Extracellular vesicles for the treatment of cancer

Abstract

The application relates to a kind of temperate cyclone center positioning method based on multi-modal data fusion and physical enhancement, the method comprises the following steps: collecting multi-source meteorological data;Three-dimensional thermal-dynamic coupling characteristic field is constructed;Dynamic adjustment of data source weight;Iterative positioning is carried out using spiral convergence search algorithm of fusion Archimedes spiral and gradient descent;Physical consistency and path anomaly detection double verification are carried out.The application has the advantages that: through three-dimensional physical field coupling perception, dynamic weight optimization and intelligent search, the positioning accuracy and efficiency are significantly improved.It can be widely used in meteorological disaster monitoring, forecasting and early warning and professional meteorological service.

Description

Technical Field

[0001] This invention relates to a method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement, belonging to the field of meteorological monitoring and forecasting. Background Technology

[0002] Extratropical cyclones are common and highly destructive weather systems in mid- to high-latitude regions. Their development, evolution, and movement significantly impact local weather patterns and daily life. Unlike the relatively symmetrical structure of tropical cyclones (typhoons), extratropical cyclones exhibit typical baroclinic behavior, resulting in a more complex structure. The surface cyclone center does not coincide with the upper-level low-pressure center; sometimes there is no corresponding low-pressure center at high altitude, or multiple subcenters exist within the surface cyclone. These complexities make it difficult to accurately determine the center location using traditional data sources such as air pressure, wind fields, and satellite cloud imagery. Existing cyclone center location methods are mostly designed for tropical cyclones (typhoons), with very few methods available for mid- to high-latitude extratropical cyclones, and they lack the application of high-precision meteorological data. Therefore, a high-precision method for locating the center of extratropical cyclones by integrating multi-source data is urgently needed.

[0003] Existing cyclone center location technology has the following limitations: 1. Insufficient data accuracy of traditional methods: Common cyclone location methods mainly rely on the analysis of traditional meteorological elements such as sea level pressure field, wind field or satellite cloud image cloud system morphology, brightness temperature, etc. (spatial resolution ≥1°, temporal resolution ≥6h). They lack the application of data from high-precision detection equipment such as millimeter-wave cloud radar (spatial resolution 30m), wind profiler radar (temporal resolution 6min), microwave radiometer, as well as radar combined reflectivity extrapolation (spatial resolution 1km) and mesoscale high spatiotemporal resolution numerical model products (such as CMA-MESO, spatiotemporal resolution 3km / 1~3h), resulting in large deviations in the location of extratropical cyclones on complex underlying surfaces.

[0004] 2. Poor interpretability of deep learning models: Using deep learning methods such as k-means clustering and convolutional neural networks (CNN) to determine the location of cyclone centers simply by classifying data samples lacks constraints on the key physical mechanisms of weather system evolution. This makes them unsuitable for different types, structures, and intensities of extratropical cyclones, resulting in low positioning accuracy.

[0005] 3. There is a lack of comparative verification of the reliability of satellite, model, and observation data, and no dynamic adjustment of the weights of different types of data in cyclone location calculations. The entire process uses a fixed-weighted average (e.g., satellite:model:ground = 1:1:1), without an adaptive weighting strategy based on different development stages and types of extratropical cyclones, such as initiation, development, maturity, and dissipation.

[0006] Therefore, this invention proposes a mid-latitude cyclone center localization technology that integrates multimodal meteorological data fusion and physical enhancement mechanisms. Through three core technologies—three-dimensional physical field coupling perception, dynamic data weight optimization, and biomimetic intelligent search—a new method for mid-latitude cyclone center localization is constructed. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a method for locating the center of a temperate cyclone based on multimodal data fusion and physical enhancement. The technical solution of this invention is as follows: A method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement, characterized by comprising the following steps: S1. Collect multi-source meteorological data, which includes at least ground observation data, numerical model reanalysis data and satellite remote sensing data; S2. Construct a three-dimensional thermodynamic-dynamic coupled feature field for the baroclinic structure of a mid-latitude cyclone. The three-dimensional thermodynamic-dynamic coupled feature field is used to characterize the thermodynamic and dynamic field features of a mid-latitude cyclone. S3. Based on the structure type and data quality assessment results of extratropical cyclones, dynamically adjust the weights of different data sources in the location calculation; S4. A spiral convergence search algorithm that integrates the Archimedes spiral trajectory and the physical gradient descent direction is used to iteratively search for the cyclone center in the three-dimensional thermo-dynamic coupling feature field. S5. Perform physical consistency verification and path anomaly detection verification on the candidate locations of the cyclone center obtained from the search, and output the final coordinates of the cyclone center.

[0008] Step S1 specifically includes: S1.1 Ground observation data is collected in real time by automatic weather stations and undergoes triple quality control, including real-time verification, extreme value verification, and spatiotemporal consistency verification based on variational assimilation method. For observation blank areas, pseudo observation values ​​are generated by virtual sensor network based on air pressure-topography relationship. S1.2 Numerical model reanalysis data extraction at four layers (1000 hPa, 850 hPa, 700 hPa, and 500 hPa) includes geopotential height, temperature, wind field, and vertical velocity, and calculates moist potential vorticity. To capture baroclinic energy, where g is gravitational acceleration, ζ is relative vorticity, and f is the geostrophic parameter. For pseudo-equivalent temperature, air pressure ; S1.3 Satellite remote sensing data is used to assist in identifying the extent and shape of cloud systems, and cloud system boundary indices are constructed through multispectral channels. This is used to highlight the boundary and wetland characteristics of temperate cyclone cloud systems; among which , These are cloud top brightness temperature (TBB) and humidity gradient, respectively. Weighting coefficients; S1.4 The multi-source meteorological data also includes millimeter-wave cloud radar data, wind profiler radar data, S-band weather radar data, microwave radiometer data, and mesoscale high spatiotemporal resolution numerical forecast data. S1.5: Data Spatiotemporal Alignment: Spatial interpolation and temporal synchronization are performed on the multi-source meteorological data from steps S1.1 to S1.4, unifying the data to a regular grid of 0.1° × 0.1°. The spatial interpolation uses inverse distance weighted interpolation, and the interpolation formula is as follows: ,in, Points to be interpolated Meteorological element values, Let i be the value of the i-th observation point. For the interpolation point and the first The distance between observation points, k=2, is a power of the distance; time synchronization uses cubic spline interpolation, the formula is... ,in, The coefficients a, b, c, and d are determined by fitting data from four adjacent time points.

