Satellite-based temperature and humidity profile and severe convective short-term nowcast product generation method and system

Abstract

The present application relates to the technical fields of meteorological disaster prediction, satellite remote sensing data processing, machine learning short-term prediction and meteorological business product generation, and particularly relates to a method and system for generating short-term severe convective weather prediction products based on satellite temperature and humidity profiles, comprising: obtaining satellite temperature and humidity profile data, minute-level precipitation real-time data and terrain data, performing coverage check, quality control and standard pressure layer interpolation calculation on the satellite temperature and humidity profile data, mapping the convective parameters to the standard grid, performing spatial aggregation statistics on the minute-level precipitation real-time data to generate a severe convective weather risk label for a future preset time window, constructing a candidate feature set based on the standard grid convective parameters and performing screening, constructing a training sample with the risk label and training a machine learning model, and generating a severe convective weather short-term prediction product based on the risk inference of the new time satellite parameter product. The present application reduces the influence of inconsistent coverage of satellite profile data and quality decline in cloud area, and outputs a severe convective weather risk product that can be deployed in business.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological disaster prediction, satellite remote sensing data processing, machine learning short-term prediction and meteorological business product generation, in particular to a satellite temperature and humidity profile-based and strong convective short-term prediction product generation method and system. BACKGROUND

[0002] Strong convective weather includes hail, short-time heavy precipitation, thunderstorm gale and other types, which has the characteristics of sudden occurrence, short life history, small spatial scale and strong disaster. For the complex terrain region of Gansu and other places in Northwest China, the strong convective process is often jointly affected by water vapor transport, low-level thermal condition, terrain lifting and local triggering mechanism. Therefore, it is necessary to generate grid risk products for the next 0 to 2 hours or the next 2 hours in business. FY-4B GIIRS infrared hyperspectral detection data can provide temperature and humidity profile inversion information, which can be used to calculate convective parameters such as CAPE, CIN, K index, total-layer precipitation, relative humidity, dew point temperature and other convective parameters, to provide environmental field basis for strong convective potential judgment. Minute-level precipitation real-time data can more accurately depict local peak values in short-time heavy precipitation and hail processes. The combination of satellite temperature and humidity profile, convective parameters, minute-level precipitation label and machine learning model can form a strong convective risk product for the next 2 hours for business application.

[0003] Chinese patent CN119620244A discloses a severe convective short-term forecast method based on deep learning and polarization vertical profile, which extracts polarization vertical profile time series features based on radar networking polarization and wind field three-dimensional grid data, and judges whether severe convection occurs in the future period through a deep learning model and a probability threshold. However, its data sources are mainly dual-polarization radar and wind field, and it does not involve the version difference, spatial coverage, quality mark and standard layer interpolation processing of FY-4B GIIRS temperature and humidity profile data. Chinese patent CN117805941B discloses a method for establishing a severe convective potential early warning method, which analyzes FY-4A and FY-4B satellite product source data and combines ground and upper air meteorological sounding and model zero field data to extract severe convective warning and potential forecast indicators. However, its technical focus is on multi-source indicator extraction, and it lacks limitations on GIIRS temperature and humidity profile quality control interpolation, convective parameter product library and business product output chain. Chinese patent CN117518299A discloses a classified severe convective near-term probability forecast method, system, device and terminal, which establishes a polarization vertical profile feature model for classifying severe convection and realizes 0 to 2 hour classified severe convective probability forecast by combining an improved extrapolation scheme. However, its label uses automatic station hourly rainfall, maximum wind and lightning record neighborhood determination, which is difficult to fully retain minute-level heavy rainfall peak and local burst signal. Chinese patent CN111796343B discloses a severe convective weather short-term near-term identification method based on artificial intelligence algorithm, which cleans, quality controls, trains and tests real-time station and radar data, and uses LightGBM for severe convective identification, training evaluation and parameter optimization. However, its main input is radar echo and ground observation, and it lacks processing of convective parameters derived from FY-4B GIIRS temperature and humidity profile. Chinese patent CN117950082A discloses a deep learning meteorological data prediction method and device based on microtopography, which encodes meteorological element data and microtopography data and inputs them into a prediction model after fusion. However, it lacks expression of terrain-adapted convective parameters based on DEM and GIIRS profile minimum effective starting layer.

[0004] The above existing technologies mainly have the following deficiencies: First, some schemes use radar, ground station or numerical model data as the main input, and lack processing of version difference, spatial coverage, quality mark, cloud mask and standard layer interpolation of FY-4B GIIRS temperature and humidity profile data, making it difficult to directly form a unified spatial reference convective parameter product library.

[0005] Second, some schemes use hourly rainfall, lightning or wind events as labels, which can weaken minute-level heavy rainfall peak and local burst signal, making it difficult to fully correspond to the rapid evolution of short-term heavy rainfall and hail processes.

[0006] Third, the fixed pressure layer index in complex terrain regions may be affected by ground height and the absence of effective layers. Existing publicly available schemes do not adequately express the terrain adaptation flow parameters based on DEM and effective starting layer.

