A convection nowcast system and method based on multi-modal data

Abstract

This application proposes a convective nowcasting system and method based on multimodal data, belonging to the field of convective nowcasting technology. This invention fundamentally solves the problem of fusion distortion caused by physical conflicts in multi-source observation data, achieving end-to-end optimization from "raw multimodal input" to "high-quality forecast output." Compared with existing technologies, this invention significantly improves the ability to capture the initial stage of convection, enhances the predictive stability of strong echo structures in the mature stage, and effectively eliminates false signals in the dissipation stage, thereby greatly improving the accuracy, physical rationality, and timeliness of the 0–2 hour radar reflectivity factor prediction field. This not only enhances the lead time and hit rate of short-term severe convective warnings but also provides more reliable technical support for disaster prevention and mitigation decision-making, possessing outstanding substantive features and significant progress.

Description

Technical Field

[0001] This invention relates to the field of convective nowcasting technology, and in particular to a convective nowcasting system and method based on multimodal data. Background Technology

[0002] Convective nowcasting aims to predict the evolution of severe convective systems within the next 0–2 hours and is the core of short-term meteorological services. While current methods attempt to integrate multi-source information such as weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground-based automatic weather station observations to improve forecasting capabilities, in practical applications, different modal data often exhibit contradictory physical states at the same spatiotemporal location due to differences in observation principles, spatiotemporal resolution, and error characteristics—for example, during the initial stage of convection, satellites show rapid cloud top cooling while radar echoes are absent, or during the dissipation stage, strong radar echoes remain while satellite brightness temperatures have risen. If such physical conflicts are used directly in modeling without processing, significant noise and bias will be introduced. However, existing technologies generally lack a systematic identification and collaborative correction mechanism for such physical conflicts. This leads to the following: the multimodal data cube before correction contains a large amount of inconsistent information, causing the feature maps extracted by the subsequent multi-branch encoder to be distorted and difficult to accurately represent the real convection structure; at the same time, due to the lack of a reliable basis for local data quality assessment, it is impossible to generate effective modal confidence weights, resulting in a blind feature fusion process, which ultimately seriously affects the accuracy and physical rationality of the radar reflectivity factor prediction field output by the spatiotemporal prediction decoder, thereby significantly reducing the reliability and availability of convection nowcasting. Therefore, there is an urgent need for a technical solution for a convection nowcasting system and method based on multimodal data. Summary of the Invention

[0003] To address the aforementioned technical problems, embodiments of the present invention provide a convection nowcasting method based on multimodal data, specifically comprising the following steps: S1. Obtain multi-source observation data of the target area at the current time and at several historical times; S2. Resample the multi-source observation data to a preset unified spatiotemporal grid to generate a multimodal synchronous data cube. The multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid. Based on preset atmospheric convection physics rules, perform condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube to obtain the corrected modal data at each location. The modal data includes: weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data; S21. Based on the physical correlation between weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data and ground automatic weather station observation data, construct a multimodal consistency criterion; S211. Based on the time lag characteristics of cloud top radiative cooling and precipitation echo occurrence in the initial stage of convection, the following conditions are used as the criteria for the initial potential of convection: For any location, if the cooling rate of two consecutive time steps in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is greater than the preset first threshold, and the reflectivity factor in the reflectivity factor data of the weather radar at the current location is less than the preset second threshold. S212. Based on the constraint effect of boundary layer thermal instability on convection triggering, the following condition is used as the thermodynamic criterion for convection triggering: For any location, if the dew point temperature in the observation data of the automatic weather station at the current location is higher than the preset third threshold, and the lifting condensation height calculated based on the temperature and dew point temperature in the observation data of the automatic weather station at the current location is lower than the preset fourth threshold. S213. Combine the convective initiation potential criterion with the convective triggering thermodynamic criterion to form a multimodal consistency criterion; S22. Based on the multimodal consistency criterion, perform joint condition judgment on the modal data at each position in the multimodal synchronization data cube to identify modal data with physical conflicts. S221. Traverse the modal data at each position in the multimodal synchronization data cube; S222. For the modal data at the current position, determine whether the convection initiation potential criterion and the convection triggering thermodynamic criterion are satisfied simultaneously. S223. If the modal data at the current position does not meet the convection inception potential criterion or the convection triggering thermodynamic criterion, then the modal data at the current position is determined to have a physical conflict, and the modal data at the current position is marked as modal data with a physical conflict; otherwise, the modal data at the current position is determined not to have a physical conflict. S23. For modal data with physical conflicts, based on the prior classification of atmospheric convection development stages, select a preset correction strategy to numerically correct the modal data with physical conflicts, and generate corrected modal data for each location. S231. Based on the spatiotemporal evolution characteristics of modal data at the current location, the atmospheric convection development stages to which the modal data with physical conflicts at the current location belong are classified as the initial stage of convection, the mature stage of convection, and the dissipation stage of convection. S232. When the classification result is the initial stage of convection, the preset first correction strategy is selected to numerically correct the modal data with physical conflicts, and the corrected modal data at each position is generated. S2321. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, the target reflectivity factor is calculated based on the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the reflectivity factor in the weather radar reflectivity factor data at the current location is replaced with the target reflectivity factor. At the same time, the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location is replaced with the arithmetic mean of the infrared brightness temperature at the current time step and the previous time step, and the corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location are generated respectively. S2322. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the humidity support coefficient based on the dew point temperature and rising condensation height in the automatic weather station observation data at the current location, and replace the reflectivity factor in the weather radar reflectivity factor data at the current location with the humidity support coefficient. At the same time, replace the dew point temperature in the automatic weather station observation data at the current location with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and replace the temperature in the automatic weather station observation data at the current location with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. S2323. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is compensated for based on the temperature in the observation data of the ground automatic weather station at the current location. The temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step. At the same time, the dew point temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step. Corrected infrared brightness temperature data of the geostationary meteorological satellite at the current location and observation data of the ground automatic weather station are generated respectively. S2324. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, an initial reflectivity factor is obtained based on the cooling rate mapping in the geostationary meteorological satellite infrared brightness temperature data at the current location. Then, the initial reflectivity factor is modulated a second time based on the dew point temperature in the ground automatic weather station observation data at the current location to update the reflectivity factor in the weather radar reflectivity factor data at the current location. At the same time, the residual between the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location and the temperature in the ground automatic weather station observation data at the current location is calculated, and half of the residual is added to the infrared brightness temperature and temperature respectively to achieve consistency adjustment. Then, the dew point temperature in the ground automatic weather station observation data at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and the corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location are generated respectively. S233. When the classification result is the convective maturity stage, select the preset second correction strategy to numerically correct the modal data with physical conflicts and generate the corrected modal data at each position. S2331. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the covariance between the reflectivity factor in the weather radar reflectivity factor data at the current location and the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and perform a proportional scaling correction on the reflectivity factor based on the covariance, while performing a reverse scaling correction on the infrared brightness temperature by the same proportion, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location respectively. S2332. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the residual between the reflectivity factor in the weather radar reflectivity factor data at the current location and the lifting condensation height in the automatic weather station observation data at the current location, and perform linear correction on the reflectivity factor based on the residual. At the same time, replace the temperature in the automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. S2333. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground-based automatic weather station, calculate the linear regression slope between the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location and the temperature in the observation data of the ground-based automatic weather station at the current location, and perform a smooth correction on the infrared brightness temperature based on the slope. At the same time, replace the temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground-based automatic weather station at the current location, respectively. S2334. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, calculate the joint covariance among the reflectivity factor in the weather radar reflectivity factor data at the current location, the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the dew point temperature in the ground automatic weather station observation data at the current location. Based on the joint covariance, normalize and correct the reflectivity factor, and correct the cooling rate by covariance projection. Then, map the corrected cooling rate back to the infrared brightness temperature to update the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location, respectively. S234. When the classification result is the convection dissipation period, select the preset third correction strategy to numerically correct the modal data with physical conflicts and generate the corrected modal data for each location. S2341. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps. Construct an exponential attenuation coefficient based on the ratio of the attenuation slope to the recovery slope, and use the exponential attenuation coefficient to attenuate and correct the reflectivity factor at the current location. At the same time, replace the infrared brightness temperature at the current location with the weighted average of the infrared brightness temperature at the current time step and the infrared brightness temperature at the previous time step, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location, respectively. S2342. When a conflict occurs between weather radar reflectivity factor data and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Based on the difference between the two slopes, perform linear compensation correction on the reflectivity factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected weather radar reflectivity factor data and ground automatic weather station observation data at the current location respectively. S2343. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, calculate the rise slope of the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location over the past two consecutive time steps, and calculate the change slope of the temperature in the observation data of the ground automatic weather station at the current location over the past two consecutive time steps. Correct the infrared brightness temperature based on the residual of the two slopes. At the same time, replace the temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the temperature of the current time step and the temperature of the previous time step, and replace the dew point temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the dew point temperature of the current time step and the dew point temperature of the previous time step. Generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground automatic weather station at the current location, respectively. S2344. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps, and the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Construct a proportional coordination factor based on the three slopes, and use the proportional coordination factor to perform nonlinear attenuation correction on the reflectivity factor and linear recovery correction on the infrared brightness temperature based on the proportional coordination factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location, respectively. S3. Input the corrected modal data at each position into the multi-branch heterogeneous encoder to extract the feature map at each position; S4. Obtain at least three feature indicators for the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. S5. Using the local confidence weights of the modal data at each location after correction, the feature maps at each location are weighted and adaptively fused to generate a fused feature tensor. The fused feature tensor is then input into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours.

