A short-time rainstorm identification and machine translation method and system
By fusing multi-source data and using interpretable machine learning models, the problems of accuracy and reliability in short-term rainstorm identification and forecasting were solved, achieving high-resolution short-term rainstorm identification and forecasting, and revealing the contribution of meteorological factors and physical mechanisms.
Patent Information
- Application Number
- CN202610661249.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-14
- Publication Date
- 2026-08-25
AI Technical Summary
Existing technologies struggle to accurately identify and forecast short-term heavy rainfall, especially in complex terrain conditions. Furthermore, existing methods lack interpretability, resulting in low prediction accuracy and insufficient reliability.
A multi-source heterogeneous data fusion method was adopted, utilizing precipitation measurement radar from FY-3GPMR satellite, precipitation data retrieved from IMERG satellite, ground meteorological observation data, and topographic data to construct a high-resolution fused precipitation dataset. Short-term rainstorm events were identified by combining a state machine model, and prediction was performed using an EBOX interpretable Bayesian optimized XGBoost model. The interpretability analysis was conducted using the SHAP method.
It has enabled the accurate identification of short-duration rainstorms and the quantification of meteorological factors, improving the accuracy and reliability of forecasts, providing a reliable basis for weather forecasting decisions, and revealing the physical mechanism of short-duration rainstorms.
Smart Images
Figure CN122632362A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of hydrological and meteorological forecasting and artificial intelligence, specifically involving a method and system for short-term rainstorm identification and mechanism interpretation based on multi-source remote sensing information and interpretable artificial intelligence technology. Background Technology
[0002] Short-duration rainstorms refer to extreme precipitation events with rainfall of at least 8 millimeters per hour. They are characterized by their suddenness, short duration, high intensity, and significant localization, often causing urban flooding, flash floods, and other disasters. Current methods for identifying and forecasting short-duration rainstorms mainly rely on numerical weather prediction, radar echo extrapolation, and traditional statistical methods, which have the following limitations:
[0003] 1) While numerical weather prediction models can provide large-scale precipitation forecasts, their spatial resolution is limited, typically ranging from a few kilometers to tens of kilometers. This makes it difficult to accurately capture the local characteristics of short-duration heavy rainfall, resulting in low accuracy in predicting heavy rainfall at the urban scale. Furthermore, numerical models exhibit systematic biases in simulating precipitation under complex terrain conditions, particularly in low mountain and hilly areas where the localized heavy rainfall caused by topographic-airflow interactions is difficult to accurately simulate.
[0004] 2) Radar echo extrapolation technology, based on weather radar observation data, performs short-term nowcasting by tracking the movement trajectory of precipitation echoes. However, this method assumes that the precipitation system remains stable in a short period of time, and its forecasting ability is significantly reduced for rapidly developing severe convective weather systems. In addition, radar observations have blind spots, and their detection effect is limited under complex terrain conditions. Furthermore, radar data quality control algorithms have limited effectiveness in eliminating false echoes.
[0005] 3) Traditional statistical methods mainly rely on historical precipitation data to establish empirical relationships, lacking in-depth analysis of the formation mechanism of short-duration rainstorms, and are difficult to adapt to the new characteristics of extreme precipitation events under the background of climate change. Existing research has insufficient understanding of the spatiotemporal variation characteristics of short-duration rainstorms at the grid scale, especially lacking joint analysis of multi-dimensional heterogeneous data, and cannot fully reveal the three-dimensional spatiotemporal differentiation patterns of short-duration rainstorms.
[0006] 4) While existing machine learning methods have made some progress in short-term heavy rainfall prediction, most models are "black box" models, lacking interpretability and failing to reveal the nonlinear influence mechanism of meteorological factors on the occurrence and development of short-term heavy rainfall, thus limiting the reliability and practical application value of the prediction results. Furthermore, existing methods do not adequately address the problem of imbalanced samples, resulting in low accuracy in identifying minority classes (heavy rainfall events).
[0007] In summary, there is an urgent need to develop an interpretable method for identifying and interpreting short-term rainstorms that can integrate multi-source remote sensing information, in order to improve the accuracy and scientific validity of short-term rainstorm forecasting. Summary of the Invention
[0008] The purpose of this invention is to provide a method and system for identifying and interpreting short-term rainstorms by integrating multi-source heterogeneous data and possessing interpretability, thereby achieving accurate identification of rainstorms and quantification of the contribution of meteorological factors, and improving forecast reliability and disaster prevention and mitigation support capabilities.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following solution:
[0010] A method for short-term rainstorm identification and mechanism interpretation includes the following steps:
[0011] Step S1: Acquisition and preprocessing of multi-source heterogeneous precipitation data: Acquire precipitation measurement radar data from FY-3GPMR satellite, precipitation data retrieved from IMERG satellite, ground meteorological observation data, and topographic data. After quality control, spatiotemporal matching, and gridding, construct a fused precipitation dataset with unified spatiotemporal resolution.
