Dual-band radar raindrop spectrum inversion method, device and electronic equipment

Abstract

The present invention provides a dual-band radar raindrop spectrum inversion method, device and electronic equipment thereof, belonging to the technical field of remote sensing image processing, comprising: inputting X-band reflectivity, Ka-band reflectivity and X / Ka-band radar dual-frequency ratio into a first learning model to obtain Ka / W-band radar dual-frequency ratio; inputting Ka / W-band radar dual-frequency ratio, X-band reflectivity and Ka-band reflectivity into a second learning model to obtain raindrop mass weighted average diameter D m ; Input the X-band reflectivity and the Ka-band reflectivity into the third learning model to obtain the liquid water content; determine the raindrop spectrum distribution. The present invention provides a raindrop spectrum inversion method based on ensemble learning, which does not require the establishment of a physical inversion model of radar parameters and raindrop spectrum parameters, and effectively solves the DFR-D problem existing in the inversion of raindrop spectrum by dual-band radar. m Double solution problem, the inverted raindrop spectrum is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a dual-band radar raindrop spectrum inversion method, device and electronic equipment thereof. Background Art

[0002] Drop Size Distribution (DSD) is a basic description of the size distribution of precipitation particles, which reflects the change in the number concentration of precipitation particles with a certain particle diameter range within a unit volume. Studying DSD is helpful to improve the accuracy of radar quantitative estimation of precipitation, improve or optimize the parameterization scheme of atmospheric climate models, and improve the effect of artificial weather modification.

[0003] Dual-band radar is widely used in raindrop spectrum inversion because it can provide a dual frequency ratio (DFR) that is only related to the raindrop spectrum distribution characteristics and has nothing to do with the number concentration of precipitation particles. However, when the precipitation particles are liquid (the shape factor of the raindrops is µ=3), when DFR>0, DFR-D m There is a unique solution that can be used to obtain the raindrop mass-weighted average diameter D through DFR inversion m value; but when DFR≤0, DFR corresponds to two D m The unique D cannot be obtained through DFR inversion. m .

[0004] Of course, to solve the DFR-D m To solve the double solution problem, relevant scholars proposed a generalized inversion method. m Dual solution problem, but DFR* is no longer D m is the only function of, but is also the normalized particle number concentration N W Therefore, this inversion method sacrifices DFR and is only D m The advantage of the function causes the problem of reduced inversion accuracy.

[0005] In addition, there is also a Ku / Ka dual-band radar that uses the rainfall rate R and D m Realization of Relationship m The inversion method, although avoiding DFR-D m The double solution problem, however, the introduced adjustment factor ε and the need to classify the precipitation process will introduce new errors and lead to inaccurate inversion results.

[0006] In view of this, it is urgent to provide a new raindrop spectrum inversion method that can effectively solve the DFR-D m Double solution problem. Summary of the invention

[0007] The present invention provides a dual-band radar raindrop spectrum inversion method, device and electronic equipment thereof, which are used to solve the drawback that the raindrop spectrum inversion cannot be performed accurately and efficiently in the prior art.

[0008] The present invention provides a dual-band radar raindrop spectrum inversion method, comprising the following steps:

[0009] Determine the X-band reflectivity, Ka-band reflectivity and X / Ka-band radar dual-frequency ratio based on X / Ka-band dual-band radar observation data within the detection range;

[0010] Inputting the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio into a first learning model, and obtaining the Ka / W-band radar dual-frequency ratio output by the first learning model;

[0011] Inputting the Ka / W band radar dual frequency ratio, the X band reflectivity and the Ka band reflectivity into a second learning model, and obtaining a raindrop mass weighted average diameter output by the second learning model;

[0012] Inputting the X-band reflectivity and the Ka-band reflectivity into a third learning model to obtain the liquid water content output by the third learning model;

[0013] Determining the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content;

[0014] The first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

[0015] The present invention also provides a dual-band radar raindrop spectrum inversion device, comprising the following modules:

[0016] A feature extraction unit is used to determine the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio based on the X / Ka-band radar observation data within the detection range;

[0017] a first identification unit, configured to input the X-band reflectivity, the Ka-band reflectivity, and the X / Ka-band radar dual-frequency ratio into a first learning model, and obtain the Ka / W-band radar dual-frequency ratio output by the first learning model;

[0018] a second identification unit, inputting the Ka / W band radar dual frequency ratio, the X band reflectivity and the Ka band reflectivity into a second learning model, and acquiring a raindrop mass weighted average diameter output by the second learning model;

[0019] a third recognition unit, inputting the X-band reflectivity and the Ka-band reflectivity into a third learning model, and acquiring the liquid water content output by the third learning model;

[0020] A result output unit, used to determine the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content;

[0021] The first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using a plurality of X / Ka dual-band radar sample data.

[0022] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the above-mentioned dual-band radar raindrop spectrum inversion methods is implemented.

[0023] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the dual-band radar raindrop spectrum inversion method as described in any one of the above is implemented.

[0024] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the dual-band radar raindrop spectrum inversion method as described in any one of the above is implemented.

[0025] The dual-band radar raindrop spectrum inversion method, device and electronic equipment provided by the present invention provide a raindrop spectrum inversion method based on ensemble learning, which does not require the establishment of a physical inversion model of radar parameters and raindrop spectrum parameters, and effectively solves the DFR-D problem existing in the dual-band radar raindrop spectrum inversion. m Double solution problem, the inverted raindrop spectrum is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 It is the X / Ka dual-band radar DFR-D obtained by existing technology during scattering simulation m A scatter diagram of .

[0028] Figure 2 is the existing generalized DFR*-D m X / Ka dual-band radar DFR-D obtained during inversionm A scatter diagram of .

[0029] Figure 3 It is a flow chart of the dual-band radar raindrop spectrum inversion method provided by the present invention.

[0030] Figure 4 It is a schematic diagram of the data collection process of the centralized training set provided by the present invention.

[0031] Figure 5 It is a training flow chart of the first learning model provided by the present invention.

[0032] Figure 6 It is a training flow chart of the second learning model provided by the present invention.

[0033] Figure 7 It is a training flow chart of the third learning model provided by the present invention.

[0034] Figure 8 It is one of the flowcharts of integrated training of any learning model provided by the present invention.

[0035] Fig. 9 This is the second flowchart of the integrated training of any learning model provided by the present invention.

[0036] Fig.10 It is a structural schematic diagram of the dual-band radar raindrop spectrum inversion device provided by the present invention.

[0037] Fig.11 It is a schematic diagram comparing the predicted value and the true value predicted by the first learning model provided by the present invention for the training set.

[0038] Fig.12 It is a schematic diagram comparing the predicted value and the true value predicted by the second learning model provided by the present invention for the training set.

