Multi-source meteorological collaborative plateau rainfall fusion method and system

Abstract

The invention discloses a plateau rainfall fusion method and a plateau rainfall fusion system based on multi-source meteorological collaboration, and relates to the technical field of plateau rainfall fusion methods, and the method comprises the steps: obtaining a multi-source rainfall product, a meteorological covariable and high-resolution digital elevation model data, carrying out the time-space alignment and normalization processing of the multi-source rainfall product and the meteorological covariable, and obtaining the high-resolution digital elevation model data; forming a multi-source meteorological data set; based on physical variables in the multi-source meteorological data set, constructing four physical process links of water vapor transportation, terrain uplift, convection triggering and cloud microphysics, generating attention weight vectors through sub attention branches corresponding to the links, obtaining a causal attention weight matrix after weighted fusion, and performing weighted fusion on multi-source rainfall products to obtain a multi-source rainfall data set; generating a causal fusion feature map; wavelet decomposition is carried out on the causal fusion feature map and meteorological variables, large-scale components and small-scale components are separated out, and meanwhile the terrain uplift rate and the included angle between the wind direction and the slope direction are calculated according to the digital elevation model data and the wind field components.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of highland precipitation fusion methods, in particular to a multi-source meteorological collaborative highland precipitation fusion method and system. BACKGROUND

[0002] The highland precipitation fusion method is a technology that comprehensively utilizes data from multiple sources (such as satellite precipitation products, reanalysis data, digital elevation models, etc.) and technical means (including machine learning algorithms, physical modeling, etc.) to accurately estimate precipitation in the special geographical environment of the highlands. Therefore, how to use advanced technical means to improve the intelligent level and security of the highland precipitation fusion method has become one of the current problems to be solved.

[0003] In the field of highland precipitation fusion methods, traditional precipitation products differ significantly in spatial and temporal resolution, accuracy, and coverage due to different sources, making it difficult to directly perform effective fusion processing. Previous methods ignored key physical processes such as water vapor transport, terrain lifting, and convective triggering that are closely related to precipitation formation, affecting the physical reasonableness and accuracy of the results. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a multi-source meteorological collaborative highland precipitation fusion method to solve the problem that traditional precipitation products differ significantly in spatial and temporal resolution, accuracy, and coverage due to different sources, making it difficult to directly perform effective fusion processing, and previous methods ignored key physical processes such as water vapor transport, terrain lifting, and convective triggering that are closely related to precipitation formation, affecting the physical reasonableness and accuracy of the results.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a multi-source meteorological collaborative highland precipitation fusion method, which includes:

[0008] Obtaining multi-source precipitation products, meteorological covariates, and high-resolution digital elevation model data, performing spatio-temporal alignment and normalization processing on the multi-source precipitation products and meteorological covariates to form a multi-source meteorological data set;

[0009] Based on the physical variables in the multi-source meteorological data set, four physical process links of water vapor transport, terrain lifting, convective triggering, and cloud microphysics are constructed, attention weight vectors are generated through corresponding sub-attention branches of each link, a causal attention weight matrix is obtained after weighted fusion, and the multi-source precipitation products are weighted and fused to generate a causal fusion feature map;

[0010] wavelet decomposition is performed on the causal fusion feature map and the meteorological variable to separate large-scale components and small-scale components, and a terrain modulation factor is generated by inputting a terrain modulation rate and an angle between a wind direction and a slope direction calculated according to the digital elevation model data and the wind field component into a terrain modulator;

[0011] The small-scale components are dynamically modulated by using the terrain modulation factor to obtain modulated small-scale features, and the modulated small-scale features are fused with the large-scale components to generate a terrain modulation fusion field;

[0012] The region containing ground observations is divided into a source domain, and the region without observations is divided into a target domain, a domain classifier and a gradient reversal layer are arranged at the end of the fusion network, a physical consistency constraint loss based on water vapor conservation is introduced into the target domain, a multi-task joint loss function including a supervised loss, a domain adversarial loss and a physical consistency constraint loss is constructed, and the network is trained to obtain a plateau precipitation fusion model;

[0013] The multi-source meteorological data set is input into the plateau precipitation fusion model, and a plateau precipitation fusion product is output.

