Training method of reflectivity factor prediction model and rainfall prediction method and device

Abstract

The invention discloses a reflectivity factor prediction model training method, a rainfall prediction method and a rainfall prediction device, and the method comprises the steps: carrying out the global feature extraction and local feature extraction of a radar image sample based on a reflectivity factor prediction model, and obtaining the global feature information and local feature information; fusing the global feature information and the local feature information based on a reflectivity factor prediction model to obtain sample feature information; performing reflectivity factor prediction based on a reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample; and iteratively training the reflectivity factor prediction model based on the labeled reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met. The rainfall prediction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a training method for a reflectivity factor prediction model, and a rainfall prediction method and device. Background Art

[0002] As an advanced meteorological detection method, radar can provide rich information on the morphology and distribution of precipitation particles, which is of great significance for improving the prediction accuracy of severe convective weather. Existing technologies generally use neural network models to extract features of radar images for rainfall prediction. However, due to the lack of rich features, the accuracy of rainfall prediction is low. Summary of the invention

[0003] The embodiments of the present application provide a reflectivity factor prediction model training method, a rainfall prediction method and a device, which can improve the accuracy of rainfall prediction.

[0004] In a first aspect, the present application provides a method for training a reflectivity factor prediction model, comprising:

[0005] Based on the reflectivity factor prediction model, global feature extraction and local feature extraction are performed on radar image samples to obtain global feature information and local feature information;

[0006] Based on the reflectivity factor prediction model, the global feature information and the local feature information are fused to obtain sample feature information;

[0007] Performing reflectivity factor prediction based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample;

[0008] The reflectivity factor prediction model is iteratively trained based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

[0009] In a second aspect, the rainfall prediction method provided by the present application comprises:

[0010] Acquire the radar image to be predicted;

[0011] Performing reflectivity factor prediction on the radar image to be predicted based on a reflectivity factor prediction model to obtain second predicted reflectivity factor information, wherein the reflectivity factor prediction model is the reflectivity factor prediction model described in any one of the first aspects;

[0012] The predicted rainfall information corresponding to the radar image to be predicted is determined based on the second predicted reflectivity factor information.

[0013] In a third aspect, the present application provides a training device for a reflectivity factor prediction model, comprising:

[0014] A feature extraction unit, used to perform global feature extraction and local feature extraction on the radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information;

[0015] A feature fusion unit, used for fusing the global feature information and the local feature information based on the reflectivity factor prediction model to obtain sample feature information;

[0016] A prediction unit, configured to perform reflectivity factor prediction based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample;

[0017] A training unit is used to iteratively train the reflectivity factor prediction model based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

[0018] In a fourth aspect, the rainfall prediction device provided by the present application includes:

[0019] An acquisition module, used for acquiring the radar image to be predicted;

[0020] a reflectivity prediction module, configured to predict a reflectivity factor of the radar image to be predicted based on a reflectivity factor prediction model to obtain second predicted reflectivity factor information, wherein the reflectivity factor prediction model is the reflectivity factor prediction model described in any one of the first aspects;

[0021] A rainfall calculation module is used to determine the predicted rainfall corresponding to the radar image to be predicted based on the second predicted reflectivity factor information.

[0022] In a fifth aspect, the electronic device provided in the present application includes a memory and a processor, the memory stores a computer program, and the processor is used to run the computer program in the memory to implement the training method of the reflectivity factor prediction model and the steps in the rainfall prediction method provided in the present application.

[0023] In a sixth aspect, the computer-readable storage medium provided in the present application stores a plurality of instructions, which are suitable for loading by a processor to implement the steps in the training method of the reflectivity factor prediction model and the rainfall prediction method provided in the present application.

[0024] In the seventh aspect, the computer program product provided in the present application includes a computer program or instructions, which, when executed by a processor, implements the steps in the training method of the reflectivity factor prediction model and the rainfall prediction method provided in the present application.

[0025] In the present application, compared with the related art, global feature extraction and local feature extraction are performed on radar image samples based on a reflectivity factor prediction model to obtain global feature information and local feature information; the global feature information and the local feature information are fused based on the reflectivity factor prediction model to obtain sample feature information; reflectivity factor prediction is performed based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample; the reflectivity factor prediction model is iteratively trained based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until the training stop condition is met. The present application performs global feature extraction and local feature extraction on radar image samples respectively, and fuses the global feature information and the local feature information to obtain sample feature information; then, based on the reflectivity factor prediction model and the sample feature information, the reflectivity factor is predicted to obtain the first predicted reflectivity factor information of the radar image sample; finally, the reflectivity factor prediction model is iteratively trained according to the annotated reflectivity factor information and the first predicted reflectivity factor information of the radar image sample; since the global feature information and the local feature information can globally and locally characterize the features of the radar image samples, sample feature information with richer features can be extracted from the radar image samples, so that a more accurate reflectivity factor prediction model can be trained to predict a more accurate radar reflectivity factor, so that a more accurate rainfall can be predicted using a more accurate radar reflectivity factor, thereby improving the accuracy of rainfall prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0027] Figure 1 is a scene schematic diagram of a rainfall prediction system provided in an embodiment of the present application;

[0028] Figure 2 It is a schematic flow chart of an embodiment of a method for training a reflectivity factor prediction model provided in an embodiment of the present application;

[0029] Figure 3 It is a schematic diagram of the structure of an embodiment of the reflectivity factor prediction model provided in the embodiment of the present application;

[0030] Figure 4 is a schematic diagram of another embodiment of the reflectivity factor prediction model provided in the embodiment of the present application;

[0031] Figure 5It is a schematic diagram of adversarial training of a reflectivity factor prediction model and a preset neural network model provided in an embodiment of the present application;

[0032] Figure 6 This is a schematic diagram of an embodiment of a rainfall prediction method provided in an embodiment of the present application;

[0033] Figure 7 It is a structural schematic diagram of an embodiment of a training device for a reflectivity factor prediction model provided in an embodiment of the present application;

[0034] Figure 8 It is a structural schematic diagram of an embodiment of a rainfall prediction device provided in an embodiment of the present application;

[0035] Fig. 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0036] It should be noted that the principles of the present application are illustrated by implementing them in an appropriate computing environment. The following description is based on the illustrated specific embodiments of the present application and should not be considered as limiting other specific embodiments of the present application that are not described in detail herein.