[0039] Fig.13 It is a schematic diagram of comparison between the predicted value and the true value of the training set predicted by the second learning model provided by the present invention under DFR (X, Ka).

[0040] Fig.14 It is a schematic diagram comparing the predicted value and the true value of the training set predicted by the third learning model provided by the present invention.

[0041] Fig.15 is the standardized particle number concentration N calculated by the present invention w Comparison diagram of the predicted value and the true value.

[0042] Fig.16 It is a schematic diagram for comparing the predicted value and the true value of DFR (Ka, W) obtained by predicting the test set data using the first learning model provided by the present invention.

[0043] Fig.17 D is obtained by predicting the test set data using the second learning model provided by the present invention. m Comparison diagram of the predicted value and the true value.

[0044] Fig.18 D is obtained by predicting the test set data using the second learning model provided by the present invention. m Comparison diagram under DFR(X,Ka).

[0045] Fig.19 It is a schematic diagram comparing the predicted value of LWC obtained by predicting the test set data using the third learning model provided by the present invention and the true value.

[0046] Fig. 20 It is a schematic diagram comparing the predicted value of Nw calculated by the method provided by the present invention for predicting the test set data and the true value.

[0047] Fig.21 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 51.6 dBZ provided by the present invention.

[0048] Fig. 22 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 41.5 dBZ provided by the present invention.

[0049] Fig.23 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 31.6 dBZ provided by the present invention.

[0050] Fig.24 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 21.7 dBZ provided by the present invention.

[0051] Fig.25 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] It should be noted that in the description of the present invention, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or position relationship indicated by the terms "upper", "lower" and the like is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0054] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0055] Combine the following Figure 1-Figure 25 Through the analysis of the prior art and in combination with the detailed elaboration of the technical solution of the present invention, it is hoped that the dual-band radar raindrop spectrum inversion method, device and electronic equipment provided by the present invention will be clearly demonstrated.

[0056] In the study of cloud precipitation microphysics, almost all problems involve particle size, so particle size is a very important and basic parameter. Among them, the drop size distribution (DSD) is the basic description of the size distribution of precipitation particles. DSD reflects the change of the number concentration of precipitation particles with a certain particle diameter range in a unit volume. Studying DSD helps to improve the accuracy of radar quantitative estimation of precipitation, improve or optimize the parameterization scheme of atmospheric climate models, and improve the effect of artificial weather modification.

[0057] The DSD of precipitation particles can be directly measured by a raindrop spectrometer, but ground-based or airborne raindrop spectrometers cannot obtain DSD information with large-scale and high temporal and spatial resolution. Therefore, the DSD of large-scale and high temporal and spatial resolution is generally obtained by inversion through radar remote sensing. Accurately obtaining the DSD of precipitation particles over a large area is one of the important research directions of radar meteorology.

[0058] Dual-band radar is widely used in raindrop spectrum inversion because it can provide a dual frequency ratio (DFR) that is only related to the raindrop spectrum distribution characteristics but has nothing to do with the number concentration of precipitation particles. DFR is defined as follows:

[0059] (1)

[0060] in, For the low frequency band, For high frequency band, is the reflectivity in the low-frequency band, is the reflectivity in the high-frequency band; dB It is a pure counting unit that represents the logarithm between the reflectivity of the low-frequency band and the reflectivity of the high-frequency band.

[0061] The raindrop spectrum model is generally expressed by the standard Gamma particle spectrum distribution, and its expression is as follows:

[0062] (2)

[0063] (3)

[0064] in, is the raindrop spectrum distribution; is the shape factor of the raindrop; is the gamma function; D is the particle diameter of the raindrop; is the intermediate variable function; It is the weighted average diameter of raindrop mass, which is the ratio of the cumulative value of the mass of all raindrops to the cumulative value of the diameter of all raindrops; The standardized particle number concentration refers to the number distribution of raindrops of different diameters per unit volume. The standardized particle number concentration can eliminate the scale effect of the raindrop spectrum in different rainfall events or different environments, so that the raindrop spectra under different conditions can be compared and analyzed.

[0065] In the case of a gamma particle spectrum distribution, DFR is simply the mass-weighted mean diameter D m and form factor µIt has been proved through experiments that when the precipitation particles are in liquid state (i.e., μ is 3), the inverted raindrop spectrum is closest to the measured raindrop spectrum. In this case, DFR is only D m function.

[0066] When the precipitation particles are liquid (µ=3), DFR>0 DFR-D m There is a unique solution, which can be obtained by DFR inversion. m However, when DFR≤0, DFR corresponds to two D m , it is impossible to obtain a unique D through DFR inversion m In the past, the method was to choose a large D m As a real solution.

[0067] Figure 1 It is the X / Ka dual-band radar DFR-D obtained by existing technology during scattering simulation m The scatter diagram of is obtained by scattering simulation using the measured raindrop spectrum data, such as Figure 1 As shown in the dashed box area, when DFR(X,Ka)≤0, two D m .

[0068] To solve the DFR-D m For the dual solution problem, relevant scholars proposed the generalized DFR*-D m The inversion method is defined as follows:

[0069] (4)

[0070] Among them, γ is the adjustment factor, γ∈[0,0.8].

[0071] Figure 2 is the existing generalized DFR*-D m X / Ka dual-band radar DFR-D obtained during inversion m Scatter diagram of , where γ=0.8 is set, Figure 2 It can be seen that The corresponding D m There is also a unique solution.

[0072] Although the above generalized DFR-D m The inversion method solves the D m The dual solution problem, but since DFR* is no longer D m The only function of , which is also the normalized number concentration N W Therefore, this inversion method sacrifices DFR and is only D m function, so the generalized DFR-D m The accuracy of the inversion method is lower than that of the conventional DFR-D mInversion method.

[0073] The existing technology also provides another dual-band raindrop spectrum inversion algorithm, using the Ku / Ka dual-band radar carried by the Global Precipitation Observation GPM satellite. The following implementation adopts the method based on the rainfall rate R and D m Relationship RD m Inversion method. However, since the rainfall rate R and D m The correlation between them will show great differences with different precipitation types. Therefore, this inversion method needs to artificially classify the precipitation types in advance, mainly into stratiform cloud precipitation and convective cloud precipitation, and then invert D m , the corresponding inversion expressions are shown in formula (5) and formula (6):

[0074] (5)

[0075] (6)

[0076] Among them, formula (5) is the RD when the precipitation type is stratiform cloud rainfall. m Inversion, formula (6) is for the RD when the precipitation type is convective cloud rainfall m Inversion, ε is the adjustment factor, R is the rainfall rate.