[0009] Step S2 specifically includes: S2.1 Hierarchical Feature Extraction: Thermal field construction: Calculation of temperature advection ,in The horizontal wind vector These are the zonal and meridional wind speed components, respectively. These represent the partial derivatives of temperature in the latitudinal and longitudinal directions, respectively; and the hydrostatic energy field. ,in For isobaric specific heat capacity, Let g be the latent heat of vaporization, z be the gravitational acceleration, and q be the specific humidity; Dynamic field construction: Calculation of vorticity field. and Q-vector divergence Where Q1 and Q2 are the Q vector components calculated according to the Hoskins quasi-geotropic Q vector definition; S2.2 Tensor Fusion and Dimensionality Reduction: This involves fusing the feature field from step S2.1 with the surface pressure gradient field. Together they are constructed into a four-dimensional feature tensor Where H is the number of grids in the longitude direction, W is the number of grids in the latitude direction, D = 4 is the number of vertical layers, corresponding to the four pressure layers of 1000hPa, 850hPa, 700hPa, and 500hPa respectively, and C = 6 is the number of characteristic channels, which specifically include: corresponding pressure layer temperature T, relative humidity RH, vorticity Q-vector divergence Moist hydrostatic energy (MSE) and surface pressure gradient ; The four-dimensional feature tensor is reduced in dimensionality using Tucker decomposition, while preserving the tensor. Where H' = 0.7H, W' = 0.7W, D' = 3, C' = 4, and the energy retention rate is not less than 95%; Step S3 specifically includes the following steps: S3.1 Data Quality Assessment: Calculating the Consistency Index of Satellite Cloud Image Features Where 0.5 in the denominator is the normalization factor, and the threshold is... ; Calculate the set discreteness of the reanalysis ,in To predict the potential height of the i-th member of the ECMWF ensemble, The mean of 10 members is 9, and the degrees of freedom are 9; calculate the ground observation continuity score (TPS). S3.2 Dynamic Update of Bayesian Weights: Dynamic updates of weights are achieved through a Bayesian posterior probability model, with the likelihood function being... Where k=1,2,3 represent data sources from satellite, reanalysis, and ground observations, respectively. For the observed values, The mean of the true values. For variance, The weights to be optimized are: the posterior distribution is sampled using the Metropolis-Hastings algorithm, the weight mean is taken after 1000 iterations, the weights of each data source are dynamically updated, and physical consistency constraints are introduced, including wind direction and cloud movement deviations of less than 30°, and the weight change rate in a single iteration not exceeding 10%. S3.3 Context-aware weight mode switching: At least three weight configuration modes are preset, corresponding to frontal cyclones, non-frontal cyclones, and explosive cyclones, respectively. The weather type of the extratropical cyclone is identified in real time through a convolutional neural network to switch weight modes; the initial weight is set to ground observation: reanalysis: satellite = 0.5:0.4:0.1.

[0010] Step S4 specifically includes: S4.1 Path Planning: Archimedes Spiral Path Parametric Method ,in, Let the initial radius be , For spiral growth rate, N = 36 represents the number of spiral rotations; and this is combined with the gradient descent direction vector of momentum acceleration. Where 0.8 and 0.2 are the momentum coefficient and the current gradient coefficient, respectively, and the three-dimensional gradient coupling factor. α=0.4, β=0.3, and γ=0.3 are the weighting coefficients for the pressure gradient term, the 850hPa temperature advection term, and the 500hPa vorticity term, respectively, and the sum of the three weights is 1. (·) indicates that the range is normalized for the entire range; S4.2 Adaptive step size adjustment: Adjust the step size according to the Armijo criterion. ,like Then accept the step size; otherwise, let Where 0.3 is the sufficient decrease coefficient; S4.3 Multiple Hypothesis Verification of Candidate Center Set: Retain candidate center points every 5 iterations The optimal solution is selected through a comprehensive scoring function: Where 0.5, 0.3, and 0.2 are the weighting coefficients for the three-dimensional gradient coupling factor, air pressure curvature, and 700 hPa vertical velocity, respectively. For the pressure Laplace operator, The vertical velocity is 700 hPa.

[0011] Step S5 specifically includes: S5.1 Physical consistency verification includes calculating the pressure gradient concentration index. , where R min Let R be the minimum radius centered at the candidate point that can detect a complete closed isobar. min =100km; R is the average pressure gradient at this closed isobar; max Let R be the outer search boundary. max =300km, The average pressure gradient within this radius is defined as follows: PGCI ≤ 0.6, and the difference between the surface wind field offset angle around the candidate point and the theoretical gradient wind direction is ≤ 45°. S5.2 Path anomaly detection examines the rationality of candidate centers from two aspects: "continuity of cyclone movement" and "regional path pattern constraints".

[0012] Step S5.2 includes: (1) Cyclone movement continuity check: Let the current candidate center position be The location result at the previous moment was The displacement distance between the two is The angle of change in the direction of movement is ,in, This represents the azimuth angle of the current time relative to the previous time. It represents the azimuth angle of the previous moment relative to the moment before that; If D>D maxOr Δθ>Δθ max And D>D min If the cyclone movement is not continuous, the candidate point is rejected and the S4 step of the re-iteration search is triggered; where D max =150km is the statistical upper limit of the maximum possible distance that a mid-latitude cyclone can travel in one hour, D min =50km is used to eliminate noise interference introduced by small-amplitude oscillations when the cyclone stagnates; Δθ max =60° is the maximum allowable turning angle in the adjacent time movement direction. If the current positioning result triggers a continuity anomaly, the candidate point is rejected and the S4 step of re-iteration search is triggered. (2) Regional path pattern constraint: Based on historical path statistics, determine whether the geographical location of the current candidate center is within the area of ​​the main movement direction of the corresponding type of historical cyclone. If it falls within this area, it passes the climate path trend verification; if it deviates from this area, it triggers a low confidence mark and prompts for manual review.

[0013] In step S2.2: For the surface pressure gradient ∇P, which is a surface variable, when constructing the four-dimensional tensor, the horizontal distribution value of ∇P is repeatedly assigned in the four vertical layers of 1000hPa, 850hPa, 700hPa and 500hPa. Wet hydrostatic energy MSE and Q vector divergence All are derived diagnostic parameters. MSE is calculated layer by layer from the temperature, potential height, and specific humidity of each layer. The horizontal wind field and temperature field of each layer are calculated layer by layer, and both are calculated separately according to the four pressure layers to form a complete four-layer vertical distribution.

[0014] The advantages of this invention are: 1. By integrating multi-source data from ground observations, numerical model reanalysis, satellite remote sensing, and various high-precision detection devices (millimeter-wave cloud radar, wind profiler radar, microwave radiometer, etc.), and constructing a three-dimensional thermodynamic-dynamic coupled field characterized by physical quantities such as wet potential vorticity, temperature advection, vorticity, and Q vector divergence, the ability to characterize the baroclinic structure of extratropical cyclones is significantly enhanced.

[0015] 2. Key physical mechanisms of weather system evolution (baroclinic energy, frontal zones, vertical motion forcing, etc.) are integrated into the entire process of feature field construction, weight optimization, and search iteration. The dynamic weight correction module introduces physical consistency constraints (wind direction and cloud system movement deviation <30°, single iteration weight change rate ≤10%). The spiral search is guided by pressure gradient, temperature advection, and vertical vorticity. The output results are verified by both the pressure gradient concentration index (PGCI≤0.6) and historical path similarity to ensure that the positioning results conform to the actual physical laws of extratropical cyclones.

[0016] 3. By switching between context-aware weighting modes, three weighting configurations are preset: frontal cyclones, non-frontal cyclones, and explosive cyclones. A convolutional neural network is used to identify weather types in real time and adjust data source weights accordingly (ground observation weight is no less than 50%). This is applicable to extratropical cyclones of different development stages, intensities, and structural types. The Bayesian dynamic update mechanism optimizes the contribution weights of satellite, reanalysis, and ground data in real time based on data quality, overcoming the shortcomings of traditional fixed-weight methods.

