Short-term rainfall prediction method and system and electronic equipment

Abstract

The invention discloses a short-term rainfall prediction method. The method comprises the following steps: S1, obtaining a time sequence radar reflectivity data set of a preset time period before a target area; carrying out the preprocessing of the time sequence radar reflectivity data set, and obtaining a preprocessed time sequence radar reflectivity data set; s2, inputting the preprocessed time sequence radar reflectivity data set into the trained radar echo extrapolation model to obtain a short-term prediction radar echo image set; s3, inputting the short-term prediction radar echo image set into a trained quantitative precipitation estimation model to obtain a short-term cumulative precipitation space distribution diagram; and S4, carrying out visualization processing on the short-term accumulated precipitation space distribution diagram to obtain a short-term precipitation prediction result.

Description

Technical Field

[0001] This invention belongs to the field of precipitation prediction technology, specifically relating to a short-term precipitation prediction method, system, and electronic equipment. Background Technology

[0002] Short-term precipitation forecasts typically refer to predictions of precipitation over the next few hours, such as precipitation forecasts for the next six hours. Short-term precipitation forecasts provide crucial information and strong data support for human production, daily life, travel, disaster prevention and mitigation, and early warning decision-making. They also provide historical data for more accurate numerical model simulations and weather forecasts.

[0003] Traditional short-term precipitation forecasting methods study atmospheric dynamics and then construct fluid dynamics and advection equations based on physical and mathematical principles to predict precipitation. However, methods using complex atmospheric physics control equations have limitations due to the quantity and quality of observational data and the understanding of atmospheric dynamic mechanisms. In contrast, radar echo extrapolation methods use faster and cheaper statistical methods and radar datasets to infer the current precipitation status.

[0004] However, there is limited research in this area, so there is an urgent need to develop a short-term precipitation prediction method and system based on radar echo extrapolation. Summary of the Invention

[0005] The primary objective of this invention is to provide a short-term precipitation forecasting method, which includes the following steps:

[0006] S1 acquires a time-series radar reflectivity dataset for a preset time period before the target area; and preprocesses the time-series radar reflectivity dataset to obtain a preprocessed time-series radar reflectivity dataset.

[0007] S2 inputs the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set.

[0008] S3 inputs the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of short-term cumulative precipitation.

[0009] S4 visualizes the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation forecast results.

[0010] Specifically, the radar echo extrapolation model in step S2 includes a first encoder, a converter, and a first decoder; the first encoder is composed of multiple stacked ConvSC modules; the ConvSC module includes a Conv2d convolutional layer, a GroupNorm normalization layer, and a SiLU activation layer; the converter is composed of multiple MetaBlock modules connected in series; the MetaBlock module includes a SpatialAttention mechanism and a MixMlp substructure; the first decoder is composed of multi-level ConvSC modules and an end Conv2d readout layer.

[0011] Specifically, the quantitative precipitation estimation model in step S3 includes a second encoder, a processor, and a second decoder; the second encoder is composed of a Stem_Block module and a multi-level ResNet_Block module stacked sequentially; the Stem_Block module includes a first convolutional branch, a second convolutional branch, and a Squeeze-Excite attention mechanism in parallel; the shallow features output by the first convolutional branch and the second convolutional branch are fused and then input to the Squeeze-Excite attention mechanism for channel weight calibration; the processor adopts an ASPP structure, which includes four parallel branches, each of which includes dilated convolutional layers with different dilation rates and a BatchNorm normalization layer; the second decoder consists of three cascaded Decoder_Block modules, each containing an Attention mechanism.

[0012] Specifically, the method for obtaining the trained radar echo extrapolation model in step S2 includes:

[0013] An initial radar echo extrapolation model is used as a generator, and a discriminator is introduced for adversarial training. The generator is trained using a time-series radar reflectivity sample set. The generator generates prediction results based on the radar reflectivity sample data of the previous time period. The discriminator distinguishes the prediction results from the radar reflectivity sample data of the next time period to generate a discrimination score. The parameters of the discriminator are updated according to the discrimination loss function, and the parameters of the generator are updated according to the joint combination loss function until the combination loss function converges. The trained generator is then used as the trained radar echo extrapolation model.

