Multi-modal radar echo extrapolation method based on Doppler dual-polarization weather radar

Abstract

The invention discloses a multi-mode radar echo extrapolation method based on a Doppler dual-polarization weather radar, and belongs to the technical field of meteorology. The method specifically comprises the following steps: firstly, acquiring Doppler dual-polarization weather-based data, and converting the Doppler dual-polarization weather-based data into multi-modal physical parameter data; then, constructing a multi-modal radar echo extrapolation model comprising a multi-modal encoder, a depth feature extraction module and a decoder; the encoder is composed of a double-branch structure, performs preliminary feature extraction on original multi-modal information, inputs the extracted features to the depth feature extraction module, further extracts depth spatial-temporal features through a self-attention mechanism, and finally, completes layer-by-layer decoding of the depth spatial-temporal features through the decoder, and introduces preliminary features to strengthen transmission of shallow feature information; and inputting newly observed radar information into the trained echo extrapolation model, and predicting a radar echo image at a future moment. According to the invention, powerful technical support is provided for real-time meteorological monitoring, rainfall forecast and extreme weather early warning.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological technologies, and particularly relates to a multimodal radar echo extrapolation method based on a Doppler dual-polarization weather radar. Background Art

[0002] Doppler weather radar is an advanced active remote sensing detection device, which is widely used in the fields of atmospheric environment monitoring and weather forecasting. Its basic principle is based on the Doppler effect. By transmitting electromagnetic waves and receiving the reflected signals, it can measure the movement speed of scatterers such as clouds, rain, and other meteorological elements relative to the radar in real time. Under certain conditions, using the phase and frequency changes of the echo signals, the radar can invert the distribution of the atmospheric wind field, the vertical velocity change of the air flow, and the local turbulence state, providing a scientific basis for monitoring and warning severe convective weather.

[0003] Traditional radar echo extrapolation methods mainly include the cross-correlation method, the centroid method of single particles, and the optical flow method. These methods rely on the correlation between signals or motion characteristics to predict the distribution of future echoes. However, due to the limitations of the algorithms themselves under complex meteorological conditions, their prediction accuracy and applicability are difficult to meet the actual needs.

[0004] In recent years, with the significant improvement in the performance of computer hardware (such as GPUs and TPUs), deep learning technologies have developed rapidly and shown superior performance to traditional methods in many fields such as image processing and spatio-temporal sequence prediction. Under this background, data-driven radar echo extrapolation models have become one of the current research hotspots and optimal solutions.

[0005] Existing research usually simplifies the radar echo extrapolation problem into a spatio-temporal sequence prediction sub-problem of deep learning, focusing on the modeling and characterization of the motion characteristics of radar echoes themselves. Although such models have achieved significant breakthroughs in short-term prediction accuracy, however, since the evolution of radar echoes is not only affected by internal dynamic processes but also closely related to other physical parameters. Most current models fail to fully integrate this additional multimodal physical parameter information, resulting in insufficient prediction capabilities when facing complex weather systems.

[0006] Through the extrapolation prediction of unobserved areas, global monitoring of large-scale weather conditions can be achieved, thereby more accurately understanding the evolution of weather systems, monitoring precipitation distribution and wind field changes, and improving the early warning ability for extreme weather. Constructing a deep learning-based model, by fully exploring and integrating historical multi-physical parameter information, not only is expected to improve the prediction accuracy of future radar echo conditions, but also provides more accurate data support for the real-time monitoring and forecasting of ground precipitation information.

[0007] However, in the existing technology, there is no method that can effectively utilize the multi-modal physical parameter information of weather radar to optimize the extrapolation effect of radar echoes. Therefore, supported by deep learning algorithms and multi-source data fusion technologies, developing a new type of radar echo extrapolation method and system has important theoretical significance and broad application prospects. Summary of the Invention

[0008] Aiming at the problem in the existing technology that only single radar echo information is relied on for extrapolation prediction, resulting in insufficient prediction accuracy and failure to fully integrate various physical parameter information, the present invention discloses a multi-modal radar echo extrapolation method based on Doppler dual-polarization weather radar, which fully integrates various physical parameter information obtained by the radar system, effectively captures the complex dynamic characteristics in the evolution process of radar echoes, significantly improves the prediction accuracy and the robustness of the system, and provides strong data support for large-scale weather system monitoring and extreme weather warning.

