A method for extrapolating short-term rainfall based on radar echo multi-scale rainfall field prediction

Abstract

The application discloses a kind of based on radar echo multiscale rainfall field prediction short-acting rainfall extrapolation method, comprising the following steps: based on the data preprocessing of meteorological observation site radar echo data by multi-modal space-time network of multi-source observation data, construct radar echo dataset;Based on the echo sequence feature extraction processing of radar echo dataset by multiscale encoder;The decomposition processing result of echo sequence feature extraction processing is handled by component decomposition module, and the basic field component, enhanced field component and detail field component are obtained;Based on adaptive fusion module, call constraint condition library to carry out feature fusion to basic field component, enhanced field component and detail field component, generate rainfall prediction result;Solve the problems, such as the lack of development prediction, short forecast period and the lack of physical mechanism expression of traditional radar extrapolation algorithm in processing complex precipitation system, to significantly improve the prediction accuracy of radar echo prediction rainfall.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar monitoring weather conditions, and particularly relates to a method for predicting short-term rainfall extrapolation based on radar echo multi-scale rainfall field. BACKGROUND

[0002] Short-time heavy rainfall is mainly caused by convective activities under strong atmospheric instability conditions. When sufficient water vapor supply, strong vertical instability stratification, and appropriate dynamic lifting mechanism (such as cold front, topographic lifting, and convergence line) jointly act, air rapidly rises to form a strong convective cloud system. The strong upward flow inside these systems makes a large amount of water vapor condense rapidly, releases a large amount of latent heat in a short time, further strengthens the convective development, and eventually leads to the rapid growth of cloud droplets and their concentration under the action of gravity, forming a high-intensity rainfall process in a short time. The organization and development of local circulation and mesoscale convective systems, as well as the maintenance of water vapor channels, are also important factors affecting the formation and maintenance of short-time heavy rainfall.

[0003] Short-term rainfall extrapolation mainly relies on the combination of multi-source data fusion and artificial intelligence algorithms. First, the intensity, range, and movement trend of the precipitation echo are obtained through radar data, and the dynamic evolution characteristics of the precipitation system can be captured by Doppler radar. The network of automatic weather stations provides real-time precipitation, temperature, humidity, air pressure, and wind direction and speed, etc. key parameters, providing ground verification and initial conditions for precipitation prediction. Numerical prediction models such as WRF and GRAPES simulate the evolution of atmospheric state by solving atmospheric dynamics equations to predict future precipitation distribution.

[0004] In terms of prediction methods, traditional extrapolation techniques such as TREC (Tracking Echo Correlation) and optical flow method predict the movement path and intensity change of the precipitation system by calculating the motion vector of the radar echo field. With the improvement of computing power, machine learning methods have become increasingly important. Convolutional neural networks (CNN) can effectively extract spatial features from radar images, and long short-term memory networks (LSTM) are good at capturing time series change rules. The combination of the two can realize the prediction of the nonlinear evolution of the precipitation system.

[0005] In addition, ensemble prediction and multi-model fusion technology can reduce the uncertainty of a single model by integrating multiple prediction results. In recent years, artificial intelligence methods based on big data such as deep learning and random forests have shown advantages in handling massive meteorological data, and can automatically learn complex precipitation patterns and rules, improving prediction accuracy and spatiotemporal resolution. The comprehensive application of these technologies makes short-term rainfall prediction develop from single data source and single model to multi-source data fusion and multi-model integration, providing more accurate decision support for disaster prevention and mitigation.

[0006] The above traditional technologies have the following disadvantages:

[0007] Defect 1: Insufficient prediction of generation and development: Traditional extrapolation algorithms are mainly based on linear motion assumptions, which are difficult to accurately capture the generation, development and extinction process of precipitation systems. The algorithm tends to translate existing echoes according to historical motion trends, but in fact, precipitation systems will undergo complex nonlinear evolution, including the sudden emergence of new convection, the strengthening or weakening of intensity, and the extinction of the system, which are difficult to accurately predict through simple extrapolation.