[0077] According to formula (5) and formula (6), the above RD m Although the inversion method avoids DFR-D m Dual solution problem, but the adjustment factor introduced ε , and the need to classify the precipitation process, which is bound to introduce new errors and lead to inaccurate inversion results.

[0078] The inventors have also recently disclosed a method based on DFR-D m The X / Ka dual-band radar raindrop spectrum inversion method needs to determine whether the X / Ka band radar dual-frequency ratio is greater than 0. When it is determined to be less than or equal to 0, D is determined according to the Ka / W band radar dual-frequency ratio. m ; When it is determined to be greater than 0, D is directly determined based on the dual-frequency ratio of the X / Ka band radar m Although this inversion method can solve the DFR-D m The dual solution problem does not require classification of precipitation processes or introduction of adjustment factors, which is better than the RD method used by the GPM satellite introduced earlier. m The raindrop spectrum inversion method has advantages, and the inverted raindrop spectrum is more accurate. However, it is still necessary to adopt different raindrop spectrum inversion strategies according to whether the dual-frequency ratio of the X / Ka band radar is greater than 0, which actually provides a dual solution method.

[0079] In view of this, the present invention provides a new raindrop spectrum inversion strategy. Taking X / Ka dual-band radar as an example, a raindrop spectrum inversion method based on ensemble learning is proposed. This method trains input variables and response variables through ensemble learning to construct DFR (Ka, W), D m and LWC learning model to solve DFR-D through ensemble learning m The double-solution problem can be solved, thereby being able to invert and obtain a more accurate raindrop spectrum. The present invention has great application value and promotion prospects, and can be applied to the inversion of raindrop spectra of various satellite-borne and ground-based dual-band radars.

[0080] Figure 3 FIG. 1 is a flow chart of the dual-band radar raindrop spectrum inversion method provided by the present invention, such as Figure 3 As shown, it mainly includes but is not limited to the following steps:

[0081] Step 301, determining the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio according to the X / Ka-band dual-band radar observation data within the detection range.

[0082] Step 302: Input the X-band reflectivity, the Ka-band reflectivity, and the X / Ka-band radar dual-frequency ratio into a first learning model to obtain the Ka / W-band radar dual-frequency ratio output by the first learning model.

[0083] Step 303: input the Ka / W band radar dual frequency ratio, the X band reflectivity, and the Ka band reflectivity into a second learning model to obtain a raindrop mass weighted average diameter output by the second learning model.

[0084] Step 304: Input the X-band reflectivity and the Ka-band reflectivity into a third learning model to obtain the liquid water content output by the third learning model.

[0085] Step 305: determining the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content.

[0086] The present invention adopts an ensemble learning method to invert the raindrop spectrum, wherein the first learning model, the second learning model and the third learning model are obtained by using multiple X / Ka dual-band radar sample data to perform ensemble training on three initial models respectively.

[0087] Ensemble learning is a research area of ​​machine learning. It improves prediction performance and generalization ability by building and combining multiple learners. When a single machine learning model is used for classification or regression tasks, its performance is often difficult to meet the requirements due to the size of the training data and the number of model parameters. The ensemble learning method integrates multiple base learners and makes category judgments or regression predictions through different fusion strategies, which has better robustness. Compared with deep learning, the ensemble learning model is smaller in size and can speed up training and reasoning through parallel computing, so it has broad application prospects.

[0088] In order to more clearly illustrate the implementation process and feasibility of the dual-band radar raindrop spectrum inversion method proposed in the present invention, a specific embodiment is described in detail below.

[0089] This embodiment can be implemented with the aid of an X / Ka dual-band radar system, which can simultaneously acquire radar observation data of the X-band (eg, 10 GHz) and the Ka-band (eg, 35 GHz).

[0090] This embodiment assumes that the first learning model, the second learning model and the third learning model have been fully trained using multiple X / Ka dual-band radar sample data through an integrated learning method, and have achieved the expected prediction accuracy and generalization ability. These learning models can accurately output prediction results such as the Ka / W band radar dual-frequency ratio, the weighted average diameter of raindrop mass and the liquid water content based on the input radar observation data. Therefore, in this embodiment, only the process of how to use the first learning model, the second learning model and the third learning model after the model training has been completed to realize the raindrop spectrum inversion is introduced.

[0091] First, the raindrop spectrum sample data within the current X / Ka dual-band radar detection range is collected in real time, and then the X-band reflectivity Z is obtained by using methods such as T matrix scattering simulation. x , Ka-band reflectivity Z ka and X / Ka band radar dual frequency ratio DFR (X, Ka).

[0092] These data can be preprocessed, including noise removal, outlier detection and elimination, data normalization, etc., to ensure the quality of the data.

[0093] The processed X-band reflectivity Z x , Ka-band reflectivity Z ka and X / Ka band radar dual frequency ratio DFR (X, Ka) are input into the trained first learning model, so that the first learning model can calculate the predicted value of Ka / W band radar dual frequency ratio DFR (Ka, W) according to these input data through its internal integrated learning mechanism.

[0094] The Ka / W band radar dual frequency ratio prediction value DFR (Ka, W) obtained in step 302 and the original X band reflectivity Z x , Ka-band reflectivity Z ka , as model inputs, are input into the second learning model, so that the second learning model can use these model inputs to perform operations through its integrated base learner (such as gradient boosting tree), and finally output the weighted average diameter D of raindrop mass. m The predicted value of .

[0095] Similarly, the original X-band reflectivity Z x , Ka-band reflectivity Z ka , are input together as model inputs into the third learning model, so that the third learning model can calculate the predicted value of liquid water content LWC based on these model inputs through its decision tree, support vector regression SVR or other integrated learning algorithms.

[0096] Finally, according to the raindrop mass weighted average diameter D obtained in step 303 and step 304, m The predicted value of liquid water content LWC is combined with the theoretical model or empirical formula of raindrop spectrum distribution to determine the raindrop spectrum distribution within the detection range. This distribution information has important application value in meteorological monitoring, weather forecasting, aviation safety assessment and other fields.

[0097] Among them, the most critical parameter for raindrop spectrum inversion is the standardized particle number concentration N W , can be calculated based on the liquid water content LWC and the raindrop mass-weighted average diameter D m To determine, the specific calculation formula can be:

[0098] (7)

[0099] in, is the water density of the raindrop.

[0100] It can be seen from the above embodiments that the dual-band radar raindrop spectrum inversion method provided by the present invention relies on a trained integrated learning model, and does not need to establish a physical inversion model of radar parameters and raindrop spectrum parameters. It can effectively, quickly and accurately extract the key parameters of the raindrop spectrum from the radar observation data, and effectively solves the DFR-D problem existing in the dual-band radar raindrop spectrum inversion. m Double solution problem, the inverted raindrop spectrum is more accurate.