[0007] Fourth, existing machine learning identification solutions mostly emphasize model classification or probability prediction itself, and do not adequately explain the joint design of low-risk sample extraction, category weights, probability thresholds, and business indicators such as TS, POD, and FAR.

[0008] Fifth, some research-based model outputs lack a standard product conversion chain with business systems, making it difficult to directly generate archival, displayable, and locally deployable high-risk convection products such as CSV, PKL, MICAPS, and PNG.

[0009] Therefore, there is a need for a method and system that can perform full version screening, coverage checks, quality control, and standard layer interpolation on satellite temperature and humidity profile data, construct severe convective risk labels based on minute-level precipitation, combine complex terrain with parameter and feature screening, and directly output operational format products for short-term severe convective forecasts. Summary of the Invention

[0010] This invention aims to provide a method and system for generating short-term forecast products for severe convection based on satellite temperature and humidity profiles and minute-level precipitation labels, in order to solve or improve problems such as inconsistent coverage of satellite profile data, degraded cloud data quality, coarse time scale of severe convection labels, insufficient representativeness of complex terrain parameters, imbalance of severe convection samples, and difficulty in operationalizing model results.

[0011] To address the aforementioned technical problems, this invention provides a method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection, comprising the following steps: Step S1: Acquire satellite temperature and humidity profile data, minute-level precipitation data, and topographic data; Step S2: Perform coverage integrity checks, quality control, and standard barosphere interpolation on the satellite temperature and humidity profile data to obtain quality-controlled temperature and humidity profile data; Step S3: Calculate the convection parameters based on the quality-controlled temperature and humidity profile data, and map the convection parameters to a standard grid; Step S4: Spatial aggregation of the minute-level precipitation data, using satellite parameter time as the starting time to statistically analyze precipitation within a future preset time window, and generate severe convection risk labels; Step S5: Construct a candidate feature set based on the convection parameters on the standard grid, perform feature screening on the candidate feature set, and construct training samples with the strong convection risk label to train the machine learning model; Step S6: Use the trained machine learning model to perform risk reasoning on the new satellite parameter products and generate short-term forecast products for severe convection.

[0012] Further, in step S2, the satellite temperature and humidity profile data is FY-4BGIIRS temperature and humidity profile inversion data; the FY-4BGIIRS temperature and humidity profile inversion data is version-screened, and valid time data with a two-dimensional spatial structure and complete spatial coverage are selected as the standard processing version; the standard processing version is subjected to a coverage integrity check, which includes checking the latitude and longitude range, number of valid points, and coverage of the operational area for data of different versions, different time periods, or different observation centers, and data that cannot completely cover the operational area is removed.

[0013] Further, in step S2, the quality control includes: removing filler values ​​and non-finite values; screening effective layers based on temperature and humidity quality indicators, retaining temperature and humidity layers with quality indicators less than a preset threshold; screening effective grid points based on cloud mask information, and outputting quality diagnostic information such as cloud mask, effective layer ratio, and interpolation ratio; for profiles that meet the minimum number of effective layers requirement, identifying missing layers within the original effective air pressure range, and performing linear interpolation on temperature and specific humidity; the standard air pressure layer interpolation includes: interpolating temperature and specific humidity to a preset number of standard air pressure layers; when the surface air pressure is greater than the minimum effective layer air pressure and the surface temperature and surface air pressure are effective, adding the surface air pressure layer to the profile, using the surface temperature for temperature, and linearly extrapolating the specific humidity based on the adjacent near-surface layer; converting the specific humidity to dew point temperature, and ensuring that the dew point temperature is not higher than the ambient temperature.

[0014] Further, in step S3, the convection parameters include at least one of CAPE, CIN, K index, total precipitation, lift index, Showalter index, TotalTotals index, relative humidity, dew point temperature, and temperature lapse rate.

[0015] Further, in step S3, terrain adaptation flow parameters are constructed based on the terrain data and the lowest effective pressure layer of the profile; the terrain adaptation flow parameters include at least one of the following: complex terrain adaptation K-type index, complex terrain adaptation Showalter-type index, complex terrain adaptation TotalTotals-type index, temperature lapse rate from the lowest effective layer to the preset pressure layer, and key layer availability indicator.

[0016] Furthermore, in step S4, the station quality is checked on the minute-level precipitation data to remove missing values, outliers, and records that do not meet the time matching requirements; the spatial aggregation is maximum value aggregation, and when the minute-level precipitation data is mapped to the standard grid, the maximum minute precipitation value of the stations within the neighborhood of the standard grid point is taken as the minute precipitation data at the corresponding time of the grid point.

[0017] Further, in step S4, taking the satellite parameter time as the reporting start time, the minute precipitation characteristics of each standard grid point within the future preset time window are statistically analyzed. The minute precipitation characteristics include at least one of the following: maximum minute precipitation, average minute precipitation, number of minutes with non-zero precipitation, number of minutes with heavy precipitation, and whether a precipitation event has occurred. Based on the maximum minute precipitation and a preset minute precipitation threshold, three categories of labels are defined: low risk, medium risk, and high risk. When the maximum minute precipitation is equal to zero, it is marked as low risk; when the maximum minute precipitation is greater than zero and less than the preset minute precipitation threshold, it is marked as medium risk; and when the maximum minute precipitation is greater than or equal to the preset minute precipitation threshold, it is marked as high risk.