[0008] Defect 2: Sample representativeness and data quality issues: The effectiveness of extrapolation algorithms is highly dependent on the representativeness of historical samples and the quality of input data. On the one hand, the training samples may not cover all weather patterns, especially extreme weather events; on the other hand, radar data itself has quality problems such as observation blind area, ground clutter, bright band effect, etc., which will be amplified in the prediction process, reducing the prediction accuracy.

[0009] Defect 3: Insufficient expression of physical mechanisms: Extrapolation algorithms can usually only provide 1-2 hours of effective prediction, and the accuracy decreases rapidly as the prediction time extends. This is mainly because the algorithm lacks sufficient consideration of the physical processes of precipitation formation, making it difficult to simulate the impact of atmospheric environmental field changes on precipitation systems and effectively integrate multi-scale meteorological elements, resulting in limited prediction ability in complex weather systems and terrain conditions. SUMMARY

[0010] In order to overcome the shortcomings of the prior art, the present application provides a radar echo multi-scale rainfall field prediction short-term rainfall extrapolation method, which solves the problems of insufficient prediction of generation and development, short prediction period and insufficient expression of physical mechanisms in traditional radar extrapolation algorithms when dealing with complex precipitation systems. Radar echo multi-scale rainfall field prediction short-term rainfall extrapolation method

[0011] In order to achieve the above-mentioned purposes of the application, the following technical solutions are adopted:

[0012] A radar echo multi-scale rainfall field prediction short-term rainfall extrapolation method, comprising the following steps:

[0013] S101, a multi-modal spatio-temporal network based on multi-source observation data is used to preprocess radar echo data of meteorological observation stations, and a radar echo dataset is constructed;

[0014] S102, a multi-scale encoder is used to extract and process echo sequence features of the radar echo dataset;

[0015] S103, a component decomposition module is used to decompose the echo sequence feature extraction and processing result, to obtain a basic field component, an enhanced field component and a detail field component;

[0016] S104, based on the adaptive fusion module call constraint condition library will be basic field component, enhanced field component and detail field component feature fusion, generate rainfall prediction results.

[0017] Further, the multi-modal spatio-temporal network construction of multi-source observation data includes the following steps:

[0018] Obtaining weather observation station and radar echo historical data for preprocessing, realizing multi-modal observation data feature cooperation, to construct training data set;

[0019] The echo sequence feature extraction processing result is processed by component decomposition processing;

[0020] The echo sequence feature extraction processing result is processed by component decomposition processing;

[0021] Based on the echo sequence feature extraction processing result and the component decomposition processing result, the radar echo prediction rainfall initial model is trained;

[0022] If the model training result is the same as the training sample result, the model training is completed; if the model training result is different from the training sample result, the model training is re-performed.

[0023] Further, obtaining weather observation station and radar echo data for data preprocessing, constructing radar echo data set includes the following steps:

[0024] The observation station with daily rainfall greater than the set daily rainfall threshold is marked;

[0025] According to the marked observation station position, the radar combined reflectivity data of the corresponding time period is extracted from the weather observation station radar echo data;

[0026] The radar combined reflectivity data is cropped and normalized.

[0027] Further, the radar echo sequence feature extraction processing based on the multi-scale encoder on the radar echo data set includes the following steps:

[0028] Add residual block connection in each down-sampling layer to alleviate the gradient vanishing problem;

[0029] Using 3D convolution kernel to capture the feature association of spatial and temporal dimensions at the same time;

[0030] Through maximum pooling, the spatial dimension is gradually compressed, and the time dimension remains unchanged.

[0031] Further, the echo sequence feature extraction processing result is processed by component decomposition module, and the basic field component, enhanced field component and detail field component are obtained, including the following steps:

[0032] Based on the basic field submodule, global features are extracted through global average pooling and fully connected layers to capture large-scale precipitation structure and overall movement trend;

[0033] Based on the enhanced field submodule, component extraction is performed through the channel attention mechanism to determine the intensity changes in the strong convection region;

[0034] Based on the detail field submodule, small-scale features and texture information of precipitation fields are extracted through a spatial attention mechanism to obtain small-scale features and texture information.

