping stage to reduce the risk of overconfidence in the model under complex terrain occlusion and extreme weather conditions. First, during business inference, a lightweight independent and identically distributed Gaussian random perturbation is applied to the input multimodal historical spatiotemporal tensor sequence. Data generated based on the same moment Each perturbation input branch with a tiny change is used to perform a single inference in parallel to obtain the corresponding set of predicted probabilities. Then calculate each pixel in Mean under sub-inference with pixel-level standard deviation : , ， in Indicates the pixel position Disaster type go through The average predicted probability value obtained after multiple perturbation inferences; Indicates the pixel position Disaster types The pixel-level standard deviation of the predicted probability; This represents the total number of small perturbation input branches generated based on test-time enhancements; This represents the index of the perturbation branch currently being executed; Indicates the first The model output in the inference branch with independent and identically distributed Gaussian random perturbations is located at Disaster types The predicted probability; This indicates a specific category of severe convective weather-related disasters.

9. A refined classification and forecasting system for severe convective weather that integrates geostationary satellite data, executing the refined classification and forecasting method for severe convective weather that integrates geostationary satellite data as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire multi-source, cross-modal meteorological data. The preprocessing module is configured to perform data preprocessing based on the acquired multi-source cross-modal meteorological data; The model building module is configured to build a joint spatiotemporal classification and forecasting network model that integrates geostationary satellite data, including building a cross-modal feature fusion model, latent space nonlinear spatiotemporal extrapolation based on SimVP, and disaster sequence decoding and multi-task classification based on ConvLSTM. The model optimization module is configured to perform heterogeneous truth-driven end-to-end joint loss and model optimization on the constructed network model; The constraint module is configured to perform post-processing consistency constraints on the optimization results based on a single forward inference and morphological filtering mechanism. The prediction module is configured to output prediction results.

Images