Precipitation nowcasting method fusing three-dimensional radar reflectivity and environmental element field

Abstract

The invention relates to a rainfall nowcasting method fusing three-dimensional radar reflectivity and an environmental element field. The rainfall nowcasting method comprises the following steps: 1) collecting radar reflectivity data, rainfall live data and environmental field element data; 2) converting the radar reflectivity data into three-dimensional data, converting rainfall live data into a two-dimensional rainfall grid point field, converting environmental field element data into a three-dimensional equidistant grid point field, and making a sample set by taking the radar reflectivity three-dimensional data and the environmental field element data as input and the two-dimensional rainfall grid point field as a label; 3) building a rainfall nowcasting model; 4) dividing a training set, a verification set and a test set, training the model on the training set, performing hyper-parameter tuning on the verification set, and performing model test on the test set; and 5) taking the radar reflectivity data and the environmental field data as model input, wherein the model output is a rainfall nowcasting field. The influence of distribution and evolution of water vapor particles in the environment field and the vertical direction on rainfall is fully considered, and the rainfall nowcasting precision is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of meteorology and machine learning technology, and specifically relates to a precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental element fields. Background Art

[0002] Precipitation nowcasting plays an important role in weather forecasting, flood prevention and other fields. Its task is to forecast the precipitation intensity at the pixel level in the forecast area based on meteorological observation data. Accurate precipitation nowcasting is the key to precipitation nowcasting business, which is of great significance for preventing disastrous severe convective precipitation, reducing social and economic losses, and protecting people's lives and property.

[0003] Radar observation data has the advantages of high temporal and spatial resolution, rich information, and wide coverage. Radar reflectivity is widely used in meteorological services, and the strength of radar reflectivity is closely related to weather changes. Therefore, radar observation data is an important basis for precipitation nowcasting. At present, the mainstream precipitation nowcasting services are mostly based on CAPPI radar reflectivity data at an altitude of two kilometers, but the single-layer radar reflectivity data can only reflect the approximate distribution of water vapor particles in the horizontal direction at an altitude of two kilometers. Since cloud monomers exist in three dimensions, the distribution of water vapor particles in the vertical direction and their evolution are also crucial to precipitation. At the same time, the formation of severe convective precipitation is related to the strong vertical movement of air. Therefore, using three-dimensional radar reflectivity data as the input of the model can enable the model to learn the changing process of convective weather in the entire three-dimensional space, thereby improving the accuracy of precipitation nowcasting.

[0004] If we only use radar observation data and ignore the impact of real-time environmental factors on convective weather changes, the accuracy of model forecasts will be limited. Storm dynamic structure and development trend depend to a large extent on thermal instability, vertical wind shear and vertical distribution of water vapor. It is difficult to accurately predict its generation, extinction, splitting and merging based on the change information of radar observation data itself.

[0005] Therefore, it is urgent to develop a precipitation nowcasting method that integrates three-dimensional radar reflectivity and environmental factor fields. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a precipitation nowcasting method that integrates three-dimensional radar reflectivity and environmental factor fields, making full use of the three-dimensional radar reflectivity and environmental field element information detected by meteorological radar, while introducing a specific structure into the model to make it pay more attention to the three-dimensional cloud clusters in the radar reflectivity data that are directly related to the precipitation, comprehensively improving the accuracy and interpretability of precipitation nowcasting, and improving the level of precipitation nowcasting.

[0007] The present invention solves the technical problem by the following technical solutions:

[0008] A precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field, the method comprising the following steps:

[0009] S1. Collect radar reflectivity data and the corresponding precipitation data of the next hour on an hourly basis, and collect environmental field element data corresponding to the radar reflectivity data, including divergence, vertical velocity, potential vorticity, potential height, specific humidity, relative humidity, and temperature.