[0017] 4. A hybrid search strategy combining Archimedes spiral and gradient descent is adopted, along with momentum acceleration and adaptive adjustment of Armijo step size, reducing the average number of iterations to 8-9. The cloud platform deployment employs region decomposition (16 subdomains × 4 CPUs), block loading (300 × 300 grid), and MPI parallel communication, with a total single positioning time of ≤15 seconds (14.3 seconds in actual measurement of the Jianghuai cyclone), and a peak memory usage of ≤64GB, meeting the requirements of real-time operational needs.

[0018] 5. It can be integrated into existing meteorological forecasting systems to output the location and movement trajectory of extratropical cyclones in real time, providing high-precision and high-reliability technical support for meteorological disaster monitoring, forecasting, early warning, and professional meteorological services for ports, shipping, etc. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0021] See Figure 1 This invention relates to a method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement, comprising the following steps: S1. Collect multi-source meteorological data, which includes at least ground observation data, numerical model reanalysis data and satellite remote sensing data; S2. Construct a three-dimensional thermodynamic-dynamic coupled feature field for the baroclinic structure of a mid-latitude cyclone. The three-dimensional thermodynamic-dynamic coupled feature field is used to characterize the thermodynamic and dynamic field features of a mid-latitude cyclone. S3. Based on the structure type and data quality assessment results of extratropical cyclones, dynamically adjust the weights of different data sources in the location calculation; S4. A spiral convergence search algorithm that integrates the Archimedes spiral trajectory and the physical gradient descent direction is used to iteratively search for the cyclone center in the three-dimensional thermo-dynamic coupling feature field. S5. Perform physical consistency verification and path anomaly detection verification on the candidate locations of the cyclone center obtained from the search, and output the final coordinates of the cyclone center.

[0022] Step S1 integrates high-precision detection data from ground automatic stations, numerical model reanalysis, satellite remote sensing, millimeter-wave cloud radar, wind profiler radar, and microwave radiometers, overcoming the limitations of low-resolution data that traditional methods rely on. This significantly improves the ability to characterize the baroclinic structure and frontal zone features of extratropical cyclones, providing rich and reliable basic information for subsequent positioning.

[0023] Triple quality control and virtual observation filling technology ensure data quality; high spatiotemporal resolution data (such as 30m cloud radar, 6min wind profiler radar, 3km / 1~3h CMA-MESO) make the positioning results closer to the real cyclone center.

[0024] Step S2 for the first time constructs physical quantities such as wet potential vorticity, temperature advection, vorticity, Q vector divergence, wet hydrostatic energy and surface pressure gradient into a four-dimensional tensor and uses Tucker decomposition to reduce dimensionality, explicitly integrating key physical mechanisms such as baroclinic energy, frontal zone location and vertical motion forcing of extratropical cyclones into the feature space.

[0025] Step S3 breaks through the limitations of traditional fixed weights by employing a context-aware weighting strategy, Bayesian MCMC dynamic updates, and CNN weather type recognition. For extratropical cyclones (surface-dominated), it automatically increases the weight of ground observations and suppresses the weight of satellite data during periods of low baroclinic energy contribution, achieving adaptive optimization of data source contributions and significantly improving positioning accuracy.

[0026] Step S4 integrates Archimedes spiral global exploration with gradient descent local fine search, combined with momentum acceleration and Armijo step size adaptation, prioritizing iteration in baroclinic energy high value region and pressure gradient dense region to avoid getting trapped in local extrema, greatly improving search efficiency (reducing the number of iterations by 60%, with an average of 8-9 convergences).

[0027] Step S5 eliminates false low-pressure centers by using the pressure gradient concentration index (PGCI≤0.6) and confirms the rationality of the trajectory by comparing it with the historical path statistical area. This dual verification mechanism ensures that the positioning results conform to the actual structure and evolution of the extratropical cyclone, minimizing the false alarm rate in business applications.

[0028] Step S1 specifically includes: S1.1 Ground observation data is collected in real time by automatic weather stations and undergoes triple quality control, including real-time verification, extreme value verification, and spatiotemporal consistency verification based on variational assimilation methods. For observation gaps, pseudo-observations are generated using a virtual sensor network based on the pressure-topography relationship. When accessing real-time data from automatic weather stations nationwide, spatiotemporal consistency verification is performed on outliers: if the pressure at a single station changes by more than 5 hPa within 10 minutes, a three-level alarm mechanism is triggered. Median filtering is used to correct abrupt data changes. For extratropical cyclones, surface air pressure and wind field data are used as core inputs, with relatively high initial weights.

[0029] S1.2 Numerical model reanalysis data extraction at four layers (1000 hPa, 850 hPa, 700 hPa, and 500 hPa) includes geopotential height, temperature, wind field, and vertical velocity, and calculates moist potential vorticity. To capture baroclinic energy, where g is gravitational acceleration, ζ is relative vorticity, and f is the geostrophic parameter. For equivalent potential temperature, It refers to air pressure; where MPV is measured in units of air pressure. ; S1.3 Satellite remote sensing data is used to assist in identifying the extent and shape of cloud systems, and cloud system boundary indices are constructed through multispectral channels. ,in , These are cloud top brightness temperature (TBB) and humidity gradient, respectively. The weighting coefficients; it should be noted that the center of the extratropical cyclone cloud system and the surface pressure center often do not coincide, and satellite data is only used to define the rough range of the search area and is not used as a direct basis for precise positioning; S1.4 The multi-source meteorological data also includes millimeter-wave cloud radar data, wind profiler radar data, S-band weather radar data, microwave radiometer data, and mesoscale high spatiotemporal resolution numerical forecast data. S1.5: Data Spatiotemporal Alignment: Spatial interpolation and temporal synchronization are performed on the multi-source meteorological data from steps S1.1 to S1.4, unifying the data to a regular grid of 0.1° × 0.1°. The spatial interpolation uses inverse distance weighted interpolation, and the interpolation formula is as follows: ,in, Points to be interpolated Meteorological element values, Let i be the value of the i-th observation point. For the interpolation point and the first The distance between observation points, where k=2 is the power of the distance; Time synchronization uses cubic spline interpolation, the formula is as follows: ,in, The coefficients a, b, c, and d are determined by fitting data from four adjacent time points.

[0030] Step S1.1 effectively eliminates outlier observations through real-time verification, extreme value verification (pressure change rate < 5 hPa / 10 min), and variational assimilation spatiotemporal consistency verification; median filtering corrects abrupt data, ensuring the reliability of surface pressure and wind field inputs. For observational gaps (such as oceans and mountains), pseudo-observations are generated using the pressure-topography relationship, significantly improving the spatial coverage of ground data. For the baroclinic characteristics of extratropical cyclones, surface pressure and wind field are initially assigned high weights (0.5), providing the most crucial physical constraints for subsequent positioning.