[0012] Step S2, Short-term rainstorm event identification and feature extraction: Based on the state machine model, short-term rainstorm events are identified according to spatiotemporal continuity, with hourly rainfall ≥8mm as the criterion. The intensity, frequency, duration, and range features are statistically analyzed to construct a feature database.
[0013] Step S3, Spatiotemporal Variation Analysis of Short-Term Rainstorms: Based on the fused precipitation dataset, conduct analysis on temporal trends, spatial patterns, and spatiotemporal heterogeneity;
[0014] Step S4: Three-dimensional structural analysis of typical short-term rainstorm events: Using FY-3GPMR satellite precipitation radar data, analyze the vertical profile of the rainstorm, the storm top height, and the characteristics of the bright band;
[0015] Step S5, Explainable Machine Learning Model Construction and Optimization: Construct the EBOX interpretable Bayesian optimized XGBoost model, and complete sample selection, feature engineering, model training, Bayesian optimization, sample imbalance handling and SHAP interpretable analysis;
[0016] Step S6: Meteorological Mechanism Interpretation of Short-Term Rainstorms: Based on the SHAP results, quantify the contribution of meteorological factors, identify key driving factors, and interpret the physical mechanisms.
[0017] Further optimization, step S1, multi-source heterogeneous precipitation data acquisition and preprocessing, specifically includes the following sub-steps:
[0018] S11: Multi-source data acquisition, simultaneously acquiring FY-3GPMR satellite precipitation measurement radar data, IMERG satellite inverted precipitation data, ground meteorological observation data, digital elevation model (DEM) data, and land use type data;
[0019] S12: Data quality control, performing extreme value checks, temporal consistency checks and spatial consistency checks on ground meteorological observation data, and eliminating abnormal observation values;
[0020] S13: Spatiotemporal unified matching, using bilinear interpolation to unify data from different sources and at different resolutions to a 0.1°×0.1° spatial grid and a 1-hour temporal resolution;
[0021] S14: Satellite data bias correction. Based on precipitation data observed by ground rain gauges, the optimal interpolation method is used to correct the systematic bias of the precipitation data retrieved from the IMERG satellite.
[0022] S15: Fusion dataset generation, which integrates corrected satellite data, ground observation data, and topographic data within a unified spatiotemporal framework to form a high-quality, high-resolution fused precipitation dataset.
[0023] Further optimization, step S2, short-term rainstorm event identification and feature extraction, specifically includes the following sub-steps:
[0024] S21: State machine initialization, setting the initial state of grid points to non-rainstorm;
[0025] S22: Rainstorm event trigger judgment. Read the rainfall on a grid-by-grid basis in hourly units. When the hourly rainfall is ≥8 mm for the first time, the status is switched to rainstorm and the start time and spatial location of the event are recorded.
[0026] S23: Continuous monitoring of rainstorm events, continuously monitoring the rainfall at grid points that have triggered rainstorm status hourly, and determining whether the termination conditions are met;
[0027] S24: Judgment of the end of a rainstorm event: When the rainfall at a grid point is less than 8 mm for two consecutive hours, the status is switched back to non-rainstorm and the end time of the event is recorded;
[0028] S25: Statistics on the characteristics of rainstorm events. For short-term rainstorm events that have been fully identified, extract and statistically analyze four characteristic indicators: rainfall intensity, duration, affected area, and cumulative precipitation, and construct a database of short-term rainstorm characteristics.
[0029] Further optimization, step S3, short-term rainstorm spatiotemporal variation analysis, specifically includes the following sub-steps:
[0030] S31: Temporal trend analysis. Based on the fused precipitation dataset, the Mann-Kendall trend test and Sen slope estimation method are used to test the significance of the temporal trend and quantitatively calculate the rate of change of the intensity, duration and range of short-term rainstorms.
[0031] S32: Spatial distribution pattern analysis. Using spatial autocorrelation analysis and hotspot analysis methods, we identify the spatial clustering characteristics and high-value clustering areas of short-term rainstorm intensity and frequency, and reveal the regional spatial differentiation pattern.
[0032] S33: Spatiotemporal heterogeneity detection, combining time series and spatial grid data, analyzes the spatiotemporal evolution differences of short-term rainstorm events in different regions and time periods, and completes the quantitative characterization of spatiotemporal heterogeneity;
[0033] S34: Spatiotemporal feature summary, which integrates time trend results, spatial pattern results, and heterogeneity detection results to form a dataset of spatiotemporal variation features of short-term rainstorms.
[0034] Further optimization, step S4, the three-dimensional structure analysis of a typical short-term rainstorm event, specifically includes the following sub-steps:
[0035] S41: Typical case selection, short-term rainstorm events that occurred during the FY-3GPMR satellite's transit were selected as the objects for three-dimensional structure analysis;
[0036] S42: Three-dimensional data extraction, based on FY-3GPMR precipitation measurement radar data, extracting precipitation vertical profile, storm top height, bright band height, and precipitation particle phase information;
[0037] S43: Three-dimensional structural feature analysis, analyzing the vertical distribution pattern of rainstorms, the correlation between the heavy precipitation area and the storm top height, the location of the bright band and phase transition characteristics, and determining the dominant area of convective precipitation;
[0038] S44: Ground observation verification, compare and verify the three-dimensional structure results retrieved by satellite radar with ground meteorological observations and rain gauge measured data to confirm the retrieval accuracy;
[0039] S45: Structural feature output, forming a typical short-term rainstorm three-dimensional structural dataset containing vertical profiles, storm top height, bright band features, and convection intensity.