[0010] S2, converting the radar reflectivity data at each radar scanning moment from the polar coordinate system into three-dimensional data in the Cartesian rectangular coordinate system, converting the matching real-time precipitation data of the next hour into a two-dimensional precipitation grid field by interpolation, converting the environmental field element data into a three-dimensional equidistant grid field by a bilinear interpolation algorithm, taking the three-dimensional radar reflectivity data and the environmental field element data as input and the corresponding two-dimensional precipitation grid field of the next hour as a label, to prepare a sample set;

[0011] S3, using the three-dimensional radar reflectivity data integrated with environmental field elements within one hour as input and the precipitation grid field within the next hour as output, to build a precipitation nowcasting model;

[0012] S4, selecting samples in the sample set with a strong radar reflectivity area and dividing them into a training set, a validation set and a test set, training the precipitation nowcasting model on the training set, performing hyperparameter tuning on the validation set, and performing model testing on the test set;

[0013] S5. Using the environmental field data and radar reflectivity data within any hour, the radar reflectivity data at each radar scanning moment is converted from the polar coordinate system to three-dimensional data under the Cartesian rectangular coordinate system using the method in S2, and the environmental field of each hour is converted from equal longitude and latitude to equal-distance three-dimensional grid field data, and the radar reflectivity data and environmental field data within the hour are used as inputs of the model, and the model output is the precipitation nowcast field corresponding to the next hour.

[0014] Moreover, the data collected by S1 include radar reflectivity data within one hour, longitude, latitude, hourly precipitation real-time data of the ground rain gauge corresponding to the radar scanning area, and environmental field data corresponding to the radar scanning area.

[0015] Moreover, the S2 converts the radar reflectivity data from the polar coordinate system into three-dimensional data in the Cartesian rectangular coordinate system, which means that the radar base data is first converted from the polar coordinate system to the Cartesian coordinate system using the bilinear interpolation method in the horizontal dimension, and then interpolated in the vertical dimension to obtain three-dimensional equidistant grid field data with a horizontal resolution of 4km×4km and a vertical resolution of 0.5km.

[0016] Moreover, the S2 converts the original environmental field data of the 0.25°×0.25° equal longitude and latitude grid field into a three-dimensional equidistant grid environmental field, which means that a three-dimensional equidistant grid field is first established in the target area, wherein the grid center is the location of the radar site, and for each grid point to be interpolated, the four nearest equal longitude and latitude grid points are selected to obtain the final equidistant grid field data using bilinear interpolation.

[0017] Moreover, the S2 converts the matching actual precipitation data into a two-dimensional precipitation grid field by using an interpolation method, which means that the precipitation record points are interpolated into a two-dimensional precipitation grid field with a horizontal resolution of 4km×4km by using an inverse distance interpolation method.

[0018] Moreover, the precipitation nowcasting model of S3 uses a three-layer ConvLSTM network based on 3D convolution to extract features from the three-dimensional radar reflectivity data every six minutes every hour, and uses the convolutional attention module CBAM to improve the spatiotemporal receptive field of the convolution module, so that the model pays more attention to the strong reflectivity area and the three-dimensional cloud clusters related to precipitation. Upsampling and 3D convolution are used to process and extract the features of environmental field elements, so that the feature size of the environmental field matches the feature size of the three-dimensional radar reflectivity output by the last layer of ConvLSTM and then fuses them. Then, upsampling is used to fuse with the output result of the previous layer, and they are processed in sequence to output a two-dimensional precipitation grid field with a resolution of 4km×4km.

[0019] Moreover, the S4 training of the precipitation nowcasting model shields the influence of precipitation in the ocean area on the results, and uses the weighted mean square loss MWMSE with mask as the loss function of the model, which is defined as follows:

[0020]

[0021] In the formula, I (x,y) Represents the precipitation intensity at (x, y) in the observed image, I' (x,y) Represents the precipitation intensity at (x, y) in the predicted image, Mask (x,y) is a land mask. The land points within the radar scanning range in the mask are 1, and the rest are 0. (x,y) is the weight corresponding to this point. Different weights are set according to the actual precipitation so that the model pays more attention to the heavy precipitation area.