[0031] Step S1.2 extracts key isobaric surface data from 1000 hPa to 500 hPa to capture the baroclinic structure of extratropical cyclones from the ground to the mid-to-high atmosphere. Calculating moist potential vorticity (MPV, unit PVU) can quantitatively characterize atmospheric convective stability and baroclinic energy distribution, highlighting frontal zones and baroclinic instability zones, and giving the characteristic field a clear physical meaning.

[0032] Step S1.3 uses weighted fusion of cloud top brightness temperature (TBB) and humidity gradient with α=0.6 and β=0.4 to automatically identify the boundary and morphology of cyclone cloud systems, assisting in confirming the approximate range of the center. However, the weight setting (the initial satellite weight is only 0.1) reflects the design concept of this invention that uses satellite data to assist in positioning of extratropical cyclones, rather than dominating it, thus avoiding interference from satellite data during the low-contribution phase of baroclinic energy.

[0033] Step S1.4 uses data from millimeter-wave cloud radar (30m spatial resolution), wind profiler radar (6min temporal resolution), S-band weather radar, microwave radiometer, and mesoscale high-resolution numerical weather prediction (CMA-MESO, 3km / 1~3h) to compensate for the shortcomings of traditional methods (≥1° / 6h) in terms of fine structure and rapid evolution. This provides high-density, high-frequency observational constraints for the three-dimensional feature field and significantly improves the positioning accuracy of complex underlying surfaces (such as the Yellow Sea).

[0034] Step S1.5 unifies heterogeneous and asynchronous multi-source data into a 0.1°×0.1° regular grid (spatial resolution much higher than the traditional 1°), and uses inverse distance weighting (k=2) to maintain the authenticity of local pressure gradient; cubic spline interpolation ensures temporal continuity, so that data with different time resolutions such as satellite (5 minutes), radar (6 minutes), and automatic station (minute level) are strictly synchronized, providing a spatiotemporally consistent three-dimensional grid field for subsequent tensor construction and spiral search.

[0035] Step S2 specifically includes: S2.1 Hierarchical Feature Extraction: Thermal field construction: Calculation of temperature advection ,in The horizontal wind vector These are the zonal and meridional wind speed components, respectively. These represent the partial derivatives of temperature in the latitudinal and longitudinal directions, respectively; and the hydrostatic energy field. ,in For isobaric specific heat capacity, Where is the latent heat of vaporization, g is the acceleration due to gravity, z is the potential height, and q is the specific humidity; Dynamic field construction: Calculation of vortex field and Q-vector divergence Where Q1 and Q2 are the Q vector components calculated according to the Hoskins quasi-geotropic Q vector definition; S2.2 Tensor Fusion and Dimensionality Reduction: The characteristic field of step S2.1 and the ground pressure gradient field Together they are constructed into a four-dimensional feature tensor Where H is the number of grids in the longitude direction, W is the number of grids in the latitude direction, D = 4 is the number of vertical layers, corresponding to the four pressure layers of 1000hPa, 850hPa, 700hPa, and 500hPa respectively, and C = 6 is the number of characteristic channels, which specifically include: corresponding pressure layer temperature T, relative humidity RH, vorticity Q-vector divergence Moist hydrostatic energy (MSE) and surface pressure gradient Of the six characteristic channels mentioned above, temperature, relative humidity, and vorticity are the fundamental variables of the standard isobaric surface analysis field. Complete data are available for the four pressure layers of 1000 hPa, 850 hPa, 700 hPa, and 500 hPa. Data for each layer can be directly extracted and filled from the reanalysis data layer by layer. The wet hydrostatic energy (MSE) and Q vector divergence (Q) are also relevant. div Both are derived diagnostic quantities. MSE is calculated layer by layer from the temperature, geopotential height, and specific humidity of each layer, while Q_div is calculated layer by layer from the horizontal wind field and temperature field of each layer. Both are calculated separately for each of the four pressure layers, forming a complete four-layer vertical distribution. For the surface pressure gradient ∇P, which is a surface variable and does not belong to any upper-level isobaric surface layer, when constructing the four-dimensional tensor, the horizontal distribution value of ∇P is repeatedly assigned in the four vertical layers of 1000hPa, 850hPa, 700hPa, and 500hPa to ensure that the surface pressure gradient information has a balanced constraint effect on the subsequent tensor decomposition and search algorithm in each layer. After the above assignment, the six channels have a complete horizontal spatial distribution in each vertical layer, which together constitute a four-dimensional feature tensor with complete dimensions. (D=4, C=6) provides mathematically complete and physically consistent input for subsequent Tucker decomposition.

[0036] The four-dimensional feature tensor is reduced in dimensionality using Tucker decomposition, while preserving the tensor. Where H' = 0.7H, W' = 0.7W, D' = 3, C' = 4, and the energy retention rate is not less than 95%.

[0037] Step S2.1 Hierarchical Feature Extraction: Thermal field construction (temperature advection and wet hydrostatic energy): Temperature advection can quantitatively identify the convergence zone of warm and cold advection, i.e., the frontal zone, which is a key area for the release of baroclinic energy in extratropical cyclones. Wet hydrostatic energy (MSE) comprehensively characterizes sensible heat energy, potential energy, and latent heat energy. Its gradient distribution can clearly indicate the wet baroclinic structure of the cyclone, especially applicable to high-humidity environments such as the Yangtze-Huaihe River cyclone and the Yellow River-Huaihe River cyclone. These thermal characteristics provide clear physical guidance for spiral search, avoiding blind searches.

[0038] Dynamic field construction (vorticity and Q-vector divergence): Vorticity reflects the degree of atmospheric rotation; the positive vorticity region is highly correlated with the cyclone center and is a core indicator of dynamic constraints. Q-vector divergence directly characterizes the vertical motion forcing mechanism; its negative value region corresponds to upward motion, precisely corresponding to the airflow convergence and lifting region near the cyclone center. By simultaneously extracting vorticity and Q-vector divergence, the dynamic structure of the cyclone (rotation + convergence) is fully incorporated into the feature space, enhancing the physical consistency of the positioning.

[0039] S2.2 Construction of Four-Dimensional Feature Tensor: The thermal field (T, RH, MSE) and the dynamic field (ζ, Q) div The system comprises six channels, including the surface pressure gradient ∇P, organized into a four-dimensional tensor across four vertical layers (1000-500 hPa). This achieves a unified representation of multi-source, multi-variable, and multi-altitude data, providing a structured platform for subsequent multimodal fusion. Specifically, T850 captures low-level thermals, RH700 indicates mid-level humidity, ζ500 reflects upper-level vortex intensity, and Q... div Diagnosing vertical motion, MSE integrates thermo-energy, and ∇P highlights changes in ground pressure.

[0040] Tucker decomposition is used to reduce the dimensionality of the four-dimensional tensor, with an energy retention rate of ≥95%. This significantly reduces the computational complexity of the subsequent spiral search algorithm while preserving core physical information. Tucker decomposition is naturally adapted to high-dimensional tensor data, maintaining the spatial, vertical, and inter-channel cross-correlation of the feature field and avoiding the loss of physical information.