[0009] The multi-modal radar echo extrapolation method based on Doppler dual-polarization weather radar includes the following steps:

[0010] Step 1: Collect basic data based on Doppler dual-polarization weather radar, and through an inversion algorithm, uniformly process the data and convert it into multi-modal physical parameter data;

[0011] The multi-modal physical parameter data includes horizontal polarization radar reflectivity Z H , differential radar reflectivity Z DR , specific differential phase K DP , co-polar correlation coefficient ρ HV and velocity spectrum width σ V etc.

[0012] Step 2: Construct a multi-modal radar echo extrapolation model of Doppler dual-polarization weather radar and train it using multi-modal data;

[0013] The multi-modal radar echo extrapolation model includes a multi-modal encoder, a deep feature extraction module and a decoder;

[0014] Among them, the encoder consists of a dual-branch structure, which preliminarily extracts features from the original multi-modal information and inputs them into the deep feature extraction module. Through the self-attention mechanism, it further extracts the deep spatio-temporal features of the multi-modal. Finally, through the decoder, it completes the layer-by-layer decoding of the deep spatio-temporal features, and introduces the preliminary features in the last layer of the decoder to strengthen the transmission of shallow feature information, and finally obtains the extrapolation result of the radar echo image;

[0015] The two core branches of the multi-modal encoder are: the inter-space feature extraction branch and the intra-space feature extraction branch.

[0016] The formula of the inter - spatial feature extraction branch is as follows:

[0017]

[0018] Among them, represents the feature of the d - th modality at the i - th (1 ≤ i ≤ 4) layer, represents the d - th (1 ≤ d ≤ 5) original modality information; SiLU represents the non - linear activation function, Conv2d represents the two - dimensional convolution operation, and σ(i) represents the number of strides of the convolution structure. If i is odd, σ(i) takes 1, and if i is even, σ(i) takes 2. M skip represents the shallow preliminary feature of the first - layer Z H information to be passed to the decoder; RFF represents the inter - factor feature extracted by the inter - spatial feature extraction branch in the multi - modal encoder. represents the feature of the 4th layer of the first modality, and Concat represents feature concatenation in the channel dimension.

[0019] The formula of the intra - spatial feature extraction branch is as follows:

[0020] E 0 = Concat(Z H , Z DR , K DP , ρ HV , σ V )

[0021]

[0022] E 0 represents the information after stacking all modality information in the channel dimension. Finally, when i (1 ≤ i ≤ 4) reaches the maximum value of 4, E 4 represents the intra - factor feature AFF extracted by the intra - spatial feature extraction branch.

[0023] Finally, the inter - factor feature RFF extracted by the inter - spatial feature extraction branch and the intra - factor feature AFF extracted by the intra - spatial feature extraction branch are stacked in the channel dimension to generate the shallow aggregation feature FF.

[0024] The formula of the deep feature extraction module is as follows:

[0025] y i' = y i + Att(BN(y i ))

[0026] y i+1 = MLP(BN(y i' )) + y i'

[0027] Att(y) = PWConv(GSSA(Mish(PWConv(y))))

[0028] MLP(y) = PWConv(DWConv(Mish(PWConv(y))))

[0029] When the initial value of i is 1, y 1 is the shallow aggregation feature FF extracted from the Multimodal Encoder, and this module is stacked L layers in total; 1 ≤ i ≤ 1 - L.