[0014] As a preferred scheme of the multi-source meteorological collaborative plateau precipitation fusion method, wherein: the multi-source precipitation products, meteorological covariants and high-resolution digital elevation model data are obtained, the multi-source precipitation products and meteorological covariants are spatio-temporally aligned and normalized to form a multi-source meteorological data set, and the specific steps are as follows:

[0015] GPMIMERG, CMORPH, PERSIANN-CCS, ERA5-Land and CHM_PRE five types of precipitation products are obtained, and five types of meteorological covariants such as 2-meter temperature, 2-meter dew point temperature, 10-meter wind speed u component and v component, and sea level pressure are synchronously obtained, and digital elevation model data with a spatial resolution of R are obtained;

[0016] The five types of precipitation products and the five types of meteorological covariants are spatio-temporally aligned, and all data are resampled to a unified spatial grid G and a time step T;

[0017] The aligned variable sequences are normalized to eliminate dimensional differences, and a multi-source meteorological data set with consistent structure is formed.

[0018] As a preferred scheme of the multi-source meteorological collaborative plateau precipitation fusion method, wherein: four physical process links of water vapor transport, terrain lifting, convective triggering and cloud microphysics are constructed based on the physical variables in the multi-source meteorological data set, an attention weight vector is generated through a corresponding sub-attention branch of each link, a causal attention weight matrix is obtained after weighted fusion, and multi-source precipitation products are weighted and fused to generate a causal fusion feature map, and the specific steps are as follows:

[0019] A water vapor transport link is constructed based on the precipitable water q, wind field components u and v in the multi-source weather data set, and the angle θ between the wind direction and the terrain slope direction and the terrain gradient A terrain lifting link is constructed, a convection triggering link is constructed with the lifting index l, and a cloud microphysical link is constructed with the cloud top temperature c;

[0020] For the water vapor transport link, q, u, and v are input into a first fully connected network, and a first attention weight vector w1 is output;

[0021] For the terrain lifting link, θ, is input into a second fully connected network, and a second attention weight vector w2 is output;

[0022] For the convection triggering link, l is input into a third fully connected network, and a third attention weight vector w3 is output;

[0023] For the cloud microphysical link, c is input into a fourth fully connected network, and a fourth attention weight vector w4 is output;

[0024] w1, w2, w3, and w4 are weighted and fused to obtain a comprehensive attention weight vector w c , and the calculation process is as follows:

[0025] w c =σ(w1+w2+w3+w4);

[0026] Where σ represents the Sigmoid activation function;

[0027] The five types of precipitation products are weighted and fused using w c to generate a causal fusion feature map f c .

[0028] As a preferred scheme of the multi-source weather collaborative plateau precipitation fusion method described in the present application, wherein: the causal fusion feature map and the weather variables are subjected to wavelet decomposition to separate large-scale components and small-scale components, and the terrain lifting rate and the angle between the wind direction and the slope direction are calculated according to the digital elevation model data and the wind field components, and input into a terrain modulator to generate a terrain modulation factor, and the specific steps are as follows:

[0029] The causal fusion feature map f c , the wind field components u and v, and the specific humidity q are subjected to wavelet decomposition, and the Daubechies wavelet basis is used for multi-scale decomposition to separate large-scale components with a spatial scale greater than L1 u L

[0030] The wind speed direction vector d=(u,v) is calculated according to the wind field components u and v, and the terrain gradient direction Calculate the angle between the wind direction and the slope direction

[0031] Calculate the terrain lifting rate w l , the expression is:

[0032]

[0033] θ and w l Input terrain modulator, output terrain modulation factor.

[0034] As a preferred scheme of the multi-source weather collaborative plateau precipitation fusion method of the application, wherein: the terrain modulation factor is used to dynamically modulate the small-scale component to obtain a modulated small-scale feature, and the modulated small-scale feature is fused with the large-scale component to generate a terrain modulation fusion field, and the specific steps are:

[0035] The field f t , the calculation method is:

[0036]

[0037] Wherein, α, β are learnable fusion weights.

[0038] As a preferred scheme of the multi-source weather collaborative plateau precipitation fusion method of the application, wherein: the region containing ground observation is divided into source domain, and the region without observation is divided into target domain, a domain classifier and a gradient reversal layer are set at the end of the fusion network, and a physical consistency constraint loss based on water vapor conservation is introduced in the target domain, a multi-task joint loss function containing supervision loss, domain adversarial loss and physical consistency constraint loss is constructed, the network is trained to obtain a plateau precipitation fusion model, and the specific steps are:

[0039] The research area is divided into source domain and target domain, wherein the source domain contains ground rain gauge observation points, and the target domain is a region without observation coverage;

[0040] A domain classifier is connected at the end of the fusion network, which takes the feature representation of the terrain modulation fusion field f t As input, and outputs the probability that the sample belongs to the source domain or the target domain;

[0041] A gradient reversal layer is set before the domain classifier, which multiplies the gradient by a negative coefficient when backpropagating, so that the network minimizes the domain classification accuracy when updating the parameters, forcing f t To be consistent in distribution in the source domain and the target domain.

