Microwave monitoring rainfall field reconstruction method based on generative model

Abstract

The invention discloses a microwave monitoring rainfall field reconstruction method based on a generative model. The method comprises the following steps: S1, obtaining rainfall data from different data sources; s2, dividing a training data set and a simulation data set, and performing data preprocessing; s3, a generative adversarial neural network model Mic-GAN is constructed; s4, iteratively training the anti-neural network model Mic-GAN by adopting the training set in the step S2 until a preset number of iterations or convergence conditions are met; and S5, reconstructing the microwave rainfall field of the target area by adopting the simulation data set in the step S2 and using the model in the step S4. According to the microwave monitoring rainfall field reconstruction method based on the generative model, optimization is continuously carried out along with the increase of data, the advantages of microwave big data are effectively utilized, and the method has good expandability.

Description

Technical Field

[0001] The present invention relates to a method for reconstructing a microwave monitoring rainfall field based on a generative model, belonging to the technical field of hydrometeorological data processing. Background Art

[0002] Rainfall monitoring, as an important part of hydrometeorology, has a profound impact on multiple fields. It is not only directly related to the production efficiency of agriculture, the effective management of water resources, and the accuracy of flood warning systems, but also plays an indispensable role in urban planning, traffic management, environmental protection and other aspects. With the development of technology, traditional rainfall monitoring methods have gradually been replaced by more accurate and efficient technical means. Among them, using a microwave station network for rainfall monitoring has emerged as a new technology in recent years.

[0003] By monitoring the attenuation phenomenon that occurs during the transmission of microwave signals in the atmosphere, scientists can invert the average rainfall situation between microwave links. Compared with traditional methods, this method provides more accurate data support and has advantages such as strong real-time performance and wide coverage. However, despite the significant progress made in microwave monitoring technology, its application still faces challenges. Specifically, since the rainfall data obtained by microwave monitoring is essentially line-type data, it is difficult to accurately describe the complex non-linear characteristics existing in the rainfall distribution when using traditional methods such as interpolation and tomography to reconstruct the rainfall field. In addition, these methods often fail to fully utilize the potential information value brought by the large amount of microwave data.

[0004] In recent years, with the rise of generative models in the field of computer vision, their excellent performance in processing image detail features has attracted wide attention. Generative models can not only effectively capture and reproduce the subtle structural changes in images, but also generate high-quality prediction results in the absence of complete information. Based on these characteristics, such models have been successfully applied to multiple fields such as image restoration, object detection, and weather forecasting. Applying generative models to reconstruct the microwave monitoring rainfall field has broad application prospects. At present, some studies have applied generative models to the downscaling task of satellite remote sensing monitoring and achieved breakthrough results. However, there are mainly two difficulties in the task of reconstructing the microwave rainfall field using generative models. On the one hand, the microwave monitoring rainfall data is line-type data, making it difficult to effectively combine with the monitoring data of existing rain gauges and radars. On the other hand, the distribution of microwave station networks varies in different regions, and the microwave monitoring accuracy is also affected by different climates, making rainfall feature extraction complex. Summary of the Invention

[0005] Objective of the Invention: To overcome the deficiencies in the prior art, the present invention provides a method for reconstructing a microwave monitoring rainfall field based on a generative model, which is continuously optimized as the data increases, effectively utilizes the advantages of microwave big data, and has good scalability.

[0006] Technical Solution: To solve the above technical problems, a method for reconstructing a microwave monitoring rainfall field based on a generative model of the present invention includes the following steps:

[0007] Step S1: Obtain precipitation data from different data sources;

[0008] Step S2: Divide the training data set and the simulation data set, and perform data preprocessing;

[0009] Step S3: Construct a generative adversarial neural network model Mic-GAN;

[0010] Step S4: Use the training set described in Step S2 to iteratively train the adversarial neural network model Mic-GAN until a preset number of iterations is reached or the loss of the generator reconstruction result meets the convergence condition;

[0011] Step S5: Use the simulation set described in Step S2 to reconstruct the microwave rainfall field of the target area using the trained model.

[0012] Preferably, the precipitation data from different data sources includes data from hydrological stations and microwave stations in the target area, as well as simultaneous data from hydrological stations, radars, and microwave station networks in at least two other regions.