[0030] Att(y) represents the attention formula, PWConv represents the Point - wise convolution with a kernel size of 1, GSSA represents the gated self - attention unit, Mish represents the self - regular non - monotonic neural activation function, MLP(y) represents the multi - layer perceptron formula, DWConv represents the depth - wise separable convolution, y i represents the spatio - temporal feature of the i - th layer, and BN represents the BatchNorm function.

[0031] The formula of the gated spatio - temporal self - attention unit GSSA is as follows:

[0032]

[0033] Among them, DWDConv represents the dilated depth convolution, y input represents the input feature, y map represents the attention map, y output represents the output feature.

[0034] Finally, the output deep spatio - temporal feature of the deep feature extraction module is y L .

[0035] The decoding module is responsible for integrating the shallow preliminary features passed from the multi - modal encoder and the deep spatio - temporal features extracted by the deep feature extraction module; through layer - by - layer decoding, it generates the final result with high - precision spatio - temporal extrapolation ability.

[0036] The data flow formula is as follows,

[0037]

[0038] Among them, and represent the features after the first and second upsamplings, represents the feature after convolution after the first upsampling. ConvTranspose2d represents the transposed convolution, and forecast is the final prediction result of the radar echo data.

[0039] Step 3: For the newly observed radar physical information, perform the same data preprocessing operations and input it into the trained multi-modal radar echo extrapolation model to obtain the radar echo prediction image at future times.

[0040] The advantages of the present invention are as follows:

[0041] By making full use of multi-modal physical parameter data and deep learning technology, the present invention breaks through the limitations of traditional single data source extrapolation methods and provides strong technical support for real-time meteorological monitoring, precipitation forecasting, and extreme weather warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flowchart of a multi-modal radar echo extrapolation method based on a Doppler dual-polarization weather radar according to the present invention;

[0043] Figure 2 is a schematic diagram of constructing a multi-modal radar echo extrapolation model of a Doppler dual-polarization weather radar according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0044] To facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail below with reference to the drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0045] The present invention discloses a multi-modal radar echo extrapolation method based on a Doppler dual-polarization weather radar. First, basic data is collected based on the Doppler dual-polarization weather radar, and quality control is performed on the radar receiving system data. Various physical parameter data is uniformly processed into multi-modal data at specified altitude levels according to the radar inversion algorithm. Secondly, a multi-modal radar echo extrapolation model based on deep learning is constructed and trained based on multi-modal grayscale images. Finally, real-time collected data of different physical parameters is input into the model through the same inversion algorithm to extrapolate the radar echo image at future times. By making full use of the unique advantages of the dual-polarization radar system and integrating multiple physical parameter information, the present invention can extract key information from historical data, obtain better radar echo extrapolation results, and thus significantly improve the accuracy and effect of echo extrapolation.

[0046] As Figure 1 shown, it includes the following steps:

[0047] Step 1: Collect basic data based on the Doppler dual-polarization weather radar, and uniformly process the data through an inversion algorithm to convert it into multi-modal physical parameter data;

[0048] First, use a Doppler dual-polarization weather radar to perform real-time scanning on precipitation, cloud clusters, and other meteorological targets in the atmosphere, and collect original basic data.

[0049] Subsequently, uniformly process the collected original basic data through a preset inversion algorithm, and convert the original data into structured multi-modal physical parameter data.

[0050] The weather radar data receiving system receives radar physical parameter information of size H×W for D at fixed time intervals, including Z H (horizontal polarization radar reflectivity), Z DR (differential radar reflectivity), K DP (specific differential phase), ρ HV (co-polar correlation coefficient), and σ V (velocity spectrum width), etc. The radar physical observation data at time t is denoted as P t ∈R D×H×W , and through unified data processing by the inversion algorithm, it is transformed into X t ∈R D×H×W multi-modal data.