[0037] In the following description of the present application, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0038] In the following description of the present application, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0040] In the prior art, neural network models are generally used to extract features of radar images for rainfall prediction. However, due to the lack of richness of the extracted features, the accuracy of rainfall prediction is low. Compared with the related art, the global feature extraction and local feature extraction are performed on radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information; the global feature information and local feature information are fused based on the reflectivity factor prediction model to obtain sample feature information; the reflectivity factor prediction is performed based on the reflectivity factor prediction model and the sample feature information to obtain the first predicted reflectivity factor information of the radar image sample; the reflectivity factor prediction model is iteratively trained based on the labeled reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until the training stop condition is met. The present application performs global feature extraction and local feature extraction on radar image samples respectively, and fuses the global feature information and the local feature information to obtain sample feature information; then, based on the reflectivity factor prediction model and the sample feature information, the reflectivity factor is predicted to obtain the first predicted reflectivity factor information of the radar image sample; finally, the reflectivity factor prediction model is iteratively trained according to the annotated reflectivity factor information and the first predicted reflectivity factor information of the radar image sample; since the global feature information and the local feature information can globally and locally characterize the features of the radar image samples, sample feature information with richer features can be extracted from the radar image samples, so that a more accurate reflectivity factor prediction model can be trained to predict a more accurate radar reflectivity factor, so that a more accurate rainfall can be predicted using a more accurate radar reflectivity factor, thereby improving the accuracy of rainfall prediction.

[0041] In order to improve the training effect of the reflectivity factor prediction model, the embodiments of the present application provide a method for training a reflectivity factor prediction model, a training device for a reflectivity factor prediction model, a rainfall prediction method, a rainfall prediction device, an electronic device, a computer-readable storage medium, and a computer program product. The method for training a reflectivity factor prediction model can be executed by the training device for the reflectivity factor prediction model, or by an electronic device that integrates the training device for the reflectivity factor prediction model.

[0042] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0043] Please refer to Figure 1The present application also provides a rainfall prediction system, which includes an electronic device. The electronic device integrates a training device for the reflectivity factor prediction model provided by the present application and a rainfall prediction device.

[0044] In order to better understand the training method of a reflectivity factor prediction model, a training device of a reflectivity factor prediction model, a rainfall prediction method, a rainfall prediction device, an electronic device, a computer-readable storage medium, and a computer program product provided in the embodiments of the present application, the application environment applicable to the embodiments of the present application is first described below.

[0045] See also Figure 1 , Figure 1 A schematic diagram of an application environment of a rainfall prediction system provided by an embodiment of the present application is shown. As an implementation mode, the training method of the reflectivity factor prediction model and the rainfall prediction method provided by the embodiment of the present application can be applied to the same electronic device. The electronic device can be as follows: Figure 1 The server 110 shown in the figure can be connected to the terminal device 120 through a network. The network is used to provide a medium for a communication link between the server 110 and the terminal device 120. The network can include various connection types, such as a wired communication link, a wireless communication link, etc., which are not limited in the embodiments of the present application. Optionally, in other embodiments, the electronic device can also be a smart phone, a laptop computer, etc.

[0046] It should be understood that Figure 1 The server 110, network and terminal device 120 are merely illustrative. Depending on the implementation requirements, there may be any number of servers, networks and terminal devices. For example, the server 110 may be a physical server or a server cluster composed of multiple servers, and the terminal device 120 may be a mobile phone, a tablet, a desktop computer, a laptop computer, and the like. It is understood that the embodiments of the present application may also allow multiple terminal devices 120 to access the server 110 at the same time.

[0047] In addition, the rainfall prediction system may also include a memory for storing raw data, intermediate data and result data in the image processing process.

[0048] In the embodiment of the present application, the memory may be a cloud memory. Cloud storage is a new concept extended and developed from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as the storage system) refers to a storage system that uses cluster applications, grid technology, and distributed storage file systems to bring together a large number of storage devices of various types (storage devices are also called storage nodes) in the network through application software or application interfaces to work together and provide external data storage and business access functions.

[0049] At present, the storage method of the storage system is: create a logical volume, and when creating a logical volume, allocate physical storage space for each logical volume. The physical storage space may be composed of disks of a storage device or several storage devices. The client stores data on a logical volume, that is, stores the data on the file system. The file system divides the data into many parts, each of which is an object. The object contains not only data but also additional information such as data identification (ID entity, ID). The file system writes each object into the physical storage space of the logical volume, and the file system records the storage location information of each object, so that when the client requests to access the data, the file system can allow the client to access the data according to the storage location information of each object.

[0050] The process of the storage system allocating physical storage space to a logical volume is as follows: based on the estimated capacity of the objects stored in the logical volume (this estimate often has a large margin relative to the actual capacity of the objects to be stored) and the grouping of the Redundant Array of Independent Disks (RAID), the physical storage space is divided into stripes in advance. A logical volume can be understood as a stripe, thereby allocating physical storage space to the logical volume.

[0051] It should be noted that Figure 1 The scenario diagram of the rainfall prediction system shown is merely an example. The rainfall prediction system and scenario described in the embodiment of the present application are intended to more clearly illustrate the technical solution of the embodiment of the present application, and do not constitute a limitation on the technical solution provided in the embodiment of the present application. A person of ordinary skill in the art can appreciate that with the evolution of the rainfall prediction system and the emergence of new business scenarios, the technical solution provided in the embodiment of the present application is equally applicable to similar technical problems.

[0052] It should be noted that the serial numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0053] Please refer to Figure 2 , Figure 2FIG. 1 is a flow chart of an embodiment of a method for training a reflectivity factor prediction model provided in an embodiment of the present application. Figure 2 As shown, the process of the training method of the reflectivity factor prediction model provided in this application is as follows:

[0054] 201. Based on the reflectivity factor prediction model, global feature extraction and local feature extraction are performed on radar image samples to obtain global feature information and local feature information.

[0055] In the embodiment of the present application, the reflectivity factor prediction model can be a convolutional neural network (CNN) or a recurrent neural network (RNN), which can be selected according to the specific situation. A convolutional neural network is an artificial neural network architecture, which is mainly used to process grid structured data. A recurrent neural network is a neural network model that can process sequence data. A recurrent neural network has a memory function, which can retain previous information when processing time series data and apply it to the current input, thereby effectively processing the time dependency in the sequence.

[0056] In the embodiment of the present application, the radar image sample may include one or more radar images. The multiple radar images in the radar image sample may be multiple radar images sequentially acquired by the radar system at a preset frequency. Of course, in other embodiments, the multiple radar images in the radar image sample may also be multiple radar images selected from radar images sequentially acquired by the radar system and arranged at unequal time intervals.