[0014] Specifically, the discriminative loss function mentioned in step S2 is as follows:

[0015] L disc =E[max(0,1-S real )]+E[max(0,1+S fake )]

[0016] Among them, S realS represents the discrimination score of the discriminator for the actual time-series radar reflectivity data; fake E[·] represents the discrimination score of the discriminator on the time-series radar reflectivity data generated by the generator; E[·] represents the expectation function.

[0017] Specifically, step S2, L gen The combined loss function includes rec_loss (reconstruction loss) and L... gan To counteract losses, the specific formula is as follows:

[0018] L gen =L gan +20·rec_loss;

[0019] Wherein, the L gan =-E[S fake The rec_loss weighted reconstruction loss includes L rec Weighted L1 reconstruction loss, L power Power loss and L perc Perceived loss, the specific formula is:

[0020] rec_loss=L rec +L power +0.01L perc ;

[0021] in, in X represents the feature map of the i-th layer; n X represents the prediction result generated by the generator; T A true short-term predicted radar echo image; in

[0022]

[0023] Where, mean h,w [] represents the average value in the image height h and width w dimensions; k is the preset intensity threshold.

[0024] The second objective of this invention is to provide a short-term precipitation forecasting system, comprising:

[0025] The preprocessing module is used to acquire a time-series radar reflectivity dataset of a target area over a preset time period; and to preprocess the time-series radar reflectivity dataset to obtain a preprocessed time-series radar reflectivity dataset.

[0026] The radar echo extrapolation module is used to input the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set.

[0027] The quantitative precipitation estimation module is used to input the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of the short-term cumulative precipitation.

[0028] The visualization module is used to visualize the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation forecast results.

[0029] A third objective of this invention is to provide an electronic device comprising:

[0030] Memory, used to store computer control instructions;

[0031] A processor, connected to the memory, is configured to retrieve and execute the computer control instructions to cause the electronic device to perform the short-term precipitation forecasting method as described in any of the preceding descriptions.

[0032] The beneficial effects of this invention are as follows:

[0033] (1) This invention seamlessly cascades two key steps: radar echo extrapolation and quantitative precipitation estimation, constructing a complete and automated prediction process that directly generates high-resolution cumulative precipitation distribution maps from raw radar data. This avoids the problems of error accumulation and inefficiency caused by traditional manual analysis or step-by-step processing, and can quickly generate refined precipitation forecast products for the short term (e.g., 0-2 hours).

[0034] (2) This invention trains a radar echo extrapolation model using an adversarial training (GAN) framework and a multi-constraint combined loss function, which integrates adversarial loss, weighted L1 reconstruction loss, power loss and perception loss. The weighted L1 loss and power loss designed for strong echo regions enable the model to focus on and accurately reconstruct the strong convective echo regions that are crucial for precipitation forecasting during training, effectively alleviating the problem of weak or smooth strong echo forecasts in traditional methods, and making the extrapolated radar echoes closer to the real physical process in terms of intensity, shape and energy distribution.

[0035] (3) The quantitative precipitation estimation model in this invention integrates the SE attention mechanism and the ASPP multi-scale perception module, which can adaptively focus on key features related to precipitation intensity and capture contextual information of different ranges at the same time, thereby more accurately converting radar reflectivity into ground precipitation and reducing the error caused by the uncertainty of ZR relationship; the residual unit and feature refinement module in the model effectively promote the propagation and utilization of deep features, alleviate the gradient vanishing problem, and enhance the model's estimation ability for complex precipitation systems;