[0051] Step 2: Construct a multi-modal radar echo extrapolation model for the Doppler dual-polarization weather radar, and use multi-modal data for training;

[0052] The multi-modal radar echo extrapolation model includes a multi-modal encoder, a deep feature extraction module, and a decoder;

[0053] As Figure 2 shown, the multi-modal data is respectively sent into a two-branch encoder for independent feature extraction. Each encoder consists of a convolution and downsampling module to extract shallow spatio-temporal features of different modalities. Then, the features extracted by the first branch and the features extracted by the second branch are stacked in the channel dimension to generate aggregated features. Subsequently, through a stacked spatio-temporal self-attention mechanism, continuously learn the continuous spatio-temporal dynamic relationship between multi-modal features to obtain deep spatio-temporal features. Finally, continuously convolve and upsample the deep spatio-temporal features for decoding to obtain the radar echo prediction image at the future moment.

[0054] The two core branches of the multi-modal encoder are: the inter-space feature extraction branch and the intra-space feature extraction branch.

[0055] The inter-space feature extraction branch focuses on cross-modal global feature interaction, generates unified inter-factor features through feature stacking to deeply explore the collaborative relationship between different modalities, and enhances the fusion effect of cross-modal information;

[0056] The spatial feature extraction branch focuses on the deep coupling relationships within the multimodality, generating in-factor features to achieve a refined representation of the internal information of the multimodality. And the features extracted by the spatial feature extraction branch for the first time are directly subjected to a feature stacking operation with the decoder module. This module integrates various physical parameters and, based on the dual-branch information fusion strategy, realizes the preliminary extraction and fusion of multimodal features, aiming to efficiently capture and integrate the key information in the multimodal data.

[0057] First, the formula of the spatial feature extraction branch of the multimodal encoder is as follows.

[0058]

[0059] Among them, represents the feature of the d-th modality at the i-th (1 ≤ i ≤ 4) layer, represents the d-th (1 ≤ d ≤ 5) original modality information; SiLU represents the non-linear activation function, Conv2d represents the two-dimensional convolution operation, σ(i) represents the number of stride steps of the convolution structure. If i is odd, σ(i) takes 1, and if i is even, σ(i) takes 2. M skip represents the shallow preliminary feature of the first layer Z H information to be passed to the decoder; RFF represents the inter-factor features extracted by the spatial feature extraction branch in the multimodal encoder. represents the feature of the 4th layer of the first modality, and Concat represents feature concatenation in the channel dimension.

[0060] Secondly, the formula of the spatial in-factor feature extraction branch of the multimodal encoder is as follows.

[0061] E 0 = Concat(Z H , Z DR , K DP , ρ HV , σ V )

[0062]

[0063] E 0 represents the information after stacking all the modality information in the channel dimension. Finally, when i (1 ≤ i ≤ 4) reaches the maximum value of 4, E 4 represents the in-factor feature AFF extracted by the spatial in-factor feature extraction branch.

[0064] Finally, the inter-factor feature RFF extracted by the spatial feature extraction branch and the in-factor feature AFF extracted by the spatial in-factor feature extraction branch are stacked in the channel dimension to generate the shallow aggregation feature FF.

[0065] The deep feature extraction module deeply learns the continuous spatio-temporal dynamic relationship between multi-modal features through a stacked spatio-temporal self-attention mechanism, strengthening the model's ability to model spatio-temporal dependencies. The formula is as follows:

[0066] y i' = y i + Att(BN(y i ))

[0067] y i+1 = MLP(BN(y i' )) + y i'

[0068] Att(y) = PWConv(GSSA(Mish(PWConv(y))))

[0069] MLP(y) = PWConv(DWConv(Mish(PWConv(y))))

[0070] When the initial value of i is 1, y 1 is the shallow aggregated feature FF extracted from the Multimodal Encoder. This module is stacked L layers; 1 ≤ i ≤ 1 - L. Att(y) represents the attention formula, PWConv represents the point-wise convolution with a kernel size of 1, GSSA represents the gated self-attention unit, Mish represents the self-regular non-monotonic neural activation function, MLP(y) represents the multi-layer perceptron formula, DWConv represents the depthwise separable convolution, and y i represents the deep spatio-temporal feature of the i-th layer, and BN represents the BatchNorm function.