[0101] Based on the content of the above embodiment, as an optional embodiment, the present invention provides a method for integrated training of the first learning model, the second learning model and the third learning model, including but not limited to:

[0102] Obtain raindrop spectrum sample data of the same X / Ka dual-band radar sample data, obtain X-band reflectivity samples, Ka-band reflectivity samples and X / Ka-band radar dual-frequency ratio samples, as well as Ka / W-band radar dual-frequency ratio labels, raindrop mass-weighted average diameter labels and liquid water content labels through T-matrix scattering simulation and rainfall characteristic parameter calculation;

[0103] The X-band reflectivity sample, the Ka-band reflectivity sample, the X / Ka-band radar dual-frequency ratio sample, and the Ka / W-band radar dual-frequency ratio label form a first training sample, and a plurality of the first training samples form a first training subset;

[0104] Using the X-band reflectivity samples, the Ka-band reflectivity samples, and the X / Ka-band radar dual-frequency ratio samples as inputs of a first learning model, obtaining the Ka / W-band radar dual-frequency ratio output by the first learning model, using the Ka / W-band radar dual-frequency ratio as a sample and the X-band reflectivity samples, the Ka-band reflectivity samples, and the raindrop mass-weighted average diameter label to form a second training sample, and a plurality of the second training samples constitute a second training subset;

[0105] The X-band reflectivity sample, the Ka-band reflectivity sample and the liquid water content label form a third training sample, and a plurality of the third training samples constitute a third training subset;

[0106] Constructing a centralized training set, wherein the centralized training set includes the first training subset, the second training subset, and the third training subset;

[0107] The first learning model, the second learning model and the third learning model are integratedly trained using the centralized training set.

[0108] Figure 4 It is a schematic diagram of the data collection process of the centralized training set provided by the present invention, such as Figure 4 As shown, before training the model, the present invention needs to construct a centralized training set (referred to as data set) first, mainly by extracting features from historically collected raindrop spectrum sample data to obtain rainfall characteristic parameters, such as the mass-weighted average diameter D m , liquid water content LWC and normalized particle number concentration N W etc., and use T matrix scattering to simulate the measured raindrop spectrum to obtain the X-band reflectivity Z X , Ka-band reflectivity Z Ka and W band reflectivity Z W , and then we can get the X / Ka band radar dual frequency ratio DFR(X,Ka) and Ka / W band radar dual frequency ratio DFR(Ka,W).

[0109] First, a large amount of raindrop spectrum sample data is collected from the historical database, which contains observation records during rainfall. Feature extraction is performed on the collected sample data to obtain key rainfall characteristic parameters, such as mass-weighted mean diameter Dm, liquid water content LWC, and standardized particle number concentration N W Etc. These parameters are important indicators for describing the characteristics of raindrop spectrum and are crucial for the training of subsequent learning models.

[0110] Among them, the weighted average diameter of raindrops is D m The calculation method of the liquid water content LWC is shown in formula (8)-formula (11), and the standardized particle number concentration N W The calculation expression is shown in formula (7) in the above embodiment.

[0111] (8)

[0112] (9)

[0113] (10)

[0114] in, D is the raindrop diameter, is the maximum diameter of a raindrop, is the minimum diameter of a raindrop, for n The definition of step distance, N(D) is the raindrop spectrum distribution.

[0115] Furthermore, the T-matrix scattering simulation technology is used to simulate the scattering characteristics of the measured raindrop spectrum. Through simulation, the X-band reflectivity Z corresponding to the raindrop spectrum can be obtained. X , Ka-band reflectivity Z Ka and W band reflectivity Z W .

[0116] Among them, the T-matrix scattering simulation solves the scattering problem of non-spherical particles by directly solving the Maxwell equations, linking the far-field scattered electromagnetic waves with the incident electromagnetic waves to obtain a 4×4 backscattering phase matrix (also called the Mueller matrix) and a 2×2 forward scattering amplitude matrix.

[0117] Radar reflectivity (mm 6 m -3 ) is defined as shown in formula (11):

[0118] (11)

[0119] Among them, λ is the wavelength, the wavelength of the X-band is 32mm, the wavelength of the Ka-band is 9mm, and the wavelength of the W-band is 3mm. mis the raindrop complex refraction index, σ H is the horizontal backscatter cross section.

[0120] The radar reflectivity is expressed by T matrix elements as shown in formula (12):

[0121] (12)

[0122] Where, 〈 〉 represents the overall volume average of the raindrop phase matrix elements, Indicates the total number of raindrops per cubic meter. , K =( m 2 -1) / ( m 2 +2), m is the raindrop complex refraction index, is the backscatter phase matrix, and the radar reflectivity factor is usually 10lg( Z ) is expressed in dBZ.

[0123] Accordingly, according to the simulated X-band reflectivity Z X , Ka-band reflectivity Z Ka and W band reflectivity Z W , calculate the X / Ka band radar dual frequency ratio DFR(X,Ka) and the Ka / W band radar dual frequency ratio DFR(Ka,W), and their expressions are shown in formula (13) and formula (14):

[0124] (13)

[0125] (14)

[0126] Then, the rainfall characteristic parameters (D m , LWC, N W etc.), simulated radar reflectivity (Z X , Z Ka , Z W ) and the calculated radar dual-frequency ratio DFR(X,Ka), DFR(Ka,W)) are combined as samples with the corresponding true labels (such as the true value of the Ka / W band radar dual-frequency ratio, the true value of the mass-weighted average diameter of raindrops, and the true value of liquid water content) to form a complete training sample.

[0127] Divide multiple such training samples according to certain rules (such as time sequence, location distribution, etc.) to construct a centralized training set. This centralized training set will be used for subsequent integrated training of the first learning model, the second learning model, and the third learning model.

[0128] Figure 5 It is a training flow chart of the first learning model provided by the present invention. The first learning model can be named DFR (Ka, W) learning model according to its output, or simply recorded as mdl_dfr_ka_w model.

[0129] Before training the first learning model, the X-band reflectivity samples, Ka-band reflectivity samples and X / Ka-band radar dual-frequency ratio samples obtained after feature extraction from raindrop spectrum sample data of N (N>1) X / Ka dual-band radar sample data are used as the data input of the model, and the Ka / W-band radar dual-frequency ratio label is used as the data output, so that the first training subset constructed by N first training samples can be obtained, and then the first learning model is centrally trained through the first training subset.

[0130] Figure 6 is a training flow chart of the second learning model provided by the present invention, such as Figure 6 As shown, the second learning model can be named D according to its output m Learning model, or simply mdl_D m Model.