[0035] Furthermore, based on the adaptive fusion module, the constraint library is invoked to perform feature fusion of the basic field components, enhanced field components, and detail field components to generate rainfall prediction results, including the following steps:

[0036] The decoder reconstructs the fundamental field components, augmented field components, and detail field components.

[0037] Based on the discriminator, the basic field features are concatenated with the features of each layer of the decoder, and multi-scale fusion is achieved through 3D convolution.

[0038] A spatial attention map is generated based on the basic field features, and the attention map is then fused with the enhancement field and detail field features in a weighted manner.

[0039] The generator produces radar echo rainfall prediction results based on the fusion results.

[0040] Furthermore, when reconstructing the fundamental field components, spatial resolution is gradually restored through residual connections and deconvolution.

[0041] Furthermore, when reconstructing the augmentation field component and detail field component, decoder branches combined with skip connections are used to preserve multi-scale information.

[0042] Furthermore, the constraint library includes loss functions for detail field components, augmentation field components, generator loss functions, and discriminator loss functions.

[0043] Furthermore, detail field component loss is used to ensure the accuracy of small-scale features and texture information.

[0044] Furthermore, the field component loss is enhanced to ensure accurate capture of intensity changes in strong convection regions.

[0045] The beneficial effects of this application are: it solves the problems of insufficient prediction of the generation and dissipation development of traditional radar extrapolation algorithms when dealing with complex precipitation systems, short prediction period and insufficient expression of physical mechanisms.

[0046] The observation station data is marked by ground observation station data, and the observation station data of a certain place with rainfall level is taken as a feature, and the corresponding radar combined reflectivity data is extracted and extracted to construct a radar echo data set, so as to ensure the prediction accuracy of the model. In the model design, a multi-scale encoder is used to extract the features of the radar echo sequence, and the features are divided into three components of the basic field, the enhanced field and the detail field through the component decomposition module, so as to improve the prediction accuracy of the radar echo prediction rainfall. Among them, the basic field captures the large-scale precipitation structure and the overall moving trend, the enhanced field focuses on the intensity change of the strong convective area, and the detail field processes the small-scale features and texture information. An adaptive parameter adjustment mechanism is designed for precipitation of different intensity levels. The extracted translation component is transmitted to the generative network, and each component is reconstructed through three independent decoder branches, and finally each component is integrated through an adaptive fusion module to output the radar echo prediction result of the next 2 hours.

[0047] The precipitation field is decomposed into three components of the basic field, the enhanced field and the detail field, and combined with the coding and decoding network structure and the generative model, the high-precision prediction of the precipitation system above the moderate rainfall level is realized, and a scientific basis is provided for disaster prevention and reduction.

[0048] The compound loss function based on the detail field and the enhanced field is realized, and the two components are taken as the constraint condition to guide the training of the GAN generator. The accuracy of the small-scale features and the texture information is ensured through the detail field loss, and the accurate capture of the intensity change of the strong convective area is ensured through the enhanced field loss. Combined with the adversarial loss, the content loss and the gradient loss, a comprehensive optimization target is formed, and the prediction accuracy of the radar echo prediction rainfall is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. 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.

[0050] Figure 1 It is a step schematic diagram of a radar echo multi-scale rainfall field prediction short-term rainfall extrapolation method. DETAILED DESCRIPTION

[0051] The embodiments of the present application will be described in detail below with reference to the drawings.