[0004] This embodiment also discloses a convection nowcasting system based on multimodal data, including the following modules: Data acquisition module: used to acquire multi-source observation data of the target area at the current time and at several historical times; Data correction module: connected to the data acquisition module, used to resample multi-source observation data to a preset unified spatiotemporal grid, generate a multimodal synchronous data cube, the multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid, and performs condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube based on preset atmospheric convection physics rules to obtain the corrected modal data at each location; Extraction module: Connected to the data correction module, it is used to input the corrected modal data at each position into the multi-branch heterogeneous encoder and extract the feature map at each position; Calculation module: Connected to the extraction module, it is used to obtain at least three feature indicators of the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. Prediction module: Connected to the calculation module, it is used to perform weighted adaptive fusion of the feature maps of each location using the local confidence weights of the modal data at each location after correction, generate a fused feature tensor, and input the fused feature tensor into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours.

[0005] The embodiments of the present invention have the following technical effects: This invention significantly improves the physical consistency and reliability of input data by establishing a multimodal physical conflict identification and stage-adaptive collaborative correction mechanism for the entire life cycle of convection. In the initial stage of convection, it effectively strengthens the multi-source consistency of weak convection signals, significantly reducing the false alarm rate. In the mature stage of convection, it accurately maintains the integrity and intensity authenticity of strong echo structures, avoiding excessive smoothing or distortion. In the dissipation stage of convection, it effectively suppresses radar residual false echoes, improving the predictability of the dissipation process. This correction mechanism ensures high physical consistency of the multimodal data cube input to the multi-branch heterogeneous encoder, significantly improving the feature map's ability to represent the real convection structure. Simultaneously, the local credibility weights calculated based on the corrected data enable the feature fusion process to be adaptive and reliable, avoiding blind weighting. Ultimately, the radar reflectivity factor prediction field output by the spatiotemporal prediction decoder is more consistent with atmospheric physical laws in terms of location, morphology, and intensity evolution. In summary, this invention fundamentally solves the problem of fusion distortion caused by physical conflicts in multi-source observation data, achieving end-to-end optimization from "raw multimodal input" to "high-quality forecast output." Compared to existing technologies, this invention significantly improves the ability to capture the initial stage of convection, enhances the predictive stability of strong echo structures in the mature stage, and effectively eliminates false signals in the dissipation stage, thereby greatly improving the accuracy, physical rationality, and timeliness of the 0–2 hour radar reflectivity factor prediction field. This not only enhances the lead time and hit rate of short-term severe convection warnings but also provides more reliable technical support for disaster prevention and mitigation decision-making, demonstrating outstanding substantive features and significant progress. Attached Figure Description

[0006] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0007] Figure 1 This is a flowchart of a convection nowcasting method based on multimodal data provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a convection nowcasting system based on multimodal data provided in an embodiment of the present invention. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0009] Example 1: As Figure 1 As shown, the present invention provides a convection nowcasting method based on multimodal data, comprising the following steps: S1. Obtain multi-source observation data of the target area at the current time and at several historical times; It is worth noting that multi-source observation data of the target area at the current time and several historical times are obtained from the meteorological operational system. This multi-source observation data includes weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground-based automatic weather station observation data. Specifically, weather radar data is obtained through a networked radar mosaic product, geostationary meteorological satellite infrared brightness temperature data comes from the infrared channel of geostationary orbit satellites, and ground-based automatic weather station observation data includes elements such as temperature and dew point temperature. These three types of observation data correspond to three modalities in subsequent processing: weather radar reflectivity factor data as the first modality, geostationary meteorological satellite infrared brightness temperature data as the second modality, and ground-based automatic weather station observation data as the third modality, collectively forming the basic input for multi-modal fusion.

[0010] S2. Resample the multi-source observation data to a preset unified spatiotemporal grid to generate a multimodal synchronous data cube. The multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid. Based on preset atmospheric convection physics rules, perform condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube to obtain the corrected modal data at each location. The modal data includes: weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data; It is worth noting that the aforementioned multi-source observation data were spatiotemporally aligned separately. First, a predefined unified spatiotemporal grid was defined with a spatial resolution of 1 km × 1 km and a time step of 6 minutes. Then, the weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data were resampled to this unified spatiotemporal grid using bilinear interpolation. The ground automatic weather station observation data were interpolated to each grid point using the inverse distance weighting method. After all modes were aligned at the same time step and spatial location, they were stacked along the time dimension to form a three-dimensional multimodal synchronous data cube. Each spatiotemporal location in the cube contains data from three channels, corresponding to three modalities: weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, providing structured input for subsequent physical consistency correction.

[0011] S21. Based on the physical correlation between weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data and ground automatic weather station observation data, construct a multimodal consistency criterion; S211. Based on the time lag characteristics of cloud top radiative cooling and precipitation echo occurrence in the initial stage of convection, the following conditions are used as the criteria for the initial potential of convection: For any location, if the cooling rate of two consecutive time steps in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is greater than the preset first threshold, and the reflectivity factor in the reflectivity factor data of the weather radar at the current location is less than the preset second threshold. It is worth noting that in the initial stage of convection, the development of deep moist convection typically begins with boundary layer thermal uplift triggering cloud tower growth. The cloud top rapidly rises to the middle and upper troposphere, where intense radiative cooling occurs, manifested as a rapid decrease in the infrared brightness temperature of geostationary meteorological satellites. Meanwhile, radar-detectable precipitation echoes, due to physical delays in the growth and descent of condensate within the cloud, often lag behind cloud top cooling by several minutes to over ten minutes, forming a typical time-series characteristic of "cloud top cooling first, then radar echo." Based on this physical mechanism, this embodiment uses "cooling rates greater than a first threshold and reflectivity factors less than a second threshold for two consecutive time steps" as the criterion for convective initiation potential, aiming to capture the critical window period where strong cloud top cooling has begun but precipitation echoes have not yet formed. This criterion has a simple structure and clear physical meaning, effectively identifying convective cells about to erupt, avoiding reliance on complex models or high computational costs. The first threshold is determined based on statistical analysis of historical cases: a large number of initial convection samples are selected, and the 90th percentile of the cloud top cooling rate is calculated, which is taken as −4 K / 6min; the second threshold is set based on the distribution of the minimum detectable radar echo and the initial weak echo, usually taken as 15 dBZ, to ensure coverage of the typical initial weak signal range, while eliminating ground clutter interference.

[0012] S212. Based on the constraint effect of boundary layer thermal instability on convection triggering, the following condition is used as the thermodynamic criterion for convection triggering: For any location, if the dew point temperature in the observation data of the automatic weather station at the current location is higher than the preset third threshold, and the lifting condensation height calculated based on the temperature and dew point temperature in the observation data of the automatic weather station at the current location is lower than the preset fourth threshold. It is worth noting that boundary layer thermal instability is a key prerequisite for triggering convection: when the surface dew point temperature is high and the temperature-dew point difference is small, the near-surface air has sufficient water vapor and a low condensation height (LCL), meaning that the air parcel can reach saturation and release latent heat under relatively small uplift disturbances, thereby initiating convection. Based on this physical constraint, this embodiment selects "dew point temperature above the third threshold and condensation height below the fourth threshold" as the thermodynamic criterion for convection triggering, because it can effectively characterize the available energy and water vapor support capacity for convection in the lower atmosphere, and relies only on observation data from surface automatic weather stations, possessing high timeliness and operational feasibility. This criterion avoids introducing complex radiosonde or numerical model outputs and is suitable for near-term warnings in areas without radiosonde coverage. The third threshold is determined by statistically analyzing the surface dew point temperature distribution one hour before a historical strong convective event and taking its 75th percentile, typically 20°C. The fourth threshold is determined by back-calculating the 90% cumulative probability of LCL in convective initiation cases and combining it with the typical convective available potential energy (CAPE) threshold, and is set at 1500 meters. This ensures that grid points with sufficient thermal instability are selected, taking into account both sensitivity and specificity.