[0042] As a preferred embodiment of the multi-source meteorological collaborative plateau precipitation fusion method of the present invention, the specific steps of inputting the multi-source meteorological dataset into the plateau precipitation fusion model and outputting the plateau precipitation fusion product are as follows:

[0043] Within the target domain, calculate the fused precipitation field f. t The resulting water vapor consumption q e Its value is equal to f t itself;

[0044] Calculate atmospheric water vapor flux convergence q c Its expression is:

[0045]

[0046] Where q is the specific humidity and v is the wind speed vector. Represents divergence operations;

[0047] Define the physical consistency constraint loss L p = / / q e -q c / / 2 Supervision loss L s f within the source domain t Mean square error of ground observations, domain adversarial loss L d The cross-entropy loss of the domain classifier;

[0048] Construct a multi-task joint loss function L = L s +γL d +ηL p Where γ and η are loss weight coefficients, the network is trained end-to-end to obtain the plateau precipitation fusion model;

[0049] The multi-source meteorological dataset is input into the plateau precipitation fusion model, and the plateau precipitation fusion product is output.

[0050] Secondly, the present invention provides a multi-source meteorological synergy plateau precipitation fusion system, comprising:

[0051] The module includes a data preprocessing module, a causal attention module, a terrain modulation module, a feature fusion module, a domain adaptive training module, and a product generation module.

[0052] The data preprocessing module is used to acquire multi-source precipitation products, meteorological covariates, and high-resolution digital elevation model data, and to perform spatiotemporal alignment and normalization processing on the multi-source precipitation products and meteorological covariates to form a multi-source meteorological dataset.

[0053] The cause-effect attention module is configured for constructing four physical process links of water vapor transport, terrain lifting, convection triggering and cloud microphysics based on the precipitable water content, the wind field component, the lifting index and the cloud top temperature in the multi-source meteorological data set, generating an attention weight vector through a corresponding sub-attention branch of each link, obtaining a cause-effect attention weight matrix after weighted fusion, and performing weighted fusion on the multi-source precipitation product to generate a cause-effect fusion feature map.

[0054] The terrain modulation module is configured for performing wavelet decomposition on the cause-effect fusion feature map and the wind field and specific humidity variable, separating out large-scale components and small-scale components, simultaneously calculating a terrain lifting rate and an angle between a wind direction and a slope direction based on digital elevation model data and the wind field component, and inputting the terrain modulation factor into a terrain modulator to generate a terrain modulation factor.

[0055] The feature fusion module is configured for performing pixel-by-pixel dynamic modulation on the small-scale components by using the terrain modulation factor to obtain modulated small-scale features, and performing weighted fusion on the modulated small-scale features and the large-scale components to generate a terrain modulation fusion field.

[0056] The domain adaptive training module is configured for dividing a region containing ground observations into a source domain and a region without observations into a target domain, setting a domain classifier and a gradient reversal layer at the end of the fusion network, introducing a physical consistency constraint loss based on water vapor conservation in the target domain, constructing a multi-task joint loss function containing a supervision loss, a domain adversarial loss and a physical consistency constraint loss, performing end-to-end training on the network, and obtaining a plateau precipitation fusion model.

[0057] The product generation module is configured for inputting the multi-source meteorological data set into the plateau precipitation fusion model and outputting a plateau precipitation fusion product.

[0058] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the multi-source meteorological collaborative plateau precipitation fusion method according to the first aspect of the present application is implemented.

[0059] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any step of the multi-source meteorological collaborative plateau precipitation fusion method according to the first aspect of the present application is implemented.

[0060] The application has the beneficial effects that: by constructing a physical process chain driven causal attention mechanism, the physical interpretability of the multi-source precipitation product fusion process is realized, the rationality of the precipitation estimation under the plateau complex weather system is improved, by introducing a terrain modulation module, the forcing effect of terrain uplift on precipitation is explicitly modeled, the spatial distribution depiction capability of precipitation on the windward slope and the leeward slope in the plateau complex terrain area is significantly enhanced, by designing an unsupervised domain adaptive training framework, the model training failure problem caused by the sparse ground observation in the plateau hinterland is effectively solved, the wide-area coverage, physically consistent and high-resolution precipitation fusion is realized, the method realizes collaborative innovation in the aspects of physical mechanism guidance, terrain response modeling and data sparsity adaptability, and the precision, reliability and application range of the precipitation fusion product in the plateau area are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0062] Figure 1 The flowchart of the multi-source meteorological collaborative plateau precipitation fusion method in embodiment 1.

