A precipitation downscaling method based on a generative adversarial network model

Abstract

This invention discloses a precipitation downscaling method based on a generative adversarial network (GAN) model, belonging to the fields of deep learning and downscaling technology. It aims to address the problems of existing methods' difficulty in modeling nonlinear relationships among multiple meteorological variables and their failure to consider the differentiated characteristics of precipitation of different intensities. The method first preprocesses multi-source data and divides it into training and test sets. Then, it constructs a downscaling model containing a generator and a discriminator. The generator is composed of multivariate feature interaction, super-resolution, and multi-level mask fusion modules in sequence. Simultaneously, a joint loss function optimization model is designed, consisting of fusion adversarial loss, weighted classification mask loss, and mean squared error loss. Finally, model training and multi-index evaluation are completed. This invention can accurately characterize the spatiotemporal features of precipitation and significantly improve the downscaling accuracy of precipitation of different intensities, providing high-resolution precipitation data support for agricultural production, disaster prevention and mitigation, and other fields.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and downscaling technology, and in particular to a precipitation downscaling method based on a generative adversarial network model. Background Technology

[0002] Precipitation is a core variable in the Earth's water cycle, making accurate prediction crucial. High spatiotemporal resolution precipitation data is a key foundation for addressing this challenge. Reanalysis precipitation data reconstructed through assimilation of historical observations and numerical simulations offers advantages such as broad coverage and long time spans, serving as an important basis for studying precipitation-related climatological characteristics. However, such data generally have low spatial resolution, making it difficult to resolve fine-scale precipitation features, leading to significant biases in simulating mesoscale and microscale precipitation processes. Therefore, downscaling methods that transform low-resolution data into high-resolution data are essential for achieving a detailed characterization of the spatiotemporal features of precipitation processes.

[0003] With the rapid development of deep learning technology, generative adversarial networks (GANs) have been introduced into the field of precipitation downscaling. Through adversarial game between the generator and discriminator, the statistical distribution of the generated results is made closer to the real distribution, which can effectively recover the high-frequency details of the precipitation field and provide a new perspective for the reconstruction of precipitation data. Although precipitation downscaling methods based on GAN models have achieved significant results, they still face two major challenges: first, the lack of accurate modeling of the complex nonlinear relationship between precipitation and multiple variables such as wind field and temperature; and second, the failure to consider the differentiated characteristics of precipitation of different intensities. Summary of the Invention

[0004] The purpose of this invention is to provide a precipitation downscaling method based on a generative adversarial network model, which can effectively fuse multivariate information and improve the downscaling accuracy of precipitation of different intensities.

[0005] To achieve the above objectives, this invention provides a precipitation downscaling method based on a generative adversarial network model, comprising the following steps: S1. Acquire low-resolution meteorological reanalysis data, high-resolution precipitation data, and high-resolution topographic data, and preprocess the data, dividing them into training and testing sets. S2. Construct a downscaling model based on generative adversarial networks: The model consists of two parts: a generator and a discriminator. The generator is composed of a multivariate feature interaction module, a super-resolution network, and a multi-level mask fusion module in sequence. The discriminator is composed of multiple convolutional layers. The downscaling model is optimized using a joint loss function during training. S3. Construct a multivariate feature interaction module: including a feature projection layer, a bidirectional cross-attention calculation layer, and a gated residual fusion layer, used to accurately model the nonlinear relationship between multiple meteorological variables; S4. Construct a super-resolution network: including a feature fusion input layer, an encoder, and a decoder; upsample multivariate interactive features through the feature fusion input layer and fuse them with terrain data; then extract features from the fused data through the encoder and decoder; S5. Construct a multi-level mask fusion module: This includes defining precipitation intensity levels, multi-level mask generation layers, and a dynamic weight learning network, which are used to learn differentiated features for precipitation of different intensities. S6. Construct the joint loss function: It consists of two parts: adversarial loss and content loss; the content loss consists of weighted classification mask loss and mean squared error loss. S7. Training and Evaluation of Downscaling Model: The downscaling model is trained on the training set until convergence, and its performance is tested on the test set. Model performance metrics include continuous numerical accuracy metrics and classification and detection performance metrics. Continuous metrics include root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (PCC). Classification and detection performance metrics include critical success index (CSI), Heidegger skill score (HSS), and false alarm rate (FAR).