[0022] The positive effects that the present invention can produce are:

[0023] The present invention fully considers the impact of the distribution and evolution of water vapor particles in the environmental field and the vertical direction on precipitation, uses 3D convolution to extract features related to the environmental field, uses ConvLSTM based on 3D convolution to extract the hidden features of the radar reflectivity changes at different times and the impact on precipitation, and uses the convolutional attention module CBAM to make the model pay more attention to the three-dimensional space cloud bodies related to precipitation. Compared with the existing precipitation nowcasting model based on the CAPPI map at an altitude of 2km, the present invention fully considers the impact of the distribution and evolution of water vapor particles in the environmental field and the vertical direction on precipitation, and improves the accuracy of precipitation nowcasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the present invention;

[0025] Figure 2 It is a structural schematic diagram of the precipitation nowcasting model of the present invention;

[0026] Figure 3 Radar station distribution map of 15 different locations in North China and East China used in the present invention;

[0027] Figure 4 This is a display diagram of individual examples of precipitation nowcasting incorporating different environmental fields into the present invention. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0029] like Figure 1 As shown, the present invention provides a precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field, comprising:

[0030] A precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field, the innovation of which is that the steps of the method are:

[0031] (1) S-band Doppler weather radar data were collected from May 2015 to October 2016 on an hourly basis, covering 15 radar stations in different locations in North China and East China. Detailed information such as the location of each station and the boundaries of its joint scanning range are as follows: Figure 3As shown; from the historical precipitation information recorded by each automatic precipitation station, the actual precipitation data within the effective area of ​​the designated radar station is obtained; the effective area of ​​the longitude and latitude information of each station is determined, and the required environmental field data is obtained from the fifth generation global atmospheric reanalysis data (ERA5) provided by the European Medium-term Weather Forecast. The effective range of the radar reflectivity data is 512km×512km. In order to make the environmental field data better surround the effective radar reflectivity, the spatial range of the environmental field is set to 1024km×1024km.

[0032] (2) The radar reflectivity data at each radar scanning moment is converted from the polar coordinate system to three-dimensional data in the Cartesian rectangular coordinate system. The corresponding equidistant grid point environment field is constructed, where the center of the grid point is the location of the radar station. The longitude and latitude positions of each grid point in the equidistant grid point environment field can be calculated based on the longitude and latitude of the radar station. The calculation formula is:

[0033]

[0034] In the formula, (x r ,y r ) is the Cartesian coordinate of the radar station in the equidistant grid field, located at the center of the grid field; (x g ,y g ) is the Cartesian coordinate of the grid point to be interpolated in the equidistant grid field; (long r ,Lat r ) are the latitude and longitude coordinates of the radar station; g ,Lat g ) is the longitude and latitude coordinates of the grid point to be interpolated. In order to match the radar reflectivity data of the equidistant three-dimensional grid point field, the original environmental field data of equal latitude is converted into an equidistant grid point environmental field using a bilinear interpolation algorithm. The bilinear interpolation method is calculated according to the longitude and latitude of the radar station as follows:

[0035]

[0036] The matching precipitation real-time data is converted into a two-dimensional precipitation grid field by interpolation, and the three-dimensional radar reflectivity data and the corresponding environmental field data are used as input, and the corresponding two-dimensional precipitation grid field is used as a label to make a sample set. There are thousands of automatic precipitation stations within the scanning range of the 15 radars selected in this embodiment. The automatic station precipitation data is interpolated into a precipitation grid field using the inverse distance weighted interpolation method. Among them, the inverse distance weighted interpolation calculation method is as follows:

[0037]

[0038] Where n represents the n automatic stations closest to the grid center. In this embodiment, n=6; Ri represents the actual precipitation value provided by the ith automatic station; k is the set power, and in this embodiment, k=2.

[0039] (3) Build a precipitation nowcasting model that takes the three-dimensional radar reflectivity data and environmental field data within one hour before the forecast time as input and the precipitation field within the next hour as output. The six-minute radar reflectivity three-dimensional data is encoded and feature extracted through the ConvLSTM network based on 3D convolution and the convolutional attention module CBAM. The convolutional attention module CBAM is used to improve the spatial receptive field of the model and guide the model to pay more attention to the position and change information of the three-dimensional cloud clusters related to precipitation; the environmental field encoder is composed of upsampling operations and 3D convolutions. The upsampling process can make the environmental field features match the radar reflectivity features in size for feature fusion. Use upsampling and convolution to build a decoder, so that the fused features are continuously fused with the features of the previous layer through convolution, and finally a two-dimensional precipitation grid field with a resolution of 4km×4km is obtained. The model structure is as follows: Figure 2 The parameter settings are shown in Table 1.