[0131] It should be noted that the Ka / W band radar dual frequency ratio samples used as the training of the second learning model are output after the first learning model recognizes the input X band reflectivity samples, Ka band reflectivity samples and the X / Ka band radar dual frequency ratio samples. The corresponding raindrop mass weighted average diameter label is obtained after the second learning model recognizes the X band reflectivity, Ka band reflectivity in the same raindrop spectrum sample data and the Ka / W band radar dual frequency ratio output by the first learning model as sample input.

[0132] In this way, before training the second learning model, N second training samples consisting of the Ka / W band radar dual frequency ratio as a sample and the X-band reflectivity sample, the Ka-band reflectivity sample and the raindrop mass weighted average diameter label can be collected to construct a second training subset. Finally, the second training subset can be used to complete the training of the second learning model.

[0133] Figure 7 is a training flow chart of the third learning model provided by the present invention, such as Figure 7As shown, similar to the training of the first learning model, when the present invention trains the third learning model, it will collect X-band reflectivity samples and Ka-band reflectivity samples related to the raindrop spectrum sample data of N historical X / Ka dual-band radar sample data, and together with the related liquid water content label LWC, form N third training samples, and construct a third training subset. Finally, the third training subset is used to complete the training of the third learning model.

[0134] The dual-band radar raindrop spectrum inversion method provided by the present invention combines the extracted characteristic parameters, the simulated radar reflectivity, and the calculated radar dual-frequency ratio with the real label to construct a centralized training set, which can ensure that the learning goal of the model during the training process is clear and help evaluate the prediction performance of the model. At the same time, through the integrated training method, the advantages of different base learners can be fully utilized to reduce the overfitting or underfitting problems that may exist in a single model. At the same time, by adjusting the training strategy and parameter settings, the prediction accuracy and generalization ability of the model can be further improved.

[0135] Figure 8 is one of the flow charts of the integrated training of any learning model provided by the present invention, such as Figure 8 As shown, for any learning model among the first learning model, the second learning model and the third learning model, this embodiment provides a method for performing integrated training on these learning models using a centralized training set (hereinafter referred to as training method 1), mainly including but not limited to:

[0136] Step 1.1, selecting a decision tree as a base learner, a plurality of the base learners constitute the learning model, setting the weight of each training sample in each of the centralized training sets to be equal, and the weight of the training sample is used to determine the probability of selecting the training sample from the centralized training set for executing this round of model training;

[0137] Step 1.2, using the centralized training set to train all the base learners, and determining a weighted prediction error for each of the training samples, wherein the weighted prediction error is determined based on the prediction result of each of the base learners for the training samples and the weight of each of the base learners, and the weight of the base learners is comprehensively determined based on their historical prediction errors;

[0138] Step 1.3, adjusting the weight of each training sample in the centralized training set according to the weighted prediction error positive feedback;

[0139] Step 1.4, iteratively execute steps 1.2 to 1.3 until the pre-training of the learning model is completed.

[0140] As an optional embodiment, the raindrop spectrum sample data is the observation data on May 11, 2022 and June 11, 2022, with a time resolution of 1 minute, a total of 2880 raindrop spectrum samples, and the variable D obtained by the calculation and T matrix scattering simulation is m , LWC, N W 、DFR(X,Ka)、Z X , Z Ka , Z W , and DFR (Ka, W) are both 2880.

[0141] First, the 2880 variables are divided into training sets and test sets. To ensure that there is no mismatch between the training sets and the test sets when training and building the first, second and third learning models, they are divided according to the time series, and every 3 samples are selected as the training set data, and the last sample is selected as the test set data. In this way, 75% of the samples (2160) are used as training sets and 25% of the samples (720) are used as test sets.

[0142] The training set data is trained using an integrated learning method to obtain the learning models mdl_dfr_ka_w, mdl_D m and mdl_LWC.

[0143] Then, the constructed learning model is used to predict the test set data, and the prediction effect is analyzed and evaluated.

[0144] For ease of description, the following embodiments will be described by taking the training of the second learning model as an example.

[0145] In the first training iteration, the weights of all training samples in the training set are initialized to equal values, usually 1 / N, where N is the total number of samples. This ensures that all samples have an equal impact on the model at the beginning of training.

[0146] In this embodiment, although self-service sampling of training samples is not involved, the effect of self-service sampling can be simulated by adjusting the weight of each training sample in each round of iteration.

[0147] The base learner is usually a weak learner, which may be a decision tree or a linear model, etc. In the radar raindrop spectrum inversion provided by the present invention, the decision tree is preferably selected as the base learner because it can handle nonlinear relationships and is easy to explain.

[0148] In each iteration, a base learner is trained according to the current sample weights. The goal of the base learner is to fit the current weighted training data as much as possible, that is, the goal of the base learner is to maximize the weighted accuracy, or minimize the weighted prediction error.

[0149] Evaluate the performance of the base learners on the centralized training set, and calculate the weighted prediction error of all base learners for each training sample, and then adjust the weight of each training sample in the centralized training set according to the weighted prediction error positive feedback. Specifically, the training samples that are predicted incorrectly will receive higher weights in the next round of iterations, while the weights of the training samples that are predicted correctly will be reduced. The greater the weight of the training sample, the greater the probability of being selected from the centralized training set for the next round of model training. In this way, the subsequent base learners will pay more attention to those training samples that were predicted incorrectly in the previous round.

[0150] As an optional embodiment, a base learner can be gradually added and the training sample weights can be adjusted to eventually obtain an integrated model consisting of multiple base learners. During the integration process, the prediction results of each base learner will be weighted and combined according to its performance (such as weight) to obtain the final prediction result.

[0151] The performance of the ensemble model can be evaluated using an independent validation set or test set. By comparing the differences between the model's predicted values ​​and the measured values, the model's accuracy, robustness, and generalization ability can be evaluated.

[0152] After each iteration, the training parameters (such as the number, type, and learning rate of base learners) can be further adjusted according to the performance of the second learning model on the validation set or test set, and iterative optimization can be performed to improve the prediction performance of the model. Finally, an integrated learning model can be effectively constructed to improve the accuracy of dual-band radar raindrop spectrum inversion. The obtained integrated model can significantly improve the robustness and accuracy of the model by combining the prediction results of multiple base learners.

[0153] Of course, in this embodiment, each base learner is independently trained on a training subset of the centralized training set, and these subsets are obtained by bootstrap sampling. Of course, the above embodiment provides a method of adjusting the weights of training samples by increasing the base learners, which does not mean that the number of base learners cannot be fixed in advance. In practical applications, the number of base learners can be pre-set according to the needs of specific problems, the limitations of computing resources, and the consideration of model performance. For example, in some cases, if it is known that a certain number of base learners is sufficient to achieve the required model performance, or in order to control the complexity and computational cost of the model, it may be chosen to pre-fix the number of base learners.