[0013] Preferably, the division of the training data and the simulation data, and the data preprocessing specifically include: using the hydrological stations and microwave signals in the target area as the simulation data set; according to the spatial range of the target area, cropping the space of other regions, and using the simultaneous hydrological stations, microwave station networks, and radar data within the corresponding spatial range of other regions as the training data set; using a general formula to invert the radar and microwave data into rainfall values using the general formula, and converting the average rainfall data of the obtained microwave link into the rainfall value of a virtual site at the center of the link; using the cumulative addition method to adjust the time resolution of the rainfall data monitored by the hydrological stations and virtual sites to the same time scale as the radar monitoring data; using the inverse distance weighting method to interpolate the rain gauges and virtual sites into the longitude and latitude grid where the radars are distributed, and normalizing the data to form a rainfall distribution matrix as the subsequent input.

[0014] Preferably, the specific structure of the constructed generative adversarial neural network model Mic-GAN includes two parts: a discriminator and a generator.

[0015] Preferably, the discriminator is composed of a multi-layer convolutional network and a linear layer. The convolutional network and the linear layer are used to extract the latent features of the input data as low-dimensional data, including the regional shape of the rainfall distribution and the distribution information of rainfall with different intensities, etc. Finally, the sigmoid function is used to output the probability that the discriminant data is real data or fake data.

[0016] Preferably, the generator is composed of a UNet convolutional network, including three parts: an encoder, a decoder, and a residual link. The input data is coupled with the feature vector of the microwave station network distribution extracted by the LSTM model in the encoder and then undergoes a downsampling operation. After multiple operations, the output data of the encoder is obtained. The data output by the encoder is input into the decoder for an upsampling operation. At the same time, skip residual connections are used to connect the feature maps in the corresponding layers of the encoder and the decoder, and finally, the rainfall field reconstructed by the generator is obtained.

[0017] Preferably, the specific steps of using the training set in step S2 to iteratively train the model until the preset number of iterations or convergence conditions are met include:

[0018] S4.1. Train the discriminator of the model;

[0019] S4.2. Train the generator of the model;

[0020] S4.3. Determine whether the preset number of iterations or convergence conditions are met. If so, stop training; otherwise, return to step S4.1 to train the discriminator first, and then train the generator according to step S4.2.

[0021] Step S4.1: Train the discriminator of the model

[0022] The task of the discriminator is to distinguish whether the input data is real data or fake data generated by the generator. The goal of training the discriminator is to maximize its ability to distinguish real and fake data. The specific steps are as follows:

[0023] Prepare real data and generated data: Randomly extract a batch of real data from the training set, and at the same time use the current generator to generate a batch of fake data.

[0024] Forward propagation: Input the real data and the fake data into the discriminator respectively to obtain the outputs of the discriminator for the real data and the fake data.

[0025] Calculate the loss: According to the output of the discriminator and the real labels (the real data label is 1, and the fake data label is 0), calculate the loss of the discriminator. Usually, the binary cross-entropy loss function is used.

[0026] Backward propagation and parameter update: Calculate the gradient of the loss with respect to the discriminator parameters, and use the optimizer to update the discriminator parameters.

[0027] Step S4.2: Train the generator of the model

[0028] The task of the generator is to generate as realistic data as possible to deceive the discriminator. The goal of training the generator is to minimize the discriminator's ability to distinguish the generated data. The specific steps are as follows:

[0029] Generate fake data: Use the current generator to generate a batch of fake data.

[0030] Forward propagation: Input the generated fake data into the discriminator to obtain the output of the discriminator.

[0031] Calculate the loss: Calculate the loss of the generator based on the output of the discriminator. The generator hopes that the discriminator will judge the generated data as real data. Therefore, the goal of the loss function is to make the discriminator's output for the generated data close to 1.

[0032] Backward propagation and parameter update: Calculate the gradient of the loss with respect to the generator's parameters and use an optimizer to update the generator's parameters.

[0033] Step S4.3: Determine whether the preset number of iterations or convergence conditions are met.

[0034] After each iteration, it is necessary to check whether the preset number of iterations or convergence conditions are met. If so, stop training; otherwise, go back to step S4.1 to continue training the discriminator, and then train the generator.