[0036] (4) The present invention outputs an intuitive spatial distribution map of cumulative precipitation, which, after visualization processing, can directly serve departments such as meteorological business forecasting, urban waterlogging early warning, transportation safety, agricultural production and emergency management, providing key decision support for impact-based short-term weather forecasting, and helping to reduce the socio-economic losses and casualties caused by extreme precipitation weather. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a short-term precipitation prediction method according to an embodiment of the present invention;

[0039] Figure 2 This is a framework diagram of the radar echo extrapolation model in an embodiment of the present invention;

[0040] Figure 3 A schematic diagram of the structure of a near-term precipitation prediction system in an embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0043] Please see Figure 1 , Figure 1 for Figure 1 This is a flowchart of a short-term precipitation prediction method according to an embodiment of the present invention; the method includes the following steps:

[0044] In this embodiment of the invention, the preprocessing in step S1 includes dataset filtering and outlier and missing value processing. Dataset filtering includes: acquiring the target area and its original radar reflectivity dataset, removing the geographic information of the target area, and obtaining a temporal radar reflectivity dataset for a preset time period. Outlier and missing value processing includes: performing outlier detection according to the time frame steps, detecting radar reflectivity frame by frame based on the time series; if the reflectivity change rate of more than 30% of pixels within a certain time frame (i.e., the difference between adjacent frames is greater than 50 dBZ / s), it is determined to be a significant outlier and the outlier is deleted entirely; otherwise, it is determined to be a non-significant outlier and outliers are deleted using Mahalanobis distance; filling missing values ​​in the temporal radar reflectivity dataset for the preset time period; after outlier processing, operating according to the time frame steps, using natural days as the statistical unit, calculating the percentage of missing values ​​in all frames each day: if the percentage of missing values ​​on a certain day is >25%, all data for that day is directly removed; if the percentage of missing values ​​is <25%, the data before and after the missing values ​​in the radar echo image data is filled using the inverse distance weighted interpolation method.

[0045] S2 inputs the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set.

[0046] Please see Figure 2 , Figure 2A framework diagram of the radar echo extrapolation model in this embodiment of the invention; in this embodiment of the invention, the radar echo extrapolation model in step S2 includes a first encoder, a converter, and a first decoder; the first encoder is composed of multiple stacked ConvSC modules; the ConvSC module includes a Conv2d convolutional layer, a GroupNorm normalization layer, and a LeakyReLU activation layer; the converter is composed of multiple Inception modules connected in series; the Inception module includes a SpatialAttention mechanism and a MixMlp substructure; the first decoder is composed of multi-level ConvSC modules and an end Conv2d readout layer. In this embodiment of the invention, the first encoder is composed of multiple stacked ConvSC modules. The BasicConv2d layer inside each ConvSC module sequentially includes a Conv2d convolutional layer, a GroupNorm normalization layer, and a LeakyReLU activation layer. It performs two-stage downsampling according to a preset sampling sequence to map the input 10 frames of single-channel radar echoes into 16-channel low-resolution features. The converter is a MidMetaNet module composed of multiple Inception modules connected in series. Each Inception module contains a GASubBlock, which embeds SpatialAttention and MixMlp substructures to realize multi-scale feature interaction between large kernel depth separable convolution and 1×1 convolution to model inter-frame spatiotemporal correlation. The first decoder is composed of multi-stage upsampling ConvSC modules and a terminal Conv2d readout layer. In the upsampling stage, PixelShuffle is used to restore spatial resolution and is fused with the feature residual of the first layer of the encoder to finally output the predicted radar echo image.