[0013] S213. Combine the convective initiation potential criterion with the convective triggering thermodynamic criterion to form a multimodal consistency criterion; S22. Based on the multimodal consistency criterion, perform joint condition judgment on the modal data at each position in the multimodal synchronization data cube to identify modal data with physical conflicts. S221. Traverse the modal data at each position in the multimodal synchronization data cube; S222. For the modal data at the current position, determine whether the convection initiation potential criterion and the convection triggering thermodynamic criterion are satisfied simultaneously. S223. If the modal data at the current position does not meet the convection inception potential criterion or the convection triggering thermodynamic criterion, then the modal data at the current position is determined to have a physical conflict, and the modal data at the current position is marked as modal data with a physical conflict; otherwise, the modal data at the current position is determined not to have a physical conflict. It is worth noting that this judgment logic is based on the necessary physical conditions for the initiation of convection: true initiation of convection must simultaneously satisfy rapid cloud top cooling (dynamic / microphysical conditions) and high humidity and low LCL in the boundary layer (thermodynamic conditions). If either criterion is not met, it indicates a physical inconsistency between the multimodal data—for example, only cloud top cooling but dry ground (no water vapor support), or favorable ground thermal conditions but no cloud top development signal, neither of which conforms to the basic physical mechanism of convection initiation. Therefore, treating "not satisfying any criterion" as a physical conflict can effectively identify false signals or observational anomalies. This judgment method has a simple structure and clear physical basis, avoiding complex threshold combinations or black-box judgments in machine learning. It ensures the reliability of conflict identification and provides clear triggering conditions for subsequent adaptive correction, significantly improving the physical consistency and forecast reliability of multimodal fusion.

[0014] S23. For modal data with physical conflicts, based on the prior classification of atmospheric convection development stages, select a preset correction strategy to numerically correct the modal data with physical conflicts, and generate corrected modal data for each location. S231. Based on the spatiotemporal evolution characteristics of modal data at the current location, the atmospheric convection development stages to which the modal data with physical conflicts at the current location belong are classified as the initial stage of convection, the mature stage of convection, and the dissipation stage of convection. It is worth noting that this classification is based on the evolution trend of modal data over time: First, the reflectivity factor of weather radar, the infrared brightness temperature of geostationary meteorological satellites, and the surface dew point temperature sequences of the past three consecutive time steps at the current location are extracted; if the reflectivity factor shows an upward trend and the infrared brightness temperature decreases rapidly, it is determined to be the initial stage of convection; if the reflectivity factor remains high (≥35 dBZ) and the brightness temperature remains stable at a low value, it is determined to be the mature stage of convection; if the reflectivity factor continues to decay, the brightness temperature rises, and the dew point temperature decreases, it is determined to be the dissipation stage of convection. This method only relies on the existing variables in the multimodal synchronous data cube, requires no external input, and can achieve stage division through simple trend criteria. It not only conforms to the physical characteristics of the convection life cycle, but also has computational efficiency and operational feasibility, providing a reliable basis for subsequent selection of corresponding correction strategies.

[0015] S232. When the classification result is the initial stage of convection, the preset first correction strategy is selected to numerically correct the modal data with physical conflicts, and the corrected modal data at each position is generated. S2321. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, the target reflectivity factor is calculated based on the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the reflectivity factor in the weather radar reflectivity factor data at the current location is replaced with the target reflectivity factor. At the same time, the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location is replaced with the arithmetic mean of the infrared brightness temperature at the current time step and the previous time step, and the corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location are generated respectively. It is worth noting that this embodiment addresses the typical missed detection scenario during the early stages of convection, where "satellites have detected rapid cloud top cooling, but radar echoes are weak." By converting the cooling rate of the infrared brightness temperature of geostationary meteorological satellites into a target reflectivity factor, it actively strengthens the weak radar signal, avoiding the omission of early convection due to insufficient radar sensitivity or beam obstruction. Simultaneously, it performs temporal smoothing of the infrared brightness temperature to suppress transient noise interference. This approach achieves numerical physical consistency between the originally isolated weak echo and the significant cloud top cooling signal, significantly enhancing the joint characterization capability of multi-source data for early convection, thereby effectively strengthening the consistency of weak signals and greatly reducing the missed detection rate. It is worth further explaining that, in this embodiment, the cooling rate of the infrared brightness temperature data of the current geostationary meteorological satellite over the past two consecutive time steps is first calculated. That is, the infrared brightness temperature of the previous time step is subtracted from the infrared brightness temperature of the current time step, and then divided by the time interval (e.g., 6 minutes) to obtain the cooling value per unit time. Then, the cooling rate is substituted into a preset mapping function, such as a linear relationship: target reflectivity factor = a × cooling rate + b, where a and b are coefficients based on the statistical fitting of historical primary convection samples to calculate the target reflectivity factor. Next, the target reflectivity factor is used to directly replace the original value in the reflectivity factor data of the weather radar at the current location. At the same time, the infrared brightness temperature data of the current geostationary meteorological satellite is smoothed, and the infrared brightness temperature of the current time step is updated to the arithmetic mean of the infrared brightness temperatures of the current time step and the previous time step. Finally, the corrected radar reflectivity factor and satellite infrared brightness temperature are output as the updated data for the corresponding mode at this location.

[0016] S2322. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the humidity support coefficient based on the dew point temperature and rising condensation height in the automatic weather station observation data at the current location, and replace the reflectivity factor in the weather radar reflectivity factor data at the current location with the humidity support coefficient. At the same time, replace the dew point temperature in the automatic weather station observation data at the current location with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and replace the temperature in the automatic weather station observation data at the current location with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. It is worth noting that this embodiment focuses on the mismatch between boundary layer thermal conditions and radar echoes. When the radar does not show a significant echo, but the ground station observes high dew point temperature and low condensation height (indicating strong convective potential), the radar reflectivity factor is reasonably enhanced using a humidity support coefficient, and the ground temperature and dew point temperature are simultaneously smoothed to eliminate abrupt changes in observation. This two-way correction ensures that favorable thermodynamic conditions are reflected in the radar data, avoiding the neglect of initial convection due to radar lag. This establishes a consistent correlation between "thermal triggering – echo response" at the multimodal level, significantly improving the reliability of capturing the initial stage of convection and reducing the risk of missed detections. It is worth further explaining that, in this embodiment, firstly, the air parchment lifting formula is used to calculate the air parchment lifting height (LCL) based on the temperature and dew point temperature observed from the automatic weather station at the current location. Then, the dew point temperature and LCL are substituted into a preset humidity support coefficient function, for example: humidity support coefficient = c × dew point temperature − d × LCL, where c and d are positive coefficients statistically calibrated based on primary convection samples, to calculate the humidity support coefficient. Next, this coefficient is used to directly replace the original reflectivity factor value in the weather radar reflectivity factor data at the current location. Simultaneously, the automatic weather station observation data is smoothed: the dew point temperature at the current time step is updated to the arithmetic mean of the dew point temperatures at the current and previous time steps, and similarly, the temperature at the current time step is updated to the arithmetic mean of the temperatures at the current and previous time steps. Finally, the corrected radar reflectivity factor data and the corrected automatic weather station observation data are output as the update results for the corresponding mode at this location.

[0017] S2323. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is compensated for based on the temperature in the observation data of the ground automatic weather station at the current location. The temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step. At the same time, the dew point temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step. Corrected infrared brightness temperature data of the geostationary meteorological satellite at the current location and observation data of the ground automatic weather station are generated respectively. It is worth noting that this embodiment addresses situations where satellite brightness temperature anomalies contradict ground thermal conditions. For example, a sudden drop in ground temperature without a corresponding change in satellite brightness temperature may stem from misjudgment of cloud phase or deviation in surface emissivity. By compensating for the deviation in infrared brightness temperature based on ground temperature and simultaneously smoothing ground temperature and dew point temperature, ground-to-space observations are made more consistent in thermodynamic state. This operation ensures the coordinated expression of near-surface unstable energy and cloud top radiation characteristics in the nascent convection environment, preventing misjudgments of no convection development due to single-source anomalies, thereby enhancing the overall reliability of multi-source weak signals, effectively supporting nascent convection identification, and reducing the false negative rate. It is worth further explaining that, in this embodiment, the current time step temperature in the observation data of the current location's automatic weather station is first obtained, and its deviation from the historical temperature of the same period or the regional background temperature is calculated. For example: deviation = current temperature - average regional temperature over the past hour. Then, the deviation is multiplied by an empirical scaling factor, such as 0.8, and calibrated through correlation analysis between satellite brightness temperature and surface temperature to obtain the infrared brightness temperature compensation amount. Next, the original infrared brightness temperature is subtracted from the compensation amount to complete the deviation compensation for the infrared brightness temperature data of the geostationary meteorological satellite. At the same time, the observation data of the automatic weather station is time-series smoothed: the temperature of the current time step is updated to the arithmetic mean of the temperature of the current time step and the temperature of the previous time step, and the dew point temperature of the current time step is updated to the arithmetic mean of the dew point temperature of the current time step and the temperature of the previous time step. Finally, the corrected geostationary meteorological satellite infrared brightness temperature data and the observation data of the automatic weather station are output as the update results of the corresponding mode at this location.