[0035] Preferably, the preset number of iterations is 5000 times; the convergence condition is that the loss BCELoss of the generator reconstruction result < 0.001, and at the same time, the mean squared error loss MSELoss between the reconstructed rainfall field and the radar-measured rainfall < 0.05.

[0036] Preferably, in step S4.1, the method for training the model discriminator is as follows:

[0037] Mark the radar data in the training set obtained in step S2 as real data and assign the data attribute target = 1; input the preprocessed data of the hydrological station and microwave virtual station in the training set into the generator to generate a batch of reconstructed data marked as fake data and assign the data attribute target = 0.

[0038] The real data and fake data are mixed and then input into the discriminator. The convolutional network and linear layer are used to extract the latent features of the input data as low-dimensional data, including the regional shape of the rainfall distribution and the distribution information of rainfall with different intensities, etc. Extract features according to the following formula:

[0039]

[0040] In the formula, h j,kis the latent feature extracted by the convolutional neural network; σ is the LeakyRelu function; L and M are the length and width of the convolutional kernel; b represents the bias parameter, and ω l,m is the weight coefficient, and both are determined by the optimization iteration of the backpropagation of the deep learning network; X l,m is the preprocessed input data matrix received by the convolutional layer.

[0041] Finally, the sigmoid function is used to output the probability of discriminating whether the data is real data or fake data. The formula for using the sigmoid function is:

[0042]

[0043] In the formula, A represents the probability that the data is real data or fake data, and x is the one-dimensional feature data finally extracted through the convolutional layer and the linear layer.

[0044] Preferably, in the step S4.1, when training the discriminator of the training model, the loss function is defined as the binary cross-entropy function, and the goal is to distinguish real data and generated fake data. During training, the parameters of the generator are fixed, and the parameters of the discriminator are updated using the backpropagation algorithm. The calculation formula of the binary cross-entropy function is as follows:

[0045] BCELoss = -(target·logA) + (1 - target)·log(1 - A) (3)

[0046] In the formula, BCELoss is the function value, target is the value marked by the input data attribute target, and A represents the probability that the input data output by the discriminator is real data or fake data.

[0047] Preferably, in the step S4.2, the method for training the generator of the model is:

[0048] Input the hydrological station and microwave virtual station data in the training set obtained in step S2 into the generator. After using the LSTM model to extract the distribution characteristics of the microwave station network and coupling with the preprocessed rainfall field data in step S2, through multiple downsamplings of the encoder and multiple upsamplings of the decoder, a reconstructed rainfall field can be generated, and then the reconstructed rainfall field is input into the discriminator to obtain the discrimination result.

[0049] Preferably, in the step S4.2, each downsampling operation in the encoder of the model generator includes: performing two convolutions on the input data using a convolutional layer with a convolutional kernel size of 3*3 and then activating through the ReLU function, and then performing max pooling using a convolutional kernel of size 2*2.

[0050] Preferably, in step S4.2, the upsampling operation of each layer of the decoder of the model generator includes: first, deconvolving the upsampling data of each layer using a convolution kernel of size 2×2, then jump-connecting the corresponding downsampled features to the data output by the deconvolution, and then performing two convolutions using a convolutional layer with a convolution kernel size of 3×3 and activating through the ReLU function.

[0051] During training, the loss function is defined as the negative of the cross-entropy loss function of the discriminator's output for fake data, and the goal is for the discriminator's judgment result for fake data to be as close as possible to real data. During training, the discriminator parameters are fixed, and the generator parameters are updated using the backpropagation algorithm.

[0052] Preferably, in step S4.3, the maximum number of iterations is set to 5000 times, and completing one round of steps S4.1 and S4.2 is counted as one training; the convergence conditions are set as BCELoss < 0.001 and MSELoss < 0.05.

[0053] Preferably, in step S5, the method for reconstructing the microwave rainfall field of the target area using the trained model with the simulation set described in step S2 is: inputting the hydrological station and microwave virtual station data in the simulation dataset obtained in step S2 into the trained model generator, and the reconstructed rainfall field can be obtained.