[0154] The dual-band radar raindrop spectrum inversion method provided by the present invention can effectively construct an integrated learning model to improve the accuracy of dual-band radar raindrop spectrum inversion. This integrated model can significantly improve the robustness and accuracy of the model by combining the prediction results of multiple base learners.

[0155] On this basis, the output of the learning model is determined by weighted majority voting or weighted averaging of the output results of all the base learners. That is, after the present invention completes the training of the first learning model, the second learning model and the third learning model in the above manner, in the actual application of the model, the integration is reflected in that the final prediction result is jointly determined by multiple base learners, which can improve the robustness and accuracy of the model.

[0156] Fig. 9 This is the second flow chart of the integrated training of any learning model provided by the present invention, such as Fig. 9 As shown, as an optional embodiment, the present invention also provides another method for integrated training of the first learning model, the second learning model and the third learning model (hereinafter referred to as training method 2), mainly including but not limited to:

[0157] Step 2.1, selecting a decision tree as a base learner, and a plurality of the base learners constitute the learning model;

[0158] Step 2.2, self-sampling a plurality of training samples from the centralized training set to obtain a plurality of sub-training sets, and using the plurality of sub-training sets to independently train to obtain a plurality of the base learners;

[0159] In step 2.3, multiple base learners are integrated and trained by voting method.

[0160] The following takes the implementation in the dual-band radar raindrop spectrum inversion method as an example to describe the entire learning model training process in detail. Also take the construction of the second learning model as an example.

[0161] Based on the method provided in the above embodiment, a centralized training set is first constructed, including: collecting raindrop spectrum sample data of multiple X / Ka dual-band radar sample data, including X-band reflectivity, Ka-band reflectivity, and X / Ka-band radar dual-frequency ratio, etc., and using them as samples. Then, features are extracted from the collected raindrop spectrum sample data, such as the weighted average diameter of raindrop mass, etc., which will be used as labels for the second learning model.

[0162] It should be noted that for the second learning model, the label of the training sample during training is D m The actual measured value, the corresponding input is the X-band reflectivity Z X , Ka-band reflectivity Z ka , Ka / W band radar dual frequency ratio , while the Ka / W band radar dual frequency ratio It is output by the first learning model.

[0163] Multiple bootstrap sampling (with replacement) is performed from the centralized training set to generate multiple training subsets. The size of each training subset is the same as the original dataset, but because of the replacement sampling, there will be overlaps between the training subsets, and there will also be unselected training samples. Generally speaking, the size of each training subset is the same as the original dataset, but it contains repeated and omitted samples to ensure the differences between the training subsets.

[0164] The base learners are usually models of the same type, which may be decision trees, linear models, etc. In the radar raindrop spectrum inversion in this embodiment, decision trees are mainly selected as the base learners.

[0165] For each training subset obtained by bootstrapping, a base learner is trained independently and in parallel. Since the training is independent, it can be processed in parallel to speed up the training. The final learning model is obtained by integrating the prediction results of all base learners according to the integration strategy (voting or averaging).

[0166] The final prediction result is obtained by voting (for classification problems) or averaging (for regression problems).

[0167] In the above two embodiments, two integrated learning methods for training learning models are provided, mainly evaluating the accuracy of prediction results through root mean square error (RMSE) and correlation coefficient (CC), and RMSE and CC are defined as follows:

[0168] (15)

[0169] (16)

[0170] in, For the i The predicted value of the training samples, For the i The true value of the training samples, For the i The covariance between the predicted value and the true value of the training samples, For the i The variance of the predicted values ​​of training samples is For the i The true value variance of the training samples.

[0171] The smaller the RMSE value, the smaller the deviation between the predicted value and the true value; CC indicates the correlation between the predicted value and the true value, and the closer it is to 1, the higher the correlation.

[0172] The evaluation results are shown in Table 1. From the evaluation results, it can be seen that the prediction results of the two ensemble learning methods have high consistency with the true value and small prediction error.

[0173] Table 1 Evaluation of learning results of centralized training set

[0174]

[0175] The dual-band radar raindrop spectrum inversion method provided by the present invention constructs multiple independent base learners and uses their prediction results to reduce the variance of the overall model, thereby improving the stability and generalization ability of the model. Specifically, multiple different training sets are generated through self-service sampling, so that each base learner can learn from slightly different data sets, thereby capturing different features in the data. The prediction results of these base learners are combined by voting or averaging, which can reduce the overfitting phenomenon that may be caused by a single learner and improve the prediction accuracy of the overall model.

[0176] Fig.10 : is a schematic diagram of the structure of the dual-band radar raindrop spectrum inversion device provided by the present invention, such as Fig.10 As shown, it mainly includes:

[0177] The feature extraction unit 111 is mainly used to determine the X-band reflectivity, Ka-band reflectivity and X / Ka-band radar dual-frequency ratio according to the X / Ka-band radar observation data within the detection range.

[0178] The first identification unit 112 is used to input the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio into a first learning model to obtain the Ka / W-band radar dual-frequency ratio output by the first learning model.

[0179] The second identification unit 113 inputs the Ka / W band radar dual frequency ratio, the X band reflectivity and the Ka band reflectivity into a second learning model to obtain a raindrop mass weighted average diameter output by the second learning model.

[0180] The third recognition unit 114 inputs the X-band reflectivity and the Ka-band reflectivity into a third learning model to obtain the liquid water content output by the third learning model.

[0181] The result output unit 115 is used to determine the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content.

[0182] The first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

[0183] It should be noted that the dual-band radar raindrop spectrum inversion device provided by the present invention can execute the dual-band radar raindrop spectrum inversion method described in any of the above embodiments during specific operation, which will not be described in detail in this embodiment.

[0184] The dual-band radar raindrop spectrum inversion device provided by the present invention relies on a trained integrated learning model, and does not need to establish a physical inversion model of radar parameters and raindrop spectrum parameters. It can effectively, quickly and accurately extract the key parameters of raindrop spectrum from radar observation data, and effectively solves the DFR-D problem existing in the dual-band radar raindrop spectrum inversion. m Double solution problem, the inverted raindrop spectrum is more accurate.

[0185] The following will be combined Figure 11-Figure 24 The feasibility of the dual-band radar raindrop spectrum inversion method and device provided by the present invention is elaborated in detail.

[0186] Fig.11 It is a schematic diagram comparing the predicted value and the true value predicted by the first learning model provided by the present invention for the training set, wherein the horizontal axis is the serial number of the training sample, and the vertical axis represents the Ka / W band radar dual frequency ratio DFR (Ka, W) predicted by the model under the certain training sample.