[0018] S2324. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, an initial reflectivity factor is obtained based on the cooling rate mapping in the geostationary meteorological satellite infrared brightness temperature data at the current location. Then, the initial reflectivity factor is modulated a second time based on the dew point temperature in the ground automatic weather station observation data at the current location to update the reflectivity factor in the weather radar reflectivity factor data at the current location. At the same time, the residual between the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location and the temperature in the ground automatic weather station observation data at the current location is calculated, and half of the residual is added to the infrared brightness temperature and temperature respectively to achieve consistency adjustment. Then, the dew point temperature in the ground automatic weather station observation data at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and the corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location are generated respectively. It is worth noting that this embodiment addresses the complex scenario of highly inconsistent information from the three sources during the initial stage of convection. An initial reflectivity factor is generated using the satellite cooling rate, and then modulated a second time using the ground dew point temperature to ensure that the radar echo responds to both cloud top dynamic cooling and boundary layer water vapor conditions. Simultaneously, the air-to-ground thermal field consistency is achieved through the equal distribution and adjustment of the residuals between infrared brightness temperature and ground temperature. This three-way collaborative mechanism comprehensively integrates the three key elements of convection initiation (cloud top cooling, water vapor support, and thermal instability), avoiding misjudgments caused by the dominance of any single mode, significantly improving the comprehensive identification capability of weak convection signals, and fundamentally reducing the false negative rate. It is worth further explaining that, in this embodiment, firstly, based on the cooling rate of the infrared brightness temperature data of the current location geostationary meteorological satellite over the past two time steps, an initial reflectivity factor is calculated through a preset mapping relationship, such as a linear function. Then, the initial value is modulated a second time using the dew point temperature from the current location's automatic weather station observation data. For example, the initial reflectivity factor is multiplied by an adjustment coefficient positively correlated with the dew point temperature. This coefficient is statistically fitted from historical primary convection samples to obtain an updated reflectivity factor, which replaces the value in the original radar data. Next, the numerical residual between the current location's infrared brightness temperature and the ground temperature is calculated, i.e., infrared brightness temperature minus ground temperature. Half of this residual is added to both the infrared brightness temperature and the ground temperature to bring them closer together, achieving thermodynamic consistency adjustment. Finally, the dew point temperature from the automatic weather station observation data is updated to the arithmetic mean of the dew point temperatures of the current time step and the previous time step. Finally, the corrected three-mode data are output. It is worth further elaborating that in the early stages of convection, the physical processes are centered on cloud top radiative cooling and boundary layer thermal uplift. Anomalies in the three modalities (A: radar echo, B: satellite brightness temperature, C: ground temperature and humidity) usually manifest as overall inconsistencies. In practice, if A and B conflict while B and C are consistent, it means that both satellite and ground indicators suggest the absence of convection (e.g., stable brightness temperature, low dew point), but the radar shows false weak echoes—such situations are highly likely to be radar clutter or non-meteorological echoes, which have already been removed during data preprocessing using conventional filtering (e.g., texture thresholding, Doppler velocity consistency). This invention focuses on the physical contradictions between real meteorological signals, rather than instrument noise. Furthermore, the signals are weak in the early stages, and any two-source conflict is sufficient to trigger correction; forcibly distinguishing whether "B and C are consistent" would introduce redundant judgment logic, increasing system complexity without providing significant gains. Therefore, adopting a strategy of four mutually exclusive categories (AB, AC, BC, and ABC) that covers all possible conflicting combinations ensures the completeness of the correction while avoiding over-engineering, thus meeting the business needs of "rapid response and simplified decision-making" in the early stages. It is worth further elaboration that the scenario of "A and B conflict and A and C conflict, but B and C are consistent" (e.g., satellite shows strong cooling, ground shows high dew point and low LCL (B and C are consistent, indicating strong nascent potential), but radar echo is abnormally weak or even missing (A and B, A and C both conflict)) is precisely the core scenario covered by S2321 and S2322 in this embodiment. In this case, this embodiment first identifies the physical conflict at the location through S223 (because the multimodal consistency criterion is not met), and then, after classifying it as nascent in S231, simultaneously triggers S2321 (A-B correction) and S2322 (A-C correction). Since the correction operations are both enhancements of A (radar) (based on the cooling rate of B and the humidity support coefficient of C respectively), they can be executed sequentially or in parallel, ultimately generating a consistent enhanced echo. In other words, "A and B conflict + A and C conflict" is not an independent new scenario, but a natural superposition of the two types of corrections, AB and AC, without the need for additional rules. Similarly, if "A and B are consistent, but A and C conflict, and B and C conflict," then the conflict is dominated by BC and is handled by S2323. All pairwise conflict patterns can be covered by pairwise corrections of AB, AC, and BC; when all three conflict, ABC correction is used for unified coordination. Therefore, the four types of cases constitute a logically complete set, and any "partial conflict" can be decomposed into combinations or special cases of existing correction steps without the need for new branches.

[0019] S233. When the classification result is the convective maturity stage, select the preset second correction strategy to numerically correct the modal data with physical conflicts and generate the corrected modal data at each position. S2331. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the covariance between the reflectivity factor in the weather radar reflectivity factor data at the current location and the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and perform a proportional scaling correction on the reflectivity factor based on the covariance, while performing a reverse scaling correction on the infrared brightness temperature by the same proportion, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location respectively. It is worth noting that this embodiment addresses the mismatch between the covariance of radar echo and satellite cooling rate in mature strong convection. By calculating their covariance and implementing proportional scaling correction, it ensures that the reflectivity factor of the strong echo core region remains physically synchronized with the cloud top cooling intensity. If the radar echo is abnormally high while satellite cooling is weak, the echo is appropriately attenuated; conversely, it is enhanced. This bidirectional dynamic adjustment avoids structural exaggeration or weakening caused by single-mode dominance, effectively maintaining the energy consistency of the strong convection system in the vertical direction, thereby accurately maintaining the integrity and intensity accuracy of the echo structure and preventing excessive smoothing or distortion. It is worth further explaining that, in this embodiment, the weather radar reflectivity factor sequence and the corresponding geostationary meteorological satellite infrared brightness temperature sequence for the past three consecutive time steps of the current location are first obtained, and the cooling rate sequence of the infrared brightness temperature is calculated; then, the sample covariance between the reflectivity factor sequence and the cooling rate sequence is calculated; if the covariance deviates from the typical range of the historical maturity period (e.g., below a threshold), the scaling factor is calculated as: typical covariance / current covariance; then, the current reflectivity factor is multiplied by this scaling factor for correction, and the current infrared brightness temperature is divided by the same scaling factor, i.e., reverse scaling, so that the covariance relationship between the two returns to a reasonable range; finally, the corrected radar and satellite data are output.

[0020] S2332. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the residual between the reflectivity factor in the weather radar reflectivity factor data at the current location and the lifting condensation height in the automatic weather station observation data at the current location, and perform linear correction on the reflectivity factor based on the residual. At the same time, replace the temperature in the automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. It is worth noting that this embodiment addresses the mismatch between strong radar echoes during the mature phase and surface lifting and condensation height. By calculating the residual between the reflectivity factor and the lifting and condensation height, the echoes are linearly corrected to ensure that precipitation intensity is consistent with the effective convective potential energy of the boundary layer. Simultaneously, weighted smoothing of surface temperature and dew point temperature eliminates local observation disturbances. This prevents the model from underestimating or overestimating convective intensity due to anomalies in surface data, ensuring that the structure of strong echoes conforms to thermodynamic constraints at both horizontal and vertical scales. This approach preserves detail while avoiding structural distortion, thus guaranteeing forecast stability. It is worth further explaining that, in this embodiment, the Lifted Condensation Level (LCL) is first calculated based on the temperature and dew point temperature observed from the current location's automatic weather station data. Then, the residual between the current radar reflectivity factor and the reference reflectivity factor expected based on the LCL is calculated; the reference value is determined by historical maturity period statistical relationships. Next, this residual is multiplied by an attenuation coefficient, such as 0.6, and then subtracted from the original reflectivity factor to complete the linear correction. Simultaneously, the ground data is weighted and smoothed: the temperature is updated to 0.7 × current temperature + 0.3 × previous time step temperature, and the dew point temperature is weighted and averaged with the same weight. Finally, the corrected radar and ground data are output.