[0041] For the surface pressure gradient ∇P, which is a surface variable, when constructing the four-dimensional tensor, the horizontal distribution value of ∇P is repeatedly assigned in four vertical layers of 1000hPa, 850hPa, 700hPa, and 500hPa. Wet hydrostatic energy MSE and Q vector divergence All are derived diagnostic parameters. MSE is calculated layer by layer from the temperature, potential height, and specific humidity of each layer. The horizontal wind field and temperature field of each layer are calculated layer by layer, and both are calculated separately according to the four pressure layers to form a complete four-layer vertical distribution.

[0042] Step S3 specifically includes the following steps: S3.1 Data Quality Assessment: Calculating the Consistency Index of Satellite Cloud Image Features Where 0.5 in the denominator is the normalization factor, and the threshold is... ; Calculate the set discreteness of the reanalysis ,in To predict the potential height of the i-th member of the ECMWF ensemble, The mean of 10 members is 9, and the degrees of freedom are 9; calculate the ground observation continuity score (TPS). S3.2 Dynamic Update of Bayesian Weights: Dynamic updates of weights are achieved through a Bayesian posterior probability model, with the likelihood function being... Where k=1,2,3 represent data sources from satellite, reanalysis, and ground observations, respectively. For the observed values, The mean of the true values. For variance, The weights to be optimized are: the posterior distribution is sampled using the Metropolis-Hastings algorithm, the weight mean is taken after 1000 iterations, the weights of each data source are dynamically updated, and physical consistency constraints are introduced, including wind direction and cloud movement deviations of less than 30°, and the weight change rate in a single iteration not exceeding 10%. S3.3 Context-aware weight mode switching: At least three weight configuration modes are preset, corresponding to frontal cyclones, non-frontal cyclones, and explosive cyclones, respectively. The weather type of the extratropical cyclone is identified in real time through a convolutional neural network to switch weight modes; the initial weight is set to ground observation: reanalysis: satellite = 0.5:0.4:0.1.

[0043] Step S3.1 quantitatively assesses the reliability of satellite data in describing cyclone cloud systems by calculating the ratio of the matching cloud boundary length to the total cloud boundary length and multiplying it by the mean of the TBB gradient / 0.5 normalization factor. A threshold CCCI ≥ 0.7 ensures that higher weights are assigned only when the satellite cloud image features are highly consistent with the actual cyclone shape, avoiding misleading positioning in cases of broken or thin clouds.

[0044] The standard deviation of geopotential height based on the 10 ECMWF ensemble forecast members reflects the uncertainty of the reanalysis data. Small dispersion indicates high reliability of the reanalysis data, allowing for increased weighting; large dispersion automatically reduces weighting, thus suppressing the impact of model bias on positioning.

[0045] Assess the temporal continuity of ground air pressure and wind fields, identify abnormal changes caused by instrument malfunctions or transmission interruptions, and ensure the reliability of core ground data.

[0046] The three categories of indicators independently assess data quality from three dimensions: satellite, model, and ground, providing objective and quantitative prior information for subsequent Bayesian updates and enabling a scientific evaluation of the contribution of data sources.

[0047] Step S3.2: Establish the likelihood function, and correlate data quality (variance σ_k^2) with weights. Direct association will assign higher weights to high-quality data (small variance). By sampling the posterior distribution using the Metropolis-Hastings algorithm and iterating 1000 times, the weights are averaged to achieve adaptive optimization of the weights without manual intervention.

[0048] Introducing a wind direction and cloud system movement deviation of <30° (ensuring coordination between low-level wind field and cloud system movement) and a single iteration weight change rate of ≤10% (preventing sudden weight changes from causing positioning jumps), the weight adjustment not only relies on statistical fitting but also conforms to the basic physical laws of atmospheric motion, enhancing the reliability and stability of the positioning results.

[0049] Bayesian dynamic updates can adjust the fusion strategy in real time according to the actual data quality, which is especially suitable for situations where the reliability of data changes significantly at different development stages of extratropical cyclones (such as incomplete satellite cloud images in the early stage and stronger surface pressure signals in the mature stage).

[0050] Step S3.3 sets different initial weights and update ranges for three typical extratropical cyclones: frontal cyclones, non-frontal cyclones, and explosive cyclones. For example, frontal cyclones increase the weight of surface frontal data, non-frontal cyclones focus on mid-to-high-level vorticity, and explosive cyclones emphasize temperature, advection, and vertical velocity.

[0051] This system utilizes CNNs to automatically classify the current cyclone type from inputs such as multispectral cloud images and geopotential height fields, and then switches to the corresponding weight mode. The classifier can complete inference in milliseconds, achieving a scenario-adaptive weight strategy.

[0052] Given the baroclinic characteristics of extratropical cyclones, the dominant role of ground-based and lower-level observation data is emphasized, with satellite data serving only as an auxiliary tool.

[0053] Step S4 specifically includes the following steps: S4.1 Path Planning: Archimedes Spiral Path Parametric Method ,in, Let the initial radius be , For spiral growth rate, N = 36 represents the number of spiral rotations; and this is combined with the gradient descent direction vector of momentum acceleration. Where 0.8 and 0.2 are the momentum term coefficient and the current gradient coefficient, respectively; the three-dimensional gradient coupling factor is a quantitative characterization of the thermodynamic-dynamic coupling strength near the center of a mid-latitude cyclone; and the constructed GCF is a weighted composite index of pressure gradient, temperature advection, and vertical vorticity, calculated using the following formula: α=0.4, β=0.3, and γ=0.3 are the weighting coefficients for the pressure gradient term, the 850hPa temperature advection term, and the 500hPa vorticity term, respectively. The sum of the three weights is 1, reflecting the physical mechanism of the pressure field dominating the positioning and the thermodynamic and dynamic fields co-constraining each other in the mid-latitude cyclone. norm(·) represents the range normalization of the entire field to eliminate the dimensional differences between different physical quantities. GCF integrates information from three aspects: the depressurization characteristics of the cyclone center, the baroclinicity of the frontal zone, and the cyclonic rotation intensity. The larger the value, the stronger the thermodynamic-dynamic coupling at that point, which directly serves as the guiding basis for gradient descent in the subsequent spiral search.

[0054] S4.2 Adaptive step size adjustment: Adjust the step size according to the Armijo criterion. ,like Then accept the step size; otherwise, let Where 0.3 is the sufficient decrease coefficient; S4.3 Multiple Hypothesis Verification of Candidate Center Set: Retain candidate center points every 5 iterations The optimal solution is selected through a comprehensive scoring function: Where 0.5, 0.3, and 0.2 are the weighting coefficients for the three-dimensional gradient coupling factor, air pressure curvature, and 700 hPa vertical velocity, respectively. For the pressure Laplace operator (pressure curvature). The vertical velocity is 700 hPa.

[0055] Step S4.1 begins with an initial radius of 0.05° (approximately 5.5 km), starting from a small area. The spiral growth rate is 0.01° / rad, rotating a total of 36 times, covering an area with a radius of approximately 1.3° (approximately 140 km), matching the typical scale of a mid-latitude cyclone. The spiral path ensures global exploration capability and avoids getting trapped in local extrema (such as false low-pressure centers).

[0056] Compared to traditional grid search or random sampling, spiral path search is more efficient and can cover key physical regions (high baroclinic energy regions and dense pressure gradient regions) in fewer iterations.