[0040] Further optimization, step S5 can be explained by the following sub-steps specifically comprising machine learning model construction and optimization:
[0041] S51: Sample construction and partitioning: Using the identified short-term rainstorm events as positive samples and selecting an equal number of non-rainstorm moments as negative samples, construct a binary classification dataset and partition it into training set, validation set and test set in a 7:2:1 ratio;
[0042] S52: Feature engineering and standardization. Temperature, air pressure, relative humidity, specific humidity, wind speed, and shortwave radiation are selected as input features, and Z-score standardization is used to eliminate the influence of dimensions.
[0043] S53: Bayesian hyperparameter optimization, with the validation set F1 score as the optimization objective, uses a Gaussian process surrogate model to automatically optimize the learning rate, maximum tree depth, subsampling ratio, column sampling ratio, and minimum sample weight;
[0044] S54: Imbalanced sample handling: SMOTE oversampling technique is used to enhance minority class samples in the training set to balance the ratio of positive to negative samples;
[0045] S55: XGBoost model training, using optimized hyperparameters to train the binary classification model, and configuring an early stopping strategy to prevent overfitting;
[0046] S56: SHAP Interpretability Analysis. The SHAP method is used to calculate the global importance and local contribution of each feature, generating interpretable analysis results for the model.
[0047] Further optimization, step S6, the interpretation of the short-term rainstorm meteorological mechanism, specifically includes the following sub-steps:
[0048] S61: Factor contribution quantification, based on SHAP global importance ranking, to determine the contribution weights of air temperature, specific humidity, shortwave radiation, relative humidity, air pressure, and wind speed to the occurrence of short-term rainstorms;
[0049] S62: Nonlinear relationship analysis, based on SHAP dependency graph, analyzes the nonlinear response relationship between each meteorological factor and the probability of rainstorm occurrence, and determines the high-risk threshold interval;
[0050] S63: Identification of key driving factors: Based on contribution weights and nonlinear response characteristics, screen and determine the dominant and secondary meteorological factors for the occurrence of short-term rainstorms.
[0051] S64: Physical mechanism interpretation, combining meteorological principles, to explain the physical mechanism of strong convection development and short-term rainstorm formation under conditions of low temperature, high humidity, low shortwave radiation, and low pressure;
[0052] S65: Output the mechanism interpretation results, forming a short-term rainstorm mechanism interpretation report that includes factor contributions, nonlinear relationships, key thresholds, and physical mechanisms.
[0053] A short-term rainstorm identification and mechanism interpretation system includes a data acquisition and preprocessing module, a short-term rainstorm event identification module, a spatiotemporal change analysis module, a three-dimensional structure analysis module, an interpretable machine learning module, and a mechanism interpretation module;
[0054] The data acquisition and preprocessing module is used to acquire multi-source remote sensing precipitation data, ground meteorological observation data and topographic data, and after quality control, spatiotemporal matching and gridding processing, fuse precipitation datasets at a resolution of 0.1°×0.1° and 1 hour.
[0055] The short-term rainstorm event identification module is used to identify short-term rainstorm events based on a state machine model and extract characteristic indicators such as intensity, duration, range, and cumulative precipitation.
[0056] The spatiotemporal variation analysis module is used to analyze the temporal variation trend, spatial distribution pattern and spatiotemporal heterogeneity of short-term rainstorms;
[0057] The three-dimensional structure analysis module is used to analyze the vertical profile, storm top height, and bright band characteristics of typical rainstorm events based on FY-3GPMR satellite precipitation radar data.
[0058] The interpretable machine learning module is used to construct the EBOX interpretable Bayesian optimized XGBoost model, complete the sample imbalance handling, and use the SHAP method to perform model interpretability analysis.
[0059] The mechanism interpretation module is used to quantify the contribution of meteorological factors, identify key driving factors, and interpret the physical mechanism of short-term rainstorm formation based on the SHAP interpretation results.
[0060] Further optimization involves the data acquisition and preprocessing module specifically used for: correcting satellite precipitation data deviations using optimal interpolation based on ground rain gauges; unifying multi-source data to a 0.1°×0.1° spatial grid and a 1-hour temporal resolution through bilinear interpolation; and configuring the state machine model in the short-term rainstorm event identification module as follows: the initial state is non-rainstorm, switching to rainstorm state when the hourly rainfall is ≥8mm for the first time, and reverting to non-rainstorm state when the rainfall is <8mm for 2 consecutive hours.