[0052] Following, the embodiments of the present application will be described in details by specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art from the disclosure of the present specification. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. The present application can also be implemented or applied by other different specific embodiments, and various modifications or changes can be made to the details in the present specification based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Embodiment one:

[0054] A radar echo multi-scale rainfall field-based short-term rainfall extrapolation method, comprising the following steps:

[0055] S101, the multi-source observation data-based multi-modal spatio-temporal network is used to perform data preprocessing on the radar echo data of the meteorological observation site, and a radar echo data set is constructed;

[0056] The radar echo data of the meteorological observation site is obtained, and the observation sites with daily rainfall exceeding a set daily rainfall threshold are marked, and the observation site positions exceeding the set daily rainfall threshold are screened out. The daily rainfall can be divided into rainfall grades according to the size of the daily rainfall, for example, daily rainfall exceeding 10 mm is defined as moderate rain. According to the marked observation site positions, radar combined reflectivity data corresponding to the time period is extracted, the radar combined reflectivity time resolution is 6 minutes, and the spatial resolution is 1 km*1 km. The radar combined reflectivity data is subjected to cutting, normalization and other processing operations to construct a radar echo data set, so as to obtain a radar echo data set with a spatial resolution of 256 km*256 km.

[0057] The multi-modal spatio-temporal network of multi-source observation data comprises the following steps:

[0058] The historical data of the meteorological observation site and the radar echo are obtained and preprocessed, the multi-modal observation data features are cooperated to construct a training data set;

[0059] The echo sequence feature extraction processing is performed on the training data set;

[0060] The echo sequence feature extraction processing result is subjected to component decomposition processing;

[0061] The radar echo prediction rainfall initial model is trained based on the echo sequence feature extraction processing result and the component decomposition processing result;

[0062] If the model training result is the same as the training sample result, the model training is completed; if the model training result is different from the training sample result, the model training is re-performed.

[0063] The radar echo data of the meteorological observation site is acquired for data preprocessing, and the radar echo dataset is constructed, including the following steps:

[0064] The observation sites with daily rainfall greater than the set daily rainfall threshold are marked for processing;

[0065] According to the marked observation site position, the radar combined reflectivity data of the corresponding time period is extracted from the meteorological observation site radar echo data;

[0066] The radar combined reflectivity data is cropped and normalized.

[0067] S102, based on a multi-scale encoder, the radar echo dataset is subjected to echo sequence feature extraction processing;

[0068] Based on the multi-scale encoder, the radar echo dataset is subjected to echo sequence feature extraction processing, and the echo sequence feature extraction processing result is obtained. The multi-scale encoder can adopt a U-Net type encoder structure, including multiple down-sampling layers, the input size is [B, 10, 256, 256, 1], B is the batch size, 10 is the time step, the time dimension information is preserved by using 3D convolution to capture the spatio-temporal features. The network hierarchy is mainly divided into three layers:

[0069] Input layer: Conv3D(64, 3x3x3, stride=1, padding=1)→BatchNorm3D→LeakyReLU(0.2)

[0070] Down-sampling layer 1: Conv3D(64, 3x3x3, stride=1, padding=1)→BatchNorm3D→LeakyReLU(0.2)→MaxPool3D(2, 2, 1)

[0071] Down-sampling layer 2: Conv3D(128, 3x3x3, stride=1, padding=1)→BatchNorm3D→LeakyReLU(0.2)→MaxPool3D(2, 2, 1)

[0072] Down-sampling layer 3: Conv3D(256, 3x3x3, stride=1, padding=1)→BatchNorm3D→LeakyReLU(0.2)→MaxPool3D(2, 2, 1)

[0073] Bottleneck: Conv3D(512, 3x3x3, stride=1, padding=1) -> BatchNorm3D -> LeakyReLU(0.2) -> Conv3D(512, 3x3x3, stride=1, padding=1) -> BatchNorm3D -> LeakyReLU(0.2).

[0074] The echo sequence feature extraction processing on the radar echo dataset based on the multi-scale encoder includes the following steps:

[0075] A residual block connection is added to each down-sampling layer to alleviate the gradient vanishing problem.

[0076] A 3D convolution kernel is used to capture the feature association of spatial and temporal dimensions at the same time.

[0077] The spatial dimension is gradually compressed by maximum pooling, and the time dimension remains unchanged.