[0057] The radar system is a dual polarization radar system. As an advanced meteorological detection method, the dual polarization radar system can provide rich precipitation particle morphology and distribution information, which is of great significance for improving the prediction accuracy of severe convective weather. In other embodiments, the radar system can also be other types of radars.

[0058] The radar image in the radar image sample may be a multi-channel image composed of multiple single-channel images, and the multiple single-channel images of the radar image include the radar reflectivity factor Z H , specific differential phase K DP and the differential reflectivity factor Z DR These data come from the dual-polarization radar system, which collects radar images at a preset frequency, which can be once every 3 minutes, once every 6 minutes, etc., depending on the specific situation.

[0059] In the embodiment of the present application, the radar image includes a single-channel image Single channel image And single channel images Single channel image Represents the radar reflectivity factor Z in the radar observation on the radar image H The value of has dimensions H×W, where H represents the height and W represents the width. Single channel image represents the differential reflectivity factor Z in the radar observation on the radar image DR The value of , its dimension is also H × W. Single channel image represents the specific differential phase K in the radar observation on the radar image DP The value of , whose dimension is also H×W.

[0060] The radar image sample may include one or more radar images. For example, the radar image sample may include 10 radar images, with a preset frequency of once every 6 minutes, and each radar image frame is Indicated by , where t represents the time step. Then the radar image sample obtained in the previous hour can be expressed as:

[0061]

[0062] Among them, X input is a radar image sample consisting of radar images from the previous hour (10 frames). The dimension of the radar image sample is 10×3×H×W. 10 represents the number of time steps. 3 represents three radar variables. H×W represents the height and width of each radar variable.

[0063] Among them, the global feature information is the global features of the radar image in the radar image sample. The global features reflect the overall properties of the image, such as color distribution, shape contour, etc., and they are usually invariant to the scale and rotation of the image. The local feature information is the local features of the radar image in the radar image sample. The local features focus on the characteristics of specific key points or regions in the image, such as corners or edges. These features help to identify specific objects and details in the image.

[0064] like Figure 3 As shown, in the embodiment of the present application, the reflectivity factor prediction model includes a local feature extraction module and a first feature fusion module. Figure 4 As shown, the local feature extraction module is a CNN model. During the training process of CNN, the network adjusts the weights of the convolution kernels through the back propagation algorithm to minimize the loss function of the classification task. CNN can indirectly adjust the weights of the input image, for example, by performing gradient ascent on the image to generate an image of a specific style. In other embodiments, the local feature extraction module can also be other networks, which can be set according to the specific situation.

[0065] In the embodiment of the present application, global feature extraction and local feature extraction are performed on radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information, including:

[0066] (1) Based on the local feature extraction module, the radar image sample is divided into at least two local images, and features are extracted from the at least two local images respectively to obtain at least two local feature maps corresponding to the at least two local images.

[0067] In an embodiment of the present application, the radar image sample can be divided into at least two local images of the same size, and feature extraction is performed on the at least two local images respectively to obtain at least two local feature maps corresponding to the at least two local images. Specifically, feature extraction can be performed through a convolutional layer in a local feature extraction module to obtain at least two local feature maps corresponding to the at least two local images.

[0068] (2) Based on the first feature fusion module, at least two local feature maps are fused to obtain local feature information.

[0069] In a specific embodiment, fusing at least two local feature maps based on a first feature fusion module to obtain local feature information includes: performing weighted fusion on at least two local feature maps based on first preset weight information corresponding to each local feature map in the first feature fusion module to obtain local feature information, wherein the first preset weight information corresponding to each local feature map in the first feature fusion module is iteratively updated during iterative training of the reflectivity factor prediction model. The first preset weight information may be a weight matrix of weight coefficients of each eigenvalue on the local feature map.

[0070] In other embodiments, a fixed weight coefficient may be pre-set for each local feature map, and multiple local feature maps may be weighted and fused based on the weight coefficient of each local feature map to obtain local feature information.

[0071] like Figure 3 As shown, in the embodiment of the present application, the reflectivity factor prediction model includes a global feature extraction module. Figure 4 As shown in the figure, the global feature extraction module is a FCNN model. FCNN is usually used for tasks such as image segmentation, where each pixel of the input image needs to be associated with a specific category or region. When training FCNN, the input image and its corresponding label are usually used as training samples, and the network parameters are adjusted through the back propagation algorithm so that the network can accurately predict the category of each pixel. Therefore, FCNN can adjust the weights of the input image variables through training so that it can better adapt to specific tasks, such as image segmentation.

[0072] In an embodiment of the present application, global feature extraction and local feature extraction are performed on radar image samples based on a reflectivity factor prediction model to obtain global feature information and local feature information, including: global feature extraction is performed on radar image samples based on a global feature extraction module to obtain global feature information.

[0073] This application introduces CNN and FCNN to evaluate the contribution of dual-polarization radar data in the nowcasting of severe convective precipitation and optimize the data fusion strategy. By introducing CNN, the observations of the dual-polarization radar are feature extracted, and the local information is modeled to extract local feature information. After CNN, FCNN is used to extract global feature information. CNN and FCNN work together to perform multi-level feature extraction on radar images, including global structural features and local detail features. Through the above feature extraction process, the model can quantitatively evaluate the contribution of dual-polarization radar data and optimize the data fusion strategy. The features extracted by CNN and FCNN can be regarded as information of different dimensions. They are integrated within the model through weighted or fusion layers to achieve efficient use of dual-polarization information and improve the accuracy of the model for nowcasting of severe convective precipitation. Finally, parameter adjustment is performed to improve training efficiency and stability to meet the needs of real-time prediction.

[0074] In the embodiment of the present application, in order to obtain radar image samples and annotated reflectivity factor information of the radar image samples, global feature extraction and local feature extraction are performed on the radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information, before, including:

[0075] (1) Acquire a plurality of first radar images and a plurality of second radar images acquired by a radar system, wherein the acquisition time of the plurality of second radar images is later than the acquisition time of the plurality of first radar images.

[0076] In the embodiment of the present application, the multiple first radar images may be a plurality of continuous radar images taken by the radar system at a preset frequency, and the multiple second radar images may be a plurality of continuous radar images taken by the radar system at a preset frequency.

[0077] In an embodiment of the present application, the radar image with the latest acquisition time among the multiple first radar images and the radar image with the earliest acquisition time among the multiple second radar images may be two adjacent frames of radar images. For example, the multiple first radar images are the 1st to 10th frames of radar images acquired by the radar system at a preset frequency, and the multiple second radar images are the 11th to 20th frames of radar images acquired by the radar system at a preset frequency. In other embodiments, the radar image with the latest acquisition time among the multiple first radar images and the radar image with the earliest acquisition time among the multiple second radar images may be two frames of radar images separated by at least one frame. For example, the multiple first radar images are the 1st to 10th frames of radar images acquired by the radar system at a preset frequency, and the multiple second radar images are the 16th to 25th frames of radar images acquired by the radar system at a preset frequency.