[0054] Beneficial effects: The microwave monitoring rainfall field reconstruction method based on the generative model of the present invention has the following advantages:

[0055] 1. By constructing a microwave monitoring rainfall field reconstruction method based on the generative model, the present invention can be continuously optimized as the data increases, effectively utilizes the advantages of microwave big data, and has good scalability.

[0056] 2. Compared with the traditional method, the rainfall field reconstruction method provided by the present invention solves the problem that the traditional method is difficult to handle the complex non-linear relationships in the data and requires manual intervention for feature extraction. Through the training of the generative adversarial model, useful features can be automatically extracted from the data, more effectively utilizes the limited observation resources, and improves the data utilization efficiency.

[0057] 3. The fusion microwave monitoring rainfall field reconstruction method of the present invention can be widely applied to rainfall field reconstruction, can adapt to different regions, different time scales, and different microwave station network layouts, can quickly process real-time data, and provides support for short-term weather forecasting and disaster warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flowchart of the microwave monitoring rainfall field reconstruction method based on the generative model of the present invention

[0059] Figure 2 Schematic diagram of the structure of the Mic-GAN based on the generative adversarial neural network model according to the present invention

[0060] Figure 3 Discriminator of the Mic-GAN generative model according to the present invention

[0061] Figure 4 Generator of the Mic-GAN generative model according to the present invention

[0062] Figure 5 Training flowchart of the Mic-GAN based on the generative adversarial neural network model according to the present invention Detailed implementation manners

[0063] The present invention will be further described in conjunction with the accompanying drawings.

[0064] As Figure 1 shown, for the rainfall field reconstruction task in an area of 572m * 572m, the microwave monitoring rainfall field reconstruction method based on the generative model specifically includes the following steps:

[0065] Step S1: Obtain precipitation data from different data sources.

[0066] In this example, the precipitation data from different data sources includes data from hydrological stations and microwave stations in the target area, as well as simultaneous data from hydrological stations, radars, and microwave station networks in at least two other regions.

[0067] Step S2: Divide the training data set and the simulation data set, and perform data preprocessing.

[0068] In this example, dividing the training data and the simulation data and performing data preprocessing specifically includes:

[0069] Taking the hydrological stations and microwave signals in the target area as the simulation data set;

[0070] According to the spatial range of the target area, cropping the space of other regions, and taking the simultaneous hydrological stations, microwave station networks, and radar data within the corresponding 572m * 572m size space range in other regions as the training data set;

[0071] Using a general formula to invert the radar and microwave data into rainfall values using the general formula, converting the average rainfall data of the obtained microwave link into the rainfall value of the virtual site at the center of the link; using the cumulative addition method to adjust the time resolution of the rainfall data monitored by the hydrological stations and virtual sites to the same 5-minute time scale as the radar monitoring data; using the inverse distance weighting method to interpolate the rain gauge stations and virtual sites into the 572 * 572 size longitude and latitude grid where the radar is distributed, and normalizing the data to form a matrix.

[0072] Step S3: Construct a generative adversarial neural network model Mic-GAN.

[0073] As Figure 2 shown, the specific structure of the constructed generative adversarial neural network model Mic-GAN includes two parts: a discriminator and a generator.

[0074] As Figure 4 shown, in this example, the generator is composed of a UNet convolutional network, including three parts: an encoder, a decoder, and a residual link. The input data is coupled with the feature vector of the microwave station network distribution extracted by the LSTM model in the encoder and then undergoes four downsampling operations. After multiple operations, the output data of the encoder is obtained. The input data is the data of the rain gauge stations and the microwave data. The data output by the encoder is input into the decoder for four upsampling operations. At the same time, skip residual connections are used to connect the feature maps in the corresponding layers of the encoder and the decoder, and finally, the rainfall field reconstructed by the generator is obtained.

[0075] As Figure 3 shown, in this example, the discriminator is composed of five convolutional layers combined with a linear layer. The convolutional layer and the linear layer have a kernel size of 5. The output of the linear layer is 1 / 2 of the input data size, and the output of the last linear layer is 1. A ReLU activation function follows each convolutional layer or linear layer. The convolutional network and the linear layer are used to extract the latent features of the input data as low-dimensional data, including the regional shape of the rainfall distribution and the distribution information of rainfall with different intensities. Finally, the sigmoid function is used to output the probability that the discriminant data is real data or fake data.