[0078] It should be noted that the core idea of the residual block connection is to introduce a skip connection to enable the network to effectively transmit information, especially when information is not easily transmitted in deep networks. It can alleviate the gradient vanishing problem and improve the efficiency of training. In this way, the deep network can maintain the input information, thereby accelerating the convergence and improving the performance of the model.

[0079] S103, decompose the echo sequence feature extraction processing result by the component decomposition module to obtain a basic field component, an enhanced field component and a detail field component;

[0080] Since most radar echo extrapolation methods based on deep learning lack the guidance of precipitation physical mechanisms and are mainly data-driven, the present application designs to decompose the precipitation field into three components to supplement the physical mechanism driving for subsequent short-term precipitation extrapolation. Therefore, the precipitation motion follows the time translation symmetry and the space translation symmetry, and its expression is as follows:

[0081]

[0082] wherein, is a given past radar field, and the future radar field x is predicted by a short-term rainfall extrapolation model parameterized by theta 1:T The model generates an ensemble forecast by integrating the random vector z of the latent space.

[0083] In the component decomposition design, a component decomposition module is designed at the bottleneck layer of the encoder, and the component decomposition module includes a base field submodule, an enhanced field submodule and a detail field submodule. A 512-channel feature map can be divided into three independent feature channel groups using a 1x1x1 convolution. The base field channel number is 170, the enhanced field channel number is 171, and the detail field channel number is 171. The decomposition mode of multiple components through the component decomposition module can comprehensively capture the multi-scale characteristics of the precipitation system, and provide more abundant information for subsequent prediction.

[0084] The echo sequence feature extraction processing result is decomposed by the component decomposition module to obtain the base field component, the enhanced field component and the detail field component, including the following steps:

[0085] Global features are extracted based on the base field submodule through global average pooling and a fully connected layer to capture large-scale precipitation structures and overall movement trends;

[0086] The component extraction is performed based on the enhanced field submodule through a channel attention mechanism to determine the intensity change of the strong convective region;

[0087] The small-scale feature and texture information of the precipitation field are extracted based on the detail field submodule through a spatial attention mechanism to obtain the small-scale feature and texture information.

[0088] For example, the base field submodule mainly captures large-scale precipitation structures and overall movement trends, and the processing flow is: Conv3D(170,1x1x1)→GlobalAvgPool3D→FC(170)→Reshape→Conv3D(170,1x1x1), and then residual connection is performed, Add(original feature)→BatchNorm3D→LeakyReLU(0.2). The enhanced field submodule represents the intensity change of the strong convective region, and the component extraction is performed through an improved channel attention mechanism. The processing flow is: Conv3D(171,1x1x1)→[MaxPool3D,AvgPool3D]→Concat→Conv2D(171,7x7,padding=3)→Sigmoid. The detail field submodule extracts small-scale features and texture information of the precipitation field, and an improved spatial attention mechanism is used. The processing flow is: Conv3D(171,1x1x1)→[MaxPool3D(channel),AvgPool3D(channel)]→Concat→Conv3D(1,7x7x1,padding=(3,3,0))→Sigmoid.

[0089] S104, based on the adaptive fusion module, the base field component, the enhanced field component and the detail field component are fused based on the constraint condition library to generate a rainfall prediction result;

[0090] The adaptive fusion module can adopt a conditional GAN architecture, and the adaptive fusion module includes a generator network structure and a discriminator network structure. The generator network structure can adopt a U-Net++ structure, which can enhance the feature extraction and fusion capability; the encoder part is designed as 4 layers of down-sampling, each layer including 2 residual blocks, and the residual block is designed as Conv3D→BatchNorm3D→LeakyReLU→Conv3D→BatchNorm3D→Add→LeakyReLU; the decoder adopts 4 layers of up-sampling, each layer including 2 residual blocks and a skip connection, and the output layer is designed according to the input, Conv3D(64, 3×3×3)→BatchNorm3D→LeakyReLU→Conv3D(20, 3×3×3)→Sigmoid.