[0078] (2) The multiple first radar images are determined as a radar sequence image, and the real reflectivity factor information in the multiple second radar images is determined as the annotated reflectivity factor information of the radar sequence image, so as to obtain the multiple radar sequence images and the corresponding annotated reflectivity factor information.

[0079] In the embodiment of the present application, the real reflectivity factor information in the plurality of second radar images includes the real reflectivity factor information of each second radar image, and the real reflectivity factor information of the second radar image is the real reflectivity factor information when the second radar image is collected. The real reflectivity factor information when the second radar image is collected can be obtained based on the single-channel image when the second radar image is collected. Determine, for example,

[0080] Single channel images of each second radar image They are Where t represents the current time. The mathematical representation of constructing the output label is:

[0081]

[0082] Among them, Y output is a Z value that contains the value of Z when the second radar image is acquired. H The sequence of data.

[0083] In the embodiment of the present application, the annotated reflectivity factor information of the radar sequence image is the real reflectivity factor information in the plurality of second radar images. Of course, the annotated reflectivity factor information of the radar sequence image may also be the real reflectivity factor information in the plurality of first radar images, which can be determined according to different prediction tasks.

[0084] (3) Determine a first preset number of radar sequence images from the plurality of radar sequence images as a first preset number of radar image samples, and obtain the first preset number of radar image samples and corresponding annotated reflectivity factor information.

[0085] In the embodiment of the present application, the first preset number may be a first preset ratio of the radar sequence image, and the first preset ratio may be 60%, 70%, or 1, which is set according to specific circumstances.

[0086] In a specific embodiment, the first preset number may be a first preset ratio of radar sequence images, and the first preset ratio is less than 1. The first preset number of radar sequence images in the plurality of radar sequence images is determined as the first preset number of radar image samples, and the first preset number of radar image samples and the corresponding annotated reflectivity factor information are obtained. The second preset number of radar sequence images in the plurality of radar sequence images is determined as the second preset number of test radar images, and the second preset number of test radar images and the corresponding annotated reflectivity factor information are obtained. Based on the second preset number of test radar images and the corresponding annotated reflectivity factor information, whether the reflectivity factor prediction model meets the preset test condition is tested. When the reflectivity factor prediction model meets the preset test condition, the reflectivity factor prediction model is output, otherwise the reflectivity factor prediction model is retrained. Among them, the ratio of the first preset number to the second preset number may be 7:3 or other values, which can be set according to specific circumstances. Among them, the preset test condition may be that the test recall rate (Recall) is greater than the preset recall rate or the test precision rate (Precision) is greater than the preset precision rate.

[0087] Sudden and localized severe convective weather refers to sudden and strong meteorological phenomena within a specific time and space range. Suddenness refers to the sudden and rapid occurrence of weather phenomena, which is often difficult to accurately predict and warn in advance. This may be caused by drastic changes in atmospheric conditions in a short period of time. Local weather phenomena usually occur in a relatively small geographical area, and weather conditions in nearby areas may not change significantly. This may be due to local influences of terrain and airflow. The suddenness and localization of such weather phenomena have a significant impact on local life, transportation, agriculture and other aspects, so timely monitoring, early warning and disaster response measures are needed to ensure public safety, but existing technologies often lack special model training for such weather.

[0088] In order to improve the prediction of sudden severe convective weather and local severe convective weather, in a specific embodiment, a first preset number of radar sequence images in a plurality of radar sequence images are determined as a first preset number of radar image samples, and the first preset number of radar image samples and corresponding annotated reflectivity factor information are obtained, including:

[0089] (1) The radar sequence images whose weather types belong to the preset weather types are determined as candidate sequence images.

[0090] Among them, the weather type can be sudden severe convective weather, local severe convective weather, non-sudden severe convective weather, and non-local severe convective weather. The conditions of different weather types can be preset. The preset weather types include sudden severe convective weather and local severe convective weather.

[0091] In the embodiment of the present application, the radar echo intensity, wind field information, temperature and humidity of the radar sequence image are obtained. If the radar sequence image meets the sudden condition, the weather type is determined to be sudden severe convective weather, and the weather type of the radar sequence image is determined to belong to the preset weather type. If the radar sequence image meets the local condition, the weather type is determined to be local severe convective weather, and the weather type of the radar sequence image is determined to belong to the preset weather type. Among them, the sudden condition and the local condition can be set according to the specific situation.

[0092] In a specific embodiment, the suddenness criterion is that the radar echo intensity of each radar image in the radar sequence image increases by more than a preset threshold within a predetermined time. The local condition is that the storm geographical range parameter of the storm impact of the radar sequence image is greater than the preset parameter, and the storm geographical range parameter of the radar sequence image can be the average value of the storm area ratio of each radar image in the radar sequence image. The storm area ratio is the area ratio of the storm impact area in the radar image. The storm impact area is the area where the storm occurs in each radar image in the radar sequence image.

[0093] (2) Randomly selecting a first preset number of candidate sequence images as a first preset number of radar image samples.

[0094] Specifically, a first preset number of candidate sequence images are randomly selected as a first preset number of radar image samples to obtain a first preset number of radar image samples and corresponding annotated reflectivity factor information, and a second preset number of candidate sequence images are randomly selected as a second preset number of test radar images to obtain a second preset number of test radar images and corresponding annotated reflectivity factor information. Based on the second preset number of test radar images and the corresponding annotated reflectivity factor information, whether the reflectivity factor prediction model meets the preset test conditions is tested, and when the reflectivity factor prediction model meets the preset test conditions, the reflectivity factor prediction model is output, otherwise the reflectivity factor prediction model is retrained.

[0095] By randomly screening the original samples for sudden and local designated samples, it is ensured that the training set and test set can fully reflect the characteristics of sudden and local severe convective weather, thereby improving the applicability of the model in practical applications.

[0096] 202. Based on the reflectivity factor prediction model, the global feature information and the local feature information are fused to obtain the sample feature information.