[0187] Fig.12 : is a schematic diagram of the comparison between the predicted value and the true value of the training set predicted by the second learning model provided by the present invention, wherein the horizontal axis is the sequence number of the training sample, and the vertical axis represents the weighted average diameter D of the raindrop mass predicted by the model under the training sample. m .

[0188] Fig.13 It is a schematic diagram of the comparison between the predicted value and the true value of the training set predicted by the second learning model provided by the present invention under DFR (X, Ka), and the ordinate is D m The horizontal axis is the X / Ka band radar dual frequency ratio DFR (X, Ka).

[0189] Fig.14 It is a schematic diagram comparing the predicted value and the true value predicted by the third learning model provided by the present invention for the training set, wherein the horizontal axis is the serial number of the training sample, and the vertical axis represents the liquid water content LWC predicted by the model under the certain training sample.

[0190] Combination Figure 11-Figure 14 As shown in the figure, the output results of the first learning model, the second learning model and the third learning model predicted in the centralized training set are compared with the true value. Figure 12-13 It can be seen that by integrating learning DFR-D m The double solution problem is solved, and the accurate D can be given when DFR≤0m , such as Fig.13 Shown in the box area.

[0191] The LWC output by the learning model and D m Substituting the predicted value of into formula (7) we can directly obtain N W Predicted value.

[0192] Fig.15 is the standardized particle number concentration N calculated by the present invention w The diagram of the comparison between the predicted value and the true value, where the horizontal axis is the serial number of the training sample, and the vertical axis represents the standardized particle number concentration N calculated under a certain training sample. W The predicted value of .

[0193] Fig.16 It is a schematic diagram comparing the predicted value and true value of DFR (Ka, W) obtained by predicting the test set data using the first learning model provided by the present invention, wherein the horizontal axis is the serial number of the training sample, and the vertical axis represents the predicted value and true value of the Ka / W band radar dual frequency ratio DFR (Ka, W) obtained under the certain training sample.

[0194] Fig.17 D is obtained by predicting the test set data using the second learning model provided by the present invention. m The diagram of the comparison between the predicted value and the true value, where the horizontal axis is the serial number of the training sample, and the vertical axis represents the weighted average diameter D of the raindrop mass under a certain training sample. m The predicted value and true value.

[0195] Fig.18 D is obtained by predicting the test set data using the second learning model provided by the present invention. m The comparison diagram under DFR (X, Ka), the horizontal axis is the X / Ka band radar dual frequency ratio DFR (X, Ka), and the vertical axis is D m .

[0196] Fig.19 It is a schematic diagram comparing the predicted value and true value of LWC obtained by predicting the test set data using the third learning model provided in the present invention, wherein the horizontal axis is the serial number of the training sample, and the vertical axis represents the predicted value and true value of the liquid water content LWC under the certain training sample.

[0197] Fig. 20 : is a schematic diagram of the comparison between the predicted value of Nw and the true value calculated by the method provided by the present invention for the test set data, wherein the horizontal axis is the sequence number of the training sample, and the vertical axis represents the standardized particle number concentration N under the training sample. w The predicted value and true value.

[0198] like Figure 16-Figure 20 The following are the DFR (Ka, W), D m , LWC and N W The comparison between the predicted value and the true value is Fig.18 It can be seen that the integrated learning method can also give the accurate D when DFR≤0 in the test set data m , solving the DFR-D m Double solution problem.

[0199] Table 2 is an evaluation of the prediction results of the test set. It can be seen from Table 2 that both ensemble learning methods can predict accurate variables, and the predicted values ​​have good consistency with the true values ​​and a small prediction error.

[0200] Table 2 Evaluation of prediction results of test set

[0201]

[0202] Fig.21 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 51.6 dBZ provided by the present invention, Fig. 22 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 41.5 dBZ provided by the present invention, Fig.23 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) under the reflectivity of 31.6 dBZ provided by the present invention, Fig.24 It is a schematic diagram of the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) at a reflectivity of 21.7 dBZ provided by the present invention.

[0203] like Figure 21-24 As shown in the figure, the inverted raindrop spectrum (DSD-Pred) and the measured raindrop spectrum (DSD-True) are compared under different reflectivity factors obtained by substituting the predicted variables into the standard Gamma spectral distribution function. The inverted raindrop spectrum shows a high consistency with the measured raindrop spectrum.

[0204] Fig.25 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Fig.25As shown, the electronic device may include: a processor 2510, a communications interface 2520, a memory 2530 and a communications bus 2540, wherein the processor 2510, the communications interface 2520 and the memory 2530 communicate with each other through the communications bus 2540. The processor 2510 may call the logic instructions in the memory 2530 to execute the dual-band radar raindrop spectrum inversion method, the method comprising: determining the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio according to the X / Ka dual-band radar observation data within the detection range; inputting the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio into a first learning model to obtain the Ka / W-band radar dual-frequency ratio output by the first learning model; inputting the Ka / W-band radar dual-frequency ratio, the X-band reflectivity and the Ka / W-band radar dual-frequency ratio into a first learning model to obtain the Ka / W-band radar dual-frequency ratio; The a-band reflectivity is input into the second learning model to obtain the weighted average diameter of raindrops output by the second learning model; the X-band reflectivity and the Ka-band reflectivity are input into the third learning model to obtain the liquid water content output by the third learning model; the raindrop spectrum distribution in the detection range is determined according to the weighted average diameter of raindrops and the liquid water content; the first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

[0205] In addition, the logic instructions in the above-mentioned memory 2530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0206] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, and when the program instructions are executed by a computer, the computer can execute the dual-band radar raindrop spectrum inversion method provided in the above-mentioned embodiments, the method comprising: determining the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio according to the X / Ka dual-band radar observation data within the detection range; inputting the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio into a first learning model, and obtaining the Ka / Ka-band reflectivity output by the first learning model. W-band radar dual-frequency ratio; inputting the Ka / W-band radar dual-frequency ratio, the X-band reflectivity and the Ka-band reflectivity into a second learning model to obtain a raindrop mass-weighted average diameter output by the second learning model; inputting the X-band reflectivity and the Ka-band reflectivity into a third learning model to obtain a liquid water content output by the third learning model; determining the raindrop spectrum distribution within the detection range according to the raindrop mass-weighted average diameter and the liquid water content; the first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using a plurality of X / Ka dual-band radar sample data.