[0040] The 3D convolution-based ConvLSTM network is derived from the 2D convolution-based ConvLSTM to extract the features of the 3D radar reflectivity data at multiple times. The 3D convolution-based ConvLSTM network replaces the convolution in ConvLSTM with a 3D convolution operation to process the 3D radar reflectivity data obtained from the radar scan at each moment, and uses the loop structure of ConvLSTM to extract and fuse the features of the 3D radar reflectivity data for six minutes every hour. The 3D convolution-based ConvLSTM network unit has three inputs and two outputs, just like the 2D convolution-based ConvLSTM network unit. The inputs are the input x at the current moment and the output x at the current moment. t , long-term memory unit c t-1 And the state variable output h at the previous moment t-1 , the output is the updated long-term memory unit c t And the current state variable output h t . Enter x t The state variable output of the previous moment is stacked on the channel dimension and then convolved. The convolution result is then sent to the forget gate f, update gate i, activation gate a, and output gate o respectively. The forget gate, update gate, activation gate, and output gate will output the corresponding feature matrix, which is Hadamard-producted with the memory unit to update the long-term memory unit. The operations performed by the above four gates are consistent with the output c t and h t The calculation method is as follows:

[0041] f = Sigmoid(h t-1 / w fh +x t *w fx +b f )

[0042] i=Sigmoid(h t-1 *w ih +x t *w ix +b i )

[0043] a=Tanh(h t-1 *w ah +x t *w ax +b a )

[0044] o=Sigmoid(h t-1 *w oh +x t *w ox +b o )

[0045]

[0046]

[0047] Among them, * represents the 3D convolution operation; w represents the corresponding convolution kernel weight; Represents Hadamard.

[0048] The CBAM convolutional attention module is composed of a CAM channel attention module and a SAM spatial attention module connected in series. The channel attention module keeps the channel dimension unchanged and compresses the spatial dimension, so that the model pays more attention to the meaningful information in the input data. The spatial attention module keeps the spatial dimension unchanged and compresses the channel dimension, so that the model pays more attention to the location information of the target. The CBAM convolutional attention module adaptively learns the channel and spatial attention weights to improve the representation ability of the neural network model.

[0049] The network structure of the present invention is as follows Figure 2As shown in the figure, it includes three parts, namely, a radar echo sequence encoder based on a 3D convolution-based ConvLSTM network and a convolutional attention module CBAM, an environmental field encoder based on 3D convolution and upsampling, and a decoder based on 3D convolution. For the encoder of the radar echo sequence, the spatial scale of each layer of feature maps is different. In order to make the model pay more attention to the three-dimensional cloud body, a convolutional attention module CBAM is added to make the model pay more attention to the area related to rainfall. The role of the environmental field encoder is to extract environmental field features and make the environmental field features consistent with the third layer feature map extracted from the radar echo sequence in spatial scale, so as to facilitate the decoding of the fused features. The fused features are finally processed by a two-dimensional convolution to obtain the precipitation nowcasting output result.

[0050] Table 1 Encoder and decoder parameter settings

[0051]

[0052]

[0053] (4) Dividing the sample set into a training set, a validation set, and a test set, training the precipitation nowcasting model on the training set, performing hyperparameter tuning on the validation set, and performing model testing on the test set.

[0054] Each radar completes a scan every six minutes on average, and environmental field data is recorded once every hour. The three-dimensional radar echo data of ten consecutive moments within one hour and the environmental field data of two adjacent moments are used as the input of the precipitation nowcasting model, and the cumulative precipitation grid field within the next hour is used as the label. A total of 7229 samples are randomly divided into training set, validation set and test set in a ratio of 4:1:1. Finally, 4811 training set samples, 1209 validation set samples and 1209 test set samples are obtained.