[0047] In an embodiment of the invention, the model is used as a generator, and a discriminator is introduced for adversarial training, utilizing the concept of generative adversarial networks. The discriminator consists of a spatial discriminator submodule and a temporal discriminator submodule connected in parallel. The spatial discriminator submodule randomly selects four time slices from the input sequence, sequentially performs 2×2 average pooling and PixelUnshuffle downsampling, expands the channels by a factor of 4, and then enters a three-layer spectral normalization residual block (the first layer DBlock amplifies the input channels to 48 times and downsamples them by half, and the subsequent two layers DBlock continue to double the number of channels in each layer while maintaining downsampling), and then passes through a DBlock layer that maintains the size to refine the features. The outputs of each time slice are summed after ReLU activation, spatial summation, BatchNorm, and linear layer transformation to obtain the spatial discriminant score. The temporal discrimination submodule first applies AvgPool3d(1,2,2) and PixelUnshuffle downsampling to the entire spatiotemporal volume, then concatenates two layers of three-dimensional DBlock (the first layer expands the channels by 48 times, and the second layer expands them by 96 times and further reduces the spatiotemporal resolution). The features are then split into time slices and passed through a two-dimensional DBlock and a DBlock that maintains the size of each slice. ReLU, spatial summation, BatchNorm, and linear mapping are performed on each slice's output. Finally, the temporal discrimination score is obtained by summing the scores along the temporal dimension. The spatial and temporal scores are concatenated along the channel dimension to form the final discrimination result, which is used to simultaneously constrain the spatial details and temporal consistency of the generated sequence.

[0048] The specific methods for obtaining the trained radar echo extrapolation model include:

[0049] An initial radar echo extrapolation model is used as a generator, and a discriminator is introduced for adversarial training. The generator is trained using a time-series radar reflectivity sample set. The generator generates prediction results based on the radar reflectivity sample data of the previous time period. The discriminator distinguishes the prediction results from the radar reflectivity sample data of the next time period to generate a discrimination score. The parameters of the discriminator are updated according to the discrimination loss function, and the parameters of the generator are updated according to the joint combination loss function until the combination loss function converges. The trained generator is then used as the trained radar echo extrapolation model.

[0050] The discriminant loss function is as follows:

[0051] L disc =E[max(0,1-S real )]+E[max(0,1+S fake )]

[0052] Among them, S real S represents the discrimination score of the discriminator for real short-term radar echo images; fakeE[·] represents the discrimination score of the discriminator for the short-term predicted radar echo image generated by the generator; E[·] represents the expectation function.

[0053] L gen The combined loss function includes rec_loss (reconstruction loss) and L... gan To counteract losses, the specific formula is as follows:

[0054] L gen =L gan +20·rec_loss;

[0055] Wherein, the L gan =-E[S fake The rec_loss weighted reconstruction loss includes L rec Weighted L1 reconstruction loss, L power Power loss and L perc Perceived loss, the specific formula is:

[0056] rec_loss=L rec +L power +0.01L perc ;

[0057] in, in X represents the feature map of the i-th layer; n This represents the short-term predicted radar echo image generated by the generator; X T This represents a true short-term radar echo image; in

[0058]

[0059] Where, mean h,w [] represents the average value in the image height h and width w dimensions; k is the preset intensity threshold.

[0060] S3 inputs the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of short-term cumulative precipitation.

[0061] In this embodiment of the invention, the quantitative precipitation estimation model in step S3 includes a second encoder, a processor, and a second decoder; the second encoder is composed of a Stem_Block module and a multi-level ResNet_Block module stacked sequentially; the Stem_Block module includes a first convolutional branch, a second convolutional branch, and a Squeeze-Excite attention mechanism in parallel; the shallow features of the outputs of the first convolutional branch and the second convolutional branch are fused and then input to the Squeeze-Excite attention mechanism for channel weight calibration; the processor adopts an ASPP structure, which includes four parallel branches, each of which includes dilated convolutional layers and BatchNorm layers with different dilation rates; the second decoder consists of three cascaded Decoder_Block modules, each containing an Attention mechanism.