[0021] S2333. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground-based automatic weather station, calculate the linear regression slope between the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location and the temperature in the observation data of the ground-based automatic weather station at the current location, and perform a smooth correction on the infrared brightness temperature based on the slope. At the same time, replace the temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground-based automatic weather station at the current location, respectively. It is worth noting that this embodiment addresses the situation where the trends of cloud top brightness temperature and ground temperature diverge during the mature stage. By calculating the linear regression slope of the two and smoothing the infrared brightness temperature accordingly, the cloud top evolution observed by the satellite is made consistent with the ground thermal feedback; at the same time, ground temperature and humidity elements are smoothed synchronously. This correction avoids false strong convection sustaining signals caused by abrupt changes in surface temperature or misjudgment of cloud facies, ensuring that multimodal data collectively reflect the true state of vigorous convection, thereby providing structurally stable input for the encoder and preventing the strong echo pattern from being distorted or broken due to data conflicts; It is worth further explaining that, in this embodiment, the infrared brightness temperature and ground temperature data of the geostationary meteorological satellite over the past three time steps at the current location are first used to fit the linear regression slope between the two using the least squares method. If the slope exceeds the reasonable range of the mature period, such as being close to 0 or negative, it is replaced by a typical historical slope, such as 0.95. Subsequently, the current ground temperature is substituted into the regression equation to calculate the expected infrared brightness temperature, and its weighted average with the original brightness temperature, with a weight of 0.6:0.4, is used as the correction value. At the same time, the ground temperature and dew point temperature are weighted and averaged separately, with the current value accounting for 0.7 and the previous value accounting for 0.3. Finally, the corrected satellite and ground data are output.

[0022] S2334. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, calculate the joint covariance among the reflectivity factor in the weather radar reflectivity factor data at the current location, the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the dew point temperature in the ground automatic weather station observation data at the current location. Based on the joint covariance, normalize and correct the reflectivity factor, and correct the cooling rate by covariance projection. Then, map the corrected cooling rate back to the infrared brightness temperature to update the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location, respectively. It is worth noting that this embodiment uses joint covariance analysis to analyze the relationship between radar reflectivity factor, satellite cooling rate, and ground dew point temperature, and normalizes and projects each mode. This mechanism ensures a high degree of coordination between the strong convective system and the three dimensions of dynamics (echo), microphysics (cloud top cooling), and thermodynamics (water vapor), avoiding overall structural imbalance caused by anomalies in any mode. For example, if the radar echo is too strong but the water vapor support is insufficient, the echo intensity is appropriately reduced. This multidimensional consistency correction ensures the three-dimensional structure of the core region of strong convection is realistic and reliable, effectively preventing excessive smoothing or false enhancement, and maintaining the physical rationality of the forecast field. It is worth further explaining that, in this embodiment, the weather radar reflectivity factor, geostationary meteorological satellite infrared brightness temperature (used to calculate the cooling rate), and dew point temperature sequence of the past three time steps at the current location are first obtained; then, a three-dimensional vector sequence composed of the three is constructed, and its 3×3 joint covariance matrix is ​​calculated; subsequently, the triplet vector at the current time is projected onto the principal component direction of the covariance matrix, and the reflectivity factor is normalized and corrected, that is, the mean is subtracted, the standard deviation is divided, and then the typical standard deviation is multiplied; the cooling rate is projected and corrected under covariance constraints to eliminate abnormal deviations; then, the corrected cooling rate is converted into the corresponding infrared brightness temperature change through a preset physical mapping relationship such as a linear inversion model, and superimposed on the infrared brightness temperature of the previous time step to obtain the updated current infrared brightness temperature; at the same time, the temperature and dew point temperature in the observation data of the ground automatic weather station are weighted and averaged respectively, with a weight of 0.7 for the current time step and a weight of 0.3 for the previous time step; finally, the corrected three-mode data are output.

[0023] It is worth further elaborating that the mature stage of convection is characterized by strong updrafts, high precipitation efficiency, and thick cloud cover. The three modalities should exhibit high synergy: strong echoes (A), rapid cloud top cooling (B), and high boundary layer water vapor (C). During this stage, if A and B conflict while B and C are consistent (e.g., continuously decreasing brightness temperature, high dew point, but a sudden drop in radar echo), it often reflects hardware limitations such as radar attenuation, blockage, or beam overshooting, rather than changes in the actual physical state. Such problems belong to observation system errors and should be pre-processed by the radar quality control module, not by the physical consistency correction mechanism of this invention. The core objective of this invention is to solve the problem of mismatch in the inherent physical relationships of multi-source meteorological signals, rather than compensating for equipment defects. Furthermore, the mature stage of convection has a compact structure, and conflicts between any two modes are usually accompanied by indirect anomalies in the third mode (e.g., weakened echoes leading to cloud top heating, which in turn affects B). Therefore, joint ABC correction can cover most real-world scenarios. Further subdividing the "partially consistent" cases will lead to rule explosion and reduce system robustness. Therefore, a four-category conflict classification is adopted to ensure both physical rationality and engineering feasibility.

[0024] It is worth further elaborating on the scenario where "A and B conflict, and A and C conflict, but B and C are consistent" (e.g., satellite brightness temperature is stable or rising, ground dew point is high and LCL is low (B and C are consistent, indicating that convection should be maintained), but radar echoes are abnormally attenuated (A and B, and A and C both conflict)). This usually stems from severe radar signal attenuation caused by heavy precipitation. However, such attenuation is usually quantitatively corrected in operational radar products using dual polarization parameters (such as differential propagation phase shift ΦDP), falling under the category of radar data preprocessing. That is, the multimodal data input in this invention has undergone basic quality control, only processing residual physical inconsistencies. More importantly, even without complete correction, the consistency of B and C in this scenario provides sufficiently strong evidence that "convection is still maintained." In this embodiment, linear compensation for the echo is performed based on the increased condensation height in S2332 (A–C correction), and the echo intensity is adjusted due to covariance anomalies in S2331 (A–B correction)—the two work together to restore a reasonable echo value. There is no need to define the "B–C consistency" condition separately, because the correction strategy itself implicitly uses the consistent mode as a reference benchmark. Similarly, in any "two conflicting and one consistent" case, the consistent two modes naturally constitute a more reliable physical constraint, and existing pairwise correction mechanisms can automatically use this constraint to correct the conflicting mode. Therefore, the design of the four cases has inherent generalization ability and does not require enumerating all sub-combinations.

[0025] S234. When the classification result is the convection dissipation period, select the preset third correction strategy to numerically correct the modal data with physical conflicts and generate the corrected modal data for each location. S2341. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps. Construct an exponential attenuation coefficient based on the ratio of the attenuation slope to the recovery slope, and use the exponential attenuation coefficient to attenuate and correct the reflectivity factor at the current location. At the same time, replace the infrared brightness temperature at the current location with the weighted average of the infrared brightness temperature at the current time step and the infrared brightness temperature at the previous time step, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location, respectively. It is worth noting that this embodiment addresses a typical problem during the dissipation phase: strong radar echoes remain while satellite brightness temperature has significantly increased. An exponential attenuation coefficient is constructed by calculating the ratio of the reflectivity factor attenuation slope to the brightness temperature increase slope, forcing the radar echo to attenuate according to physical laws. Simultaneously, the brightness temperature is weighted and smoothed to suppress noise. This effectively identifies and suppresses false radar signals that do not conform to the dissipation trend, preventing the model from misjudging that convection is still ongoing, thereby significantly improving the accuracy of the dissipation process forecast and ensuring that the predicted field reflects the weakening state of the system in a timely manner. It is worth further explaining that, in this embodiment, the radar reflectivity factor values ​​of the past two consecutive time steps at the current location are first extracted, and its attenuation slope is calculated. That is, the reflectivity factor of the previous time step is subtracted from the reflectivity factor of the current time step, and then divided by the time interval (e.g., 6 minutes) to obtain the attenuation per unit time. At the same time, the infrared brightness temperature values ​​of the corresponding location in the past two time steps are extracted, and its recovery slope is calculated. That is, the brightness temperature of the current time step is subtracted from the brightness temperature of the previous time step and then divided by the same time interval. Subsequently, the absolute value of the attenuation slope is divided by the absolute value of the recovery slope. If the denominator is zero, the ratio is set to 1. The resulting ratio is truncated to the interval [0,1] and used as the exponential attenuation coefficient. Then, the current reflectivity factor is multiplied by the exponential function value of this coefficient (e.g., exp(−exponential attenuation coefficient)) to complete the nonlinear attenuation correction. At the same time, the current infrared brightness temperature is updated to a weighted average value, with a weight of 0.6 for the current value and 0.4 for the previous time step value to smooth instantaneous fluctuations. Finally, the corrected radar reflectivity factor and infrared brightness temperature data are output.

