Complex rainfall type accurate identification method and system
By integrating multimodal feature learning and physical constraints through the PI-SwinT model, the issues of accuracy and rationality in the identification of complex precipitation types were resolved. This enabled high-precision identification of multiple mixed precipitation types, improving the accuracy and interpretability of weather forecasts.
Patent Information
- Application Number
- CN202511668930.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-10
AI Technical Summary
Existing precipitation type identification methods are not very accurate when facing complex weather conditions, especially mixed precipitation. They are unable to accurately distinguish between precipitation types with similar physical properties but different disaster impacts, such as freezing rain and ice pellets. Furthermore, they lack the ability to comprehensively model the deep nonlinear correlation between multidimensional meteorological elements and precipitation types.
The PI-SwinT model is adopted, which combines physical information neural network constraints and the Swin-Transformer architecture. By combining multimodal feature learning and physical constraints, physical constraints such as static equilibrium, pressure gradient and temperature vertical lapse rate are constructed, and the model training process is optimized to achieve high-precision identification of complex precipitation types.
It significantly improves the accuracy and physical plausibility of identifying complex mixed precipitation types such as rain, snow, sleet, freezing rain, and ice pellets, and expands the application scope of meteorological forecasting models.
Smart Images
Figure CN121502215A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological detection technology, and in particular to a method, system, device and storage medium for accurate identification of complex precipitation types. Background Technology
[0002] In the field of atmospheric science and meteorological forecasting, the identification of complex precipitation types has received widespread attention as a core technical problem with significant application value. However, due to the low accuracy of identifying mixed precipitation types under complex weather conditions (such as winter precipitation, transitional season precipitation, and severe convective precipitation such as hail), existing precipitation phase identification technologies often perform poorly when faced with problems such as single data sources and simple threshold rules.
[0003] Traditional precipitation type identification methods primarily rely on single meteorological radar reflectivity factors, ground observation station data, or numerical model products, making logical judgments based on empirical thresholds (such as the zero-degree bright band and temperature thresholds). These methods perform poorly when dealing with mixed precipitation types involving complex temperature stratification and the coexistence of various precipitation particles. They often struggle to accurately identify precipitation types with similar physical properties but vastly different hazardous impacts, such as freezing rain, ice pellets, and wet snow. Furthermore, traditional machine learning-based identification methods heavily depend on feature engineering and domain expert knowledge. When faced with complex nonlinear meteorological models, these models are prone to insufficient feature extraction or inadequate generalization ability. Moreover, most existing identification methods lack the comprehensive modeling capability to understand the complex meteorological conditions behind precipitation events, making it difficult to accurately capture the deep nonlinear correlations between multidimensional meteorological elements and precipitation types. Summary of the Invention
[0004] The purpose of this invention is to provide a method, system, device and storage medium for accurate identification of complex precipitation types, aiming to solve or improve at least one of the above-mentioned technical problems.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for accurate identification of complex precipitation types includes: Obtain the original dataset; the dataset includes the crowdsourced dataset and the ERA5 reanalysis dataset; A PI-SwinT model is constructed; the PI-SwinT model consists of physical information neural network constraints and a Swin-Transformer architecture; the physical information neural network constraints include static equilibrium constraints, pressure gradient constraints, and temperature vertical lapse rate constraints. The PI-SwinT model is trained using the original dataset, and the trained PI-SwinT model is used to identify complex precipitation types in the test data.
[0006] Optionally, the process of training the PI-SwinT model using the original dataset includes: The total loss function is constructed based on the physical information neural network constraints and multi-class cross-entropy loss. The original dataset is input into the PI-SwinT model, and the total loss function is used for iterative training until the maximum number of iterations is reached, resulting in a trained PI-SwinT model.
[0007] Optionally, the static equilibrium constraint is specifically expressed as: in, p For pressure field, r air density, g This is the acceleration due to gravity.
[0008] Optionally, the pressure gradient constraint is specifically expressed as: in, p For pressure field, T For temperature field, u , v For wind field components, n is the normal vector of the spatial grid, used to describe the constraint direction of the pressure gradient in the spatial direction.
[0009] Optionally, the temperature vertical lapse rate constraint is specifically expressed as: in, T For temperature field, c The adiabatic lapse rate is denoted as .
[0010] This invention also provides a system for accurately identifying complex precipitation types, comprising: A data acquisition unit is used to acquire raw datasets; the datasets include crowdsourced datasets and ERA5 reanalysis datasets. The model building unit is used to build the PI-SwinT model; the PI-SwinT model consists of physical information neural network constraints and a Swin-Transformer architecture; the physical information neural network constraints include static equilibrium constraints, pressure gradient constraints, and temperature vertical lapse rate constraints. The model training and application unit is used to train the PI-SwinT model using the original dataset, and to use the trained PI-SwinT model to identify complex precipitation types in the test data.
