Melting layer top height machine learning inversion method and system
Through machine learning inversion method, combined with satellite and precipitation radar data, a BP neural network model is established, which solves the problem of insufficient inversion accuracy of the melt layer in the traditional method and achieves higher inversion accuracy.
Patent Information
- Application Number
- CN202411725834.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-04-08
AI Technical Summary
When traditional methods identify the height of the melt layer, the inversion accuracy is limited, making it difficult to accurately obtain the top height of the melt layer.
Using machine learning inversion method, a BP neural network model is established by combining satellite group one, satellite group two and GPM dual-frequency precipitation radar data, and the data is time-space matching, preprocessing and training, and a melt layer top height inversion model is generated.
The inversion accuracy of the top height of the melt layer is effectively improved, and the accuracy of the data content is improved through data preprocessing and model training, and the problem of insufficient accuracy in the existing technology is optimized.
Smart Images
Figure CN119783503A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method and system for machine learning inversion of melt layer top height, and belongs to the field of meteorological satellite remote sensing design. Background Art
[0002] The melt layer is the height layer where precipitation particles are transformed from solid to liquid, which is meaningful for analyzing the relevant characteristics of particles in different states and the particle transformation process. The melt layer is an important component of the cloud and precipitation system, and its height change directly affects the growth of cloud droplets and the formation of precipitation. In the melt layer, the density, falling speed, shape and size of ice crystal particles will change, and these changes directly affect the radar reflectivity factor and other microphysical parameters. Traditional methods mainly identify the height of the melt layer by enhancing the reflectivity factor, and the inversion accuracy is limited. Summary of the invention
[0003] The technical problem solved by the present invention is: in view of the problem that traditional means in the current prior art easily affect the inversion accuracy of the melting layer height, a method and system for machine learning inversion of the melting layer top height are proposed.
[0004] The present invention solves the above technical problems by the following technical solutions:
[0005] A method for inverting the height of the top of a melt layer using machine learning, comprising:
[0006] Collect satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data;
[0007] Perform spatial and temporal matching on the data of satellite group 1 and satellite group 2, preprocess the matched data, obtain the preprocessed data of satellite group 1 and satellite group 2, and generate a matching data set with the GPM dual-frequency precipitation radar data;
[0008] The preprocessed data is input into the learning model for training to obtain the inversion model of the melt layer top height.
[0009] The matching time used for spatiotemporal matching of the collected data is based on the GPM dual-frequency precipitation radar time. The detection time of satellite group 1 and satellite group 2 and the time interval of the GPM dual-frequency precipitation radar are less than the preset range.
[0010] The method for preprocessing the matching data set is:
[0011] Determine the data of satellite group 1 and satellite group 2 in the matching data set as preprocessing objects;
[0012] After performing radiometric calibration and geometric correction on a set of satellite data, brightness temperature data and visible spectrum data are obtained;
[0013] Orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction are performed on the two sets of satellite data to obtain the dual-polarization backscatter coefficients.
[0014] After the learning model is activated by the activation function, the preprocessed data is input into the learning model for training. The learning model uses a BP neural network. The BP neural network includes an input layer, a hidden layer and an output layer. The number of layers is 1. The optimization target of the BP neural network is the number of neurons in the hidden layer.
[0015] The method for data training of the BP neural network is:
[0016] After receiving the preprocessed data and performing normalization processing, the normalized data is used as input data; the BP neural network generates and outputs model training values based on the normalized data, and the model training values are denormalized after output.
[0017] When the number of neurons in the hidden layer of the BP neural network is optimized within the preset range, the data training is considered completed. The BP neural network after data training is used as the inversion model for the top height of the melting layer. The normalized data is used as input data to generate the model inversion value of the top height of the melting layer and output it.
[0018] The preprocessed data is randomly divided into a training set and a validation set, with a division ratio of 7:3.
[0019] The method for normalizing the preprocessed data is:
[0020]
[0021] In the formula, x input is the data set for the normalized input model, x i Includes all input features and output features of the dataset, x max and x min Indicates the maximum and minimum values of the feature.