[0018] Further, in step S5, a candidate feature pool is constructed, which includes basic convection parameters, terrain adaptation parameters, temperature-derived features, physical combination features, time-varying features, time-lag features, and spatial statistical features. Core features are selected from the candidate feature pool based on feature importance, preset feature quantity sensitivity, and business indicators. All medium-risk and high-risk samples are retained, and low-risk samples are extracted according to a preset ratio. Higher category weights are set for high-risk categories. The machine learning model is a multi-class probability model that outputs three probabilities: low-risk, medium-risk, and high-risk. The machine learning model includes at least one of XGBoost, LightGBM, random forest, gradient boosting tree, neural network, or a combination thereof. During training, the TS index is used as the main business evaluation index, and the high-risk probability threshold is searched and optimized with reference to POD, FAR, and neighborhood TS indexes.

[0019] Further, in step S6, the severe convective short-term forecast product includes at least one of CSV format product, PKL format product, MICAPS format product and PNG format product; the severe convective short-term forecast product includes fields such as forecast start time, forecast start time, forecast end time, longitude, latitude, risk level, low risk probability, medium risk probability, high risk probability and risk score.

[0020] The system for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection includes a data acquisition module, a satellite data quality control and coverage screening module, a convection parameter generation module, a minute-level precipitation label construction module, a feature construction and screening module, a model training and optimization module, a risk reasoning module, and a product output module. The data acquisition module acquires satellite temperature and humidity profile data, minute-level precipitation data, and topographic data; The satellite data quality control and coverage screening module performs coverage integrity checks, quality control, and standard barosphere interpolation on the satellite temperature and humidity profile data to obtain quality-controlled temperature and humidity profile data. The convection parameter generation module calculates convection parameters based on the quality-controlled temperature and humidity profile data, and maps the convection parameters to a standard grid. The minute-level precipitation label construction module spatially aggregates the minute-level precipitation data to generate severe convective risk labels. The feature construction and filtering module constructs a candidate feature set based on the convection parameters on the standard grid and filters the core features; The model training and optimization module trains a machine learning model based on the core features and the strong convection risk label, and optimizes the class weights and probability thresholds. The risk reasoning module uses a trained machine learning model to perform risk reasoning on new satellite parameter products and outputs risk probability, risk level and risk score. The product output module converts the risk probability, the risk level, and the risk score into a short-term forecast product for severe convection.

[0021] The advantages of this invention compared to the prior art are: This invention reduces spatial reference inconsistencies caused by differences in coverage between different data versions, observation centers, or time periods by screening satellite temperature and humidity profile data by version, observation time, and coverage integrity checks, thereby improving the spatiotemporal consistency of the convective parameter product library.

[0022] This invention identifies and corrects low-quality profiles, missing layers, and anomalous layers through processing such as quality marking, cloud masking, effective layer ratio, interpolation ratio, and standard barosphere interpolation. This reduces the impact of infrared hyperspectral profile data on the calculation of convective parameters such as CAPE, CIN, and K-index in cloud areas or low-quality regions. Furthermore, by using minute-level spatial maximum aggregation of precipitation and statistical analysis of future preset time windows, it retains the peak values ​​of short-term, localized heavy precipitation, reducing the smoothing effect on sudden strong convective signals and making the training labels more closely aligned with short-term strong convective operational objectives.

[0023] This invention improves the problem of insufficient representativeness of fixed pressure layer indicators in complex terrain areas by introducing topographic data and the lowest effective pressure layer to construct terrain-adaptive flow parameters, thereby enhancing the ability to express grid stability, water vapor conditions and stratification characteristics in high-altitude and mountainous areas.

[0024] This invention alleviates the class imbalance problem caused by the scarcity of high-risk samples in strong convection by extracting low-risk samples, setting class weights, and optimizing probability thresholds. It also makes model training and threshold selection more closely aligned with business evaluation metrics such as TS, POD, and FAR. Through a standardized business output process, it converts model inference results into products in formats such as CSV, PKL, MICAPS, and PNG, realizing a continuous process from satellite data processing and model inference to display in the business system, which facilitates localized deployment, business consultation, and product archiving. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0026] In the attached diagram: Figure 1 This is a flowchart of the method for generating short-term forecast products of severe convection based on satellite temperature and humidity profiles and the present invention.

[0027] Figure 2 This is a diagram illustrating the overall architecture of the system for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection, as described in this invention. Detailed Implementation

[0028] The detailed description of the following embodiments is used to illustrate the principles of this application, but should not be used to limit the scope of this application. That is, the method and system for generating satellite temperature and humidity profiles and strong convection short-term forecast products are not limited to the described embodiments.

[0029] The present invention will be further described below with reference to embodiments.