[0057] A momentum coefficient of 0.8 preserves the historical search direction, while a current gradient coefficient of 0.2 introduces new information, effectively suppressing local oscillations and accelerating convergence.

[0058] The gradient direction is guided by the three-dimensional gradient coupling factor (GCF), which integrates pressure gradient, temperature advection and vertical vorticity, so that the search path always points to the region with the strongest physical properties in the center of the cyclone, rather than simply the lowest pressure point.

[0059] Step S4.2 sets an adaptive step size adjustment (Armijo criterion). If the step size is satisfied, it is accepted; otherwise, it is halved. The step size is automatically adjusted to avoid skipping the cyclone center due to an excessively large step size or slow convergence due to an excessively small step size. A sufficient descent coefficient of 0.3 ensures that the objective function (GCF) is sufficiently improved in each iteration, mathematically guaranteeing convergence. No manual setting of a fixed step size is required, making it suitable for extratropical cyclones of different intensities and scales (e.g., strong explosive cyclones require fine-grained searches with small step sizes, while weak cyclones can be approximated quickly with large step sizes).

[0060] Step S4.3 retains candidate points every 5 iterations to avoid prematurely discarding potential optimal solutions and to preserve multiple high-quality candidate centers during the search process. By using multi-factor weighting, it avoids misjudgments caused by relying on only a single indicator (such as the lowest pressure point, for example, the lowest pressure point may not be in the true main center when there are multiple sub-centers). The final selected optimal solution simultaneously satisfies thermodynamic, dynamic, and vertical motion conditions, exhibiting stronger physical consistency.

[0061] It also includes constructing a three-dimensional gradient coupling factor (GCF) to quantify the thermodynamic-dynamic coupling strength near the cyclone center. This GCF integrates pressure gradient, temperature advection, and vertical vorticity, serving as a guiding principle for gradient descent in the spiral convergence search algorithm. The GCF integrates three independent physical quantities, making it more sensitive to the cyclone center.

[0062] Step S5 specifically includes: S5.1 Physical consistency verification includes calculating the pressure gradient concentration index. , where R min Let R be the minimum radius centered at the candidate point that can detect a complete closed isobar. min =100km). R is the average pressure gradient at this closed isobar; max Let R be the outer search boundary. max =300km). The average pressure gradient within this radius is used. Based on historical case statistics, PGCI ≤ 0.6 is determined as the criterion for the true cyclone center. At the same time, the presence of counterclockwise rotational convergence of the wind field around the candidate point (the difference between the surface wind field offset angle and the theoretical gradient wind direction ≤ 45°) is used as an additional verification condition. Both conditions must be met to pass the consistency test. S5.2 Path anomaly detection examines the rationality of candidate centers from two aspects: "continuity of cyclone movement" and "regional path pattern constraints".

[0063] (1) Cyclone movement continuity check: The movement of a temperate cyclone between adjacent positioning times (interval of 1 to 3 hours) should meet the basic physical constraint of a smooth path, and there should be no situation where the displacement distance is too large in a short period of time and the movement direction changes drastically. Let the current candidate center position be The location result at the previous moment was The displacement distance between the two is The angle of change in the direction of movement is ,in, This represents the azimuth angle of the current time relative to the previous time. It represents the azimuth angle of the previous moment relative to the moment before that; If D>D max Or Δθ>Δθ max And D>D min If the cyclone movement is not continuous, the candidate point is rejected and the S4 step of the re-iteration search is triggered; where D max =150km is the statistical upper limit of the maximum possible distance that a mid-latitude cyclone can travel in one hour, D min =50km is used to ensure that the steering angle constraint only takes effect when the displacement is large enough, in order to eliminate noise interference introduced by small-amplitude oscillations when the cyclone is stationary; Δθ max =60° is the maximum allowable turning angle in the direction of movement in adjacent time steps. If the current positioning result triggers a continuity anomaly, that is, the candidate point has "long distance jump" or "sharp turn" relative to the previous time step, the candidate point is rejected and the S4 step is triggered to re-iterate the search.

[0064] (2) Regional path pattern constraint: Based on historical path statistics, determine whether the geographical location of the current candidate center is within the area of ​​the main movement direction of the corresponding type of historical cyclone. If it falls within this area, it passes the climate path trend verification; if it deviates from this area, it triggers a low confidence mark and prompts for manual review.

[0065] In step S5.1, a true extratropical cyclone center exhibits the typical structural characteristic of "a large pressure gradient at the periphery and a relatively gentle gradient at the core." This means that the pressure gradient decreases and isobars become sparser as one approaches the center, resulting in a significantly smaller PGCI value (much less than 1). In contrast, false low-pressure centers (such as those caused by topographic disturbances or weak fluctuations) do not possess this radially decreasing gradient structure from the periphery to the core. The gradient difference between their interior and periphery is not significant, and their PGCI values ​​are close to or even greater than 1. Through historical case analysis and optimization, a threshold of PGCI ≤ 0.6 can effectively eliminate false centers and reduce the false alarm rate.

[0066] Step S5.2 employs a two-layer anomaly detection mechanism: cyclone center continuity check and regional path pattern constraint. This mechanism constrains the path rationality of the positioning results from both kinematic and climatological dimensions. Leveraging the inherent smoothness of extratropical cyclone movement paths and the regional climatic path patterns of different types of extratropical cyclones, the cyclone center position generally does not experience significant jumps or drastic changes in direction within adjacent time intervals. By setting upper limits for displacement distance and turning angle, the mechanism effectively eliminates abnormal jumps caused by data noise or algorithms getting trapped in local extrema. It also allows the current cyclone to deviate from its typical path within a reasonable range, triggering a low-confidence flag only when the direction of movement significantly exceeds the historical statistical distribution area. This design respects the climatic reality of the diversity of extratropical cyclone paths, retains the ability to detect extreme anomalies, and has low computational costs, making it suitable for real-time operational scenarios.

[0067] PGCI verifies the physical authenticity of the cyclone center by analyzing its horizontal spatial structure, while path anomaly detection verifies its rationality by analyzing the temporal and spatial macroscopic scale of the cyclone's trajectory. The combination of these two methods eliminates instantaneous false centers and avoids path deviations caused by anomalies in single-point data.

[0068] The following example, a Jianghuai cyclone affecting the Yellow Sea in 2025, illustrates in detail the extratropical cyclone center localization method based on multimodal data fusion and physical enhancement described in this invention. This embodiment is only used to illustrate the technical solution of this invention and does not constitute a limitation on the scope of protection.

[0069] I. Data Acquisition and Preprocessing (Step S1) For this Jianghuai cyclone event, step S1 was executed: S1.1: Integrate real-time ground observation data from over 3,000 automatic weather stations along the Yellow Sea coast and surrounding areas, including air pressure, wind direction, wind speed, and temperature. The raw data undergoes real-time verification (delay <2 min), extreme value verification (air pressure change rate <5 hPa / 10 min), and spatiotemporal consistency verification based on variational assimilation methods. For sparsely observed sea areas, pseudo-observations are generated using a virtual sensor network based on air pressure-topography relationships to fill in the gaps.