[0022] The method for denormalizing the model training values after output is:
[0023] y real =y pred *(y max -y min )+y min
[0024] In the formula, y real Indicates the height of the top of the melt layer in the model inversion, y pred Represents the model prediction value, y max and min Indicates the maximum and minimum values of the output feature.
[0025] The machine learning inversion model of the melt layer top height includes a data acquisition module, a time-space matching module, a preprocessing module, a learning model generation and training module, among which:
[0026] Data acquisition module, collecting satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data;
[0027] The space-time matching module performs space-time matching on the collected satellite group 1 and satellite group 2 data, and sends the matched data to the preprocessing module after obtaining it;
[0028] The preprocessing module preprocesses the space-time matching data sent by the space-time matching module. The preprocessing includes: radiation calibration and geometric correction of satellite group 1 data, orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction of satellite group 2 data. The brightness temperature data and visible spectrum data obtained after the preprocessing of satellite group 1 data are sent to the learning model generation and training module, and the dual-polarization backscattering coefficient obtained after the preprocessing of satellite group 2 data are sent to the learning model generation and training module.
[0029] The learning model generation and training module generates the learning model to be trained, adjusts the model according to the dual-polarization backscattering coefficient, and uses the preprocessed data as the training set and verification set to train the learning model. After the training is completed, the inversion model of the top height of the melt layer is obtained.
[0030] The advantages of the present invention compared with the prior art are:
[0031] The present invention provides a method and system for machine learning inversion of the melt layer top height. By combining two groups of satellites and satellite-borne synthetic aperture radar related products with GPM dual-frequency precipitation radar data, a machine learning inversion model is established, which effectively improves the inversion accuracy of the melt layer top height. By data preprocessing and normalization and denormalization of model input and output data, the accuracy of data content is improved, further optimizing the insufficient accuracy of the melt layer top height in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 A flow chart of the method for machine learning inversion of melt layer top height provided by the present invention. DETAILED DESCRIPTION
[0033] A method and system for machine learning inversion of melt layer top height obtains two sets of satellite and precipitation radar data, matches the three sets of data in time and space based on time and space information to obtain a matching data set, preprocesses the three sets of data to obtain brightness temperature, visible spectrum, and synthetic aperture radar dual-polarization backscattering coefficient of the input model, and inputs the obtained input data as a training data set into a learning model. Through training, an inversion model for the melt layer top height is obtained as a calculation model for the melt layer top height data.
[0034] The machine learning inversion method for the height of the top of the melt layer includes the following specific steps:
[0035] Collect satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data;
[0036] Perform spatial and temporal matching on the data of satellite group 1 and satellite group 2, preprocess the matched data, obtain the preprocessed data of satellite group 1 and satellite group 2, and generate a matching data set with the GPM dual-frequency precipitation radar data;
[0037] The preprocessed data is input into the learning model for training to obtain the inversion model of the melt layer top height.
[0038] The matching time used for spatiotemporal matching of the collected data is based on the GPM dual-frequency precipitation radar time. The detection time of satellite group 1 and satellite group 2 and the time interval of the GPM dual-frequency precipitation radar are less than the preset range.
[0039] The method for preprocessing the matching data set is:
[0040] Determine the data of satellite group 1 and satellite group 2 in the matching data set as preprocessing objects;
[0041] After performing radiometric calibration and geometric correction on a set of satellite data, brightness temperature data and visible spectrum data are obtained;
[0042] Orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction are performed on the two sets of satellite data to obtain the dual-polarization backscatter coefficients.
[0043] After the learning model is activated by the activation function, the preprocessed data is input into the learning model for training. The learning model uses the BP neural network, which includes an input layer, a hidden layer, and an output layer. The number of layers is 1. The optimization target of the BP neural network is the number of neurons in the hidden layer.
[0044] The method of BP neural network for data training is:
[0045] After receiving the preprocessed data and performing normalization processing, the normalized data is used as input data; the BP neural network generates and outputs model training values based on the normalized data, and the model training values are denormalized after output.
[0046] When the number of neurons in the hidden layer of the BP neural network is optimized within the preset range, the data training is considered completed. The BP neural network after data training is used as the inversion model for the top height of the melting layer. The normalized data is used as input data to generate the model inversion value of the top height of the melting layer and output it.