[0063] Figure 2 The schematic diagram of the multi-source meteorological collaborative plateau precipitation fusion system in embodiment 1. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0065] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0066] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0067] Embodiment, refer to Figure 1 and Figure 2For the embodiment of the application, the embodiment provides a multi-source meteorological collaborative plateau precipitation fusion method, comprising the following steps:

[0068] S1, obtaining multi-source precipitation products, meteorological covariants and high-resolution digital elevation model data, performing spatio-temporal alignment and normalization processing on the multi-source precipitation products and meteorological covariants to form a multi-source meteorological data set;

[0069] Further, five types of precipitation products, GPM IMERG, CMORPH, PERSIANN-CCS, ERA5-Land and CHM_PRE, are obtained, and five types of meteorological covariants, 2-meter temperature, 2-meter dew point temperature, u component and v component of 10-meter wind speed, and sea level pressure, are obtained synchronously, and digital elevation model data with a spatial resolution of R are obtained;

[0070] The five types of precipitation products and the five types of meteorological covariants are subjected to spatio-temporal alignment processing, and all the data are resampled to a unified spatial grid G and a time step T;

[0071] The normalized each variable sequence is subjected to normalization processing to eliminate dimensional differences and form a multi-source meteorological data set with consistent structure;

[0072] It should be noted that the spatio-temporal alignment process realizes spatial resampling by a bilinear interpolation method, and time step alignment is completed by linear interpolation of time series, so as to ensure that each data source has comparability on the same geographical grid and time node, and the normalization processing adopts Z-score standardization by using the mean and standard deviation of each variable in the time series, so as to eliminate the interference of different physical dimensions on model training and improve the numerical stability of input data.

[0073] S2, based on the physical variables in the multi-source meteorological data set, four physical process links of water vapor transport, terrain lifting, convection triggering and cloud microphysics are constructed, attention weight vectors are generated through corresponding sub-attention branches of each link, a causal attention weight matrix is obtained after weighted fusion, and the multi-source precipitation products are subjected to weighted fusion to generate a causal fusion feature map;

[0074] Further, based on the precipitable water q, wind field components u and v in the multi-source meteorological data set, the water vapor transport link is constructed, the angle θ between the wind direction and the terrain slope direction and the terrain gradient The terrain lifting link is constructed, the convection triggering link is constructed by using the lifting index l, and the cloud microphysics link is constructed by using the cloud top temperature c;

[0075] For the water vapor transport link, q, u and v are input into the first fully connected network, and the first attention weight vector w1 is output;

[0076] For the terrain lifting link, θ, The input second full connection network outputs a second attention weight vector w2;

[0077] The input l is input into a third full connection network for the convection triggering link, and a third attention weight vector w3 is output;

[0078] The input c is input into a fourth full connection network for the cloud microphysical link, and a fourth attention weight vector w4 is output;

[0079] The w1, w2, w3 and w4 are weighted and fused to obtain a comprehensive attention weight vector w c , and the calculation process is as follows:

[0080] w c =σ(w1+w2+w3+w4);

[0081] wherein σ represents a Sigmoid activation function;

[0082] The w c is used to perform weighted fusion on the five types of precipitation products to generate a causal fusion feature map f c .

[0083] It should be noted that the full connection network in the sub-attention branch is composed of two layers of neurons, the activation function adopts ReLU, and the output layer is connected to the Sigmoid function to ensure that the attention weight is between 0 and 1. The input variables of each physical process link are directly obtained through physical formula derivation or reanalysis data before input, ensuring that their physical meaning is clear and there is a clear causal relationship with the precipitation formation process.