[0076] Step S4: Use the training set described in Step S2 to iteratively train the model until the preset number of iterations or convergence conditions are met;

[0077] As Figure 5 shown, the specific steps of using the training set described in Step S2 to iteratively train the model until the preset number of iterations or convergence conditions are met in this example include:

[0078] S4.1: Train the discriminator of the model;

[0079] S4.2: Train the generator of the model;

[0080] S4.3: Determine whether the preset number of iterations or convergence conditions are met. If so, stop training; otherwise, return to Step S4.1 to train the discriminator first, and then train the generator according to Step S4.2.

[0081] Step S4.1: In this example, the method for training the discriminator of the model is:

[0082] Mark the radar data in the training set obtained in step S2 as real data, and assign the data attribute target = 1; input the preprocessed data of the hydrological station and the microwave virtual station in the training set into the generator to generate a batch of reconstructed data marked as fake data, and assign the data attribute target = 0.

[0083] The real data and the fake data are mixed and then input into the discriminator. After the input data passes through five convolutional layers combined with a linear layer in sequence, the latent features of the input data are extracted as low-dimensional data, including the regional shape of the rainfall distribution and the distribution information of rainfall with different intensities, etc. Extract features according to the following formula:

[0084]

[0085] In the formula, h j,k is the latent feature extracted by the convolutional neural network; σ is the LeakyRelu function; L and M are the length and width of the convolutional kernel; b represents the bias parameter, ω l,m is the weight coefficient, and both are determined by the optimization iteration of the backpropagation of the deep learning network; X l,m is the preprocessed input data matrix received by the convolutional layer.

[0086] The output data of the last linear layer is one-dimensional data x. Use the sigmoid function to output the probability that the discriminant data is real data or fake data. The formula of the sigmoid function is:

[0087]

[0088] In the formula, A represents the probability that the data is real data or fake data.

[0089] When training the discriminator of the training model, define the loss function as the binary cross-entropy function, and the goal is to distinguish real data from generated fake data. Fix the generator parameters during training and use the backpropagation algorithm to update the discriminator parameters. The calculation formula of the binary cross-entropy function is as follows:

[0090] BCELoss = -(target·logA) + (1 - target)·log(1 - A) (3)

[0091] In the formula, BCELoss is the function value, target is the value marked by the input data attribute target, and A represents the probability that the input data output by the discriminator is real data or fake data.

[0092] In step S4.2, the method for training the generator of the training model in this example is:

[0093] Input the hydrological stations and microwave virtual station data in the training set obtained in step S2 into the generator. After using the LSTM model to extract the microwave station network distribution characteristics and coupling them with the preprocessed rainfall field data in step S2, through multi-layer downsampling of the encoder and multi-layer upsampling of the decoder, a reconstructed rainfall field can be generated. Then, input the reconstructed rainfall field into the discriminator to obtain the discrimination result.

[0094] The operations of each of the four layers of downsampling in the encoder include: performing two convolutions on the input data using a convolutional layer with a convolutional kernel size of 3*3, activating through the ReLU function, and then performing max pooling using a convolutional kernel of size 2*2. For example, when the input rainfall stations and microwave rainfall field are connected in the channel dimension to form data of size 572*572*2, during the first downsampling operation, it is coupled and connected with the 572*572*1 feature vector obtained from the LSTM model to form data of size 572*572*3. After activation through 2 convolutional layers, the size becomes 568*568*64, and then through a max pooling layer, the data size becomes 284*284*64; after the second downsampling operation, the data size becomes 140*140*128. After four downsampling operations, data of size 32*32*512 can be finally obtained. After performing two more convolutions on the data, the size becomes 28*28*1024 and is input into the decoder.