[0030] Example 1 like Figure 1 , 2 As shown, this embodiment first acquires FY-4BGIIRS data and performs coverage checks, quality control, and interpolation. Then, it calculates convection parameters and performs standard gridding. On the other hand, it acquires minute-level precipitation data and performs spatial maximum aggregation. It constructs three-class labels according to the next 2-hour window. Then, it uses the standard gridded convection parameters as model input features and uses them together with the label samples to train the XGBoost three-class business model. Finally, it performs risk inference on the new time data to generate a strong convection risk product for the next 2 hours.

[0031] This embodiment uses FY-4B GIIRS L2 temperature and humidity profile data is used as the main input. The original data is in NetCDF format, with the file name containing the AVP product identifier. The spatial resolution of the data is 12km nadir resolution, and it contains 101 standard pressure layers in the vertical direction. The main variables include pressure profile, temperature profile, specific humidity profile, geopotential height, temperature quality indicator, humidity quality indicator, longwave latitude and longitude, cloud mask, digital elevation, and land-sea mask. In the data screening stage, this embodiment compares two types of GIIRS files, V1 and V0001. The V1 file mainly shows a single time period and one-dimensional spatial structure, with the coverage concentrated in local strips. The V0001 file adopts a two-dimensional spatial structure and contains residence point or time information, with more complete spatial coverage and is more suitable for building a standard product library of convective parameters in Gansu region. Therefore, this embodiment prioritizes the use of valid time period data of V0001 as the standard processing version, and performs a coverage integrity check on the data with the first 00 time period that is too far south or has insufficient valid points, and removes the files that cannot completely cover the business area.

[0032] In the quality control phase, this embodiment employs a grid-by-grid profile inspection method. First, filled values ​​and non-finite values ​​are eliminated. Then, effective layers are selected based on temperature and humidity quality indicators. In the current processing configuration, layers with quality indicators less than 3 are prioritized for retention; that is, layers of better quality (classes 0, 1, and 2) are used. Layers with quality indicators of 3 or higher are treated as low-quality layers. Each grid point must retain at least 3 effective pressure layers, and the number of effective layers below 100 hPa must meet the parameter calculation requirements. For profiles that meet the minimum effective layer requirement, this embodiment first identifies missing layers within the original effective pressure range and performs linear interpolation on temperature and specific humidity. Subsequently, the profiles are uniformly interpolated to 24 standard pressure layers, including 100, 125, 150, 175, 200, 225, 250, 300, 350, 400, 450, 500, 550, 600, 650, 700, 750, 800, and 850. For 900, 925, 950, 975, and 1000 hPa, if the original data provides surface pressure and surface temperature, and the surface pressure is greater than the current lowest effective layer pressure, then the surface pressure layer is added to the profile. The temperature is the surface temperature, and the specific humidity is linearly extrapolated based on the adjacent near-surface layers and limited to a physically reasonable range. The interpolated specific humidity is further converted to dew point temperature and corrected by the physical constraint that the dew point temperature should not be higher than the ambient temperature. The process also outputs quality diagnostic variables such as cloud mask CLM, effective layer ratio valid_ratio, and interpolation ratio interp_ratio. Among them, valid_ratio is used to represent the proportion of usable layers after quality control and interpolation processing to the original number of layers, and interp_ratio is used to represent the proportion of interpolated supplementary layers to the final calculated layers. Parameters such as CAPE, CIN, and K index are all calculated based on the processed temperature, dew point temperature, and pressure profiles.

[0033] To address the issue of insufficient representativeness of fixed pressure layers in the complex terrain of Gansu, this embodiment further combines DEM and profile minimum effective pressure layer to construct supplementary parameters such as complex terrain-adapted K-type index, complex terrain-adapted Showalter-type index, complex terrain-adapted TotalTotals-type index, temperature lapse rate from minimum effective layer to 500 hPa, and key layer availability indicator. These parameters are then regrinded to the Gansu standard grid, enabling the use of a unified spatial reference for convection parameters, precipitation labels, model samples, and operational products.

[0034] The FY-4BGIIRS convection parameter product generated in this embodiment mainly includes categories such as basic parameters, stability index, humidity parameters, temperature parameters, and complex terrain adaptation parameters, among which: In the basic parameters, CAPE is calculated based on the integral of the air parcel uplift profile and the ambient temperature and humidity profile to determine the positive buoyancy energy, CIN is calculated based on the integral of the air parcel uplift profile and the ambient temperature and humidity profile to determine the negative buoyancy energy, and the K-index is calculated in the key layers at 850, 700, and 500 hPa using the following formula: in, The temperature is 850 hPa. The temperature is 700 hPa. At a temperature of 500 hPa, The dew point temperature is 850 hPa. The dew point temperature is 700 hPa.

[0035] The total precipitation capacity of the entire floor is calculated by integrating along the entire profile: in, For wetness, For air pressure, This is the acceleration due to gravity.

[0036] In the stability index, the uplift index LI uplifts a representative low-level gas parcel to 500 hPa and is calculated using the following formula: in, An ambient temperature of 500 hPa. The temperature at which the gas parcel rises to 500 hPa.