[0084] S3, wavelet decomposition is performed on the causal fusion feature map and the meteorological variables to separate the large-scale component and the small-scale component, and the terrain lifting rate and the angle between the wind direction and the slope direction are calculated according to the digital elevation model data and the wind field component, and input into the terrain modulator to generate a terrain modulation factor;

[0085] Further, the causal fusion feature map f c , the wind field components u and v, and the specific humidity q are wavelet decomposed, the Daubechies wavelet basis is used for multi-scale decomposition, and the large-scale component with a spatial scale greater than L1 is separated

[0086] The wind speed direction vector d=(u,v) is calculated according to the wind field components u and v, combined with the terrain gradient direction The angle between the wind direction and the slope direction is calculated

[0087] The terrain lifting rate w l is calculated, and its expression is as follows:

[0088]

[0089] θ is calculated by w l Input terrain modulator, output terrain modulation factor;

[0090] It should be noted that the wavelet decomposition is performed frame by frame in the two-dimensional spatial domain, and the selection of the Daubechies wavelet basis is based on its good time-frequency localization characteristics, which can effectively separate the large-scale background field and the small-scale terrain response signal in the precipitation field. The calculation result of the terrain lifting rate is used to quantify the vertical motion intensity of the wind field under the terrain forcing, which is the key physical quantity determining the rain enhancement effect on the windward slope.

[0091] S4, dynamically modulating the small-scale component by using the terrain modulation factor to obtain a modulated small-scale feature, and fusing the modulated small-scale feature with the large-scale component to generate a terrain modulation fusion field;

[0092] The fusion field f t , and the calculation method is:

[0093]

[0094] Wherein, α and β are learnable fusion weights;

[0095] It should be noted that the pixel-by-pixel multiplication modulation realizes the fine regulation and control of the terrain factor on the local precipitation characteristics, ensures that the spatial distribution of precipitation in complex terrain areas (such as valleys and ridges) is more physically reasonable, and the learnable fusion weights are automatically optimized through back propagation, so that the model can adaptively balance the contributions of large-scale circulation background and local terrain enhancement effect.

[0096] S5, dividing the area containing ground observations into a source domain and the area without observations into a target domain, setting a domain classifier and a gradient reversal layer at the end of the fusion network, introducing a physical consistency constraint loss based on water vapor conservation in the target domain, and constructing a multi-task joint loss function containing a supervised loss, a domain adversarial loss and a physical consistency constraint loss. Train the network to obtain a plateau precipitation fusion model;

[0097] Further, the research area is divided into a source domain and a target domain, wherein the source domain contains ground rain gauge observation points, and the target domain is an area without observation coverage;

[0098] A domain classifier is connected at the end of the fusion network, which takes the feature representation of the terrain modulation fusion field f t as input and outputs the probability that the sample belongs to the source domain or the target domain;

[0099] A gradient reversal layer is set before the domain classifier, which multiplies the gradient by a negative coefficient when back propagating, so that the network minimizes the domain classification accuracy when updating the parameters, forcing f tThe distribution is consistent on the source domain and the target domain;

[0100] It should be noted that the introduction of the gradient inversion layer makes the optimization direction of the domain classifier opposite to the feature extraction network, thereby forcing the network to extract domain-invariant features that are independent of the observation distribution, improving the generalization ability of the model in the unobserved area. The mechanism cooperates with the physical consistency constraint to ensure that the target domain fusion result remains physically self-consistent even in the absence of true value supervision.

[0101] S6, inputting the multi-source meteorological data set into the plateau precipitation fusion model to output a plateau precipitation fusion product;

[0102] Further, in the target domain, the fused precipitation field f t The water vapor consumption q e is equal to f t Itself;

[0103] The atmospheric water vapor flux convergence q c is calculated, and its expression is:

[0104]

[0105] Wherein, q is specific humidity, v is wind speed vector, Indicates the divergence operation;

[0106] The physical consistency constraint loss L p = / / q e -q c / 2 , the supervision loss L s is the mean square error of f t and the ground observation in the source domain, and the domain adversarial loss L d is the cross-entropy loss of the domain classifier;

[0107] A multi-task joint loss function L = L s + γL d + ηL p is constructed, wherein γ and η are loss weight coefficients, the network is trained end-to-end, and a plateau precipitation fusion model is obtained;

[0108] The multi-source meteorological data set is input into the plateau precipitation fusion model to output a plateau precipitation fusion product;

[0109] It should be noted that the physical consistency constraint loss participates in gradient update during model training and is used to constrain the water vapor balance in the target domain to prevent non-physical water vapor sources and sinks in the fusion result. The weight coefficients of the multi-task joint loss function are determined through the performance of the validation set to ensure that the supervision signal, domain adaptation target and physical law are optimally balanced.