[0026] S2342. When a conflict occurs between weather radar reflectivity factor data and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Based on the difference between the two slopes, perform linear compensation correction on the reflectivity factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected weather radar reflectivity factor data and ground automatic weather station observation data at the current location respectively. It is worth noting that this embodiment addresses the contradiction between the lack of radar echo attenuation during the dissipation phase and the rapid decrease in surface dew point temperature. By comparing the difference between the attenuation slope of the reflectivity factor and the dew point decrease slope, the echo is linearly compensated and corrected to synchronize its attenuation rate with boundary layer water vapor dissipation; simultaneously, surface temperature and humidity data are smoothed. This prevents residual radar echoes from misleading the model into continuing strong precipitation forecasts, ensuring that the precipitation intensity during the dissipation phase is consistent with the near-surface dry and cold air intrusion process, and improving the rationality of the timing and intensity forecasts for the dissipation process. It is worth further explaining that, in this embodiment, the attenuation slope of the radar reflectivity factor at the current location over the past two consecutive time steps is first calculated (previous time value minus current value divided by the time interval); simultaneously, based on the dew point temperature sequence in the ground automatic weather station observation data, its decreasing slope in the same time period is calculated (previous time dew point temperature minus current value divided by the time interval); then, the difference between the two slopes is calculated (attenuation slope − decreasing slope). If the difference is positive, it indicates that the radar attenuation is faster than the water vapor dissipation, and there may be false residual echoes; this difference is multiplied by an empirical compensation coefficient (e.g., 0.5, calibrated through historical dissipation case statistics), and the resulting compensation is subtracted from the current reflectivity factor to complete the linear compensation correction; at the same time, the temperature and dew point temperature in the ground automatic weather station observation data are time-series smoothed respectively: the current temperature is updated to 0.6 × current temperature + 0.4 × previous time step temperature, and the dew point temperature is weighted averaged with the same weight; finally, the corrected radar reflectivity factor and ground observation data are output.

[0027] S2343. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, calculate the rise slope of the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location over the past two consecutive time steps, and calculate the change slope of the temperature in the observation data of the ground automatic weather station at the current location over the past two consecutive time steps. Correct the infrared brightness temperature based on the residual of the two slopes. At the same time, replace the temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the temperature of the current time step and the temperature of the previous time step, and replace the dew point temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the dew point temperature of the current time step and the dew point temperature of the previous time step. Generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground automatic weather station at the current location, respectively. It is worth noting that this embodiment addresses the issue of slow satellite brightness temperature recovery during the dissipation period while ground temperature has already changed rapidly. The infrared brightness temperature is corrected by calculating the residual between the brightness temperature recovery slope and the ground temperature change slope, and ground temperature and humidity data are simultaneously smoothed to ensure that the air-to-ground thermal states remain consistent in the dissipation trend. This avoids misjudging the cloud body's continued existence due to satellite lag or abnormal surface radiation, ensuring that multiple modes jointly indicate convective dissipation, thereby improving the overall forecast rationality and physical consistency of the dissipation process. It is worth further explaining that, in this embodiment, the rise slope of the infrared brightness temperature at the current location between two consecutive time steps is first calculated. Specifically, the infrared brightness temperature at the current time step is subtracted from the infrared brightness temperature at the previous time step, and then divided by the time interval between the two time steps (e.g., 6 minutes) to obtain the brightness temperature rise rate per unit time. Simultaneously, the slope of temperature change in the ground automatic weather station observation data between the same two time steps is calculated, i.e., the temperature at the current time step is subtracted from the temperature at the previous time step, and then divided by the same time interval. Subsequently, the residual between the two slopes is calculated, i.e., the infrared brightness temperature rise rate minus the ground temperature change rate. If the residual... If the absolute value exceeds the preset tolerance (e.g., 1 Kelvin every 6 minutes), it indicates an inconsistency between satellite and ground observations in terms of thermal evolution trends. This residual is multiplied by an empirical adjustment coefficient (e.g., 0.8, calibrated through correlation analysis of satellite brightness temperature and ground temperature during historical dissipation phases), and the resulting correction is added to the infrared brightness temperature at the current time step to align it with the ground thermal change trend. Simultaneously, the temperature and dew point temperatures from the automatic weather station observation data are weighted and averaged, with the current time step data weighted at 0.6 and the previous time step data weighted at 0.4. Finally, the corrected infrared brightness temperature and the automatic weather station observation data are output.

[0028] S2344. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps, and the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Construct a proportional coordination factor based on the three slopes, and use the proportional coordination factor to perform nonlinear attenuation correction on the reflectivity factor and linear recovery correction on the infrared brightness temperature based on the proportional coordination factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location, respectively. It is worth noting that this embodiment integrates three types of slopes—radar attenuation, satellite rise, and ground dew point decline—to construct a proportional coordination factor and implement nonlinear linkage correction for the three modes. If radar echo attenuation is too slow, brightness temperature rise is insufficient, or dew point decline is lagging, the system will coordinately adjust the values ​​of these three factors to force them to conform to the physical laws of the dissipation phase. This multi-source consistency constraint completely eliminates false signals maintained by isolated modes, ensuring that the predicted field accurately reflects the overall dissipation status of the convective system and significantly improving the reliability and timeliness of the dissipation phase forecast. It is worth further explaining that, in this embodiment, the following steps are first taken: the attenuation slope of the weather radar reflectivity factor at the current location over the past two consecutive time steps (i.e., the previous value minus the current value divided by the time interval); the rise slope of the infrared brightness temperature of the geostationary meteorological satellite (i.e., the current value minus the previous value divided by the time interval); and the fall slope of the dew point temperature of the automatic weather station (i.e., the previous value minus the current value divided by the time interval). Then, the absolute values ​​of the three slopes are taken and normalized to construct a proportional coordination factor. For example, the coordination factor α is set to the ratio of the rise slope + fall slope to (attenuation slope + rise slope + fall slope), limited to between 0 and 1. Next, a nonlinear attenuation correction is applied to the current reflectivity factor, i.e., multiplied by the exponential term exp(−α). Simultaneously, a linear rise correction is applied to the infrared brightness temperature, i.e., α is added and multiplied by a typical rise amplitude (e.g., 2K). Finally, the temperature and dew point temperature in the automatic weather station observation data are updated to a weighted average of 0.6 times the current value plus 0.4 times the previous time step value. The corrected three-mode data are then output. It is worth further elaborating that the typical characteristics of the convection dissipation period are echo attenuation, cloud top warming, and the intrusion of dry and cold boundary layer air. All three modes should simultaneously exhibit a "weakening" trend. If A and B conflict while B and C are consistent (e.g., brightness temperature rises, dew point falls, but radar echoes do not attenuate), this almost certainly corresponds to residual false radar echoes (e.g., three-body scattering, ground object echoes). Such signals are widely recognized in meteorological operations and can be filtered out at the front end using simple criteria such as echo persistence and vertical structure. The correction mechanism of this invention focuses on the problem of mismatched multi-source evolution rates during the actual dissipation process (e.g., excessively slow echo attenuation), rather than eliminating non-meteorological echoes. More importantly, the system structure is loose during the dissipation period, and the coupling between modes is weakened. Allowing "B and C are consistent while A is abnormal" as an independent case will lead to fragmentation of the correction strategy, potentially harming genuine weak echoes. In the current four-type conflict design, ABC joint correction can effectively suppress isolated anomalies through multi-slope coordination; AB, AC, and BC cover two dominant conflicts. This design strikes a balance between suppressing false signals and preserving genuine residual convection, avoiding over-correction and meeting the forecast objective of "robust decay and prevention of false clearing" during the dissipation period.