[0091] The discriminator network structure can adopt a PatchGAN discriminator, which focuses on the authenticity of the local area, and splices the real radar echo sequence and the basic field feature (or splices the generator network structure output sequence and the basic field feature). The main structure is 5 layers of convolution down-sampling, each layer being Conv3D→BatchNorm3D→LeakyReLU; the final output is Conv3D(1, 3×3×3)→Sigmoid, judging the authenticity of each spatio-temporal patch.

[0092] In order to make the short-term extrapolation model have a physical mechanism of precipitation field, the basic field feature of the corresponding resolution is innovatively integrated into each decoding layer of the generator, and the feature fusion mode is Concat([decoding layer feature, basic field feature])→Conv3D→BatchNorm3D→LeakyReLU, and then an attention map is generated based on the basic field feature through an attention mechanism guide, guiding the generator to focus on the key area.

[0093] The adaptive fusion module calls a constraint condition library to perform feature fusion on the basic field component, the enhanced field component and the detail field component to generate a rainfall prediction result, including the following steps:

[0094] Reconstructing the basic field component, the enhanced field component and the detail field component based on the decoder;

[0095] Splicing the basic field feature and the features of each layer of the decoder based on the discriminator, and realizing multi-scale fusion through 3D convolution;

[0096] Generating a spatial attention map based on the basic field feature, and performing weighted fusion on the attention map, the enhanced field and the detail field feature;

[0097] Generating a radar echo rainfall prediction result based on the fusion result by the generator.

[0098] It should be noted that based on the decoder reconstructing the base field component, the enhanced field component and the detail field component, when reconstructing the base field component, the spatial resolution is gradually restored through residual connection and deconvolution, and when reconstructing the enhanced field component and the detail field component, the multi-scale information is reserved by using the decoder branch combined with the jump connection. After reconstructing each component by using the independent decoder branch, the contribution weight of each component is dynamically adjusted by using the weighted average or the gating mechanism, and finally the future 2-hour radar echo prediction sequence is output. Through the combination of the component decomposition module driven by the physical mechanism and the adaptive fusion module, the multi-scale feature modeling (large-scale translation, strong convective evolution, small-scale texture) of the precipitation system is realized, and through the constraint condition library compound loss function and the adaptive fusion mechanism, the accuracy and the prediction period (up to 2 hours) of the short-term precipitation extrapolation are significantly improved.

[0099] The constraint condition library is constructed by using the loss function to constrain the enhanced field component and the detail field component. The constraint condition library includes a detail field component loss function, an enhanced field component loss function, a generator loss function and a discriminator loss function.

[0100] Based on the compound loss function design of the detail field and the enhanced field, the two components are used as constraint conditions to guide the training of the GAN generator. The accuracy of small-scale features and texture information is ensured through the detail field component loss, and the accurate capture of the intensity change of the strong convective area is ensured through the enhanced field component loss. Combined with the adversarial loss, the content loss and the gradient loss, a comprehensive optimization target is formed, and the prediction accuracy is significantly improved.

[0101] It should be noted that in order to better capture the birth and death changes of the precipitation field, the detail field and the enhanced field are used as loss function constraints to guide the GAN generator learning.

[0102] For example, the detail field loss function is set as L_detail=MSE(Detail(G(z)),Detail(y));

[0103] Enhanced detail loss: L_detail_enhanced=0.5*MSE(Detail(G(z)),Detail(y))+0.5*SSIM_Loss(Detail(G(z)),Detail(y));

[0104] Gradient loss: Capture edge and texture features;

[0105] Enhanced field loss design: L_enhance=MSE(G(z)*W,y*W), wherein W is an intensity weight matrix, the intensity weight matrix W is dynamically generated according to the reflectivity value, and higher weight is given to the high reflectivity area;

[0106] Weight calculation: W = exp((y-threshold) / scale) for the region of y>threshold;

[0107] GAN loss function design:

[0108] Adversarial loss: L_adv = E[log(D(y))] + E[log(1-D(G(z)))];

[0109] Content loss: L_content = MSE(G(z), y);

[0110] Detail field loss: L_detail = MSE(Detail(G(z)), Detail(y));

[0111] Enhancement field loss: L_enhance 0.5*MSE(Detail(G(z)), Detail(y)) + 0.5*SSIM_Loss(Detail(G(z)), Detail(y));

[0112] Generator total loss: L_G = λ1L_adv + λ2L_content + λ3L_detail + λ4L_enhance + λ5*L_gradient (λ1, λ2, λ3, λ4, λ5 are weight coefficients, which can be adjusted according to the training result).