[0110] The embodiment also provides a multi-source meteorological collaborative plateau precipitation fusion system, comprising:

[0111] a data preprocessing module, a causal attention module, a terrain modulation module, a feature fusion module, a domain adaptive training module and a product generation module;

[0112] The data preprocessing module is configured to acquire multi-source precipitation products, meteorological covariants and high-resolution digital elevation model data, perform spatio-temporal alignment and normalization processing on the multi-source precipitation products and the meteorological covariants, and form a multi-source meteorological data set.

[0113] The causal attention module is configured to construct four physical process links of water vapor transport, terrain lifting, convection triggering and cloud microphysics based on the precipitable water content, wind field component, lifting index and cloud top temperature in the multi-source meteorological data set, generate an attention weight vector through a sub-attention branch corresponding to each link, obtain a causal attention weight matrix after weighted fusion, and perform weighted fusion on the multi-source precipitation products to generate a causal fusion feature map.

[0114] The terrain modulation module is configured to perform wavelet decomposition on the causal fusion feature map and the wind field and specific humidity variables, separate large-scale components and small-scale components, calculate a terrain lifting rate and an angle between a wind direction and a slope direction based on the digital elevation model data and the wind field component, and input the terrain modulation module to generate a terrain modulation factor.

[0115] The feature fusion module is configured to use the terrain modulation factor to dynamically modulate the small-scale components pixel by pixel to obtain modulated small-scale features, and perform weighted fusion on the modulated small-scale features and the large-scale components to generate a terrain modulation fusion field.

[0116] The domain adaptive training module is configured to divide a region containing ground observations as a source domain and a region without observations as a target domain, set a domain classifier and a gradient reversal layer at the end of the fusion network, introduce a physical consistency constraint loss based on water vapor conservation in the target domain, construct a multi-task joint loss function comprising a supervision loss, a domain adversarial loss and a physical consistency constraint loss, perform end-to-end training on the network, and obtain a plateau precipitation fusion model.

[0117] The product generation module is configured to input the multi-source meteorological data set into the plateau precipitation fusion model and output a plateau precipitation fusion product.

[0118] The embodiment also provides a computer device suitable for the multi-source meteorological collaborative plateau precipitation fusion method, comprising a memory and a processor.

[0119] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, a carrier network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the shell of the computer device, or an external keyboard, a touchpad or a mouse, etc.

[0120] The embodiment also provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the highland precipitation fusion method for realizing multi-source meteorological collaboration provided in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0121] In conclusion, the application realizes the physical interpretability of the multi-source precipitation product fusion process by constructing a physical process chain driven causal attention mechanism, improves the rationality of precipitation estimation under the plateau complex weather system, introduces a terrain modulation module to explicitly model the forcing effect of terrain uplift on precipitation, significantly enhances the ability to depict the spatial distribution of precipitation on the windward slope and leeward slope in the complex terrain area of the plateau, and effectively solves the problem of model training failure caused by sparse ground observation in the plateau hinterland by designing an unsupervised domain adaptive training framework, realizing wide-area coverage, physically consistent and high-resolution precipitation fusion. The method realizes collaborative innovation in the aspects of physical mechanism guidance, terrain response modeling and data sparsity adaptability, significantly improves the precision, reliability and application range of the precipitation fusion product in the plateau area.

[0122] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for fusion of plateau precipitation data from multiple meteorological sources, characterized in that: include: Acquire multi-source precipitation products, meteorological covariates, and high-resolution digital elevation model data; perform spatiotemporal alignment and normalization on the multi-source precipitation products and meteorological covariates to form a multi-source meteorological dataset. Based on the physical variables in the multi-source meteorological dataset, four physical process links are constructed: water vapor transport, orographic lifting, convection triggering, and cloud microphysics. Attention weight vectors are generated through the sub-attention branches corresponding to each link. After weighted fusion, a causal attention weight matrix is ​​obtained. Multi-source precipitation products are then weighted and fused to generate a causal fusion feature map. Wavelet decomposition is performed on the causal fusion feature map and meteorological variables to separate large-scale and small-scale components. At the same time, the terrain lifting rate and the angle between wind direction and slope aspect are calculated based on the digital elevation model data and wind field components, and then input into the terrain modulator to generate terrain modulation factor. The small-scale component is dynamically modulated using the terrain modulation factor to obtain the modulated small-scale features, which are then fused with the large-scale component to generate a terrain modulation fusion field. The region containing ground observations is divided into the source domain and the region without observations is the target domain. A domain classifier and a gradient reversal layer are set at the end of the fusion network. A physical consistency constraint loss based on water vapor conservation is introduced into the target domain. A multi-task joint loss function including supervision loss, domain adversarial loss and physical consistency constraint loss is constructed. The network is trained to obtain the plateau precipitation fusion model. The multi-source meteorological dataset is input into the plateau precipitation fusion model, and the plateau precipitation fusion product is output.