[0061] Further optimizations include an interpretable machine learning module comprising a sample partitioning unit, a feature standardization unit, a Bayesian hyperparameter optimization unit, a SMOTE sample balancing unit, an XGBoost training unit, and a SHAP interpretation unit; the machine interpretation module is also used to output a report on the importance ranking of meteorological factors, nonlinear response relationships, high-risk threshold intervals, and physical mechanism interpretation of short-term rainstorms.
[0062] Compared with the prior art, the present invention has the following beneficial effects:
[0063] 1. This invention employs a multi-source heterogeneous data fusion method, integrating multiple data sources such as satellite remote sensing and ground observations. This overcomes the limitations of single data sources, improves the spatiotemporal resolution and accuracy of precipitation data, and provides a high-quality data foundation for short-term heavy rainfall identification. In particular, by utilizing FY-3GPMR satellite precipitation measurement radar data, it is possible to obtain three-dimensional vertical structure information of precipitation, compensating for the shortcomings of traditional two-dimensional observations.
[0064] 2. This invention constructs a short-term rainstorm event identification algorithm based on a state machine model, which can accurately capture the start and end time and spatial range of short-term rainstorm events, realize the automated identification and feature extraction of short-term rainstorm events, and provide reliable event samples for subsequent analysis.
[0065] 3. The EBOX interpretable Bayesian optimized XGBoost model constructed in this invention organically combines Bayesian optimization, imbalance handling, and SHAP interpretation. It not only improves the accuracy of short-term rainstorm identification but also realizes the interpretability of model prediction results. It can quantify the contribution of each meteorological factor to the occurrence of short-term rainstorms, providing a basis for meteorological forecasters to make decisions.
[0066] 4. Through SHAP interpretability analysis, this invention reveals that air temperature and specific humidity are the key meteorological factors that dominate short-duration rainstorms. It finds that rainstorms are most easily triggered under low temperature and high humidity conditions. At this time, shortwave radiation is blocked by cumulus clouds, and low pressure promotes the development of strong convection, providing new insights into the physical mechanism of short-duration rainstorms.
[0067] 5. The method of this invention has good versatility and scalability, and can be applied to the study of short-term rainstorms in different regions and under different climatic backgrounds, providing technical support for urban disaster prevention and mitigation, meteorological early warning and water resource management. Attached Figure Description
[0068] Figure 1 This is a schematic diagram of the overall process of the short-term rainstorm identification and mechanism interpretation method of the present invention;
[0069] Figure 2 This is a flowchart of the short-term rainstorm event identification algorithm based on the state machine model of this invention;
[0070] Figure 3 This is a schematic diagram of the EBOX interpretable machine learning model architecture of the present invention;
[0071] Figure 4 This is a spatial distribution map of the intensity of short-duration summer rainstorms in Nanjing City, according to Embodiment 1 of the present invention.
[0072] Figure 5 This is a schematic diagram of the three-dimensional vertical structure of a typical short-term rainstorm event according to the present invention;
[0073] Figure 6 This is a ranking diagram of the importance of SHAP features in this invention;
[0074] Figure 7 This is a SHAP dependency graph of meteorological factors on the impact of short-term heavy rainfall in this invention. Detailed Implementation
[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0076] Example 1:
[0077] like Figure 1 As shown, a method for short-term rainstorm identification and mechanism interpretation specifically includes:
[0078] Step S1: Acquisition and Preprocessing of Multi-Source Heterogeneous Precipitation Data
[0079] This embodiment selects Nanjing City as the study area, and the study period is the summer (June-August) of 2018-2023. The multi-source data obtained include: (1) FY-3GPMR satellite precipitation measurement radar data, with a spatial resolution of about 5 kilometers and a temporal resolution determined according to the satellite orbit; (2) IMERG satellite inverted precipitation data, with a spatial resolution of 0.1° and a temporal resolution of 30 minutes; (3) China Meteorological Administration ground meteorological observation station data, including elements such as temperature, air pressure, relative humidity, wind speed, and wind direction, with a temporal resolution of 1 hour; (4) SRTM digital elevation model, with a spatial resolution of 90 meters; (5) land use type data, with a spatial resolution of 30 meters.
[0080] Data preprocessing includes: (1) Data quality control: extreme value checks, temporal consistency checks and spatial consistency checks are performed on the ground observation data to remove outliers; (2) Spatiotemporal matching: bilinear interpolation is used to unify data of different resolutions to a 0.1°×0.1° grid and the time resolution is unified to 1 hour; (3) Deviation correction: based on the ground rain gauge observation data, the optimal interpolation method is used to correct the deviation of the IMERG data. The correlation coefficient between the corrected IMERG data and the ground observation is increased from 0.72 to 0.89, and the root mean square error is reduced from 3.2 mm to 1.8 mm.
[0081] Step S2: Identification and Feature Extraction of Short-Term Rainstorm Events
[0082] A short-term rainstorm event identification algorithm is constructed based on a state machine model, such as... Figure 2As shown, the workflow of the state machine model is as follows: Initialize the state to "non-rainstorm". When the hourly rainfall of a certain grid reaches or exceeds 8 mm for the first time, the state changes to "rainstorm", and the start time and location of the event are recorded; continuously monitor the hourly rainfall of the grid. When the rainfall is less than 8 mm for two consecutive hours, the state changes to "non-rainstorm", and the end time of the event is recorded; statistically analyze the duration of the event, the maximum hourly rainfall intensity, the cumulative rainfall, and the affected area.