[0029] It is worth further elaboration that the core application scenario for the S2341 (A–B) and S2342 (A–C) corrections of this invention is the case where "A and B conflict, and A and C conflict, but B and C are consistent" (e.g., satellite brightness temperature rises significantly, ground dew point drops rapidly (B and C are consistent, clearly indicating dissipation), but radar echo attenuation is slow or even maintained (A and B, A and C both conflict)). This embodiment calculates the attenuation slope of A and the rise slope of B (S2341), and the attenuation slope of A and the dew point drop slope of C (S2342), and performs dual attenuation correction on A accordingly. Because the consistency between B and C enhances the confidence of the dissipation judgment, and the two corrections are in the same direction, their effects are superimposed, resulting in a more thorough elimination of false echoes. This is not a new situation, but rather a synergistic manifestation of AB and AC corrections. Furthermore, if there are instances of "A and B being consistent, A and C conflicting, and B and C conflicting" (e.g., radar and satellite both show attenuation, but ground temperature and humidity are abnormally high), it may reflect local thermal disturbances (such as urban heat islands). However, such disturbances are usually small in scale and short in duration, with limited impact in 0–2 hour nowcasts, and will be smoothed by S2343 (B–C correction) through slope residual correction. In summary, all "partially conflicting" situations either fall within the scope of preprocessing or can be naturally covered by existing pairwise or trimodal corrections without increasing rule complexity. The four types of conflict design are particularly important during the dissipation period because they avoid excessive interference with weak real signals while ensuring effective suppression of typical spurious echoes.

[0030] S3. Input the corrected modal data at each position into the multi-branch heterogeneous encoder to extract the feature map at each position; It is worth noting that in this embodiment, the multi-branch heterogeneous encoder adopts an existing mature deep learning architecture, consisting of three parallel but structurally different coding branches: the radar branch uses a 3D convolutional neural network to capture spatiotemporal echo dynamics, the satellite branch uses a fusion of 2D convolution and attention mechanisms to extract cloud top texture and evolution features, and the ground station branch uses a fully connected network to embed spatial location coding to characterize the thermal field distribution. The three modal data at each location after correction are input into the corresponding branch, and after layer-by-layer downsampling and nonlinear transformation, feature maps for each location are output in a unified feature dimension. These feature maps include: a spatiotemporal feature map reflecting the radar echo structure, a brightness temperature feature map characterizing the cloud top cooling trend, and a ground thermodynamic feature map depicting boundary layer water vapor and stability. These three correspond one-to-one on the spatial grid, providing a foundation for subsequent reliability assessment and fusion.

[0031] S4. Obtain at least three feature indicators for the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. It is worth noting that in this embodiment, at least three feature indicators are extracted from the modal data at each corrected location: the spatiotemporal continuity of the radar reflectivity factor (measured by the rate of change of adjacent time steps), the local gradient consistency of the infrared brightness temperature of the geostationary meteorological satellite (calculated by the standard deviation of the 3×3 neighborhood), and the degree of deviation between the observation data of the ground automatic weather station and the regional background field (such as the absolute difference between temperature and dew point temperature relative to the mean of the 5 km neighborhood). Subsequently, each feature indicator is normalized and assigned a confidence contribution weight based on its physical meaning—for example, the higher the continuity, the smoother the gradient, and the smaller the deviation, the higher the local confidence of the corresponding mode. Finally, the indicators are weighted and fused to calculate the local confidence weights of the three modes—weather radar, satellite, and ground station—at the current location, with values ​​ranging from 0 to 1, for subsequent adaptive weighted fusion of feature maps.

[0032] S5. Using the local confidence weights of the modal data at each location after correction, the feature maps at each location are weighted and adaptively fused to generate a fused feature tensor. The fused feature tensor is then input into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours. It is worth noting that in this embodiment, firstly, spatial interpolation is performed on the local confidence weights of the three modalities—weather radar, geostationary meteorological satellite, and automatic weather station data—to ensure accurate matching of each modality in terms of geographical location. Then, based on each adaptive weighting factor, a weighted fusion operation is performed on the feature maps of different modalities at the same location to generate a fusion feature tensor that comprehensively reflects multiple observational information. Specifically, for each geographical location, its fusion feature is obtained by multiplying the features of each modality at that location by their corresponding local confidence weights and then summing the results. Next, the obtained fusion feature tensor is input into a pre-trained spatiotemporal prediction decoder. This decoder is composed of a Convolutional Long Short-Term Memory (ConvLSTM) network or a similar spatiotemporal sequence model, capable of capturing and utilizing the spatiotemporal correlations in multi-source data. The spatiotemporal prediction decoder learns effective spatiotemporal patterns by encoding the fusion features of historical time steps, and based on this, predicts radar reflectivity factors for multiple time steps within the next 0-2 hours on a time-by-time basis, outputting a series of continuous prediction fields for subsequent early warning and analysis. This process not only enhances the accuracy and stability of forecasts, but also improves the speed of response to sudden weather events.

[0033] Example 2: Figure 2 As shown, this invention also discloses a convection nowcasting system based on multimodal data, comprising the following modules: Data acquisition module: used to acquire multi-source observation data of the target area at the current time and at several historical times; Data correction module: connected to the data acquisition module, used to resample multi-source observation data to a preset unified spatiotemporal grid, generate a multimodal synchronous data cube, the multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid, and performs condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube based on preset atmospheric convection physics rules to obtain the corrected modal data at each location; Extraction module: Connected to the data correction module, it is used to input the corrected modal data at each position into the multi-branch heterogeneous encoder and extract the feature map at each position; Calculation module: Connected to the extraction module, it is used to obtain at least three feature indicators of the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. Prediction module: Connected to the calculation module, it is used to perform weighted adaptive fusion of the feature maps of each location using the local confidence weights of the modal data at each location after correction, generate a fused feature tensor, and input the fused feature tensor into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours.

[0034] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0035] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0036] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A convection nowcasting method based on multimodal data, characterized in that, Includes the following steps: S1. Obtain multi-source observation data of the target area at the current time and at several historical times; S2. Resample the multi-source observation data to a preset unified spatiotemporal grid to generate a multimodal synchronous data cube. The multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid. Based on preset atmospheric convection physics rules, perform condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube to obtain the corrected modal data at each location. S3. Input the corrected modal data at each position into the multi-branch heterogeneous encoder to extract the feature map at each position; S4. Obtain at least three feature indicators for the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. S5. Using the local confidence weights of the modal data at each location after correction, the feature maps at each location are weighted and adaptively fused to generate a fused feature tensor. The fused feature tensor is then input into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours.

2. The convection nowcasting method based on multimodal data according to claim 1, characterized in that, The modal data includes: weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data.

3. The convection nowcasting method based on multimodal data according to claim 2, characterized in that, The modal data at each location in the multimodal synchronization data cube are subjected to conditional judgment and correction operations based on preset atmospheric convection physics rules to obtain corrected modal data at each location, including: S21. Based on the physical correlation between weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data and ground automatic weather station observation data, construct a multimodal consistency criterion; S22. Based on the multimodal consistency criterion, perform joint condition judgment on the modal data at each position in the multimodal synchronization data cube to identify modal data with physical conflicts. S23. For modal data with physical conflicts, based on the prior classification of atmospheric convection development stages, select a preset correction strategy to numerically correct the modal data with physical conflicts, and generate corrected modal data for each location.

4. The convection nowcasting method based on multimodal data according to claim 3, characterized in that, The multimodal consistency criterion is constructed based on the physical correlation between weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, including: S211. Based on the time lag characteristics of cloud top radiative cooling and precipitation echo occurrence in the initial stage of convection, the following conditions are used as the criteria for the initial potential of convection: For any location, if the cooling rate of two consecutive time steps in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is greater than the preset first threshold, and the reflectivity factor in the reflectivity factor data of the weather radar at the current location is less than the preset second threshold. S212. Based on the constraint effect of boundary layer thermal instability on convection triggering, the following condition is used as the thermodynamic criterion for convection triggering: For any location, if the dew point temperature in the observation data of the automatic weather station at the current location is higher than the preset third threshold, and the lifting condensation height calculated based on the temperature and dew point temperature in the observation data of the automatic weather station at the current location is lower than the preset fourth threshold. S213. Combine the convective initiation potential criterion with the convective triggering thermodynamic criterion to form a multimodal consistency criterion.

5. The convection nowcasting method based on multimodal data according to claim 4, characterized in that, The method based on multimodal consistency criteria, which performs joint conditional judgment on modal data at each location in the multimodal synchronization data cube to identify modal data with physical conflicts, includes: S221. Traverse the modal data at each position in the multimodal synchronization data cube; S222. For the modal data at the current position, determine whether the convection initiation potential criterion and the convection triggering thermodynamic criterion are satisfied simultaneously. S223. If the modal data at the current position does not meet the convection inception potential criterion or the convection triggering thermodynamic criterion, then the modal data at the current position is determined to have a physical conflict, and the modal data at the current position is marked as modal data with a physical conflict; otherwise, the modal data at the current position is determined not to have a physical conflict.