[0047] The method for normalizing the preprocessed data is:
[0048]
[0049] In the formula, x input is the data set for the normalized input model, x i Includes all input features and output features of the dataset, x max and x min Indicates the maximum and minimum values of the feature.
[0050] The method for denormalizing the model training values after output is:
[0051] y real =y pred *(y max -y min )+y min
[0052] In the formula, y real Indicates the height of the top of the melt layer in the model inversion, y pred Represents the model prediction value, y max and min Indicates the maximum and minimum values of the output feature.
[0053] The machine learning inversion model of the melt layer top height includes a data acquisition module, a time-space matching module, a preprocessing module, a learning model generation and training module, among which:
[0054] Data acquisition module, collecting satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data;
[0055] The space-time matching module performs space-time matching on the collected satellite group 1 and satellite group 2 data, and sends the matched data to the preprocessing module after obtaining it;
[0056] The preprocessing module preprocesses the space-time matching data sent by the space-time matching module. The preprocessing includes: radiation calibration and geometric correction of satellite group 1 data, orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction of satellite group 2 data. The brightness temperature data and visible spectrum data obtained after the preprocessing of satellite group 1 data are sent to the learning model generation and training module, and the dual-polarization backscattering coefficient obtained after the preprocessing of satellite group 2 data are sent to the learning model generation and training module.
[0057] The learning model generation and training module generates the learning model to be trained, adjusts the model according to the dual-polarization backscattering coefficient, and uses the preprocessed data as the training set and verification set to train the learning model. After the training is completed, the inversion model of the top height of the melt layer is obtained.
[0058] The following is further described in conjunction with the accompanying drawings and preferred embodiments of the specification:
[0059] In the current embodiment, Fengyun-4 and Sentinel-1 are used as satellite group 1 and satellite group 2 respectively, and GPM dual-frequency precipitation radar data is selected;
[0060] The data of Fengyun-4A and B are obtained from the Fengyun Satellite Remote Sensing Data Service Network; the data of Sentinel-1A and B are obtained from the official website of ESA, and the data of GPM dual-frequency precipitation radar are obtained from the official website of NASA precipitation measurement.
[0061] Based on the spatiotemporal information, the three are matched in time and space to obtain a matching data set. The three data are preprocessed to obtain the brightness temperature, visible spectrum, synthetic aperture radar dual-polarization backscatter coefficient and melt layer top height data of the input model.
[0062] The data are matched based on the observation time and space of the GPM dual-frequency precipitation radar. The spatial resolution of the data is 5 km. The observation time of the Sentinel-1 satellite of the Fengyun-4 and the observation time interval of the GPM dual-frequency precipitation radar shall not exceed ±15 minutes.
[0063] ENVI software was used to preprocess the FY-4 data, including radiation calibration and geometric correction, to obtain brightness temperature and visible spectrum data.
[0064] The Sentinel-1 satellite data were preprocessed using SNAP software, including orbit correction, thermal noise removal, radiation calibration, filtering, and geometric correction to obtain dual-polarization backscatter coefficients.
[0065] A dataset of Fengyun-4, Sentinel-1 and GPM dual-frequency precipitation radar data was established by combining spatiotemporal information.
[0066] The above data set is input into the BP machine learning model to train and obtain the melt layer top height inversion model.
[0067] The machine learning model constructed is a BP neural network, and the activation function used is Sigmod. The established BP neural network model includes an input layer, a hidden layer, and an output layer. The number of layers of the three layers is 1, and the number of neurons in the hidden layer is optimized during the model training process;
[0068] In order to ensure the reliability of the model, the model data set is randomly divided into a training set and a validation set, with a ratio of 7:3.
[0069] Because different eigenvalues have different dimensions, the data set needs to be normalized before being input into the model; finally, the predicted value obtained by the model is denormalized to obtain the height of the top of the melt layer inverted by the model. Normalization and denormalization can be expressed as:
[0070]
[0071] y real =y pred *(y max -y min )+y min
[0072] Among them, x input is the data set for the normalized input model, x i Includes all input features and output features of the dataset, x max and x min Represents the maximum and minimum values of the feature; y real Indicates the height of the top of the melt layer in the model inversion, y pred Represents the model prediction value, y max and min Indicates the maximum and minimum values of the output feature.