[0095] The operations of each layer of the decoder's four - layer upsampling include: First, the upsampled data of each layer is de - convolved using a convolution kernel of size 2*2, and then the corresponding downsampled features are connected to the data output by the de - convolution. After that, two convolutions are performed using a convolution layer with a convolution kernel size of 3*3 and then activated through the ReLU function. Using the data of 28*28*1024 output by the encoder, through skip - residual connections, the features obtained by activation before the maximum pooling after convolution in each layer of downsampling are cropped and then connected for upsampling operations. Similarly, four upsampling layers are set. After the upsampled data of each layer is de - convolved using a convolution kernel of size 2*2, the size of the input data in terms of length and width can be expanded to twice the original size. Then, the corresponding downsampled features are connected to the data output by the de - convolution, and the number of channels also becomes twice the original. Finally, two convolutions are performed using a convolution layer with a convolution kernel size of 3*3 and then activated through the ReLU function. Set the padding to 2 to keep the shape unchanged, and the number of channels becomes 1 / 2 of the original. For example, in the bottom layer, the data of size 28*28*1024 is de - convolved to obtain data of size 56*56*512, which is patched with the data of size 64*64*512 obtained in the corresponding bottom - layer downsampling after filling the boundaries, and then activated through two convolution layers to obtain data of size 64*64*512. After four downsampling operations, data of size 572*572*128 can be finally obtained. The data is then convolved twice to become 572*572*64, and then a convolution layer with a convolution kernel size of 1*1 is used to reduce the number of channels to 1, obtaining a reconstructed rainfall field of size 572*572*1.

[0096] During training, the loss function is defined as the negative of the cross - entropy loss function of the discriminator's output result for fake data, and the goal is for the discriminator's judgment result for fake data to be as close as possible to real data. During training, the discriminator parameters are fixed, and the existing backpropagation algorithm is used to update the generator parameters.

[0097] Step S4.3: Set the maximum number of iterations to 5000 times. Completing one round of Step S4.1 and Step S4.2 counts as one training; set the convergence condition to BCELoss < 0.001.

[0098] Step S5: Use the simulation set described in Step S2 to reconstruct the microwave rainfall field of the target area using the trained model.

[0099] In this example, the hydrological station and microwave virtual station data in the simulation dataset obtained in Step S2 are input into the trained model generator, and the reconstructed rainfall field can be obtained.

[0100] In summary, the microwave rainfall field reconstruction method based on the generative model provided by the present invention can effectively utilize the data volume advantage of microwave monitoring, reconstruct the rainfall field to achieve the accuracy of the same spatio-temporal resolution as radar monitoring, and make up for the deficiency that non-linear transformation is difficult to accurately describe. In addition, by referring to this specification and in the practice of the present disclosure, those skilled in the art should be clear that, without departing from the scope of the claimed protection of the present disclosure, modifications and changes can be made to the disclosed model.

[0101] The present invention aims to overcome the limitations of the prior art, make full use of the potential of microwave data by combining the advantages of microwave big data with the generative model, and combine the rainfall monitoring data of rain gauges and radars to achieve a more accurate and detailed reconstruction of the rainfall field. This new method not only has the potential to improve the accuracy and reliability of rainfall monitoring, but may also bring new research perspectives and technological breakthroughs to related fields.

[0102] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A microwave monitoring rainfall field reconstruction method based on a generative model, characterized in that: The steps include: Step S1: Obtain precipitation data from different data sources; Step S2: Divide the training data set and the simulation data set, and perform data preprocessing; Step S3: construct a generative adversarial neural network model Mic-GAN; Step S4: Iteratively train the anti-neural network model Mic-GAN using the training set described in step S2 until a preset number of iterations is reached or the loss of the generator reconstruction result meets the convergence condition; Step S5: Reconstruct the microwave rainfall field in the target area using the simulated data set in step S2 and the model in step S4.

2. The microwave monitoring rainfall field reconstruction method based on the generative model according to claim 1 is characterized in that: In step S1, the precipitation data from different data sources include data from hydrological stations and microwave station networks in the target area, and data from hydrological stations, radars and microwave station networks in at least two other areas during the same period.