[0097] In an embodiment of the present application, the reflectivity factor prediction model fusion includes a second feature fusion module, which fuses global feature information and local feature information based on the reflectivity factor prediction model to obtain sample feature information, including: weighted fusion of global feature information and local feature information based on second preset weight information in the second feature fusion module to obtain sample feature information. The second preset weight information in the second feature fusion module is iteratively updated during iterative training of the reflectivity factor prediction model. The second preset weight information includes a first weight matrix for each eigenvalue on the global feature information and a second weight matrix for each eigenvalue on the local feature information.

[0098] In other embodiments, fixed weight coefficients may be pre-set for the global feature information and the local feature information, and the global feature information and the local feature information may be weighted and summed based on the weight coefficients of the global feature information and the local feature information to obtain the sample feature information.

[0099] 203. Perform reflectivity factor prediction based on the reflectivity factor prediction model and sample feature information to obtain first predicted reflectivity factor information of the radar image sample.

[0100] In the embodiment of the present application, the first predicted reflectivity factor information may be based on a plurality of predicted reflectivity factor information predicted by the radar image sample. The predicted reflectivity factor information may be a single channel image Z H , single channel image Z H The dimension is H×W, where H represents the height and W represents the width. The single-channel image Z H Includes H×W predicted radar reflectivity factors.

[0101] like Figure 3 As shown, in the embodiment of the present application, the reflectivity factor prediction model includes a time recurrent neural network and a residual network, such as Figure 4 As shown, the residual network includes two convolutional layers, and the two convolutional layers are Conv2d layers.

[0102] In an embodiment of the present application, reflectivity factor prediction is performed based on sample feature information and a reflectivity factor prediction model to obtain first predicted reflectivity factor information, including: extracting features from sample feature information based on a residual network to obtain extracted features; fusing the extracted features with the sample feature information to obtain residual feature information; and predicting reflectivity factors of radar image samples based on a time recurrent neural network and the residual feature information to obtain first predicted reflectivity factor information.

[0103] The introduction of residual connection can effectively slow down the loss of feature information, which is especially important for prediction problems such as long-term series in severe convection forecast. Based on the residual network, the sample feature information is extracted to obtain the extracted features; the extracted features and sample feature information are fused to obtain the residual feature information C. The calculation formula is as follows:

[0104] C=σ(W*X+b ′ )+X,

[0105] Among them, C is the residual feature information after residual connection, W is the convolution kernel, X is the sample feature information, b′ is the bias term, σ is the activation function, (W*X+b ′ ) is the extracted feature.

[0106] In an embodiment of the present application, the reflectivity factor prediction model may include two or more residual networks. By introducing the residual network, the deep network can generally capture higher-level features and abstract patterns in the data. In addition, the representation ability of the model is improved, enabling it to handle complex tasks. The skip connection of the residual network helps to reduce the risk of overfitting of the model. The model is easier to learn the identity mapping and improves the generalization of the model.

[0107] like Figure 3 As shown, in the embodiment of the present application, the time recurrent neural network includes an encoder, a self-attention unit, and a decoder. The encoder includes at least two long short-term memory artificial neural networks connected in series, and the decoder includes at least two long short-term memory artificial neural networks connected in series. Figure 4 As shown, the encoder includes 3 serially connected ConvLSTM modules, and the decoder includes 2 serially connected ConvLSTM modules.

[0108] In the embodiment of the present application, the reflectivity factor of the radar image sample is predicted based on the time recurrent neural network and the residual feature information to obtain the first predicted reflectivity factor information, including:

[0109] (1) Based on the encoder, the residual feature information is mapped to the first hidden state vector.

[0110] Specifically, the encoder maps the residual feature information C to the first hidden state vector H1 in the representation space. The specific formula is:

[0111] H1=Encoder(C).

[0112] Among them, the calculation process of the multi-layer ConvLSTM module can be expressed by the following formula:

[0113]

[0114] Among them, ConvLSTM lrepresents the ConvLSTM unit of layer l, H l and C l are the hidden state and cell state of the lth layer respectively.

[0115] (2) Based on the self-attention unit, the weights of each parameter in the first hidden state vector are adjusted to obtain the second hidden state vector.

[0116] Specifically, the second hidden state vector is denoted as H2.

[0117] (3) Based on the decoder, the second hidden state vector is mapped into a mapping feature vector.

[0118] The decoder maps the second hidden state vector H2 back to the output sequence Y, that is, the specific formula is,

[0119] Y=Decoder(H2)).

[0120] The Time Recurrent Neural Network is a Sequence-to-Sequence (Seq2Seq, S2S) model. The Seq2Seq model is a deep learning model architecture used to handle sequence-to-sequence tasks. It is also known as a special case of the Encoder-Decoder model. Using this Seq2Seq model as the generator of GAN means that it is part of the GAN network, which is responsible for generating prediction results similar to real data. During the training process, its goal is to simulate the distribution of real data as much as possible to deceive the discriminator. Finally, by introducing the discriminator, the model can improve the feature learning ability of the Seq2Seq model through adversarial training.

[0121] (4) Based on the output layer processing mapping feature vector, the first predicted reflectivity factor information is obtained.

[0122] 204. Iteratively train a reflectivity factor prediction model based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

[0123] In a specific embodiment, the training stop condition is that the first loss value determined based on the labeled reflectivity factor information and the first predicted reflectivity factor information is not greater than a first preset value or the number of iterations of the reflectivity factor prediction model is greater than a preset number.

[0124] Specifically, a first loss value is determined based on the labeled reflectivity factor information and the first predicted reflectivity factor information of the radar image sample using a first preset loss function. When the first loss value is greater than the first preset loss value, the network parameters of the reflectivity factor prediction model are gradient updated to obtain an updated reflectivity factor prediction model. The new first predicted reflectivity factor information of the radar image sample is calculated based on the updated reflectivity factor prediction model. A new first loss value is determined based on the labeled reflectivity factor information and the new first predicted reflectivity factor information of the radar image sample using the first preset loss function. When the first loss value is not greater than the first preset loss value or the number of iterations of the reflectivity factor prediction model is greater than the preset number, the reflectivity factor prediction model is determined to meet the training stop condition. For example, the first preset loss function can be the mean square error. The mean square error is one of the commonly used loss functions in regression problems. It measures the average square difference between the predicted value and the actual value. Therefore, the loss function selected during the model training process is the mean square error. Controlling the change of the learning rate during the training process helps to improve the stability and efficiency of the training.

[0125] In another specific embodiment, in order to improve the accuracy of the prediction of the reflectivity factor prediction model, the reflectivity factor prediction model is iteratively trained based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until the training stop condition is met, including:

[0126] (1) A generative adversarial network model is constructed using the reflectivity factor prediction model as the generator and the preset neural network model as the discriminator.