2. The plateau precipitation fusion method based on multi-source meteorological synergy as described in claim 1, characterized in that: The steps for acquiring multi-source precipitation products, meteorological covariates, and high-resolution digital elevation model data, and performing spatiotemporal alignment and normalization on the multi-source precipitation products and meteorological covariates to form a multi-source meteorological dataset are as follows: Five types of precipitation products were acquired: GPM IMERG, CMORPH, PERSIANN-CCS, ERA5-Land, and CHM_PRE. Simultaneously, five types of meteorological covariates were acquired: 2-meter temperature, 2-meter dew point temperature, 10-meter wind speed (u and v components), and sea level pressure. Digital elevation model data with a spatial resolution of R were also acquired. The five types of precipitation products and five types of meteorological covariates were spatiotemporally aligned, and all data were resampled to a unified spatial grid G ​​and time step T. The aligned variable sequences are normalized to eliminate dimensional differences and form a multi-source meteorological dataset with consistent structure.

3. The plateau precipitation fusion method based on multi-source meteorological synergy as described in claim 2, characterized in that: The process involves constructing four physical processes—moisture transport, orographic lifting, convection triggering, and cloud microphysics—based on the physical variables in the multi-source meteorological dataset. Attention weight vectors are generated through sub-attention branches corresponding to each process. These vectors are then weighted and fused to obtain a causal attention weight matrix. Finally, multi-source precipitation products are weighted and fused to generate a causal fusion feature map. The specific steps are as follows: Based on the precipitable water q, wind field components u and v in the multi-source meteorological dataset, a water vapor transport link is constructed, using the angle θ between wind direction and terrain slope aspect and the terrain gradient. Construct a topographic lifting stage, use the lifting index l to construct a convection trigger stage, and use the cloud top temperature c to construct a cloud microphysics stage; For the water vapor transport process, q, u, and v are input into the first fully connected network, and the first attention weight vector w1 is output. For the terrain uplift stage, θ, Input the second fully connected network and output the second attention weight vector w2; For the convection triggering stage, l is input into the third fully connected network, and the third attention weight vector w3 is output. For the cloud microphysics component, input c into the fourth fully connected network and output the fourth attention weight vector w4; We then weight and fuse w1, w2, w3, and w4 to obtain the comprehensive attention weight vector w. c The calculation process is as follows: w c =σ(w1+w2+w3+w4); Where σ represents the Sigmoid activation function; Using w c Weighted fusion of five types of precipitation products generates a causal fusion feature map f. c .

4. The plateau precipitation fusion method based on multi-source meteorological coordination as described in claim 3, characterized in that: The steps involve performing wavelet decomposition on the causal fusion feature map and meteorological variables to separate large-scale and small-scale components, and simultaneously calculating the terrain uplift rate and the wind direction / slope aspect angle based on the digital elevation model data and wind field components. These are then input into a terrain modulator to generate a terrain modulation factor. For the causal fusion feature map f c The wind field components u and v, and specific humidity q are decomposed using wavelet decomposition. Then, the Daubechies wavelet basis is used for multi-scale decomposition to separate large-scale components with spatial scales greater than L1. u L v L q L Small-scale components with a spatial scale smaller than L2 u S v S q S ; Calculate terrain gradient based on digital elevation model data in h represents the terrain height; The wind speed direction vector d = (u, v) is calculated based on the wind field components u and v, and then combined with the terrain gradient direction. Calculate the angle between wind direction and slope aspect Calculate the rate of terrain uplift w l Its expression is: Let θ and w l Input the terrain modulator, output the terrain modulation factor.

5. The plateau precipitation fusion method based on multi-source meteorological coordination as described in claim 4, characterized in that: The specific steps are as follows: Dynamically modulate the small-scale component using the terrain modulation factor to obtain modulated small-scale features, and then fuse these features with the large-scale component to generate a terrain modulation fusion field. Using terrain modulation factors for small-scale components u S v S q S Pixel-wise multiplication modulation is performed to obtain the modulated small-scale features. u S′ v S′ q S′ ; Modulated small-scale features With large-scale components Perform weighted fusion to generate a terrain modulation fusion field f t The calculation method is as follows: Where α and β are learnable fusion weights.