[0011] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described method for accurately identifying complex precipitation types.
[0012] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for accurately identifying complex precipitation types as described above.
[0013] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method, system, device, and storage medium for accurate identification of complex precipitation types. The method includes acquiring a raw dataset; the dataset includes a crowdsourced dataset and an ERA5 reanalysis dataset; constructing a PI-SwinT model; the PI-SwinT model consists of physical information neural network constraints and a Swin-Transformer architecture; the physical information neural network constraints include static equilibrium constraints, pressure gradient constraints, and vertical temperature lapse rate constraints; training the PI-SwinT model using the raw dataset; and using the trained PI-SwinT model to identify complex precipitation types in the test data. This invention can solve the identification difficulties caused by the complexity of precipitation phases and the limitations of traditional identification methods, thereby improving the accuracy of precipitation type identification. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a flowchart of the precipitation type identification process in this embodiment; Figure 2 This is a schematic diagram of the data flow during the training of the PI-SwinT model in this embodiment. Detailed Implementation
[0016] The technical solutions of the embodiments 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, and 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.
[0017] The purpose of this invention is to provide a method, system, device and storage medium for accurate identification of complex precipitation types, aiming to solve or improve at least one of the above-mentioned technical problems.
[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] like Figure 1-Figure 2 As shown, this invention provides a method for accurate identification of complex precipitation types. By combining multimodal feature learning and PINN constraints, it achieves high-precision automated identification of complex mixed precipitation types such as rain, snow, sleet, freezing rain, and ice pellets, significantly improving the model's accuracy and physical rationality in identifying complex precipitation weather.
[0020] This invention employs a crowdsourced dataset and ERA5 meteorological data for research. The crowdsourced dataset originates from the "TahoeRain or Snow" and "Mountain Rain or Snow" citizen science projects, covering precipitation observation areas in multiple mountainous regions of the United States. It contains 38,500 geotagged and timestamped precipitation phase observation records, collected from January 8, 2020 to July 24, 2023. While the crowdsourced dataset provides abundant precipitation observation data, the accuracy of identifying complex mixed precipitation types remains a critical issue. Therefore, this study first explores a multimodal feature fusion method, integrating multi-source meteorological data and extracting hierarchical features based on the Swin-Transformer architecture. The Swin-Transformer, a type of visual Transformer, uses a sliding window attention mechanism and a hierarchical feature extraction process to enable the model to fully learn spatial feature information at different scales. This allows it to capture deep features closely related to precipitation types from multidimensional meteorological elements, achieving the identification of complex precipitation types from a feature learning perspective. However, in the research task of precipitation type identification based on multi-source meteorological data, simply extracting data-driven features still cannot guarantee the physical rationality of the prediction results. Therefore, it is necessary to develop an identification method that integrates data features and physical constraints while ensuring the inherent consistency between the two.
[0021] Accordingly, this invention first constructs a Swin-Transformer model with physical information constraints (Physics-Informed Swin-Transformer, PI-SwinT). By jointly modeling the correlation between multi-source meteorological features and physical constraints, and simultaneously extracting deep features with physical consistency, it achieves high-precision identification of complex precipitation types. PI-SwinT not only retains the hierarchical feature learning capability of the Swin-Transformer but also introduces a physical information neural network constraint module into the model structure, ensuring that the identified precipitation types are consistent with atmospheric physical laws. The specific structure is as follows: Figure 2 As shown. The core idea of PI-SwinT is based on physical constraints. It uses Patch Partition and multi-layered Swin-Transformer Blocks to represent the coupling relationships of multi-dimensional meteorological elements as part of feature learning. It combines multiple physical constraint loss functions, including physical consistency constraints, temperature, humidity, and classification, to obtain precipitation type results that match real atmospheric conditions. Its implementation process is as follows: First, input multi-source meteorological data X m The data is standardized through a data preprocessing and resizing module to obtain preprocessed meteorological data. X'm Subsequently, the Swin-Transformer feature extractor was used. F ( X'm The preprocessed data is transformed into a multi-scale feature representation. This hierarchical feature extraction ensures that meteorological information at different spatial scales can be effectively captured and modeled. In the Patch Embedding stage, the input data is converted into a patch sequence.