[0127] Generative Adversarial Network (GAN) is a deep learning model consisting of two neural networks: the generator and the discriminator. They compete with each other and learn from each other. Through continuous adversarial training, the generator's ability to generate realistic samples is improved, while the discriminator's ability to identify true and false samples is improved. The discriminator is an independent neural network whose goal is to distinguish between generated predictions and real data. During the training process, the discriminator receives a pair of inputs: generated predictions and real data. It learns how to distinguish between the two. The loss function of the discriminator includes the cross entropy loss when it correctly classifies generated results and real data. At the same time, the goal of the generator is to deceive the discriminator so that it cannot accurately distinguish between generated predictions and real data. This leads to an adversarial training process, in which an adversarial network is formed between the generator and the discriminator, with the goal of making the generator generate more realistic predictions. Through this adversarial training mechanism, the generator will continuously improve its feature learning ability, thereby better simulating the real data distribution and improving the accuracy and reliability of predictions.

[0128] In the embodiment of the present application, the reflectivity factor prediction model is used as a generator, and the preset neural network model is used as a discriminator to construct a generative adversarial network model. The network structure of the discriminator can be set according to the specific situation. The discriminator can be a convolutional long short-term memory network (Convolutional Long Short-Term Memory, ConvLSTM), a long short-term memory network (Long Short-Term Memory, LSTM) or other networks.

[0129] Specifically, Figure 5 As shown, the preset neural network model includes three Conv2d layers and four ConvLSTM modules connected in series.

[0130] (2) Iteratively train the generative adversarial network model based on the annotated reflectivity factor information of the radar image samples and the first predicted reflectivity factor information until the training stop condition is met.

[0131] In the embodiment of the present application, the training stop condition is determined to be satisfied when the second loss value of the marked reflectivity factor information and the first predicted reflectivity factor information is not greater than the second preset loss value or the number of iterations of the reflectivity factor prediction model is greater than the preset number.

[0132] In an embodiment of the present application, a second loss value is determined based on the annotated reflectivity factor information and the first predicted reflectivity factor information of the radar image sample using a second preset loss function. When the second loss value is greater than the second preset loss value, the network parameters of the generator are gradient updated to obtain an updated generator. New first predicted reflectivity factor information is generated based on the updated generator. A new second loss value is determined based on the new first predicted reflectivity factor information and the annotated reflectivity factor information. When the second loss value is greater than the second preset loss value, the network parameters of the discriminator are gradient updated to obtain an updated discriminator. The generator and the discriminator are updated alternately and iteratively until the new second loss value is not greater than the second preset loss value or the number of iterations of the reflectivity factor prediction model is greater than the preset number of times, it is determined that the training stop condition is met.

[0133] Specifically, Figure 5 As shown, the reflectivity factor prediction model and the preset neural network model in the generative adversarial network model are respectively used as a generator and a discriminator for adversarial training until the training stop condition is met to obtain the reflectivity factor prediction model.

[0134] Further, in order to improve the training effect, the first preset number of radar image samples is divided into multiple batches, each batch includes a third preset number of radar image samples. For example, the third preset number is 64. The reflectivity factor prediction model is trained multiple times using multiple batches of radar image samples.

[0135] Furthermore, in order to make the training process more stable and efficient, and to adapt to the needs of more different scenarios. In model training, the gradients of multiple batches of radar image samples are accumulated before the reflectivity factor prediction model is updated. Usually used in memory-constrained situations, the gradient accumulation of multiple small batches can be used to simulate the training process of a large batch. Through gradient accumulation, the number of samples in a single batch can be reduced, thereby reducing memory usage, so that large models can be trained on memory-constrained devices.

[0136] Please refer to Figure 6 , Figure 6 FIG. 1 is a flow chart of an embodiment of a rainfall prediction method provided in an embodiment of the present application. Figure 6 As shown, the process of the rainfall prediction method provided by this application is as follows:

[0137] 401. Obtain a radar image to be predicted.

[0138] In the embodiment of the present application, the number of radar images to be predicted may be one or multiple radar images collected by the radar system, and may be set according to specific circumstances.

[0139] 402. Predict the reflectivity factor of the radar image to be predicted based on the reflectivity factor prediction model to obtain second predicted reflectivity factor information.

[0140] The reflectivity factor prediction model is the reflectivity factor prediction model of any one of the above embodiments.

[0141] Specifically, the radar image to be predicted is input into the reflectivity factor prediction model to obtain the second predicted reflectivity factor information. The second predicted reflectivity factor information can be based on multiple predicted reflectivity factor information predicted by the radar image to be predicted, and each predicted reflectivity factor information in the second predicted reflectivity factor information is the predicted reflectivity factor information at multiple different time points after the end time of the acquisition of the radar image to be predicted. The predicted reflectivity factor information can be a single channel image Z H , single channel image Z H The dimension is H×W, where H represents the height and W represents the width. The single-channel image Z H Includes H×W predicted radar reflectivity factors.

[0142] 403. Determine predicted rainfall information corresponding to the radar image to be predicted based on the second predicted reflectivity factor information.

[0143] The predicted rainfall information includes predicted rainfall feature maps at multiple different time points after the end time of the radar image acquisition to be predicted. The dimension of the predicted rainfall feature map is H×W, where H represents height and W represents width. The predicted rainfall feature map includes H×W predicted rainfall values.

[0144] In an embodiment of the present application, the predicted rainfall information corresponding to the radar image to be predicted is determined based on the first predicted reflectivity factor information, including: determining the predicted rainfall information corresponding to the radar image to be predicted based on a preset reflectivity-rainfall relationship and the first predicted reflectivity factor information, the reflectivity-rainfall relationship including a mapping relationship between the radar reflectivity factor and the rainfall.

[0145] The relationship between reflectivity and rainfall can be expressed by the following formula:

[0146] R=a·Z b ,

[0147] Among them, R represents the precipitation intensity, usually in millimeters / hour or inches / hour. Z represents the radar reflectivity factor, in decibels dBZ. a and b are empirical parameters that usually need to be calibrated according to specific regions and precipitation types. The values ​​of these parameters depend on the characteristics of different places and different precipitation types.

[0148] Substitute each predicted radar reflectivity factor in the first predicted reflectivity factor information into the reflectivity-rainfall relationship to obtain predicted rainfall information.

[0149] The present application predicts the predicted rainfall information after the acquisition end time through the radar image to be predicted, and can implement the subsequent rainfall information based on the acquired radar image to be predicted.