[0070] S1.2: Download ECMWF ERA5 reanalysis data and extract the geopotential height, temperature, wind field, and vertical velocity at four layers: 1000 hPa, 850 hPa, 700 hPa, and 500 hPa. Calculate the moist potential vortex (MPV) to capture the baroclinic energy distribution of the cyclone.

[0071] S1.3: Receives infrared (10.8μm) and water vapor (6.7μm) channel data from the Fengyun-4 satellite, with a spatial resolution of 0.5km and a temporal resolution of 5min. The cloud top brightness temperature (TBB) and humidity gradient field are extracted using the cloud system boundary index algorithm to assist in identifying the range and morphology of cyclone cloud systems.

[0072] S1.4: Simultaneously access millimeter-wave cloud radar (spatial resolution 30m), wind profiler radar (temporal resolution 6min), S-band weather radar, microwave radiometer, and CMA-MESO mesoscale numerical forecast products (temporal and spatial resolution 3km / 1~3h) as supplementary high-precision data sources.

[0073] S1.5: Data Spatiotemporal Alignment. All data are unified to a 0.1°×0.1° regular grid using inverse distance weighted interpolation. The interpolation formula is... k=2. Time synchronization uses cubic spline interpolation. This aligns asynchronous data to the same timeline.

[0074] II. Construction and Fusion of Three-Dimensional Feature Fields (Step S2) Execute step S2: S2.1 Hierarchical Feature Extraction: Thermal field: Calculation of temperature advection ; Calculate wet hydrostatic energy ,in .

[0075] Dynamic field: Calculating vorticity ; Calculate the Q-vector divergence .

[0076] S2.2 Tensor Fusion and Dimensionality Reduction: The above characteristic fields and the surface pressure gradient ∇P are constructed into a four-dimensional characteristic tensor. Where H and W represent the number of latitude and longitude grids, D=4 (vertical layers), and C=6 (channels: T, RH, ζ, Qdiv, MSE, ∇P). Tucker decomposition and dimensionality reduction are used to preserve the core tensors. H'=0.7H, W'=0.7W, D'=3, C'=4, energy retention rate ≥95%.

[0077] III. Dynamic Weight Adjustment (Step S3) Execute step S3: S3.1 Data quality assessment: Calculate the satellite cloud image feature consistency index CCCI (threshold ≥ 0.7), reanalysis set dispersion σENS (based on ECMWF 10 members), and ground observation continuity score TPS.

[0078] S3.2 Dynamic Update of Bayesian Weights: Establishing the Likelihood Function k=1, 2, 3 correspond to satellite, reanalysis, and ground, respectively. The posterior distribution is sampled using the Metropolis-Hastings algorithm, and the weighted mean is taken after 1000 iterations. A physical consistency constraint is introduced: wind direction and cloud movement deviation <30°, and the weight change rate in a single iteration ≤10%.

[0079] S3.3 Context-Aware Weighting Mode Switching: The convolutional neural network identifies the current cyclone type as a "Jianghuai Cyclone" and automatically switches to the frontal cyclone weighting mode. The initial weights are set to Ground:Reanalysis:Satellite = 0.5:0.4:0.1. After dynamic updates, the weight of ground data increases to 0.58, the weight of reanalysis data is 0.38, and the weight of satellite data decreases to 0.04.

[0080] IV. Spiral Convergence Iterative Localization (Step S4) Execute step S4: S4.1 Path Planning: Using Archimedes Spiral Parametric Method Combined with the momentum acceleration gradient direction GCF is a three-dimensional gradient coupling factor (integrating pressure gradient, temperature advection and vertical vorticity).

[0081] S4.2 Adaptive Step Size: Adjusting the step size according to the Armijo criterion ,like Then accept the step size, otherwise .

[0082] S4.3 Multiple Hypothesis Testing: Candidate center points are retained every 5 iterations. Through a comprehensive scoring function Select the optimal solution.

[0083] In this Jianghuai cyclone location, the algorithm converged in only 8 iterations, reducing the computational load by 60% compared to the traditional gradient descent method (average 20 iterations).

[0084] V. Verification and Output (Step S5) Execute step S5: S5.1 Physical Consistency Verification: Calculation of Pressure Gradient Concentration Index Where R is taken min =100km, R max =300km). This represents the average pressure gradient within this radius. The positioning result is PGCI=0.41, with a wind direction deviation of 30°, thus excluding false low-pressure centers.

[0085] S5.2 Path Anomaly Detection: Cyclone movement continuity checks and regional path pattern constraints are employed. The path is compared with the historical Jianghuai cyclone path database. The calculated center displacement distance D is between 30 and 50 km, and the movement direction angle Δθ is between 10° and 50°, which passes verification. The final output is the cyclone center coordinates (e.g., 34.2°N, 122.8°E). Implementation effect

[0086] Tested during a cyclone weather event in 2025, this invention was effective at 110... ° E~126 ° E, 30 ° N~41 ° Within the N-range, the positioning was performed 1342 times, with 1100 correct positionings, achieving a success rate of 82%. This technical approach, through refined process design and enhanced physical constraints, achieves high precision, high efficiency, and high reliability in the positioning of extratropical cyclone centers, meeting the needs of meteorological service operations.

[0087] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement. Its features are, Includes the following steps: S1. Collect multi-source meteorological data, which includes at least ground observation data, numerical model reanalysis data and satellite remote sensing data; S2. Construct a three-dimensional thermodynamic-dynamic coupled feature field for the baroclinic structure of a mid-latitude cyclone. The three-dimensional thermodynamic-dynamic coupled feature field is used to characterize the thermodynamic and dynamic field features of a mid-latitude cyclone. S3. Based on the structure type and data quality assessment results of extratropical cyclones, dynamically adjust the weights of different data sources in the location calculation; S4. A spiral convergence search algorithm that integrates the Archimedes spiral trajectory and the physical gradient descent direction is used to iteratively search for the cyclone center in the three-dimensional thermo-dynamic coupling feature field. S5. Perform physical consistency verification and path anomaly detection verification on the candidate locations of the cyclone center obtained from the search, and output the final coordinates of the cyclone center.

2. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Ground observation data is collected in real time by automatic weather stations and undergoes triple quality control, including real-time verification, extreme value verification, and spatiotemporal consistency verification based on variational assimilation method. For observation blank areas, pseudo observation values ​​are generated by virtual sensor network based on air pressure-topography relationship. S1.2 Numerical model reanalysis data extraction at four layers (1000 hPa, 850 hPa, 700 hPa, and 500 hPa) includes geopotential height, temperature, wind field, and vertical velocity, and calculates moist potential vorticity. To capture baroclinic energy, where g is gravitational acceleration, ζ is relative vorticity, and f is the geostrophic parameter. For pseudo-equivalent temperature, For air pressure, ; S1.3 Satellite remote sensing data is used to assist in identifying the extent and shape of cloud systems, and cloud system boundary indices are constructed through multispectral channels: ,in , These are cloud top brightness temperature (TBB) and humidity gradient, respectively. Weighting coefficients; S1.4 The multi-source meteorological data also includes millimeter-wave cloud radar data, wind profiler radar data, S-band weather radar data, microwave radiometer data, and mesoscale high spatiotemporal resolution numerical forecast data. S1.5: Data Spatiotemporal Alignment: Spatial interpolation and temporal synchronization are performed on the multi-source meteorological data from steps S1.1 to S1.4, unifying the data to a regular grid of 0.1° × 0.1°. The spatial interpolation uses inverse distance weighted interpolation, and the interpolation formula is as follows: ,in, Points to be interpolated Meteorological element values, Let i be the value of the i-th observation point. For the interpolation point and the first The distance between observation points, where k=2 is the power of the distance; Time synchronization uses cubic spline interpolation, the formula is as follows: ,in, The coefficients a, b, c, and d are determined by fitting data from four adjacent time points.

3. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 1, characterized in that, Step S2 specifically includes: S2.1 Hierarchical Feature Extraction: Thermal field construction: Calculation of temperature advection ,in The horizontal wind vector These are the zonal and meridional wind speed components, respectively. These represent the partial derivatives of temperature in the latitudinal and longitudinal directions, respectively; and the hydrostatic energy field. ,in For isobaric specific heat capacity, Where is the latent heat of vaporization, g is the acceleration due to gravity, z is the potential height, and q is the specific humidity; Dynamic field construction: Calculation of vortex field and Q-vector divergence Where Q1 and Q2 are the Q vector components calculated according to the Hoskins quasi-geotropic Q vector definition; S2.2 Tensor Fusion and Dimensionality Reduction: The characteristic field of step S2.1 and the ground pressure gradient field Together they are constructed into a four-dimensional feature tensor Where H is the number of grids in the longitude direction, W is the number of grids in the latitude direction, D = 4 is the number of vertical layers, corresponding to the four pressure layers of 1000hPa, 850hPa, 700hPa, and 500hPa respectively, and C = 6 is the number of characteristic channels, which specifically include: corresponding pressure layer temperature T, relative humidity RH, vorticity Q-vector divergence Moist hydrostatic energy (MSE) and surface pressure gradient ; The four-dimensional feature tensor is reduced in dimensionality using Tucker decomposition, while preserving the tensor. Where H' = 0.7H, W' = 0.7W, D' = 3, C' = 4, and the energy retention rate is not less than 95%.

4. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1 Data Quality Assessment: Calculating the Consistency Index of Satellite Cloud Image Features Where 0.5 in the denominator is the normalization factor, and the threshold is... ; Calculate the set discreteness of the reanalysis ,in To predict the potential height of the i-th member of the ECMWF ensemble, The mean of 10 members is 9, and the degrees of freedom are 9; calculate the ground observation continuity score (TPS). S3.2 Dynamic Update of Bayesian Weights: Dynamic updates of weights are achieved through a Bayesian posterior probability model, with the likelihood function being... Where k=1,2,3 represent data sources from satellite, reanalysis, and ground observations, respectively. For the observed values, The mean of the true values. For variance, The weights to be optimized are: the posterior distribution is sampled using the Metropolis-Hastings algorithm, the weight mean is taken after 1000 iterations, the weights of each data source are dynamically updated, and physical consistency constraints are introduced, including wind direction and cloud movement deviations of less than 30°, and the weight change rate in a single iteration not exceeding 10%. S3.3 Context-aware weight mode switching: At least three weight configuration modes are preset, corresponding to frontal cyclones, non-frontal cyclones, and explosive cyclones, respectively. The weather type of the extratropical cyclone is identified in real time through a convolutional neural network to switch weight modes; the initial weight is set to ground observation: reanalysis: satellite = 0.5:0.4:0.

1.

5. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 1, characterized in that, Step S4 specifically includes: S4.1 Path Planning: Archimedes Spiral Path Parametric Method ,in, Let the initial radius be , For spiral growth rate, N = 36 represents the number of spiral rotations; and this is combined with the gradient descent direction vector of momentum acceleration. Where 0.8 and 0.2 are the momentum coefficient and the current gradient coefficient, respectively, and the three-dimensional gradient coupling factor. α=0.4, β=0.3, and γ=0.3 are the weighting coefficients of the pressure gradient term, the 850hPa temperature advection term, and the 500hPa vorticity term, respectively. The sum of the three weights is 1. norm(·) indicates that the range is normalized for the entire field. S4.2 Adaptive step size adjustment: Adjust the step size according to the Armijo criterion. ,like Then accept the step size; otherwise, let Where 0.3 is the sufficient decrease coefficient; S4.3 Multiple Hypothesis Verification of Candidate Center Set: Retain candidate center points every 5 iterations The optimal solution is selected through a comprehensive scoring function: Where 0.5, 0.3, and 0.2 are the weighting coefficients for the three-dimensional gradient coupling factor, air pressure curvature, and 700 hPa vertical velocity, respectively. For the pressure Laplace operator, The vertical velocity is 700 hPa.

6. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 1, characterized in that, Step S5 specifically includes: S5.1 Physical consistency verification includes calculating the pressure gradient concentration index. , where R min Let R be the minimum radius centered at the candidate point that can detect a complete closed isobar. min =100km; R is the average pressure gradient at this closed isobar; max Let R be the outer search boundary. max =300km, The average pressure gradient within this radius is defined as follows: PGCI ≤ 0.6, and the difference between the surface wind field offset angle around the candidate point and the theoretical gradient wind direction is ≤ 45°. S5.2 Path anomaly detection examines the rationality of candidate centers from two aspects: the continuity of cyclone movement and the constraints of regional path patterns.

7. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 6, characterized in that, Step S5.2 includes: (1) Cyclone movement continuity check: Let the current candidate center position be The location result at the previous moment was The displacement distance between the two is The angle of change in the direction of movement is ,in, This represents the azimuth angle of the current time relative to the previous time. It represents the azimuth angle of the previous moment relative to the moment before that; If D>D max Or Δθ>Δθ max And D>D min If the cyclone movement is not continuous, the candidate point is rejected and the S4 step of the re-iteration search is triggered; where D max =150km is the statistical upper limit of the maximum possible distance that a mid-latitude cyclone can travel in one hour, D min =50km is used to eliminate noise interference introduced by small-amplitude oscillations when the cyclone stagnates; Δθ max =60° is the maximum allowable turning angle in the adjacent time movement direction. If the current positioning result triggers a continuity anomaly, the candidate point is rejected and the S4 step of re-iteration search is triggered. (2) Regional path pattern constraint: Based on historical path statistics, determine whether the geographical location of the current candidate center is within the area of ​​the main movement direction of the corresponding type of historical cyclone. If it falls within this area, it passes the climate path trend verification; if it deviates from this area, it triggers a low confidence mark and prompts for manual review.

8. The method for locating the center of a mid-latitude cyclone based on multimodal data fusion and physical enhancement according to claim 6, characterized in that, In step S2.2: For the surface pressure gradient ∇P, which is a surface variable, when constructing the four-dimensional tensor, the horizontal distribution value of ∇P is repeatedly assigned in four vertical layers of 1000hPa, 850hPa, 700hPa, and 500hPa. Wet hydrostatic energy MSE and Q vector divergence All are derived diagnostic parameters. MSE is calculated layer by layer from the temperature, potential height, and specific humidity of each layer. The horizontal wind field and temperature field of each layer are calculated layer by layer, and both are calculated separately according to the four pressure layers to form a complete four-layer vertical distribution.

Images