[0022] In the formula H 0 It is by X'm The initial feature representation is obtained through patch embedding. During feature extraction, hierarchical features are extracted layer by layer using a multi-stage stacking structure. In the formula H l It is the first l The characteristic representation of a stage. SwinStage l Indicates the first l The Swin-Transformer module of the layer, L This represents the total number of layers. After feature extraction using the Swin-Transformer, the final feature representation... Hl The samples are fed into a classification head for probability prediction of precipitation type: in W c and b c These are the weights and bias parameters of the classification head, respectively. In parallel, the Physical Constraint Loss Calculation Module (PINN) simultaneously receives the preprocessed input data. X'm and Swin-Transformer Extracted features H l The loss due to three physical constraints is calculated. These physical constraints include static equilibrium constraints, pressure gradient constraints, and vertical temperature lapse rate constraints. In the formula, p For pressure field, T For temperature field, u , v For wind field components, r air density, g It is the acceleration due to gravity. c Let be the adiabatic lapse rate. The total loss due to physical constraints is: in l 1. l 2 and l 3 represents the weighting coefficients of each physical constraint.
[0023] During training, the total loss function of PI-SwinT is designed through a loss fusion module to simultaneously optimize classification accuracy and physical consistency, ensuring that the prediction results conform to atmospheric physical laws. Among them, classification loss L cls For multi-class cross-entropy loss: In the formula N For the sample size, C This represents the number of precipitation type categories. y i,c For real labels, P cls,i,c Predict probabilities for the model 。 l' This is the balance coefficient for physical constraint losses. It is achieved by minimizing the total loss function. L totalAt the same time, optimize model parameters i To ensure that the predicted precipitation type is consistent with physical laws: The final output is obtained through optimization after loss fusion: Through this mechanism of parallel computation of feature extraction and physical constraints, and loss fusion optimization, PI-SwinT can learn the complex nonlinear relationship between meteorological elements and precipitation types in a multi-scale feature space. At the same time, it ensures that the prediction results conform to the laws of atmospheric physics, and achieves high-precision physical consistency identification of complex precipitation types such as rain, snow, sleet, freezing rain, and ice pellets, thereby improving the model's identification accuracy and physical interpretability.
[0024] The technical solution of this invention solves the problems of insufficient feature extraction and lack of physical rationality in the identification of complex precipitation types by integrating multimodal feature learning and physical constraint mechanisms through the PI-SwinT model. This enables accurate identification of various mixed precipitation types such as rain, snow, sleet, freezing rain, and ice pellets. This method not only improves the model's accuracy and physical consistency in identifying complex precipitation weather, but also expands the application scope of existing meteorological prediction models in precipitation identification. It has broad application prospects in the field of complex precipitation weather prediction and is expected to provide scientific basis and technical support for scientific research and practical applications in areas such as meteorological forecasting and early warning, transportation safety, and agricultural meteorological services.
[0025] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0026] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for accurate identification of complex precipitation types, characterized in that, include: Obtain the original dataset; the dataset includes the crowdsourced dataset and the ERA5 reanalysis dataset; A PI-SwinT model is constructed; the PI-SwinT model consists of physical information neural network constraints and a Swin-Transformer architecture; the physical information neural network constraints include static equilibrium constraints, pressure gradient constraints, and temperature vertical lapse rate constraints. The PI-SwinT model is trained using the original dataset, and the trained PI-SwinT model is used to identify complex precipitation types in the test data.
2. The method for accurate identification of complex precipitation types according to claim 1, characterized in that, The specific process of training the PI-SwinT model using the original dataset includes: The total loss function is constructed based on the physical information neural network constraints and multi-class cross-entropy loss. The original dataset is input into the PI-SwinT model, and the total loss function is used for iterative training until the maximum number of iterations is reached, resulting in a trained PI-SwinT model.
3. The method for accurate identification of complex precipitation types according to claim 1, characterized in that, The static equilibrium constraint is specifically expressed as follows: in, p For pressure field, ρ air density, g This is the acceleration due to gravity.
4. The method for accurate identification of complex precipitation types according to claim 1, characterized in that, The pressure gradient constraint is specifically expressed as follows: in, p For pressure field, T For the temperature field, u , v For wind field components, n is the normal vector of the spatial grid, used to describe the constraint direction of the pressure gradient in the spatial direction.
5. The method for accurate identification of complex precipitation types according to claim 1, characterized in that, The vertical temperature lapse rate constraint is specifically expressed as follows: in, T For the temperature field, γ The adiabatic lapse rate is denoted as .
6. A system for accurately identifying complex precipitation types, characterized in that, include: A data acquisition unit is used to acquire raw datasets; the datasets include crowdsourced datasets and ERA5 reanalysis datasets. The model building unit is used to build the PI-SwinT model; the PI-SwinT model consists of physical information neural network constraints and a Swin-Transformer architecture; the physical information neural network constraints include static equilibrium constraints, pressure gradient constraints, and temperature vertical lapse rate constraints. The model training and application unit is used to train the PI-SwinT model using the original dataset, and to use the trained PI-SwinT model to identify complex precipitation types in the test data.
7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform the method for accurate identification of complex precipitation types according to any one of claims 1-5.
8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method for accurate identification of complex precipitation types as described in any one of claims 1-5.