[0037] The Showalter index raises the pressure of an 850 hPa gas parcel to 500 hPa and is calculated using the following formula: in, An ambient temperature of 500 hPa. The temperature at which the gas parcel rises to 500 hPa.

[0038] The Totals index at 850 and 500 hPa is calculated using the following formula: in, The temperature is 850 hPa. At a temperature of 500 hPa, The dew point temperature is 850 hPa.

[0039] Humidity parameters include relative humidity at 700 hPa and 850 hPa, average relative humidity from 700 to 850 hPa, and dew point temperature at 700 hPa and 850 hPa. The relative humidity is calculated based on the corresponding layer temperature and dew point temperature, the average relative humidity within the calculation range, or directly extracted.

[0040] Temperature parameters include 700 hPa and 850 hPa temperatures and the temperature lapse rate from the lower to the middle layers, where the temperature lapse rate from the lower to the middle layers is calculated by the following formula: in, For lower temperature, The temperature is 850 hPa. The height difference is between the two floors.

[0041] Among the parameters for adapting to complex terrain, the K-type index for adapting to complex terrain is calculated by the following formula: ; in, The lowest effective starting layer temperature, The lowest effective starting layer dew point temperature, The temperature is 700 hPa. At a temperature of 500 hPa, The dew point temperature is 700 hPa.

[0042] The Showalter-type index for complex terrain adaptation raises the lowest effective air parcel to 500 hPa and compares the ambient temperature with the air parcel temperature. The TotalTotals-type index for complex terrain adaptation is calculated using the following formula: in, The lowest effective starting layer temperature, The lowest effective starting layer dew point temperature, The temperature is 500 hPa.

[0043] The temperature lapse rate from the lowest effective layer to 500 hPa is calculated by the following formula: in, The lowest effective starting layer temperature, At a temperature of 500 hPa, From average temperature and Estimate.

[0044] Among the aforementioned convection parameters, each parameter has corresponding requirements for the quality of the input profile data. Specifically, CAPE, CIN, and total precipitable water require complete profiles and tolerate local omissions; total precipitable water is calculated by integrating the entire profile. The K-index requires at least two key layers and can be interpolated. The lifting index and Showalter index strictly require complete profiles. The TotalTotals index requires 850 and 500 hPa data. The 700 hPa and 850 hPa relative humidity, dew point temperature, and temperature are single-layer data and can be interpolated. The average relative humidity from 700 to 850 hPa requires at least one layer to be valid within the range. The temperature lapse rate from the lower to middle layers... Temperatures of 850 and 500 hPa are required. The K-type index for complex terrain is designed for complex terrain and machine learning features and is not equivalent to the traditional K-index. The Showalter-type index for complex terrain is used to supplement features to replace the fixed 850 hPa starting layer and is not equivalent to the traditional Showalter index. The TotalTotals-type index for complex terrain is used to supplement the representation of complex terrain areas and is not equivalent to the traditional TotalTotals. The temperature lapse rate from the lowest effective layer to 500 hPa requires the original effective profile to be complete. The recommended range for product-level applications is 0 to 20 K / km. Values ​​outside this range should be set to NaN.

[0045] In addition to the main convection parameters mentioned above, this embodiment also synchronously outputs quality control and auxiliary diagnostic variables, including the lowest effective starting layer pressure ml_terrain_start_pressure, the lowest effective starting layer temperature ml_terrain_start_temp, the lowest effective starting layer dew point temperature ml_terrain_start_td, the 850hPa key layer availability flag ml_terrain_850_available, cloud mask clm, terrain elevation dem, the effective layer ratio after quality control valid_ratio, and the interpolation supplementary layer ratio interp_ratio. Among them, the cloud mask is used to identify the cloud influence and profile inversion quality, the terrain elevation is used to characterize the complex terrain background, the effective layer ratio is used to diagnose the profile availability, and the interpolation ratio is used to diagnose the interpolation dependence.

[0046] In the construction of severe convective weather risk labels, this embodiment uses minute-level station precipitation data as the basis for the construction of severe convective weather risk labels. Minute-level precipitation data can retain the peak changes in short-term heavy precipitation processes and is more suitable for describing the suddenness and locality of severe convective weather than hourly cumulative precipitation. In the data preprocessing stage, this embodiment first performs a quality check on the minute-level station precipitation records, removing missing values, outliers, and records that do not meet the time matching requirements, and counts the stations and times of positive precipitation to ensure the usability of the real-time data. Subsequently, the station precipitation data is mapped to a standard grid consistent with the convection parameters of FY-4BGIIRS. Considering that minute-level station precipitation has the characteristics of uneven spatial distribution and local concentration of heavy precipitation signals, this embodiment adopts the spatial maximum aggregation method, that is, taking the maximum minute precipitation value of the station within the neighborhood of the standard grid point as the precipitation real-time of the corresponding grid point. This method can retain the local heavy precipitation peak and avoid the weakening of short-term heavy precipitation signals by average aggregation.