[0006] Preferably, S1 specifically includes the following: Low-resolution meteorological reanalysis data includes two categories: precipitation and ancillary meteorological data. Precipitation data includes total precipitation, large-scale precipitation, and convective precipitation variables. Ancillary meteorological data includes the 10-meter wind field U / V component and 2-meter temperature. High-resolution precipitation and topographic data are both publicly available. The collected data are divided into training and testing sets according to time, with earlier data used as the training set and more recent data as the testing set. Data preprocessing specifically involves performing a logarithmic transformation on precipitation data and standardization on ancillary meteorological data, with the following calculation formulas: ; ; in, This is the original precipitation value; To supplement the raw values ​​of meteorological data, , These are the mean and standard deviation of the corresponding variables in the training set, respectively.

[0007] Preferably, in S2, the multivariate feature interaction module obtains interaction features by modeling the complex relationships between meteorological variables; the super-resolution network first upsamples the interaction features, then fuses them with high-resolution terrain data, and finally extracts features to obtain a high-resolution feature map; the multi-level mask fusion module learns and fuses differentiated features of different precipitation intensities on the high-resolution feature map to generate the final downscaled precipitation field; the discriminator uses 4 layers of convolution to output a two-dimensional authenticity scoring matrix for judging the authenticity of the precipitation field.

[0008] Preferably, the construction method of the multivariate feature interaction module in S3 is as follows: S31. Constructing a feature projection layer: This involves applying the preprocessed precipitation data... With auxiliary meteorological data Precipitation projection features are obtained by projecting them onto a shared subspace using linear transformations. and auxiliary meteorological projection characteristics The transformation formula is: , ; in, , For the projection matrix, , For bias terms; S32. Construct a bidirectional cross-attention computation layer: [This is followed by a seemingly unrelated sentence about a cross- , Generate query matrices respectively Key matrix Value matrix And calculate bidirectional cross attention: ; ; ; ; in For the learnable parameter matrix determined through training, This is the scaling factor; S33. Constructing a gated residual fusion layer: Defining learnable gated scalars , and output transformation matrix , Focusing attention on preprocessed precipitation data Auxiliary meteorological data The data are fused and then normalized before being output to obtain the precipitation output characteristics. and auxiliary meteorological output characteristics : ; ; Ultimately, and The features are concatenated and the resulting multivariate interaction features are output. .

[0009] Preferably, the construction of the super-resolution network in S4 specifically includes the following: S41. Construct a feature fusion input layer: integrate multivariate interaction features. Upsampling is performed to match the spatial resolution of the high-resolution terrain data; then, the upsampled features and terrain data are stitched together along the channel dimension to output a preliminary fused feature map. S42. Constructing the encoder: The encoder consists of three cascaded downsampling modules. Each module performs the following operations in sequence: First, the feature map resolution is reduced by max pooling; then, the pooled features are input into the residual block, which contains two consecutive "convolutional layer-batch normalization layer-ReLU" operation combinations, and the input and output of the residual block are added together through skip connections. S43. Decoder Construction: This decoder has a symmetrical structure with the encoder, consisting of three cascaded upsampling modules. Each module performs the following operations in sequence: First, the input features are upsampled using transposed convolution; then, the upsampled feature map is concatenated with the corresponding feature map from the encoder through skip connections; finally, the concatenated features are input into a residual block with the same structure as the encoder for feature fusion; after three levels of upsampling, a high-resolution feature map is finally output. .

[0010] Preferably, the construction process of the multi-level mask fusion module in S5 includes: S51. Define precipitation intensity levels: Divide precipitation intensity into four levels: no rain, light rain, heavy rain, and rainstorm. S52. Construct a multi-level mask generation layer: Receive high-resolution feature maps. And for each intensity level The initial spatial mask is generated using a separate convolutional layer. : ; The kernel size used With precipitation intensity level The increase decreases; S53. Construct a dynamic weight learning network: Use an attention module to process the feature maps. Perform global average pooling and calculate the adaptive fusion weights corresponding to each level of mask accordingly. The masks at each level are summed according to their corresponding weights. After further refinement by convolutional layers, a downscaled precipitation field is output.

[0011] Preferably, in S6, the joint loss function Expressed as: ; in To counteract the loss, least squares generative adversarial network loss is employed; The weighted classification mask loss is expressed as follows: , Represents the total number of spatial locations. For the first The weights of each precipitation intensity category, the weights It increases with increasing precipitation intensity. This indicates the true high-resolution precipitation field in the th... At the spatial location, belonging to the first The probability values ​​of each category, The first term represents the downscaling precipitation field. The spatial location belongs to the first The probability of each category; Calculate the mean square error between the downscaled precipitation field and the true high-resolution precipitation field; , , , These are the adversarial loss weights, content loss weights, mask loss weights, and MSE loss weights, respectively.