[0083] Using the methods described above, a total of 2,847 short-duration rainstorm events were identified in Nanjing during the study period. Statistical results show that the average duration of these events was 2.3 hours, the average maximum hourly rainfall intensity was 15.6 mm, the average cumulative rainfall was 28.4 mm, and the average affected area was 156 square kilometers.
[0084] Step S3: Spatiotemporal Variation Analysis of Short-Term Rainstorms
[0085] Based on a fused precipitation dataset, this study analyzes the spatiotemporal distribution characteristics of short-duration summer rainstorms in Nanjing. Temporal variation analysis employs the Mann-Kendall trend test and Sen slope estimation method, while spatial analysis utilizes spatial autocorrelation analysis and hotspot analysis. Figure 4 As shown, the intensity of short-term rainstorms in Nanjing during the summer exhibits significant spatial differentiation characteristics.
[0086] The results showed that: (1) the short-term rainstorms in Nanjing during the summer showed a significant increasing trend, with the average hourly rainfall intensity increasing at a rate of 0.42 mm per year, which passed the significance test at the 95% confidence level; (2) the duration of short-term rainstorms showed a lengthening trend, with the average duration increasing at a rate of 0.08 hours per year; (3) the affected area of short-term rainstorms showed a decreasing trend, indicating that the localization of rainstorm events was more obvious; (4) the spatial distribution showed a significant north-south difference, with the intensity of short-term rainstorms increasing in areas south of the Yangtze River and slightly decreasing in areas north of the Yangtze River, with the Yangtze River becoming a natural dividing line for the change in rainstorm intensity.
[0087] Step S4: Three-dimensional structural analysis of typical short-term rainstorm events
[0088] A short-duration rainstorm event that occurred during the FY-3GPMR satellite's transit on July 15, 2023, was selected as a typical case for analysis. This event occurred between 14:00 and 17:00 in the afternoon, with a maximum hourly rainfall intensity of 35.4 mm, mainly affecting Pukou District, Yuhuatai District, and Jiangning District of Nanjing City.
[0089] The three-dimensional vertical structure of this event was analyzed using precipitation measurement radar data from the FY-3GPMR satellite, such as... Figure 5As shown. The results show that: (1) the vertical structure of the rainstorm is inverted cone-shaped and mainly occurs below the frozen layer (about 5 km in height); (2) the rainfall is distributed horizontally in strips, and the heavy rainfall area is significantly correlated with the convective characteristics of the high storm top; (3) the storm top height reaches up to 12 km, indicating strong convective activity; (4) the bright band analysis shows that the height of the bright band in the rainstorm area is about 4.5 km, the area above the bright band is ice phase precipitation particles, and the area below the bright band is liquid precipitation. The bright band characteristics of the high-intensity rainfall area are not obvious, indicating that convective precipitation is the main feature.
[0090] Step S5, EBOX: Explainable Machine Learning Model Construction and Optimization
[0091] The EBOX interpretable Bayesian optimized XGBoost model is used for short-term heavy rainfall identification and prediction. The model architecture is as follows: Figure 3 As shown. The specific steps are as follows:
[0092] (1) Sample construction: The training dataset was constructed by using the 2847 identified short-term rainstorm events as positive samples and randomly selecting an equal number of non-rainstorm moments as negative samples. The dataset was divided into training set, validation set and test set in a ratio of 7:2:1.
[0093] (2) Feature engineering: Six near-surface meteorological factors, namely air temperature, air pressure, relative humidity, specific humidity, wind speed, and shortwave radiation, were selected as input features. The Z-score method was used for feature standardization to eliminate the influence of dimensions.
[0094] (3) Bayesian optimization: A Gaussian process is used as the surrogate model, and the optimization objective is the F1 score on the validation set. The optimization parameters include: learning rate (range 0.01-0.3), maximum tree depth (range 3-10), subsampling ratio (range 0.6-1.0), column sampling ratio (range 0.6-1.0), and minimum sample weight (range 1-10). After 50 iterations of optimization, the optimal parameter combination is obtained: learning rate 0.15, maximum tree depth 6, subsampling ratio 0.85, column sampling ratio 0.8, and minimum sample weight 3.
[0095] (4) Imbalance handling: SMOTE oversampling technology is used to oversample the minority class samples in the training set so that the ratio of positive to negative samples reaches 1:1.
[0096] (5) Model training: The XGBoost binary classification model was trained using the optimized parameters. The number of training rounds was set to 200 rounds, and the number of early stopping rounds was set to 20 rounds.