3. The microwave monitoring rainfall field reconstruction method based on the generative model according to claim 1 is characterized in that: The step S2 specifically includes: The hydrological stations and microwave station network data of the target area are used as simulation data sets; According to the spatial range of the target area, the space of other regions is cropped, and the hydrological stations, microwave station networks and radar data of other regions in the corresponding spatial range are used as training data sets; Using a general formula, the radar and microwave data are inverted into rainfall values ​​using a general formula, and the average rainfall data of the microwave link is converted into the rainfall value of a virtual station at the center of the link; The temporal resolution of rainfall data monitored by hydrological stations and virtual stations was adjusted to the same time scale as radar monitoring data using the accumulation method; The inverse distance weighted method is used to interpolate the rainfall stations and virtual stations into the latitude and longitude grid of the radar distribution, and the data is normalized to form a rainfall distribution matrix as subsequent input;.

4. The microwave monitoring rainfall field reconstruction method based on a generative model according to claim 1 is characterized in that: In step S3, the specific structure of the adversarial neural network model Mic-GAN is: It consists of two parts: the generator and the discriminator; The generator is composed of a UNet convolutional network, including an encoder, a decoder, and a residual link: the input data is coupled with the feature vector of the microwave station network distribution extracted by the LSTM model in the encoder and then downsampled, and the encoder output data is obtained after multiple layers of operation; the encoder output data is input into the decoder for upsampling, and the feature maps in the corresponding layers of the encoder and decoder are connected using skip residual connections, and finally the rainfall field reconstructed by the generator is obtained; The discriminator consists of multiple layers of convolutional networks and linear layers. Convolutional networks and linear layers are used to extract the potential features of the input data into low-dimensional data, including the regional shape of rainfall distribution and the distribution information of rainfall of different intensities. Finally, the sigmoid function is used to output the probability that the discriminant data is real data or false data.

5. The microwave monitoring rainfall field reconstruction method based on a generative model according to claim 1 is characterized in that: In step S4, the iterative training model using the training set in step S2 until a preset number of iterations or convergence condition is met specifically includes: S4.1, training model discriminator; S4.2, training model generator; S4.

3. Determine whether the preset number of iterations or convergence conditions are met. If so, stop training. Otherwise, return to step S4.1 to train the discriminator first, and then train the generator according to step S4.

2.

6. The microwave monitoring rainfall field reconstruction method based on the generative model according to claim 5 is characterized in that: In step S4.1, the method for training the model discriminator is: Mark the radar data in the training set obtained in step S2 as real data, and assign the data attribute target=1; Input the rainfall distribution matrix of the hydrological stations and microwave station network in the training set obtained in step S2 and the coordinate sequence of the microwave station network into the generator to generate a batch of reconstructed data marked as false data, and assign the data attribute target = 0; The real data and fake data are mixed and input into the discriminator, and compared with the output of the discriminator for training. The loss function is defined as the binary cross entropy function during training. The goal is to distinguish the real data from the generated fake data. The generator parameters are fixed during training, and the back propagation algorithm is used to update the discriminator parameters. The calculation formula of the binary cross entropy function is as follows: BCELoss=-(target·logA)+(1-target)·log(1-A) (1) In the formula, BCELoss is the function value, target is the value of the input data attribute target label, and A represents the probability that the input data output by the discriminator is real data or false data.

7. The microwave monitoring rainfall field reconstruction method based on the generative model according to claim 6 is characterized in that: In step S4.2, the method for training the model generator is: The rainfall distribution matrix of the hydrological station and microwave station network in the training set obtained in step S2 and the coordinate sequence of the microwave station network are input into the generator, and the distribution features of the microwave station network are extracted using the LSTM model and coupled with the rainfall field data preprocessed in step S2. After multi-layer downsampling by the encoder and multi-layer upsampling by the decoder, a reconstructed rainfall field is generated, and then the reconstructed rainfall field is input into the discriminator to obtain the discrimination result; During training, the loss function is defined as the inverse of the cross entropy loss function of the discriminator's output of fake data. During training, the discriminator parameters are fixed and the generator parameters are updated using the back-propagation algorithm.

8. The microwave monitoring rainfall field reconstruction method based on a generative model according to claim 1 is characterized in that: In step S5, the method of reconstructing the microwave rainfall field in the target area using the trained model using the simulated data set in step S2 is: inputting the rainfall distribution matrix of the hydrological station and the microwave station network in the simulated data set obtained in step S2 and the coordinate sequence of the microwave station network into the trained model generator to obtain the reconstructed rainfall field.

Images