6. The convection nowcasting method based on multimodal data according to claim 5, characterized in that, For modal data with physical conflicts, based on the prior classification of atmospheric convection development stages, a preset correction strategy is selected to numerically correct the modal data with physical conflicts, generating corrected modal data for each location, including: S231. Based on the spatiotemporal evolution characteristics of modal data at the current location, the atmospheric convection development stages to which the modal data with physical conflicts at the current location belong are classified as the initial stage of convection, the mature stage of convection, and the dissipation stage of convection. S232. When the classification result is the initial stage of convection, the preset first correction strategy is selected to numerically correct the modal data with physical conflicts, and the corrected modal data at each position is generated. S233. When the classification result is the convective maturity stage, select the preset second correction strategy to numerically correct the modal data with physical conflicts and generate the corrected modal data at each position. S234. When the classification result is the convection dissipation period, select the preset third correction strategy to numerically correct the modal data with physical conflicts, and generate the corrected modal data for each location.

7. A convection nowcasting method based on multimodal data according to claim 6, characterized in that, When the classification result indicates the initial stage of convection, a preset first correction strategy is selected to numerically correct the modal data with physical conflicts, generating corrected modal data for each location, including: S2321. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, the target reflectivity factor is calculated based on the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the reflectivity factor in the weather radar reflectivity factor data at the current location is replaced with the target reflectivity factor. At the same time, the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location is replaced with the arithmetic mean of the infrared brightness temperature at the current time step and the previous time step, and the corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location are generated respectively. S2322. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the humidity support coefficient based on the dew point temperature and rising condensation height in the automatic weather station observation data at the current location, and replace the reflectivity factor in the weather radar reflectivity factor data at the current location with the humidity support coefficient. At the same time, replace the dew point temperature in the automatic weather station observation data at the current location with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and replace the temperature in the automatic weather station observation data at the current location with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. S2323. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location is compensated for based on the temperature in the observation data of the ground automatic weather station at the current location. The temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the temperature at the current time step and the temperature at the previous time step. At the same time, the dew point temperature in the observation data of the ground automatic weather station at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step. Corrected infrared brightness temperature data of the geostationary meteorological satellite at the current location and observation data of the ground automatic weather station are generated respectively. S2324. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data simultaneously, an initial reflectivity factor is obtained based on the cooling rate mapping in the geostationary meteorological satellite infrared brightness temperature data at the current location. Then, the initial reflectivity factor is modulated a second time based on the dew point temperature in the ground automatic weather station observation data at the current location to update the reflectivity factor in the weather radar reflectivity factor data at the current location. At the same time, the residual between the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location and the temperature in the ground automatic weather station observation data at the current location is calculated, and half of the residual is added to the infrared brightness temperature and temperature respectively to achieve consistency adjustment. Then, the dew point temperature in the ground automatic weather station observation data at the current location is replaced with the arithmetic mean of the dew point temperature at the current time step and the dew point temperature at the previous time step, and the corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location are generated respectively.

8. A convection nowcasting method based on multimodal data according to claim 7, characterized in that, When the classification result is convective maturity, a preset second correction strategy is selected to numerically correct the modal data with physical conflicts, generating corrected modal data for each location, including: S2331. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the covariance between the reflectivity factor in the weather radar reflectivity factor data at the current location and the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and perform a proportional scaling correction on the reflectivity factor based on the covariance, while performing a reverse scaling correction on the infrared brightness temperature by the same proportion, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location respectively. S2332. When a conflict occurs between weather radar reflectivity factor data and automatic weather station observation data, calculate the residual between the reflectivity factor in the weather radar reflectivity factor data at the current location and the lifting condensation height in the automatic weather station observation data at the current location, and perform linear correction on the reflectivity factor based on the residual. At the same time, replace the temperature in the automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate corrected weather radar reflectivity factor data and automatic weather station observation data at the current location respectively. S2333. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground-based automatic weather station, calculate the linear regression slope between the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location and the temperature in the observation data of the ground-based automatic weather station at the current location, and perform a smooth correction on the infrared brightness temperature based on the slope. At the same time, replace the temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the observation data of the ground-based automatic weather station at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground-based automatic weather station at the current location, respectively. S2334. When a conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data simultaneously, calculate the joint covariance among the reflectivity factor in the weather radar reflectivity factor data at the current location, the cooling rate in the geostationary meteorological satellite infrared brightness temperature data at the current location, and the dew point temperature in the ground automatic weather station observation data at the current location. Based on the joint covariance, normalize and correct the reflectivity factor, and perform covariance projection correction on the cooling rate. Then, map the corrected cooling rate back to the infrared brightness temperature to update the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location, respectively.

9. A convection nowcasting method based on multimodal data according to claim 8, characterized in that, When the classification result indicates a convection dissipation period, a preset third correction strategy is selected to numerically correct the modal data with physical conflicts, generating corrected modal data for each location, including: S2341. When a conflict occurs between weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps. Construct an exponential attenuation coefficient based on the ratio of the attenuation slope to the recovery slope, and use the exponential attenuation coefficient to attenuate and correct the reflectivity factor at the current location. At the same time, replace the infrared brightness temperature at the current location with the weighted average of the infrared brightness temperature at the current time step and the infrared brightness temperature at the previous time step, and generate corrected weather radar reflectivity factor data and geostationary meteorological satellite infrared brightness temperature data at the current location, respectively. S2342. When a conflict occurs between weather radar reflectivity factor data and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, and calculate the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Based on the difference between the two slopes, perform linear compensation correction on the reflectivity factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step, and generate the corrected weather radar reflectivity factor data and ground automatic weather station observation data at the current location respectively. S2343. When a conflict occurs between the infrared brightness temperature data of a geostationary meteorological satellite and the observation data of a ground automatic weather station, calculate the rise slope of the infrared brightness temperature in the infrared brightness temperature data of the geostationary meteorological satellite at the current location over the past two consecutive time steps, and calculate the change slope of the temperature in the observation data of the ground automatic weather station at the current location over the past two consecutive time steps. Correct the infrared brightness temperature based on the residual of the two slopes. At the same time, replace the temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the temperature of the current time step and the temperature of the previous time step, and replace the dew point temperature in the observation data of the ground automatic weather station at the current location with the weighted average of the dew point temperature of the current time step and the dew point temperature of the previous time step. Generate the corrected infrared brightness temperature data of the geostationary meteorological satellite and the observation data of the ground automatic weather station at the current location, respectively. S2344. When the conflict involves weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data, calculate the attenuation slope of the reflectivity factor in the weather radar reflectivity factor data at the current location over the past two consecutive time steps, the recovery slope of the infrared brightness temperature in the geostationary meteorological satellite infrared brightness temperature data at the current location over the past two consecutive time steps, and the decrease slope of the dew point temperature in the ground automatic weather station observation data at the current location over the past two consecutive time steps. Construct a proportional coordination factor based on the three slopes, and use the proportional coordination factor to perform nonlinear attenuation correction on the reflectivity factor and linear recovery correction on the infrared brightness temperature based on the proportional coordination factor. At the same time, replace the temperature in the ground automatic weather station observation data at the current location with the weighted average of the temperature at the current time step and the temperature at the previous time step, and replace the dew point temperature in the ground automatic weather station observation data at the current location with the weighted average of the dew point temperature at the current time step and the dew point temperature at the previous time step. Generate corrected weather radar reflectivity factor data, geostationary meteorological satellite infrared brightness temperature data, and ground automatic weather station observation data at the current location.

10. A convective nowcasting system based on multimodal data, used to execute the convective nowcasting method based on multimodal data as described in any one of claims 1-9, characterized in that, Includes the following modules: Data acquisition module: used to acquire multi-source observation data of the target area at the current time and at several historical times; Data correction module: connected to the data acquisition module, used to resample multi-source observation data to a preset unified spatiotemporal grid, generate a multimodal synchronous data cube, the multimodal synchronous data cube contains modal data at each location in the unified spatiotemporal grid, and performs condition judgment and correction operations on the modal data at each location in the multimodal synchronous data cube based on preset atmospheric convection physics rules to obtain the corrected modal data at each location; Extraction module: Connected to the data correction module, it is used to input the corrected modal data at each position into the multi-branch heterogeneous encoder and extract the feature map at each position; Calculation module: Connected to the extraction module, it is used to obtain at least three feature indicators of the modal data at each position after correction, and calculate the local confidence weight of the modal data at each position after correction based on the at least three feature indicators. Prediction module: Connected to the calculation module, it is used to perform weighted adaptive fusion of the feature maps of each location using the local confidence weights of the modal data at each location after correction, generate a fused feature tensor, and input the fused feature tensor into the spatiotemporal prediction decoder to output the radar reflectivity factor prediction field for multiple time steps in the next 0–2 hours.

Images