[0012] Preferably, in S7, the formulas for each test indicator are as follows: ; ; ; ; ; ; in For spatial location index, The total number of spatial locations. To obtain a true high-resolution precipitation field in the first Precipitation values ​​at a spatial location To downscale the precipitation field in the first Precipitation values ​​at a spatial location , This represents the average of the true high-resolution precipitation field and the downscaled precipitation field. , , , These represent the number of samples that are true positive, false positive, true negative, and false negative, respectively.

[0013] The present invention also provides a computer program product that, when run on a computer, causes the computer to execute the precipitation downscaling method based on a generative adversarial network model.

[0014] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the precipitation downscaling method based on a generative adversarial network model.

[0015] Therefore, the precipitation downscaling method based on the generative adversarial network model described above has the following beneficial effects: 1) The multivariate feature interaction module fully models the complex nonlinear relationship between precipitation and multiple variables such as wind field and temperature; 2): The multi-level mask fusion module and the joint loss function work together to achieve differentiated feature learning for precipitation of different intensities.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention; Figure 2 This is a structural framework diagram of the precipitation downscaling model according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example 1 This invention provides a precipitation downscaling method based on a generative adversarial network model, the process of which is as follows: Figure 1 As shown, the constructed model is as follows Figure 2 As shown, the specific implementation steps are as follows: S1. Acquire low-resolution meteorological reanalysis data, high-resolution precipitation data, and high-resolution topographic data, and preprocess the data to divide them into training and testing sets.

[0021] Low-resolution meteorological reanalysis data includes two categories: precipitation and ancillary meteorological data. Precipitation data includes total precipitation, large-scale precipitation, and convective precipitation variables. Ancillary meteorological data includes the 10-meter wind field U / V component and 2-meter temperature. In this embodiment, the low-resolution meteorological reanalysis data was collected from the ERA5 reanalysis dataset. High-resolution precipitation data was collected from the MSWEP precipitation dataset; high-resolution topographic data was collected from the ETOPO1 global topographic model. All data were collected over a period from 1980 to 2020. Data from 1980 to 2015 was used for the training set, and data from 2016 to 2020 was reserved for the test set. Data preprocessing specifically involved performing a logarithmic transformation on the precipitation data and standardization on the ancillary meteorological data.

[0022] The calculation formula for precipitation data preprocessing is as follows: ; The calculation formula for auxiliary meteorological data preprocessing is as follows: ; in, This is the original precipitation value; To supplement the raw values ​​of meteorological data, , These are the mean and standard deviation of the corresponding variables in the training set, respectively.

[0023] S2. Constructing a downscaling model based on generative adversarial networks: The model consists of a generator and a discriminator. The generator is composed of a multivariate feature interaction module, a super-resolution network, and a multi-level mask fusion module. The discriminator consists of multiple convolutional layers. The downscaling model is optimized using a joint loss function during training. Specifically, the multivariate feature interaction module obtains interaction features by modeling the complex relationships between meteorological variables. The super-resolution network first upsamples the interaction features, then fuses them with high-resolution terrain data, and finally extracts features to obtain a high-resolution feature map. The multi-level mask fusion module learns and fuses differentiated features of different precipitation intensities from the high-resolution feature map to generate the final downscaled precipitation field. The discriminator uses four convolutional layers to output a two-dimensional authenticity scoring matrix for judging the authenticity of the precipitation field.

[0024] S3. Construct a multivariate feature interaction module: This module includes a feature projection layer, a bidirectional cross-attention calculation layer, and a gated residual fusion layer, used to accurately model the nonlinear relationships between multiple meteorological variables. The specific construction process is as follows: S31. Constructing a feature projection layer: This involves applying the preprocessed precipitation data... With auxiliary meteorological data Precipitation projection features are obtained by projecting them onto a shared subspace using linear transformations. and auxiliary meteorological projection characteristics The transformation formula is: , ; in, , For the projection matrix, , This is a bias term.

[0025] S32. Construct a bidirectional cross-attention computation layer: [This is followed by a seemingly unrelated sentence about a cross- , Generate query matrices respectively Key matrix Value matrix And calculate bidirectional cross attention: ; ; ; ; in For the learnable parameter matrix determined through training, This is the scaling factor.