[0097] (6) Model evaluation: The model performance was evaluated on the test set. The accuracy was 0.87, the precision was 0.82, the recall was 0.76, the F1 score was 0.79, and the AUC was 0.91. The model performed well in majority class recognition, but there were still some shortcomings in minority class recognition, which may be due to the large differences in environmental parameter characteristics under different weather conditions.
[0098] Step S6: Interpretation of Short-Term Rainstorm Meteorological Mechanisms
[0099] The SHAP method was used to analyze the interpretability of the model prediction results. The SHAP value can quantify the contribution of each feature to the model prediction; a positive value indicates an increase in the probability of heavy rain, while a negative value indicates a decrease in the probability of heavy rain.
[0100] The results of the global feature importance analysis are as follows: Figure 6 As shown: (1) Temperature is the most important predictor, with an SHAP importance value of 0.42; (2) Specific humidity is the second most important predictor, with an SHAP importance value of 0.38; (3) Shortwave radiation ranks third, with an SHAP importance value of 0.21; (4) Relative humidity ranks fourth, with an SHAP importance value of 0.18; (5) Air pressure ranks fifth, with an SHAP importance value of 0.15; (6) Wind speed has a relatively small impact, with an SHAP importance value of 0.08.
[0101] SHAP dependency graph analysis revealed the nonlinear influence mechanism of various meteorological factors on short-duration heavy rainfall, such as... Figure 7 As shown: (1) Temperature is negatively correlated with the probability of rainstorms. Rainstorms are more likely to be triggered under low temperature conditions because low temperature usually corresponds to the development of strong convective weather systems; (2) Specific humidity is positively correlated with the probability of rainstorms. High humidity conditions provide sufficient water vapor for rainstorms; (3) Shortwave radiation is negatively correlated with the probability of rainstorms. Low shortwave radiation indicates that the sky is covered by cumulus clouds, which is conducive to the development of convection; (4) Air pressure is negatively correlated with the probability of rainstorms. Low air pressure promotes upward motion, which is conducive to the development of strong convection.
[0102] Comprehensive analysis shows that short-duration rainstorms in Nanjing during the summer are mainly driven by temperature and specific humidity, and tend to form under conditions of low temperature and high humidity. At this time, shortwave radiation is blocked by cumulus clouds, low air pressure promotes the development of strong convection, while wind speed has a relatively small impact. This finding provides new insights into the physical mechanisms of short-duration rainstorms and offers important reference data for weather forecasting.
[0103] Example 2:
[0104] A short-term rainstorm identification and mechanism interpretation system includes a data acquisition and preprocessing module, a short-term rainstorm event identification module, a spatiotemporal change analysis module, a three-dimensional structure analysis module, an interpretable machine learning module, and a mechanism interpretation module;
[0105] The data acquisition and preprocessing module is used to acquire multi-source remote sensing precipitation data, ground meteorological observation data and topographic data, and construct a fused precipitation dataset with unified spatiotemporal resolution through quality control, spatiotemporal matching and gridding processing.
[0106] The short-term rainstorm event identification module is used to identify short-term rainstorm events based on a state machine model and extract characteristic indicators such as intensity, duration, range, and cumulative precipitation.
[0107] The spatiotemporal variation analysis module is used to analyze the temporal variation trend, spatial distribution pattern and spatiotemporal heterogeneity of short-term rainstorms;
[0108] The three-dimensional structure analysis module is used to analyze the vertical profile, storm top height, and bright band characteristics of typical rainstorm events based on FY-3GPMR satellite precipitation radar data.
[0109] The interpretable machine learning module is used to construct the EBOX interpretable Bayesian optimized XGBoost model, complete the sample imbalance handling, and use the SHAP method to perform model interpretability analysis.
[0110] The mechanism interpretation module is used to quantify the contribution of meteorological factors, identify key driving factors, and interpret the physical mechanism of short-term rainstorm formation based on the SHAP interpretation results.
[0111] The data acquisition and preprocessing module is specifically used to: correct satellite precipitation data deviations using the optimal interpolation method based on ground rain gauges, and unify multi-source data to a 0.1°×0.1° spatial grid and a 1-hour temporal resolution through bilinear interpolation; the state machine model in the short-term rainstorm event identification module is configured as follows: the initial state is non-rainstorm, the state switches to rainstorm when the hourly rainfall is ≥8mm for the first time, and the state returns to non-rainstorm when the rainfall is <8mm for 2 consecutive hours.
[0112] The interpretable machine learning module includes a sample partitioning unit, a feature standardization unit, a Bayesian hyperparameter optimization unit, a SMOTE sample balancing unit, an XGBoost training unit, and a SHAP interpretation unit; the machine interpretation module is also used to output a report on the importance ranking of meteorological factors, nonlinear response relationships, high-risk threshold intervals, and physical mechanism interpretation of short-term rainstorms.