[0073] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.
[0074] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.
Claims
1. A method for inverting the height of the top of a melt layer by machine learning, characterized in that include: Collect satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data; Perform spatial and temporal matching on the data of satellite group 1 and satellite group 2, preprocess the matched data, obtain the preprocessed data of satellite group 1 and satellite group 2, and generate a matching data set with the GPM dual-frequency precipitation radar data; The preprocessed data is input into the learning model for training to obtain the inversion model of the melt layer top height.
2. The method for machine learning inversion of melt layer top height according to claim 1, characterized in that: The matching time used for spatiotemporal matching of the collected data is based on the GPM dual-frequency precipitation radar time. The detection time of satellite group 1 and satellite group 2 and the time interval of the GPM dual-frequency precipitation radar are less than the preset range.
3. The method for machine learning inversion of melt layer top height according to claim 1, characterized in that: The method for preprocessing the matching data set is: Determine the data of satellite group 1 and satellite group 2 in the matching data set as preprocessing objects; After performing radiometric calibration and geometric correction on a set of satellite data, brightness temperature data and visible spectrum data are obtained; Orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction are performed on the two sets of satellite data to obtain the dual-polarization backscatter coefficients.
4. The method for machine learning inversion of melt layer top height according to claim 1, characterized in that: After the learning model is activated by the activation function, the preprocessed data is input into the learning model for training. The learning model uses a BP neural network. The BP neural network includes an input layer, a hidden layer and an output layer. The number of layers is 1. The optimization target of the BP neural network is the number of neurons in the hidden layer.
5. A method for machine learning inversion of melt layer top height according to claim 4, characterized in that: The method for data training of the BP neural network is: After receiving the preprocessed data and performing normalization processing, the normalized data is used as input data; the BP neural network generates and outputs model training values based on the normalized data, and the model training values are denormalized after output.
6. A method for machine learning inversion of melt layer top height according to claim 5, characterized in that: When the number of neurons in the hidden layer of the BP neural network is optimized within the preset range, the data training is considered completed. The BP neural network after data training is used as the inversion model for the top height of the melting layer. The normalized data is used as input data to generate the model inversion value of the top height of the melting layer and output it.
7. The method for machine learning inversion of melt layer top height according to claim 4, characterized in that: The preprocessed data is randomly divided into a training set and a validation set, with a division ratio of 7:
3.
8. The method for machine learning inversion of melt layer top height according to claim 4, characterized in that: The method for normalizing the preprocessed data is: In the formula, x input is the data set for the normalized input model, x i Includes all input features and output features of the dataset, x max and x min Indicates the maximum and minimum values of the feature.
9. The method for machine learning inversion of melt layer top height according to claim 4, characterized in that: The method for denormalizing the model training values after output is: and real =and pred *(and max -and min )+y min In the formula, y real Indicates the height of the top of the melt layer in the model inversion, y pred Represents the model prediction value, y max and min Indicates the maximum and minimum values of the output feature.
10. A melt layer top height machine learning system for implementing the inversion method of claim 1, characterized in that: It includes data acquisition module, space-time matching module, preprocessing module, learning model generation and training module, among which: Data acquisition module, collecting satellite group 1, satellite group 2 and GPM dual-frequency precipitation radar data; The space-time matching module performs space-time matching on the collected satellite group 1 and satellite group 2 data, and sends the matched data to the preprocessing module after obtaining it; The preprocessing module preprocesses the space-time matching data sent by the space-time matching module. The preprocessing includes: radiation calibration and geometric correction of satellite group 1 data, orbit correction, thermal noise removal, radiation calibration, filtering and geometric correction of satellite group 2 data. The brightness temperature data and visible spectrum data obtained after the preprocessing of satellite group 1 data are sent to the learning model generation and training module, and the dual-polarization backscattering coefficient obtained after the preprocessing of satellite group 2 data are sent to the learning model generation and training module. The learning model generation and training module generates the learning model to be trained, adjusts the model according to the dual-polarization backscattering coefficient, and uses the preprocessed data as the training set and verification set to train the learning model. After the training is completed, the inversion model of the top height of the melting layer is obtained.