[0026] S33. Constructing a gated residual fusion layer: Defining learnable gated scalars , and output transformation matrix , Focusing attention on preprocessed precipitation data Auxiliary meteorological data The data are fused and then normalized before being output to obtain the precipitation output characteristics. and auxiliary meteorological output characteristics : ; ; Ultimately, and The features are concatenated and the resulting multivariate interaction features are output. .

[0027] S4. Constructing a super-resolution network: This includes a feature fusion input layer, an encoder, and a decoder. The feature fusion input layer upsamples multivariate interactive features and fuses them with terrain data. Subsequently, the encoder and decoder extract features from the fused data. Specifically, this includes the following: S41. Construct a feature fusion input layer: integrate multivariate interaction features. Upsampling is performed to match the spatial resolution of the high-resolution terrain data; then, the upsampled features and terrain data are stitched together along the channel dimension to output a preliminary fused feature map.

[0028] S42. Constructing the encoder: The encoder consists of three cascaded downsampling modules. Each module performs the following operations in sequence: First, the feature map resolution is reduced by max pooling; then, the pooled features are input into the residual block, which contains two consecutive "convolutional layer-batch normalization layer-ReLU" operation combinations, and the input and output of the residual block are added together by skip connections.

[0029] S43. Decoder Construction: This decoder has a symmetrical structure with the encoder, consisting of three cascaded upsampling modules. Each module performs the following operations in sequence: First, the input features are upsampled using transposed convolution; then, the upsampled feature map is concatenated with the corresponding feature map from the encoder through skip connections; finally, the concatenated features are input into a residual block with the same structure as the encoder for feature fusion; after three levels of upsampling, a high-resolution feature map is finally output. .

[0030] S5. Construct a multi-level mask fusion module: This includes defining precipitation intensity levels, multi-level mask generation layers, and a dynamic weight learning network, used for learning differentiated features for precipitation of different intensities. The specific process is as follows: S51. Define precipitation intensity levels: Divide precipitation intensity into four levels: no rain, light rain, heavy rain, and torrential rain; this embodiment defines precipitation intensity levels. Daily precipitation The function (unit: mm) is defined piecewise as follows: .

[0031] S52. Construct a multi-level mask generation layer: Receive high-resolution feature maps. And for each intensity level The initial spatial mask is generated using a separate convolutional layer. : ; The kernel size used With precipitation intensity level The value decreases as it increases; in this embodiment, .

[0032] S53. Construct a dynamic weight learning network: Use an attention module to process the feature maps. Perform global average pooling and calculate the adaptive fusion weights corresponding to each level of mask accordingly. The masks at each level are summed according to their corresponding weights. After further refinement by convolutional layers, a downscaled precipitation field is output.

[0033] S6. Construct the joint loss function: It consists of two parts: adversarial loss and content loss; the content loss comprises weighted classification mask loss and mean squared error loss; the joint loss function... Expressed as: ; in To counteract the loss, least squares generative adversarial network loss is employed; The weighted classification mask loss is expressed as follows: , Represents the total number of spatial locations. For the first The weights of each precipitation intensity category, the weights It increases with increasing precipitation intensity. This indicates the true high-resolution precipitation field in the th... At the spatial location, belonging to the first The probability values ​​of each category, The first term represents the downscaling precipitation field. The spatial location belongs to the first The probability of each category; Calculate the mean square error between the downscaled precipitation field and the true high-resolution precipitation field; , , , These are the adversarial loss weights, content loss weights, mask loss weights, and MSE loss weights, respectively.

[0034] S7. Training and Evaluation of the Downscaling Model: The downscaling model is trained on the training set until convergence, and its performance is tested on the test set. Model performance metrics include continuous numerical accuracy metrics and classification / detection performance metrics. Continuous metrics include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Pearson Correlation Coefficient (PCC). Classification / detection performance metrics include Critical Success Index (CSI), Heidegger Skill Score (HSS), and False Alarm Rate (FAR). The thresholds for classification / detection metrics are 0.1 mm, 25 mm, and 50 mm. The formulas for each test metric are as follows: ; ; ; ; ; ; in For spatial location index, The total number of spatial locations. To obtain a true high-resolution precipitation field in the first Precipitation values ​​at a spatial location To downscale the precipitation field in the first Precipitation values ​​at a spatial location , This represents the average of the true high-resolution precipitation field and the downscaled precipitation field. , , , The numbers represent the sample numbers for true positives, false positives, true negatives, and false negatives, respectively. In this embodiment, the proposed method was further compared with existing methods in the field, and the results are shown in Tables 1 and 2, respectively. Table 1 compares the results of the continuous numerical accuracy index, and Table 2 compares the results of the classification detection performance index.