[0113] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. A method for identifying and interpreting short-term rainstorms, characterized in that, The steps include the following: Step S1: Acquisition and preprocessing of multi-source heterogeneous precipitation data: Acquire precipitation measurement radar data from FY-3GPMR satellite, precipitation data retrieved from IMERG satellite, ground meteorological observation data, and topographic data. After quality control, spatiotemporal matching, and gridding, construct a fused precipitation dataset with unified spatiotemporal resolution. Step S2, Short-duration heavy rain event identification and feature extraction: Based on a state machine model, using hourly rainfall ≥8mm as the criterion, short-duration heavy rain events are identified according to spatiotemporal continuity. Intensity, frequency, duration, and range features are statistically analyzed to construct a feature database. ； Step S3, Spatiotemporal Variation Analysis of Short-Term Rainstorms: Based on the fused precipitation dataset, conduct analysis on temporal trends, spatial patterns, and spatiotemporal heterogeneity; Step S4: Three-dimensional structural analysis of typical short-term rainstorm events: Using FY-3GPMR satellite precipitation radar data, analyze the vertical profile of the rainstorm, the storm top height, and the characteristics of the bright band; Step S5: Construction and optimization of interpretable machine learning model: Construct an EBOX interpretable Bayesian optimized XGBoost model, and complete sample selection, feature engineering, model training, Bayesian optimization, sample imbalance handling and SHAP interpretable analysis; Step S6: Meteorological Mechanism Interpretation of Short-Term Rainstorms: Based on the SHAP results, quantify the contribution of meteorological factors, identify key driving factors, and interpret the physical mechanisms.
2. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, Step S1, multi-source heterogeneous precipitation data acquisition and preprocessing, specifically includes the following sub-steps: S11: Multi-source data acquisition, simultaneously acquiring FY-3GPMR satellite precipitation measurement radar data, IMERG satellite inverted precipitation data, ground meteorological observation data, digital elevation model (DEM) data, and land use type data; S12: Data quality control, performing extreme value checks, temporal consistency checks and spatial consistency checks on ground meteorological observation data, and eliminating abnormal observation values; S13: Spatiotemporal unified matching, using bilinear interpolation to unify data from different sources and at different resolutions to a 0.1°×0.1° spatial grid and a 1-hour temporal resolution; S14: Satellite data bias correction. Based on precipitation data observed by ground rain gauges, the optimal interpolation method is used to correct the systematic bias of the precipitation data retrieved from the IMERG satellite. S15: Fusion dataset generation, which integrates corrected satellite data, ground observation data, and topographic data within a unified spatiotemporal framework to form a high-quality, high-resolution fused precipitation dataset.
3. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, Step S2, short-term rainstorm event identification and feature extraction, specifically includes the following sub-steps: S21: State machine initialization, setting the initial state of grid points to non-rainstorm; S22: Rainstorm event trigger judgment. Read the rainfall on a grid-by-grid basis in hourly units. When the hourly rainfall is ≥8 mm for the first time, the status is switched to rainstorm and the start time and spatial location of the event are recorded. S23: Continuous monitoring of rainstorm events, continuously monitoring the rainfall at grid points that have triggered rainstorm status hourly, and determining whether the termination conditions are met; S24: Judgment of the end of a rainstorm event: When the rainfall at a grid point is less than 8 mm for two consecutive hours, the status is switched back to non-rainstorm and the end time of the event is recorded; S25: Statistics on the characteristics of rainstorm events. For short-term rainstorm events that have been fully identified, extract and statistically analyze four characteristic indicators: rainfall intensity, duration, affected area, and cumulative precipitation, and construct a database of short-term rainstorm characteristics.
4. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, The step S3, short-term rainstorm spatiotemporal variation analysis, specifically includes the following sub-steps: S31: Temporal trend analysis. Based on the fused precipitation dataset, the Mann-Kendall trend test and Sen slope estimation method are used to test the significance of the temporal trend and quantitatively calculate the rate of change of the intensity, duration and range of short-term rainstorms. S32: Spatial distribution pattern analysis. Using spatial autocorrelation analysis and hotspot analysis methods, we identify the spatial clustering characteristics and high-value clustering areas of short-term rainstorm intensity and frequency, and reveal the regional spatial differentiation pattern. S33: Spatiotemporal heterogeneity detection, combining time series and spatial grid data, analyzes the spatiotemporal evolution differences of short-term rainstorm events in different regions and time periods, and completes the quantitative characterization of spatiotemporal heterogeneity; S34: Spatiotemporal feature summary, which integrates time trend results, spatial pattern results, and heterogeneity detection results to form a dataset of spatiotemporal variation features of short-term rainstorms.
5. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, Step S4, the three-dimensional structure analysis of a typical short-duration rainstorm event, specifically includes the following sub-steps: S41: Typical case selection, short-term rainstorm events that occurred during the FY-3GPMR satellite's transit were selected as the objects for three-dimensional structure analysis; S42: Three-dimensional data extraction, based on FY-3GPMR precipitation measurement radar data, extracting precipitation vertical profile, storm top height, bright band height, and precipitation particle phase information; S43: Three-dimensional structural feature analysis, analyzing the vertical distribution pattern of rainstorms, the correlation between the heavy precipitation area and the storm top height, the location of the bright band and phase transition characteristics, and determining the dominant area of convective precipitation; S44: Ground observation verification, compare and verify the three-dimensional structure results retrieved by satellite radar with ground meteorological observations and rain gauge measured data to confirm the retrieval accuracy; S45: Structural feature output, forming a typical short-term rainstorm three-dimensional structural dataset containing vertical profiles, storm top height, bright band features, and convection intensity.
6. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, Step S5 can be explained by the following sub-steps in the construction and optimization of the machine learning model: S51: Sample construction and partitioning: Using the identified short-term rainstorm events as positive samples and selecting an equal number of spatiotemporally matched non-rainstorm samples as negative samples, construct a binary classification dataset and partition it into training set, validation set and test set in a 7:2:1 ratio; S52: Feature engineering and standardization. Temperature, air pressure, relative humidity, specific humidity, wind speed, and shortwave radiation are selected as input features, and Z-score standardization is used to eliminate the influence of dimensions. S53: Bayesian hyperparameter optimization, with the validation set F1 score as the optimization objective, uses a Gaussian process surrogate model to automatically optimize the learning rate, maximum tree depth, subsampling ratio, column sampling ratio, and minimum sample weight; S54: Imbalanced sample handling: SMOTE oversampling technique is used to enhance minority class samples in the training set to balance the ratio of positive to negative samples; S55: XGBoost model training, using optimized hyperparameters to train the binary classification model, and configuring an early stopping strategy to prevent overfitting; S56: SHAP Interpretability Analysis. The SHAP method is used to calculate the global importance and local contribution of each feature, generating interpretable analysis results for the model.
7. The short-term rainstorm identification and mechanism interpretation method according to claim 1, characterized in that, The short-term rainstorm meteorological mechanism interpretation step S6 specifically includes the following sub-steps: S61: Factor contribution quantification, based on SHAP global importance ranking, to determine the contribution weights of air temperature, specific humidity, shortwave radiation, relative humidity, air pressure, and wind speed to the occurrence of short-term rainstorms; S62: Nonlinear relationship analysis, based on SHAP dependency graph, analyzes the nonlinear response relationship between each meteorological factor and the probability of rainstorm occurrence, and determines the high-risk threshold interval; S63: Key driving factor identification: Based on contribution weight and nonlinear response characteristics, screen and determine the dominant and secondary meteorological factors for the occurrence of short-term rainstorms. S64: Physical mechanism interpretation, combining meteorological principles, to explain the physical mechanism of strong convection development and short-term rainstorm formation under conditions of relatively low temperature, high humidity, low shortwave radiation, and low pressure; S65: Output the mechanism interpretation results, forming a short-term rainstorm mechanism interpretation report that includes factor contributions, nonlinear relationships, key thresholds, and physical mechanisms.
8. A short-term rainstorm identification and machine interpretation system, characterized in that, It includes modules for data acquisition and preprocessing, short-term rainstorm event identification, spatiotemporal change analysis, three-dimensional structure analysis, interpretable machine learning, and mechanism interpretation. The data acquisition and preprocessing module is used to acquire multi-source remote sensing precipitation data, ground meteorological observation data and topographic data, and construct a fused precipitation dataset with unified spatiotemporal resolution through quality control, spatiotemporal matching and gridding processing. The short-term rainstorm event identification module is used to identify short-term rainstorm events based on a state machine model and extract characteristic indicators such as intensity, duration, range, and cumulative precipitation. The spatiotemporal variation analysis module is used to analyze the temporal variation trend, spatial distribution pattern and spatiotemporal heterogeneity of short-term rainstorms; The three-dimensional structure analysis module is used to analyze the vertical profile, storm top height, and bright band characteristics of typical rainstorm events based on FY-3GPMR satellite precipitation radar data. The interpretable machine learning module is used to construct the EBOX interpretable Bayesian optimized XGBoost model, complete the sample imbalance handling, and use the SHAP method to perform model interpretability analysis. The mechanism interpretation module is used to quantify the contribution of meteorological factors, identify key driving factors, and interpret the physical mechanism of short-term rainstorm formation based on the SHAP interpretation results.
9. The system according to claim 8, characterized in that, The data acquisition and preprocessing module is specifically used to: correct satellite precipitation data deviations using the optimal interpolation method based on ground rain gauges, and unify multi-source data to a 0.1°×0.1° spatial grid and a 1-hour temporal resolution through bilinear interpolation; the state machine model in the short-term rainstorm event identification module is configured as follows: the initial state is non-rainstorm, the state switches to rainstorm when the hourly rainfall is ≥8mm for the first time, and the state returns to non-rainstorm when the rainfall is <8mm for 2 consecutive hours.
10. The system according to claim 8, characterized in that, The interpretable machine learning module includes a sample partitioning unit, a feature standardization unit, a Bayesian hyperparameter optimization unit, a SMOTE sample balancing unit, an XGBoost training unit, and a SHAP interpretation unit; the machine interpretation module is also used to output a report on the importance ranking of meteorological factors, nonlinear response relationships, high-risk threshold intervals, and physical mechanism interpretation of short-term rainstorms.