[0055] The precipitation nowcasting model is trained using the masked weighted squared error (MWMSE) as the model loss function to distinguish land precipitation from ocean precipitation. The MWMSE is defined as follows:

[0056]

[0057] In the formula, I (x,y) Represents the precipitation intensity at (x, y) in the observed image, I' (x,y) Represents the precipitation intensity at (x, y) in the predicted image, Mask (x,y) is a land mask, where the land points within the radar scanning range are 1 and the rest are 0. (x,y)is the weight corresponding to the point. Different weights can be set according to the actual precipitation, so that the model pays more attention to the heavy precipitation area. (x,y) The settings are as follows:

[0058]

[0059] (5) This embodiment uses the critical success index CSI, the fairness skill score ETS, and the Heidke skill score HSS to measure the consistency between the proportion of predicted precipitation above the threshold and the proportion of actual precipitation above the threshold under a certain precipitation threshold, and uses the root mean square error RMSE to measure the difference between the predicted precipitation and the actual precipitation. The larger the values ​​of CSI, ETS, and HSS, the better the model performance, and the smaller the value of RMSE, the better the model performance. CSI, ETS, HSS, and RMSE are defined as follows:

[0060]

[0061]

[0062] In the formula, y i represents the actual grid point precipitation, Represents the predicted grid point precipitation, N represents the total number of grid points in the precipitation grid map, and the calculation formula of REF is as follows:

[0063]

[0064] Among them, TP, FN, FP, and TN are defined by the confusion matrix shown in Table 2:

[0065] Table 2 Confusion matrix

[0066]

[0067] In this embodiment, the corresponding precipitation nowcasting model is trained by integrating radar echo data of environmental fields such as divergence, vertical velocity, potential vorticity, potential height, specific humidity, relative humidity, and temperature. The training process is implemented based on the Pytorch architecture, and the Adams optimizer is used for parameter update. The learning rate is 0.0001, the attenuation coefficient is (0.9, 0.999), the batch size is 4, and the training rounds of each model are 50. The model with the lowest MWMSE corresponding to the round on the validation set is selected as the final model for testing. The test results are shown in Table 3, and the individual examples are shown as follows Figure 4 shown.

[0068] Table 3 Pixel-level prediction scores under different precipitation thresholds in different environmental fields

[0069]

[0070]

[0071]

[0072] From Table 3 and Figure 4 It can be seen that:

[0073] (1) Due to the characteristics of severe convective weather, such as strong suddenness, fast movement speed and irregular changes, in general, whether there is no environmental field or there is an environmental field, with the increase of precipitation threshold, the RMSE shows an upward trend, the critical success index CSI gradually decreases, and the HSS score and ETS score gradually decrease; with the increase of precipitation threshold, the error of precipitation nowcasting also increases significantly.

[0074] (2) According to the comparison of the above experimental results, in addition to the potential height, the introduction of environmental field elements can improve the accuracy of the precipitation nowcasting model. Among them, if the precipitation threshold is set between 1mm and 20mm, the root mean square error RMSE is relatively reduced by 5%-27.1%, the critical success index CSI is relatively increased by 3%-34.7%, the HSS score is relatively increased by 2.2%-21.8%, and the ETS score is relatively increased by 3.6%-35.1%; if the precipitation threshold exceeds 20mm, the root mean square error RMSE is relatively reduced by 4.6%-37.1%, the critical success index CSI is relatively increased by 41.1%-1989%, the HSS score is relatively increased by 2.2%-141.1%, and the ETS score is relatively increased by 10.4%-198.9%. Therefore, the introduction of environmental field elements can improve the accuracy of precipitation nowcasting in heavy precipitation areas.

[0075] (3) When the precipitation threshold is below 20 mm, the precipitation nowcasting model with vertical velocity has a significantly better improvement on the accuracy of precipitation nowcasting than the precipitation nowcasting model with other environmental field elements, followed by the precipitation nowcasting model with divergence. The precipitation nowcasting model with geopotential height has the worst performance, and its test results are not much different from those of the precipitation nowcasting model without environmental field. When the precipitation threshold exceeds 20 mm, the precipitation nowcasting model with divergence performs outstandingly, still maintaining a relatively low error and high prediction accuracy, with RMSE and CSI of 5.6959, 7.8406, 0.5307 and 0.4111, respectively.

[0076] (4) By Figure 4 It can be seen from the individual case studies that the precipitation nowcasting model that introduces divergence and vertical velocity is relatively accurate in predicting heavy precipitation areas compared to other models, and the precipitation intensity is roughly the same.