[0035] Table 1 Comparison of continuous numerical accuracy results between this method and existing methods in this field

[0036] Table 2 Comparison of classification and detection performance indicators between this method and existing methods in the field

[0037] Therefore, this invention adopts a precipitation downscaling method based on a generative adversarial network model, which accurately integrates precipitation with variables such as wind field and temperature through a multivariate feature interaction module, and achieves differentiated learning from no rain, light rain to heavy rain through a multi-level mask fusion module. The model is optimized by combining a joint loss function. This method helps to solve the problems of existing technologies that are difficult to model nonlinear relationships of multiple meteorological variables and do not consider the differentiated characteristics of precipitation of different intensities. It accurately depicts the spatiotemporal characteristics of fine-scale precipitation and improves the overall accuracy of precipitation downscaling.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A precipitation downscaling method based on a generative adversarial network model, characterized in that, Includes the following steps: S1. Acquire low-resolution meteorological reanalysis data, high-resolution precipitation data, and high-resolution topographic data, and preprocess the data, dividing them into training and testing sets. S2. Construct a downscaling model based on generative adversarial networks: The model consists of two parts: a generator and a discriminator. The generator is composed of a multivariate feature interaction module, a super-resolution network, and a multi-level mask fusion module in sequence. The discriminator is composed of multiple convolutional layers. The downscaling model is optimized using a joint loss function during training. S3. Construct a multivariate feature interaction module: including a feature projection layer, a bidirectional cross-attention calculation layer, and a gated residual fusion layer, used to accurately model the nonlinear relationship between multiple meteorological variables; S4. Construct a super-resolution network: including a feature fusion input layer, an encoder, and a decoder; upsample multivariate interactive features through the feature fusion input layer and fuse them with terrain data; then extract features from the fused data through the encoder and decoder; S5. Construct a multi-level mask fusion module: This includes defining precipitation intensity levels, multi-level mask generation layers, and a dynamic weight learning network, which are used to learn differentiated features for precipitation of different intensities. S6. Construct the joint loss function: It consists of two parts: adversarial loss and content loss; the content loss consists of weighted classification mask loss and mean squared error loss. S7. Training and Evaluation of Downscaling Model: The downscaling model is trained on the training set until convergence, and its performance is tested on the test set. Model performance metrics include continuous numerical accuracy metrics and classification and detection performance metrics. Continuous metrics include root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (PCC). Classification and detection performance metrics include critical success index (CSI), Heidegger skill score (HSS), and false alarm rate (FAR).

2. The precipitation downscaling method based on a generative adversarial network model according to claim 1, characterized in that, S1 specifically includes the following: Low-resolution meteorological reanalysis data includes two categories: precipitation and ancillary meteorological data. Precipitation data includes total precipitation, large-scale precipitation, and convective precipitation variables. Ancillary meteorological data includes the 10-meter wind field U / V component and 2-meter temperature. High-resolution precipitation and topographic data are both publicly available. The collected data are divided into training and testing sets according to time, with earlier data used as the training set and more recent data as the testing set. Data preprocessing specifically involves performing a logarithmic transformation on precipitation data and standardization on ancillary meteorological data, with the following calculation formulas: ； ； in, This is the original precipitation value; To supplement the raw values ​​of meteorological data, , These are the mean and standard deviation of the corresponding variables in the training set, respectively.

3. The precipitation downscaling method based on a generative adversarial network model according to claim 2, characterized in that, In S2, the multivariate feature interaction module obtains interaction features by modeling the complex relationships between meteorological variables; the super-resolution network first upsamples the interaction features, then fuses them with high-resolution terrain data, and finally extracts features to obtain a high-resolution feature map; the multi-level mask fusion module learns and fuses differentiated features of different precipitation intensities on the high-resolution feature map to generate the final downscaled precipitation field; the discriminator uses 4 layers of convolution to output a two-dimensional authenticity scoring matrix for judging the authenticity of the precipitation field.