[0047] In the time label construction phase, this embodiment uses the GIIRS parameter product time as the starting time and statistically analyzes the minute precipitation characteristics of each standard grid point within the next 2-hour window, including the maximum minute precipitation, average minute precipitation, number of minutes with non-zero precipitation, number of minutes with heavy precipitation, and whether precipitation events have occurred. These statistics are used to describe whether precipitation occurs at the grid point within the next 2 hours, the duration of precipitation, and whether short-term heavy precipitation peaks occur, in order to establish the correspondence between the parameter field and the forecast target. This embodiment constructs a three-class risk label based on the maximum minute precipitation within the next 2-hour window, with the threshold set to [0.0, 0.3] mm / min. When the maximum minute precipitation within the next 2-hour window is equal to 0, it is marked as low risk; when the maximum minute precipitation is greater than 0 and less than 0.3 mm / min, it is marked as medium risk; and when the maximum minute precipitation is greater than or equal to 0.3 mm / min, it is marked as high risk. This label definition retains the information on the presence or absence of precipitation while highlighting short-term heavy precipitation events, thus forming a model training target, suitable for training strong convective potential models within the next 2 hours.

[0048] Since the number of high-risk samples for severe convection is significantly less than that of low-risk samples, this embodiment retains all medium- and high-risk samples during the sample construction phase and extracts low-risk samples proportionally to reduce the dominant influence of samples with no precipitation or weak risk on model training. During the training phase, higher class weights are further assigned to high-risk categories so that the model pays more attention to samples with severe precipitation and severe convection during the learning process, thereby alleviating the class imbalance problem. The above-mentioned site quality check, maximum spatial aggregation, statistics for the next 2 hours, three-class label division, and sample balancing and weighting processing together constitute the construction process of minute-level precipitation labels for the next 2 hours.

[0049] In the model feature construction stage, this embodiment uses the standard meshed FY-4BGIIRS convection parameter product as the model input basis, and constructs candidate features based on the unstable energy, water vapor conditions, vertical stratification, complex terrain influence, and spatial neighborhood changes required for the occurrence of strong convection. A total of 123 candidate features are constructed, mainly including basic convection parameters, terrain adaptation parameters, temperature-derived features, physical combination features, time-varying features, time-lag features, and spatial statistical features. These features retain the physical meaning of traditional strong convection diagnostic indicators while introducing spatial neighborhood statistics. The model is adapted to complex terrain information, enabling it to learn the differences in water vapor, humidity, and stability between different grid points. In the feature selection stage, this embodiment first eliminates variables with severe missing data, weak physical meaning, or poor stability. Then, it uses XGBoost feature importance to perform initial screening of candidate features and conducts Top-K feature quantity sensitivity experiments to compare the model's TS, POD, FAR, and neighborhood TS performance under different feature quantity combinations. Finally, it comprehensively considers validation set TS, test set stability, feature physical interpretability, and business application maintainability to determine 9 core features for model input.

[0050] The final model features show that the model mainly retains variables related to water vapor, humidity, and adaptation to complex terrain. TPW and its spatial statistics are used to describe overall water vapor conditions, 700 hPa and 850 hPa relative humidity are used to describe mid-to-lower level water vapor configuration, and the terrain adaptation K-index is used to supplement stability diagnosis under complex terrain. This result indicates that under the data conditions of this embodiment, water vapor sufficiency, spatial differences in water vapor, and stability information under terrain background are important inputs for judging the risk of severe convection in the next two hours. Among the nine core model features finally determined, rh_700_spatial_zscore represents the spatial anomaly of 700 hPa relative humidity, reflecting the degree of deviation of mid-level humidity relative to the surrounding area; moisture_flux is the water vapor flux index of the combination of overall water vapor and lower-level humidity, characterizing the water vapor transport background; and rh_850_spatial_zscore represents the 850 hPa relative humidity. Spatial anomalies in relative humidity reflect local increases or decreases in lower-level humidity. `tpw_spatial_max` is the maximum precipitable water content within the neighborhood, representing the maximum water vapor conditions in the surrounding area. `ml_terrain_k_index_spatial_mean` is the spatial average of the terrain-adaptive K-index, representing stability conditions under complex terrain conditions. `tpw_spatial_mean` is the average precipitable water content within the neighborhood, representing the regional water vapor background. `rh_700_spatial_max` is the spatial maximum relative humidity at 700 hPa, representing high humidity conditions in the middle layer. `rh_850_spatial_min` is the spatial minimum relative humidity at 850 hPa, assisting in describing the spatial gradient of lower-level humidity. `tpw_spatial_min` is the spatial minimum precipitable water content within the entire layer, which, together with the maximum and mean values ​​of `TPW`, describes the spatial differences in water vapor.