[0150] In order to facilitate better implementation of the training method of the reflectivity factor prediction model provided in the embodiment of the present application, the embodiment of the present application also provides a training device for the reflectivity factor prediction model based on the training method of the reflectivity factor prediction model. The meanings of the terms are the same as those in the training method of the reflectivity factor prediction model. For specific implementation details, please refer to the description in the above method embodiment.

[0151] Please refer to Figure 7 , Figure 7 70 is a schematic diagram of a structure of an embodiment of a training device for a reflectivity factor prediction model provided in an embodiment of the present application. The training device for a reflectivity factor prediction model may include a feature extraction unit 701, a feature fusion unit 702, a prediction unit 703, and a training unit 704, wherein:

[0152] A feature extraction unit 701 is used to perform global feature extraction and local feature extraction on the radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information;

[0153] A feature fusion unit 702 is used to fuse the global feature information and the local feature information based on the reflectivity factor prediction model to obtain sample feature information;

[0154] A prediction unit 703 is used to perform reflectivity factor prediction based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample;

[0155] The training unit 704 is used to iteratively train the reflectivity factor prediction model based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

[0156] The specific implementation of each of the above modules can be found in the previous embodiments and will not be described in detail here.

[0157] Please refer to Figure 8 , Figure 8 8 is a schematic diagram of the structure of an embodiment of a rainfall prediction device provided in an embodiment of the present application. The rainfall prediction device may include a feature extraction unit 801, a reflectivity prediction module 802, and a rainfall calculation module 803, wherein:

[0158] The feature extraction unit 801 is used for an acquisition module, which is used to acquire the radar image to be predicted;

[0159] A reflectivity prediction module 802 is used to predict the reflectivity factor of the radar image to be predicted based on the reflectivity factor prediction model to obtain second predicted reflectivity factor information;

[0160] The rainfall calculation module 803 is used to determine the predicted rainfall corresponding to the radar image to be predicted based on the second predicted reflectivity factor information.

[0161] The specific implementation of each of the above modules can be found in the previous embodiments and will not be described in detail here.

[0162] In the present application, compared with the related art, global feature extraction and local feature extraction are performed on radar image samples based on a reflectivity factor prediction model to obtain global feature information and local feature information; the global feature information and the local feature information are fused based on the reflectivity factor prediction model to obtain sample feature information; reflectivity factor prediction is performed based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample; the reflectivity factor prediction model is iteratively trained based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until the training stop condition is met. The present application performs global feature extraction and local feature extraction on radar image samples respectively, and fuses the global feature information and the local feature information to obtain sample feature information; then, based on the reflectivity factor prediction model and the sample feature information, the reflectivity factor is predicted to obtain the first predicted reflectivity factor information of the radar image sample; finally, the reflectivity factor prediction model is iteratively trained according to the annotated reflectivity factor information and the first predicted reflectivity factor information of the radar image sample; since the global feature information and the local feature information can globally and locally characterize the features of the radar image samples, sample feature information with richer features can be extracted from the radar image samples, so that a more accurate reflectivity factor prediction model can be trained to predict a more accurate radar reflectivity factor, so that a more accurate rainfall can be predicted using a more accurate radar reflectivity factor, thereby improving the accuracy of rainfall prediction.

[0163] Please refer to Fig. 9 , Fig. 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0164] The electronic device may include components such as a processor 101 with one or more processing cores, a memory 102 with one or more computer-readable storage media, a power supply 103, and an input unit 104. Those skilled in the art will appreciate that the electronic device structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0165] The processor 101 is the control center of the electronic device, which uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 102, and calling data stored in the memory 102. Optionally, the processor 101 may include one or more processing cores; optionally, the processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 101.

[0166] The memory 102 can be used to store software programs and modules. The processor 101 executes various functional applications and the training of the reflectivity factor prediction model by running the software programs and modules stored in the memory 102. The memory 102 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 102 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 102 may also include a memory controller to provide the processor 101 with access to the memory 102.

[0167] The electronic device also includes a power supply 103 for supplying power to each component. Optionally, the power supply 103 can be logically connected to the processor 101 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 103 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.

[0168] The electronic device may further include an input unit 104, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0169] Although not shown, the electronic device may also include a display unit, an image acquisition element, etc., which will not be described in detail here. Specifically in this embodiment, the processor 101 in the electronic device will load the executable code corresponding to one or more computer programs into the memory 102 according to the following instructions, and the processor 101 will execute the steps of the method provided in this application, such as:

[0170] Based on the reflectivity factor prediction model, global feature extraction and local feature extraction are respectively performed on the radar image samples to obtain global feature information and local feature information; based on the reflectivity factor prediction model, the global feature information and the local feature information are fused to obtain sample feature information; based on the reflectivity factor prediction model and the sample feature information, reflectivity factor prediction is performed to obtain first predicted reflectivity factor information of the radar image samples; based on the annotated reflectivity factor information of the radar image samples and the first predicted reflectivity factor information, the reflectivity factor prediction model is iteratively trained until a training stop condition is met;

[0171] Alternatively, obtaining a radar image to be predicted;

[0172] Based on the reflectivity factor prediction model, the reflectivity factor of the radar image to be predicted is predicted to obtain second predicted reflectivity factor information;

[0173] The predicted rainfall information corresponding to the radar image to be predicted is determined based on the second predicted reflectivity factor information.

[0174] It should be noted that the electronic device provided in the embodiment of the present application belongs to the same concept as the reflectivity factor prediction in the above embodiment, and its specific implementation process is detailed in the above related embodiments and will not be repeated here.

[0175] The present application also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program stored thereon is executed on a processor of an electronic device provided in an embodiment of the present application, the processor of the electronic device executes the steps in the reflectivity factor prediction provided in the present application. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0176] The present application also provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs various optional implementations of the above-mentioned reflectivity factor prediction.

[0177] The above is a detailed introduction to the training method of a reflectivity factor prediction model, a rainfall prediction method and a device provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application.

[0178] It should be noted that when the above embodiments of the present application are applied to specific products or technologies, the relevant data of the user is involved, and the user's permission or consent is required, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant region.

Claims

1. A method for training a reflectivity factor prediction model, characterized in that: The training method of the reflectivity factor prediction model includes: Based on the reflectivity factor prediction model, global feature extraction and local feature extraction are performed on radar image samples to obtain global feature information and local feature information; Based on the reflectivity factor prediction model, the global feature information and the local feature information are fused to obtain sample feature information; Performing reflectivity factor prediction based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample; The reflectivity factor prediction model is iteratively trained based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

2. The method for training a reflectivity factor prediction model according to claim 1, characterized in that: The reflectivity factor prediction model includes a local feature extraction module and a first feature fusion module. The global feature extraction and local feature extraction are performed on the radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information, including: Based on the local feature extraction module, the radar image sample is divided into at least two local images, and features are extracted from the at least two local images respectively to obtain at least two local feature maps corresponding to the at least two local images; Based on the first feature fusion module, at least two of the local feature maps are fused to obtain local feature information.