[0071] A gated spatio-temporal self-attention unit (Gatedspatiotemporal self-attention, GSSA) is designed for the complex multi-modal spatio-temporal features in this task, aiming to further extract deep features effective for the extrapolation task. The formula is as follows:

[0072]

[0073] Among them, DWDConv represents the dilated depth convolution, y input represents the input feature, y map represents the attention map, and y output represents the output feature.

[0074] Finally, the output feature of the deep feature extraction module is y L .

[0075] The decoding module is responsible for integrating the shallow radar features transmitted from the multi-modal encoder and the multi-modal features extracted by the deep feature extraction module. Through layer-by-layer decoding, this module generates the final result with high-precision spatio-temporal extrapolation ability. The overall architecture design effectively combines the shallow and deep features of multi-modal information, aiming to improve the prediction performance of the model in complex dynamic environments. The data flow formula is as follows,

[0076]

[0077] where, and represent the features after the first and second upsamplings, represents the feature after convolution after the first upsampling. ConvTranspose2d represents transposed convolution, and forecast is the final prediction result of the radar echo data.

[0078] Step 3: For the real-time observed radar physical information, perform the same data preprocessing operations and input them into the trained multi-modal radar echo extrapolation model to obtain the prediction result of the future radar echo state at the current moment.

[0079] For the current observation time t, the input of the multi-modal radar echo extrapolation model is represents the observation data of the past T historical moments, is the radar echo data predicted for the next T' moments.

[0080] The present invention uses a variety of physical parameter data such as reflectivity, differential reflectivity, correlation coefficient, and differential phase collected by a Doppler dual-polarization weather radar. Through preprocessing and feature extraction, the multi-modal data is converted into input features suitable for deep learning modeling. Then, a multi-modal radar echo extrapolation model based on deep learning is constructed to achieve high-precision extrapolation of the radar echo image at future moments. By fully integrating various physical parameter information obtained by the radar system, this method effectively captures the complex dynamic characteristics in the evolution process of radar echoes, significantly improves the prediction accuracy and system robustness, and provides strong data support for large-scale weather system monitoring and extreme weather warning.

[0081] Finally, it should be noted that: Although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of what is claimed, but are mainly used to describe the features of specific embodiments of a particular invention. Certain features described in multiple embodiments within this specification can also be implemented in combination in a single embodiment. On the other hand, the various features described in a single embodiment can also be implemented separately in multiple embodiments or in any suitable sub-combination. Additionally, although features may operate in certain combinations as described above and were even initially claimed as such, one or more features from the claimed combination can in some cases be removed from that combination, and the claimed combination can be directed to a sub-combination or a variation of a sub-combination.

[0082] Similarly, although operations are depicted in the drawings in a particular order, this should not be construed as requiring that these operations be performed in the particular order shown or sequentially, or that all illustrated operations be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Additionally, the separation of the various system modules and components in the above embodiments should not be construed as required in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0083] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the acts recited in the claims can be performed in a different order and still achieve the desired result. Additionally, the processes depicted in the drawings are not necessarily in the particular order or sequential order shown to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0084] As described above, only the preferred specific implementation manners of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope of the conclusion of the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A multimodal radar echo extrapolation method based on Doppler dual-polarization weather radar, characterized in that, It includes the following steps: First, based on the Doppler dual-polarization weather radar, basic data is collected, and through an inversion algorithm, the data is uniformly processed and transformed into multi-modal physical parameter data; Then, a multi-modal radar echo extrapolation model of the Doppler dual-polarization weather radar is constructed and trained using multi-modal data; For the radar physical information observed in real time, after it is transformed into multi-modal physical parameter data, it is input into the trained multi-modal radar echo extrapolation model to obtain a radar echo prediction image at a future time; The multi-modal radar echo extrapolation model includes a multi-modal encoder, a deep feature extraction module, and a decoder; Among them, the encoder consists of a dual-branch structure: an inter-space feature extraction branch and an intra-space feature extraction branch. Preliminary feature extraction is performed on the original multi-modal physical parameter data, and the inter-factor feature RFF extracted by the inter-space feature extraction branch and the intra-factor feature AFF extracted by the intra-space feature extraction branch are superimposed in the channel dimension to generate a shallow aggregation feature FF and input it into the deep feature extraction module; The deep feature extraction module further extracts the deep spatio-temporal features y of multi-modal through the self-attention mechanism L ; Finally, the deep spatio-temporal features are decoded layer by layer through the decoder, and the preliminary feature M is introduced in the last layer of the decoder skip , to strengthen the transmission of shallow feature information, and finally obtain the extrapolation result of the radar echo image.