[0062] In an embodiment of the invention, the second encoder is composed of a Stem_Block and a multi-level ResNet_Block stacked sequentially. The Stem_Block contains two parallel branches (Conv2D convolutional layer (3×3), BatchNorm & ReLU normalization layer, Conv2D convolutional layer (3×3)) and incorporates Squeeze-Excite attention. Each ResNet_Block module consists of a dual convolutional main branch (BatchNorm layer → ReLU layer → 3×3 convolutional layer) and a 1×1 shortcut branch. After residual merging, the channel weights are enhanced through the Squeeze-Excite module, thereby extracting deep features of the input multi-source raster at multiple scales. The processor adopts an ASPP structure, composed of 3×3 convolutions with different dilation rates and BatchNorm parallel branches. After aggregating wide receptive field information, it is fused through 1×1 convolution to achieve multi-scale modeling of high-level semantics. The second decoder consists of three cascaded Decoder_Block layers. Each layer contains an Attention mechanism (Attention_Block) for explicitly selecting skip features on the encoding side, and restores spatial resolution step by step through upsampling and ResNet_Block. The decoded output is refined by ASPP and a 1×1 convolutional readout layer to generate precipitation rate results, and ReLU is applied at the end to ensure that the precipitation estimate is non-negative.

[0063] S4 visualizes the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation forecast results.

[0064] In this embodiment of the invention, the spatial distribution map of short-term cumulative precipitation is interpolated onto a higher resolution grid using methods such as Kriging interpolation or bilinear interpolation. The interpolated data is then unified with basic geographic information (such as provincial, municipal, and county boundaries, rivers, and topography) to the same map projection coordinate system (such as WGS-84). Subsequently, the data is converted into the standard GeoTIFF format (a raster image format containing geographic coordinate information). According to operational regulations, precipitation intensity is divided into different levels such as light rain (<10mm / h), moderate rain, heavy rain, and torrential rain, and each level is matched with a specific color (such as blue, green, yellow, and red).

[0065] Overlay geographic layers such as administrative boundaries, major roads, river systems, and important facilities (e.g., reservoirs) onto the precipitation distribution map. Based on the risk levels defined in Phase Two, design an intuitive and industry-standard color map. Typically, a gradient from cool to warm colors to warning colors (e.g., blue → green → yellow → orange → red) is used to represent precipitation from weak to strong. The visualization product is an animated sequence: string together multiple static images from consecutive moments (e.g., one image every 10 minutes for the next 0-6 hours) to generate GIFs or video animations, visually demonstrating the movement, development, and evolution of precipitation. The rendered image or animation files are saved to a designated server directory and published externally via a web service API or internal business platform interface.

[0066] Please see Figure 3 , Figure 3 This is a schematic diagram of a short-term precipitation forecasting system according to an embodiment of the present invention; an embodiment of the present invention provides a short-term precipitation forecasting system 100, comprising:

[0067] The preprocessing module 101 is used to acquire a time-series radar reflectivity dataset of a target area over a preset time period; and to preprocess the time-series radar reflectivity dataset to obtain a preprocessed time-series radar reflectivity dataset.

[0068] The radar echo extrapolation module 102 is used to input the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set.

[0069] The quantitative precipitation estimation module 103 is used to input the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of the short-term cumulative precipitation.

[0070] The visualization module 104 is used to visualize the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation prediction results.

[0071] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. An embodiment of the present invention also provides an electronic device 40, including: at least one processor 401, and at least one memory 402 and a bus 403 connected to the processor; wherein the processor 401 and the memory 402 communicate with each other via the bus 403; the processor 401 is used to call program instructions in the memory 402; the memory 402 is used to store at least one computer program, which, when executed by the at least one processor 401, causes the electronic device to implement the short-term precipitation prediction method provided in the above embodiments. It should be noted that... Figure 4 The illustrated electronic device structure 40 is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0072] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A short-term precipitation forecasting method, characterized in that, The method includes the following steps: S1 acquires a time-series radar reflectivity dataset for a preset time period before the target area; and preprocesses the time-series radar reflectivity dataset to obtain a preprocessed time-series radar reflectivity dataset. S2 inputs the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set. S3 inputs the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of short-term cumulative precipitation. S4 visualizes the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation forecast results.