6. The plateau precipitation fusion method based on multi-source meteorological coordination as described in claim 5, characterized in that: The region containing ground observations is designated as the source domain, and the region without observations is designated as the target domain. A domain classifier and a gradient reversal layer are set at the end of the fusion network. A physical consistency constraint loss based on water vapor conservation is introduced into the target domain. A multi-task joint loss function including supervised loss, domain adversarial loss, and physical consistency constraint loss is constructed. The network is then trained to obtain the plateau precipitation fusion model. The specific steps are as follows: The study area was divided into a source domain and a target domain, where the source domain included ground rain gauge observation points and the target domain was the area without observation coverage. A domain classifier is accessed at the end of the fusion network. This classifier uses terrain-modulated fusion field f t The features are represented as inputs, and the output is the probability that a sample belongs to the source domain or the target domain. A gradient inversion layer is placed before the domain classifier. This layer multiplies the gradient by a negative coefficient during backpropagation, causing the network to minimize the domain classification accuracy when updating parameters, thus forcing f... t The distribution is consistent across the source and target domains.

7. The plateau precipitation fusion method based on multi-source meteorological synergy as described in claim 6, characterized in that: The specific steps for inputting the multi-source meteorological dataset into the plateau precipitation fusion model and outputting the plateau precipitation fusion product are as follows: Within the target domain, calculate the fused precipitation field f. t The resulting water vapor consumption q e Its value is equal to f t itself; Calculate atmospheric water vapor flux convergence q c Its expression is: Where q is the specific humidity and v is the wind speed vector. • Represents divergence operation; Define the physical consistency constraint loss L p = / / q e -q c / / 2 Supervision loss L s f within the source domain t Mean square error of ground observations, domain adversarial loss L d The cross-entropy loss of the domain classifier; Construct a multi-task joint loss function L = L s +γL d +ηL p Where γ and η are loss weight coefficients, the network is trained end-to-end to obtain the plateau precipitation fusion model; The multi-source meteorological dataset is input into the plateau precipitation fusion model, and the plateau precipitation fusion product is output.

8. A multi-source meteorological synergy plateau precipitation fusion system, based on the multi-source meteorological synergy plateau precipitation fusion method according to any one of claims 1 to 7, characterized in that: include: The module includes a data preprocessing module, a causal attention module, a terrain modulation module, a feature fusion module, a domain adaptive training module, and a product generation module. The data preprocessing module is used to acquire multi-source precipitation products, meteorological covariates, and high-resolution digital elevation model data, and to perform spatiotemporal alignment and normalization processing on the multi-source precipitation products and meteorological covariates to form a multi-source meteorological dataset. The causal attention module is used to construct four physical process links based on the precipitable water, wind field component, lifting index and cloud top temperature in the multi-source meteorological dataset: water vapor transport, orographic lifting, convection triggering and cloud microphysics. It generates attention weight vectors through the sub-attention branches corresponding to each link, and obtains the causal attention weight matrix after weighted fusion. It also performs weighted fusion on the multi-source precipitation products to generate a causal fusion feature map. The terrain modulation module is used to perform wavelet decomposition on the causal fusion feature map and wind field and specific humidity variables to separate large-scale components and small-scale components. At the same time, it calculates the terrain lifting rate and the angle between wind direction and slope aspect based on digital elevation model data and wind field components, and inputs them into the terrain modulator to generate terrain modulation factors. The feature fusion module is used to dynamically modulate the small-scale component pixel by pixel using the terrain modulation factor to obtain the modulated small-scale features, and then perform weighted fusion with the large-scale component to generate a terrain modulation fusion field. The domain adaptive training module is used to divide the region containing ground observations into the source domain and the region without observations into the target domain. A domain classifier and gradient reversal layer are set at the end of the fusion network, and a physical consistency constraint loss based on water vapor conservation is introduced into the target domain. A multi-task joint loss function including supervision loss, domain adversarial loss and physical consistency constraint loss is constructed. The network is trained end-to-end to obtain the plateau precipitation fusion model. The product generation module is used to input multi-source meteorological datasets into the plateau precipitation fusion model and output plateau precipitation fusion products.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the plateau precipitation fusion method with multi-source meteorological coordination as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the plateau precipitation fusion method with multi-source meteorological coordination as described in any one of claims 1 to 7.

Images