3. The method for training a reflectivity factor prediction model according to claim 2, characterized in that: The fusing at least two of the local feature maps based on the first feature fusion module to obtain local feature information includes: Based on the first preset weight information corresponding to each of the local feature maps in the first feature fusion module, at least two of the local feature maps are weightedly fused to obtain the local feature information, wherein the first preset weight information corresponding to each of the local feature maps in the first feature fusion module is iteratively updated during the iterative training of the reflectivity factor prediction model.

4. The method for training a reflectivity factor prediction model according to claim 1, characterized in that: The iterative training of the reflectivity factor prediction model based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is satisfied includes: Using the reflectivity factor prediction model as a generator and the preset neural network model as a discriminator, a generative adversarial network model is constructed; The generative adversarial network model is iteratively trained based on the annotated reflectivity factor information and the first predicted reflectivity factor information of the radar image sample until a training stop condition is met.

5. The method for training a reflectivity factor prediction model according to claim 1, characterized in that: The reflectivity factor prediction model includes a time recurrent neural network and a residual network, and the reflectivity factor prediction is performed based on the reflectivity factor prediction model and the sample feature information to obtain the first predicted reflectivity factor information of the radar image sample, including: Performing feature extraction on the sample feature information based on the residual network to obtain extracted features; Fusion of the extracted features and the sample feature information to obtain residual feature information; The reflectivity factor of the radar image sample is predicted based on the time recurrent neural network and the residual feature information to obtain first predicted reflectivity factor information.

6. The method for training a reflectivity factor prediction model according to claim 5, characterized in that: The time recurrent neural network includes an encoder, a self-attention unit, a decoder, and an output layer. The method predicts the reflectivity factor of the radar image sample based on the time recurrent neural network and the residual feature information to obtain first predicted reflectivity factor information, including: Mapping the residual feature information into a first hidden state vector based on the encoder; Adjusting the weights of each parameter in the first hidden state vector based on the self-attention unit to obtain a second hidden state vector; Mapping the second hidden state vector to a mapping feature vector based on the decoder; The mapping feature vector is processed based on the output layer to obtain first predicted reflectivity factor information.

7. The method for training a reflectivity factor prediction model according to claim 6, characterized in that: The encoder includes at least two serially connected long short-term memory artificial neural networks, and the decoder includes at least two serially connected long short-term memory artificial neural networks.

8. The method for training a reflectivity factor prediction model according to claim 1, characterized in that: The training method of the reflectivity factor prediction model also includes: Acquire a plurality of first radar images and a plurality of second radar images acquired by a radar system, wherein the acquisition time of the plurality of second radar images is later than the acquisition time of the plurality of first radar images; Determine the plurality of first radar images as one radar sequence image, and determine the real reflectivity factor information in the plurality of second radar images as the annotated reflectivity factor information of the radar sequence images, to obtain the plurality of radar sequence images and the corresponding annotated reflectivity factor information; A first preset number of the radar sequence images among the plurality of radar sequence images are determined as a first preset number of radar image samples, and the first preset number of radar image samples and corresponding annotated reflectivity factor information are obtained.

9. The method for training a reflectivity factor prediction model according to claim 8, characterized in that: The step of determining a first preset number of the radar sequence images in the plurality of radar sequence images as a first preset number of radar image samples, and obtaining the first preset number of radar image samples and corresponding annotated reflectivity factor information, comprises: Determine the radar sequence images whose weather types belong to preset weather types as candidate sequence images; A first preset number of the candidate sequence images are randomly selected as a first preset number of the radar image samples.

10. A rainfall prediction method, characterized in that: The rainfall prediction method comprises: Acquire the radar image to be predicted; Performing reflectivity factor prediction on the radar image to be predicted based on the reflectivity factor prediction model to obtain second predicted reflectivity factor information, wherein the reflectivity factor prediction model is the reflectivity factor prediction model according to any one of claims 1 to 9; The predicted rainfall information corresponding to the radar image to be predicted is determined based on the second predicted reflectivity factor information.

11. The rainfall prediction method according to claim 10, characterized in that: The determining, based on the second predicted reflectivity factor information, the predicted rainfall information corresponding to the to-be-predicted radar image comprises: The predicted rainfall information corresponding to the radar image to be predicted is determined based on a preset reflectivity-rainfall relationship and the second predicted reflectivity factor information, wherein the reflectivity-rainfall relationship includes a mapping relationship between a radar reflectivity factor and rainfall.

12. A training device for a reflectivity factor prediction model, characterized in that: The training device of the reflectivity factor prediction model comprises: A feature extraction unit, used to perform global feature extraction and local feature extraction on the radar image samples based on the reflectivity factor prediction model to obtain global feature information and local feature information; A feature fusion unit, used for fusing the global feature information and the local feature information based on the reflectivity factor prediction model to obtain sample feature information; A prediction unit, configured to perform reflectivity factor prediction based on the reflectivity factor prediction model and the sample feature information to obtain first predicted reflectivity factor information of the radar image sample; A training unit is used to iteratively train the reflectivity factor prediction model based on the annotated reflectivity factor information of the radar image sample and the first predicted reflectivity factor information until a training stop condition is met.

13. A rainfall prediction device, characterized in that: The rainfall prediction device comprises: An acquisition module, used for acquiring the radar image to be predicted; A reflectivity prediction module, used to predict the reflectivity factor of the radar image to be predicted based on a reflectivity factor prediction model to obtain second predicted reflectivity factor information, wherein the reflectivity factor prediction model is the reflectivity factor prediction model according to any one of claims 1 to 9; A rainfall calculation module is used to determine the predicted rainfall corresponding to the radar image to be predicted based on the second predicted reflectivity factor information.

14. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the training method of the reflectivity factor prediction model described in any one of claims 1 to 9 and the steps in the rainfall prediction method described in any one of claims 10 to 11.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, which are suitable for loading by a processor to execute the steps in the training method of the reflectivity factor prediction model described in any one of claims 1 to 9 and the rainfall prediction method described in any one of claims 10 to 11.

16. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the training method of the reflectivity factor prediction model described in any one of claims 1 to 9 and the steps in the rainfall prediction method described in any one of claims 10 to 11 are implemented.

Images