[0207] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the computer program is implemented to perform the following steps provided in the above embodiments: determining an X-band reflectivity, a Ka-band reflectivity, and an X / Ka-band radar dual-frequency ratio according to X / Ka dual-band radar observation data within a detection range; inputting the X-band reflectivity, the Ka-band reflectivity, and the X / Ka-band radar dual-frequency ratio into a first learning model to obtain a Ka / W-band radar dual-frequency ratio output by the first learning model; and inputting the Ka / W-band radar dual-frequency ratio, the X-band reflectivity, the Ka-band reflectivity, and the X / Ka-band radar dual-frequency ratio into a first learning model to obtain a Ka / W-band radar dual-frequency ratio output by the first learning model. The X-band reflectivity and the Ka-band reflectivity are input into a second learning model to obtain a raindrop mass-weighted average diameter output by the second learning model; the X-band reflectivity and the Ka-band reflectivity are input into a third learning model to obtain a liquid water content output by the third learning model; according to the raindrop mass-weighted average diameter and the liquid water content, the raindrop spectrum distribution in the detection range is determined; the first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

[0208] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0209] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dual-band radar raindrop spectrum inversion method, characterized in that: include: Determine the X-band reflectivity, Ka-band reflectivity and X / Ka-band radar dual-frequency ratio based on X / Ka-band dual-band radar observation data within the detection range; Inputting the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio into a first learning model, and obtaining the Ka / W-band radar dual-frequency ratio output by the first learning model; Inputting the Ka / W band radar dual frequency ratio, the X band reflectivity and the Ka band reflectivity into a second learning model, and obtaining a raindrop mass weighted average diameter output by the second learning model; Inputting the X-band reflectivity and the Ka-band reflectivity into a third learning model to obtain the liquid water content output by the third learning model; Determining the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content; The first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

2. The dual-band radar raindrop spectrum inversion method according to claim 1, characterized in that: The method further includes performing integrated training on the first learning model, the second learning model and the third learning model, specifically: Obtain X-band reflectivity samples, Ka-band reflectivity samples and X / Ka-band radar dual-frequency ratio samples from raindrop spectrum sample data of any X / Ka dual-band radar sample data, as well as Ka / W-band radar dual-frequency ratio labels, raindrop mass-weighted average diameter labels and liquid water content labels; The X-band reflectivity sample, the Ka-band reflectivity sample, the X / Ka-band radar dual-frequency ratio sample, and the Ka / W-band radar dual-frequency ratio label form a first training sample, and a plurality of the first training samples form a first training subset; Using the X-band reflectivity samples, the Ka-band reflectivity samples, and the X / Ka-band radar dual-frequency ratio samples as inputs of a first learning model, obtaining the Ka / W-band radar dual-frequency ratio output by the first learning model, using the Ka / W-band radar dual-frequency ratio as a sample and the X-band reflectivity samples, the Ka-band reflectivity samples, and the raindrop mass-weighted average diameter label to form a second training sample, and a plurality of the second training samples constitute a second training subset; The X-band reflectivity sample, the Ka-band reflectivity sample and the liquid water content label form a third training sample, and a plurality of the third training samples constitute a third training subset; Constructing a centralized training set, wherein the centralized training set includes the first training subset, the second training subset, and the third training subset; The first learning model, the second learning model and the third learning model are integratedly trained using the centralized training set.

3. The dual-band radar raindrop spectrum inversion method according to claim 2, characterized in that: For any one of the first learning model, the second learning model and the third learning model, using the centralized training set to perform integrated training on the learning model includes: Step 1.1, selecting a decision tree as a base learner, a plurality of the base learners constitute the learning model, setting the weight of each training sample in each of the centralized training sets to be equal, and the weight of the training sample is used to determine the probability of selecting the training sample from the centralized training set for executing this round of model training; Step 1.2, using the centralized training set to train all the base learners, and determining a weighted prediction error for each of the training samples, wherein the weighted prediction error is determined based on the prediction result of each of the base learners for the training samples and the weight of each of the base learners, and the weight of the base learners is comprehensively determined based on their historical prediction errors; Step 1.3, adjusting the weight of each training sample in the centralized training set according to the weighted prediction error positive feedback; Step 1.4, iteratively execute steps 1.2 to 1.3 until the pre-training of the learning model is completed.

4. The dual-band radar raindrop spectrum inversion method according to claim 3, characterized in that: The output of the learning model is determined based on weighted majority voting or weighted averaging of the output results of all the base learners.

5. The dual-band radar raindrop spectrum inversion method according to claim 2, characterized in that: For any one of the first learning model, the second learning model and the third learning model, using the centralized training set to perform integrated training on the learning model includes: Step 2.1, selecting a decision tree as a base learner, and a plurality of the base learners constitute the learning model; Step 2.2, self-sampling a plurality of training samples from the centralized training set to obtain a plurality of sub-training sets, and using the plurality of sub-training sets to independently train to obtain a plurality of the base learners; In step 2.3, multiple base learners are integrated and trained by voting method.

6. The dual-band radar raindrop spectrum inversion method according to any one of claims 2 to 5, characterized in that: The raindrop mass weighted average diameter label and liquid water content label in the raindrop spectrum sample data of any X / Ka dual-band radar sample data are obtained by moment calculation, and the specific calculation expression is: ； ； ； in, D is the raindrop diameter; is the maximum diameter of raindrops; is the minimum diameter of a raindrop; for n Definition of step distance; N(D) is the raindrop spectrum distribution; is the mass-weighted mean diameter of raindrops; is the liquid water content.

7. A dual-band radar raindrop spectrum inversion device, characterized in that: include: A feature extraction unit is used to determine the X-band reflectivity, the Ka-band reflectivity and the X / Ka-band radar dual-frequency ratio according to the X / Ka-band radar observation data within the detection range; a first identification unit, configured to input the X-band reflectivity, the Ka-band reflectivity, and the X / Ka-band radar dual-frequency ratio into a first learning model, and obtain the Ka / W-band radar dual-frequency ratio output by the first learning model; a second identification unit, inputting the Ka / W band radar dual frequency ratio, the X band reflectivity and the Ka band reflectivity into a second learning model, and acquiring a raindrop mass weighted average diameter output by the second learning model; a third recognition unit, inputting the X-band reflectivity and the Ka-band reflectivity into a third learning model, and acquiring the liquid water content output by the third learning model; A result output unit, used to determine the raindrop spectrum distribution in the detection range according to the raindrop mass-weighted average diameter and the liquid water content; The first learning model, the second learning model and the third learning model are obtained by integrating and training three initial models respectively using multiple X / Ka dual-band radar sample data.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the dual-band radar raindrop spectrum inversion method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the dual-band radar raindrop spectrum inversion method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the dual-band radar raindrop spectrum inversion method according to any one of claims 1 to 6 is implemented.

Images