[0113] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and the unit described above can refer to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0114] The terms "first", "second", and "third" and the like in the specification of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0115] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields, characterized in that, Includes the following steps: S101. Based on a multi-modal spatiotemporal network of multi-source observation data, radar echo data from meteorological observation stations are preprocessed to construct a radar echo dataset. S102. Based on a multi-scale encoder, perform echo sequence feature extraction processing on the radar echo dataset; S103. The echo sequence feature extraction processing result is decomposed by the component decomposition module to obtain the basic field component, the enhanced field component and the detail field component. S104. Based on the adaptive fusion module, the constraint library is called to perform feature fusion of the basic field components, enhanced field components and detailed field components to generate rainfall prediction results. Step S103 includes the following steps: Based on the basic field submodule, global features are extracted through global average pooling and fully connected layers to capture large-scale precipitation structure and overall movement trend; Based on the enhanced field submodule, component extraction is performed through the channel attention mechanism to determine the intensity changes in the strong convection region; Based on the detail field submodule, small-scale features and texture information of precipitation field are extracted through spatial attention mechanism to obtain small-scale features and texture information; Step S104 includes the following steps: The decoder reconstructs the fundamental field components, augmented field components, and detail field components. Based on the discriminator, the basic field features are concatenated with the features of each layer of the decoder, and multi-scale fusion is achieved through 3D convolution. A spatial attention map is generated based on the basic field features, and the attention map is then fused with the enhancement field and detail field features in a weighted manner. The generator produces radar echo rainfall prediction results based on the fusion results.

2. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, The construction of the multimodal spatiotemporal network of the multi-source observation data includes the following steps: Historical data from meteorological observation stations and radar echoes were acquired and preprocessed to achieve feature synergy of multimodal observation data in order to construct a training dataset; Echo sequence feature extraction is performed on the training dataset; Component decomposition is performed on the feature extraction results of the echo sequence. The initial model for radar echo prediction of rainfall was trained based on the results of echo sequence feature extraction and component decomposition. If the model training result is the same as the training sample result, the model training is complete; if the model training result is different from the training sample result, the model training is repeated.

3. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, Step S101 includes the following steps: Observation stations with daily rainfall exceeding a set daily rainfall threshold will be marked. Based on the marked locations of observation stations, extract the radar combined reflectivity data for the corresponding time period from the radar echo data of meteorological observation stations; The radar composite reflectivity data is cropped and normalized.

4. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, Step S102 includes the following steps: Add residual block connections to each downsampling layer to mitigate the vanishing gradient problem; Use 3D convolution kernels to capture feature correlations in both spatial and temporal dimensions; Max pooling is used to gradually compress the spatial dimension while keeping the temporal dimension unchanged.

5. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, When reconstructing the fundamental field components, spatial resolution is gradually restored through residual connections and deconvolution.

6. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, When reconstructing the enhanced field components and detail field components, decoder branches combined with skip connections are used to preserve multi-scale information.

7. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 1, characterized in that, The constraint library includes detail field component loss functions, augmentation field component loss functions, generator loss functions, and discriminator loss functions.

8. The method for extrapolating short-term rainfall prediction based on radar echo multi-scale rainfall fields according to claim 7, characterized in that, The detailed field component loss is used to ensure the accuracy of small-scale features and texture information, while the enhanced field component loss is used to ensure the accurate capture of intensity changes in strong convection regions.

Images