[0077] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field, characterized by: The steps of the method are: S1. Collect radar reflectivity data and the corresponding precipitation data of the next hour on an hourly basis, and collect environmental field element data corresponding to the radar reflectivity data, including divergence, vertical velocity, potential vorticity, potential height, specific humidity, relative humidity, and temperature. S2, converting the radar reflectivity data at each radar scanning moment from the polar coordinate system into three-dimensional data in the Cartesian rectangular coordinate system, converting the matching real-time precipitation data of the next hour into a two-dimensional precipitation grid field by interpolation, converting the environmental field element data into a three-dimensional equidistant grid field by a bilinear interpolation algorithm, taking the three-dimensional radar reflectivity data and the environmental field element data as input and the corresponding two-dimensional precipitation grid field of the next hour as a label, to prepare a sample set; S3, using the three-dimensional radar reflectivity data integrated with environmental field elements within one hour as input and the precipitation grid field within the next hour as output, to build a precipitation nowcasting model; S4, selecting samples in the sample set with a strong radar reflectivity area and dividing them into a training set, a validation set and a test set, training the precipitation nowcasting model on the training set, performing hyperparameter tuning on the validation set, and performing model testing on the test set; S5. Using the environmental field data and radar reflectivity data within any hour, the radar reflectivity data at each radar scanning moment is converted from the polar coordinate system to three-dimensional data under the Cartesian rectangular coordinate system using the method in S2, and the environmental field of each hour is converted from equal longitude and latitude to equal-distance three-dimensional grid field data, and the radar reflectivity data and environmental field data within the hour are used as inputs of the model, and the model output is the precipitation nowcast field corresponding to the next hour.

2. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The data collected by S1 include radar reflectivity data within one hour, longitude, latitude, hourly precipitation data of the ground rain gauge corresponding to the radar scanning area, and environmental field data corresponding to the radar scanning area.

3. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The S2 converts the radar reflectivity data from the polar coordinate system into three-dimensional data in the Cartesian rectangular coordinate system, which means that the radar base data is first converted from the polar coordinate system to the Cartesian coordinate system using the bilinear interpolation method in the horizontal dimension, and then interpolated in the vertical dimension to obtain three-dimensional equidistant grid field data with a horizontal resolution of 4km×4km and a vertical resolution of 0.5km.

4. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The S2 converts the original environmental field data of the 0.25°×0.25° equal longitude and latitude grid field into a three-dimensional equidistant grid environmental field, which means first establishing a three-dimensional equidistant grid field in the target area, wherein the grid center is the location of the radar site, and for each grid point to be interpolated, selecting the four nearest equal longitude and latitude grid points nearby and using bilinear interpolation to obtain the final equidistant grid field data.

5. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The S2 converts the matched actual precipitation data into a two-dimensional precipitation grid field by using an interpolation method, which means that the precipitation record points are interpolated into a two-dimensional precipitation grid field with a horizontal resolution of 4km×4km by using an inverse distance interpolation method.

6. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The precipitation nowcasting model of S3 uses a three-layer ConvLSTM network based on 3D convolution to extract features from the three-dimensional radar reflectivity data every six minutes every hour, and uses the convolutional attention module CBAM to improve the spatiotemporal receptive field of the convolution module, so that the model pays more attention to the strong reflectivity area and the three-dimensional cloud clusters related to precipitation. Upsampling and 3D convolution are used to process and extract the features of environmental field elements, so that the feature size of the environmental field matches the feature size of the three-dimensional radar reflectivity output by the last layer of ConvLSTM and then fuses them. Then, upsampling is used to fuse with the output result of the previous layer, and they are processed in sequence to finally output a two-dimensional precipitation grid field with a resolution of 4km×4km.

7. The precipitation nowcasting method integrating three-dimensional radar reflectivity and environmental factor field according to claim 1 is characterized in that: The S4 training of the precipitation nowcasting model shields the influence of precipitation in the ocean area on the results, and uses the weighted mean square loss MWMSE with mask as the loss function of the model, which is defined as follows: In the formula, I (x,y) Represents the precipitation intensity at (x, y) in the observed image, I' (x,y) Represents the precipitation intensity at (x, y) in the predicted image, Mask (x,y) is a land mask. The land points within the radar scanning range in the mask are 1, and the rest are 0. (x,y) is the weight corresponding to this point. Different weights are set according to the actual precipitation so that the model pays more attention to the heavy precipitation area.

Images