2. The method according to claim 1, characterized in that, The radar echo extrapolation model in step S2 includes a first encoder, a converter, and a first decoder; the first encoder is composed of multiple stacked ConvSC modules; the ConvSC module includes a Conv2d convolutional layer, a GroupNorm normalization layer, and a LeakyReLU activation layer; the converter is composed of multiple Inception modules connected in series; the Inception module includes a SpatialAttention mechanism and a MixMlp substructure; the first decoder is composed of multi-level ConvSC modules and an end Conv2d readout layer.

3. The method according to claim 1, characterized in that, The quantitative precipitation estimation model in step S3 includes a second encoder, a processor, and a second decoder. The second encoder is composed of a Stem_Block module and a multi-level ResNet_Block module stacked sequentially. The Stem_Block module includes a first convolutional branch, a second convolutional branch, and a Squeeze-Excite attention mechanism in parallel. The shallow features output by the first and second convolutional branches are fused and then input to the Squeeze-Excite attention mechanism for channel weight calibration. The processor adopts an ASPP structure, which includes four parallel branches. Each parallel branch includes dilated convolutional layers with different dilation rates and a BatchNorm normalization layer. The second decoder consists of three cascaded Decoder_Block modules, each containing an Attention mechanism.

4. The method according to claim 1, characterized in that, The specific method for obtaining the trained radar echo extrapolation model in step S2 includes: An initial radar echo extrapolation model is used as a generator, and a discriminator is introduced for adversarial training. The generator is trained using a time-series radar reflectivity sample set. The generator generates prediction results based on the radar reflectivity sample data of the previous time period. The discriminator distinguishes the prediction results from the radar reflectivity sample data of the next time period to generate a discrimination score. The parameters of the discriminator are updated according to the discrimination loss function, and the parameters of the generator are updated according to the joint combination loss function until the combination loss function converges. The trained generator is then used as the trained radar echo extrapolation model.

5. The method according to claim 4, characterized in that, The discriminant loss function mentioned in step S2 is specifically as follows: L disc =E[max(0,1-S real )]+E[max(0,1+S fake )] Among them, S real S represents the discrimination score of the discriminator for real short-term radar echo images; fake E[·] represents the discrimination score of the discriminator for the short-term predicted radar echo image generated by the generator; E[·] represents the expectation function.

6. The method according to claim 4, characterized in that, Step S2 L gen The combined loss function includes rec_loss (reconstruction loss) and L... gan To counteract losses, the specific formula is as follows: L gen =L gan +20·rec_loss; Wherein, the L gan =-E[S fake The rec_loss weighted reconstruction loss includes L rec Weighted L1 reconstruction loss, L power Power loss and L perc Perceived loss, the specific formula is: rec_loss=L rec +L power +0.01L perc ; in, in X represents the feature map of the i-th layer; n This represents the short-term predicted radar echo image generated by the generator; X T This represents a true short-term radar echo image; in Where, mean h,w [] represents the average value in the image height h and width w dimensions; k is the preset intensity threshold.

7. A short-term precipitation forecasting system, characterized in that, include: The preprocessing module is used to acquire the temporal radar reflectivity dataset of the target area over a preset time period. The time-series radar reflectivity dataset is then preprocessed to obtain a preprocessed time-series radar reflectivity dataset. The radar echo extrapolation module is used to input the preprocessed time-series radar reflectivity dataset into the trained radar echo extrapolation model to obtain a short-term predicted radar echo image set. The quantitative precipitation estimation module is used to input the short-term predicted radar echo image set into the trained quantitative precipitation estimation model to obtain the spatial distribution map of the short-term cumulative precipitation. The visualization module is used to visualize the spatial distribution map of the short-term cumulative precipitation to obtain the short-term precipitation forecast results.

8. An electronic device, characterized in that, include: Memory, used to store computer control instructions; A processor, connected to the memory, is configured to retrieve and execute the computer control instructions to cause the electronic device to perform the short-term precipitation forecasting method as described in any one of claims 1-6.

Images