2. The multimodal radar echo extrapolation method based on Doppler dual-polarization weather radar according to claim 1, characterized in that The multi-modal physical parameter data includes horizontal polarization radar reflectivity Z H , differential radar reflectivity Z DR , specific differential phase K DP , co-polar correlation coefficient ρ HV and velocity spectrum width σ V .

3. A multi-modal radar echo extrapolation method based on a Doppler dual-polarization weather radar according to claim 2, characterized in that The formula of the inter-space feature extraction branch is as follows, Represents the feature of the d-th modality at the i-th layer, where 1 ≤ i ≤ 4; Represents the d-th original modality information, where 1 ≤ d ≤ 5; SiLU represents the non-linear activation function, Conv2d represents the two-dimensional convolution operation, and σ(i) represents the number of stride steps of the convolution structure. If i is odd, σ(i) is taken as 1, and if i is even, σ(i) is taken as 2; M skip Represents the first layer Z to be passed to the decoder H Shallow preliminary features of the information; RFF represents the inter-factor features extracted by the inter-space feature extraction branch in the multi-modal encoder; represents the features of the 4th layer of the first modality, and Concat represents feature concatenation in the channel dimension; The formula of the intra-space feature extraction branch is as follows: E 0 = Concat(Z H , Z DR , K DP , ρ HV , σ V ) E 0 represents the information after stacking all modal information in the channel dimension. Finally, when i is the maximum value of 4, E 4 represents the factor-internal feature AFF extracted by the branch of feature extraction in space.

4. A multimodal radar echo extrapolation method based on a Doppler dual-polarization weather radar according to claim 1, wherein, The formula of the deep feature extraction module is as follows, y i' = y i + Att(BN(y i )) y i+1 = MLP(BN(y i' )) + y i' Att(y) = PWConv(GSSA(Mish(PWConv(y)))) MLP(y) = PWConv(DWConv(Mish(PWConv(y)))) When the initial value of i is 1, y 1 is the shallow aggregation feature FF extracted from the Multimodal Encoder, and this module is stacked L layers in total; 1 ≤ i ≤ 1 - L; Att(y) represents the attention formula, PWConv represents the point-wise convolution with a kernel size of 1, GSSA represents the gated self-attention unit, Mish represents the self-regularized non-monotonic neural activation function, MLP(y) represents the multi-layer perceptron formula, DWConv represents the depthwise separable convolution, and y i represents the depth spatio-temporal feature of the i-th layer, and BN represents the BatchNorm function; Finally, the output deep spatio-temporal feature of the deep feature extraction module is y L .

5. A multimodal radar echo extrapolation method based on a Doppler dual-polarization weather radar according to claim 4, characterized in that The formula of the gated spatio-temporal self-attention unit GSSA is as follows: Among them, DWDConv represents dilated depthwise convolution, y input represents the input feature, y map represents the attention map, y output represents the output feature.

6. The multimodal radar echo extrapolation method based on a Doppler dual-polarization weather radar according to claim 1, characterized in that The data flow formula of the decoding module is as follows: Among them, and represent the features after the first and second upsamplings, represents the features after convolution after the first upsampling. ConvTranspose2d represents transposed convolution, and forecast is the final prediction result for radar echo data.

Images