[0051] The model training uses an XGBoost tri-class classification model with a multi:softprob objective function to output the probabilities of low-risk, medium-risk, and high-risk classes. The main model parameters include num_class=3, n_estimators=500, max_depth=5, min_child_weight=3, learning_rate=0.06, subsample=0.85, colsample_bytree=0.85, reg_alpha=0.05, and reg_lambda=2.0. The evaluation function is mlogloss. To alleviate the class imbalance problem caused by the small number of high-risk samples, this embodiment introduces class weights during the training phase and assigns higher weights to high-risk classes. The final class weights are set to {0:1.0, 1:1.0, ..., ... 2:6.0}, where category 0, category 1 and category 2 correspond to low risk, medium risk and high risk respectively. After the model outputs the probability, it further uses TS as the main business evaluation indicator and refers to POD, FAR and neighborhood TS to perform a combined search for the high risk probability threshold. Finally, the high risk probability threshold is determined to be 0.45. Through the above design, the model can not only output three classification results but also output high risk probability and risk score, providing a basis for the generation of business products.

[0052] During the operational phase, this embodiment applies the same processing flow to the new FY-4BGIIRS data as during the training phase, including data coverage checks, quality control, standard barosphere interpolation, convective parameter calculation, standard gridding, and model feature reproduction. The processed gridded convective parameter products are input into the trained XGBoost three-class classification model. The model outputs three probabilities—low risk, medium risk, and high risk—for each standard grid point. In the risk assessment phase, this embodiment generates gridded risk results based on the model's output category probabilities, the high-risk probability threshold, and the operational scoring rules. Output fields include pred_class, prob_low, prob_medium, prob_high, and risk_score_2h, where pred_class represents the final risk level, prob_high represents the high-risk category probability, and risk_score_2h represents the operational risk score for the strong convective potential over the next two hours. By simultaneously outputting the risk level, category probability, and risk score, operational personnel can view both discrete risk levels and assess risk strength and spatial continuity based on probability magnitude.

[0053] The product output module converts inference results into multiple formats, including CSV, PKL, MICAPS, and PNG. CSV products are used for tabular viewing and business archiving; PKL products are used for offline analysis and evaluation; MICAPS products are used for display in meteorological business systems; and PNG product maps are used for consultation presentations and rapid quality checks. Product content includes forecast start time, forecast start time, forecast end time, latitude and longitude, risk level, low-risk probability, medium-risk probability, high-risk probability, risk score, and related diagnostic information. The business-oriented product package encapsulates the data processing, model inference, and product output processes, including runtime code, model files, configuration files, sample products, official product output directory, plotting resources, and log files. This product package supports one-click execution from newly downloaded FY-4BGIIRS raw data and direct inference from already gridded convection parameter products. It also supports latest time-based execution, specified time-based inference, overwrite reruns, and batch product library generation, facilitating deployment, operation, and maintenance in local business environments.

[0054] In optional embodiments of the present invention, satellite temperature and humidity profile data are not limited to FY-4BGIIRS data, but may also be other satellite inversion data with temperature profiles, humidity profiles, quality indicators and latitude and longitude positioning information. The machine learning model is not limited to the XGBoost model, but may also be random forest, gradient boosting tree, LightGBM, CatBoost, neural network or multi-model ensemble method. The forecast time window is not limited to the next 2 hours, but may also be set to the next 1 hour, 3 hours, 6 hours or other short-term forecast time windows according to operational needs. The risk label threshold and probability judgment threshold are not limited to specific values, but may be adjusted according to regional climate characteristics, observation data density, operational early warning standards and model validation results. The operational product format is not limited to CSV, PKL, MICAPS and PNG, but may also be extended to gridded text, binary files, database records, web service interfaces or graphic product formats required by other meteorological operational systems.

[0055] It should be noted that the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0056] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection, characterized in that, Includes the following steps: Step S1: Acquire satellite temperature and humidity profile data, minute-level precipitation data, and topographic data; Step S2: Perform coverage integrity checks, quality control, and standard barosphere interpolation on the satellite temperature and humidity profile data to obtain quality-controlled temperature and humidity profile data; Step S3: Calculate the convection parameters based on the quality-controlled temperature and humidity profile data, and map the convection parameters to a standard grid; Step S4: Spatial aggregation of the minute-level precipitation data, using satellite parameter time as the starting time to statistically analyze precipitation within a future preset time window, and generate severe convection risk labels; Step S5: Construct a candidate feature set based on the convection parameters on the standard grid, perform feature screening on the candidate feature set, and construct training samples with the strong convection risk label to train the machine learning model; Step S6: Use the trained machine learning model to perform risk reasoning on the new satellite parameter products and generate short-term forecast products for severe convection.

2. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection as described in claim 1, characterized in that, In step S2, the satellite temperature and humidity profile data is FY-4B GIIRS temperature and humidity profile inversion data; the FY-4BGIIRS temperature and humidity profile inversion data is version-screened, and valid time data with a two-dimensional spatial structure and complete spatial coverage are selected as the standard processing version; the standard processing version is subjected to a coverage integrity check, which includes checking the latitude and longitude range, number of valid points, and coverage of the operational area for data of different versions, different time periods, or different observation centers, and data that cannot completely cover the operational area is removed.

3. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection as described in claim 1, characterized in that, In step S2, the quality control includes: removing filler values ​​and non-finite values; screening effective layers based on temperature and humidity quality indicators, and retaining temperature and humidity layers with quality indicators less than a preset threshold; screening effective grid points based on cloud mask information, and outputting quality diagnostic information such as cloud mask, effective layer ratio, and interpolation ratio; for profiles that meet the minimum number of effective layers, identifying missing layers within the original effective air pressure range, and performing linear interpolation on temperature and specific humidity; the standard air pressure layer interpolation includes: interpolating temperature and specific humidity to a preset number of standard air pressure layers; when the surface air pressure is greater than the minimum effective layer air pressure and the surface temperature and surface air pressure are effective, adding the surface air pressure layer to the profile, using the surface temperature for temperature, and linearly extrapolating the specific humidity based on the adjacent near-surface layer; converting the specific humidity to dew point temperature, and ensuring that the dew point temperature is not higher than the ambient temperature.

4. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection as described in claim 1, characterized in that, In step S3, the convection parameters include at least one of CAPE, CIN, K index, total precipitation, lift index, Showalter index, Total Totals index, relative humidity, dew point temperature, and temperature lapse rate.

5. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection as described in claim 1, characterized in that, In step S3, terrain adaptation flow parameters are constructed based on the terrain data and the lowest effective pressure layer of the profile. The terrain adaptation flow parameters include at least one of the following: complex terrain adaptation K-type index, complex terrain adaptation Showalter type index, complex terrain adaptation Total Totals type index, temperature lapse rate from the lowest effective layer to the preset pressure layer, and key layer availability indicator.

6. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection according to claim 1, characterized in that, In step S4, the station quality is first checked on the minute-level precipitation data to remove missing values, outliers, and records that do not meet the time matching requirements. The spatial aggregation is maximum value aggregation. When mapping the minute-level precipitation data to the standard grid, the maximum minute precipitation value of the stations within the neighborhood of the standard grid point is taken as the minute precipitation data at the corresponding time of the grid point.

7. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection as described in claim 1, characterized in that, In step S4, using the satellite parameter time as the reporting start time, the minute precipitation characteristics of each standard grid point within a future preset time window are statistically analyzed. The minute precipitation characteristics include at least one of the following: maximum minute precipitation, average minute precipitation, number of minutes with non-zero precipitation, number of minutes with heavy precipitation, and whether a precipitation event has occurred. Based on the maximum minute precipitation and a preset minute precipitation threshold, three categories of labels are assigned: low risk, medium risk, and high risk. When the maximum minute precipitation is equal to zero, it is marked as low risk; when the maximum minute precipitation is greater than zero and less than the preset minute precipitation threshold, it is marked as medium risk; and when the maximum minute precipitation is greater than or equal to the preset minute precipitation threshold, it is marked as high risk.

8. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection according to claim 1, characterized in that, In step S5, a candidate feature pool is constructed, which includes basic convection parameters, terrain adaptation parameters, temperature-derived features, physical combination features, time-varying features, time-lag features, and spatial statistical features. Core features are selected from the candidate feature pool based on feature importance, sensitivity to the preset number of features, and business metrics. All medium-risk and high-risk samples are retained, and low-risk samples are drawn according to a preset ratio. Higher category weights are assigned to high-risk categories; the machine learning model is a multi-class probability model that outputs three probabilities: low risk, medium risk, and high risk; the machine learning model includes at least one of XGBoost, LightGBM, random forest, gradient boosting tree, neural network, or a combination thereof; during training, the TS index is used as the main business evaluation index, and the high-risk probability threshold is searched and optimized with reference to POD, FAR, and neighborhood TS indexes.

9. The method for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection according to claim 1, characterized in that, In step S6, the severe convective short-term forecast product includes at least one of CSV format, PKL format, MICAPS format, and PNG format products; the severe convective short-term forecast product includes fields such as forecast start time, forecast start time, forecast end time, longitude, latitude, risk level, low risk probability, medium risk probability, high risk probability, and risk score.

10. A system for generating short-term forecast products based on satellite temperature and humidity profiles and severe convection, characterized in that, It includes a data acquisition module, a satellite data quality control and coverage screening module, a convection parameter generation module, a minute-level precipitation label construction module, a feature construction and screening module, a model training and optimization module, a risk reasoning module, and a product output module; The data acquisition module acquires satellite temperature and humidity profile data, minute-level precipitation data, and topographic data; The satellite data quality control and coverage screening module performs coverage integrity checks, quality control, and standard barosphere interpolation on the satellite temperature and humidity profile data to obtain quality-controlled temperature and humidity profile data. The convection parameter generation module calculates the convection parameters based on the quality-controlled temperature and humidity profile data, and maps the convection parameters to a standard grid. The minute-level precipitation label construction module spatially aggregates the minute-level precipitation data to generate severe convection risk labels. The feature construction and filtering module constructs a candidate feature set based on the convection parameters on the standard grid and filters the core features; The model training and optimization module trains a machine learning model based on the core features and the strong convection risk label, and optimizes the class weights and probability thresholds. The risk reasoning module uses a trained machine learning model to perform risk reasoning on new satellite parameter products and outputs risk probability, risk level and risk score. The product output module converts the risk probability, the risk level, and the risk score into a short-term forecast product for severe convection.

Images