4. The precipitation downscaling method based on a generative adversarial network model according to claim 3, characterized in that, The specific method for constructing the multivariate feature interaction module in S3 is as follows: S31. Constructing a feature projection layer: This involves applying the preprocessed precipitation data... With auxiliary meteorological data Precipitation projection features are obtained by projecting them onto a shared subspace using linear transformations. and auxiliary meteorological projection characteristics The transformation formula is: ， ； in, , For the projection matrix, , For bias terms; S32. Construct a bidirectional cross-attention computation layer: [This is followed by a seemingly unrelated sentence about a cross- , Generate query matrices respectively Key matrix Value matrix And calculate bidirectional cross attention: ； ； ； ； in For the learnable parameter matrix determined through training, This is the scaling factor; S33. Constructing a gated residual fusion layer: Defining learnable gated scalars , and output transformation matrix , Focusing attention on preprocessed precipitation data Auxiliary meteorological data The data are fused and then normalized before being output to obtain the precipitation output characteristics. and auxiliary meteorological output characteristics : ； ； Ultimately, and The features are concatenated and the resulting multivariate interaction features are output. .

5. The precipitation downscaling method based on a generative adversarial network model according to claim 4, characterized in that, The construction of the super-resolution network in S4 specifically includes the following: S41. Construct a feature fusion input layer: integrate multivariate interaction features. Upsampling is performed to match the spatial resolution of the high-resolution terrain data; then, the upsampled features and terrain data are stitched together along the channel dimension to output a preliminary fused feature map. S42. Constructing the encoder: The encoder consists of three cascaded downsampling modules. Each module performs the following operations in sequence: First, the feature map resolution is reduced by max pooling; then, the pooled features are input into the residual block, which contains two consecutive "convolutional layer-batch normalization layer-ReLU" operation combinations, and the input and output of the residual block are added together through skip connections. S43. Decoder Construction: The decoder has a symmetrical structure with the encoder and consists of three cascaded upsampling modules. Each module performs the following operations in sequence: First, the input features are upsampled through transposed convolution; then, the upsampled feature map is concatenated with the feature map of the corresponding layer in the encoder through skip connections; finally, the concatenated features are input into a residual block with the same structure as the encoder for feature fusion. After three levels of upsampling, the final output is a high-resolution feature map. .

6. The precipitation downscaling method based on a generative adversarial network model according to claim 5, characterized in that, The construction process of the multi-level mask fusion module in S5 includes: S51. Define precipitation intensity levels: Divide precipitation intensity into four levels: no rain, light rain, heavy rain, and rainstorm. S52. Construct a multi-level mask generation layer: Receive high-resolution feature maps. And for each intensity level The initial spatial mask is generated using a separate convolutional layer. : ； The kernel size used With precipitation intensity level The increase decreases; S53. Construct a dynamic weight learning network: Use an attention module to process the feature maps. Perform global average pooling and calculate the adaptive fusion weights corresponding to each level of mask accordingly. The masks at each level are summed according to their corresponding weights. After further refinement by convolutional layers, a downscaled precipitation field is output.

7. The precipitation downscaling method based on a generative adversarial network model according to claim 6, characterized in that, In S6, the joint loss function Expressed as: ； in To counteract the loss, least squares generative adversarial network loss is employed; The weighted classification mask loss is expressed as follows: , Represents the total number of spatial locations. For the first The weights of each precipitation intensity category, the weights It increases with increasing precipitation intensity. This indicates the true high-resolution precipitation field in the th... At the spatial location, belonging to the first The probability values ​​of each category, The first term represents the downscaling precipitation field. The spatial location belongs to the first The probability of each category; Calculate the mean square error between the downscaled precipitation field and the true high-resolution precipitation field; , , , These are the adversarial loss weights, content loss weights, mask loss weights, and MSE loss weights, respectively.

8. The precipitation downscaling method based on a generative adversarial network model according to claim 7, characterized in that, In S7, the formulas for each test indicator are as follows: ； ； ； ； ； ； in For spatial location index, The total number of spatial locations. To obtain a true high-resolution precipitation field in the first Precipitation values ​​at a spatial location To downscale the precipitation field in the first Precipitation values ​​at a spatial location , This represents the average of the true high-resolution precipitation field and the downscaled precipitation field. , , , These represent the number of samples that are true positive, false positive, true negative, and false negative, respectively.

9. A computer program product, characterized in that, When the computer program product is run on a computer, it causes the computer to perform the precipitation downscaling method based on a generative adversarial network model as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the precipitation downscaling method based on a generative adversarial network model as described in any one of claims 1-8.

Images