A Deep Learning-Based Downscaling Method for ERA5 Precipitation Products

Abstract

This invention discloses a deep learning-based method for downscaling ERA5 precipitation products. First, ERA5 reanalysis data, GPM satellite data, and daily precipitation observation data from ground stations are acquired and preprocessed. The correlation between meteorological factors and ground-observed precipitation is measured using the Pearson correlation coefficient method, and meteorological factors with high correlation are selected as effective features. Data are cropped to the study area and normalized. A precipitation downscaling model is established, and a loss function and model parameters are designed. The optimal model is trained, and the downscaling results are evaluated. A two-branch structure is adopted, using GPM satellite data as labels. One branch uses low-resolution precipitation as input, and the other branch uses effective meteorological features as input for learning assistance. The trained model not only captures temporal and spatial data variations but also effectively captures extreme cases, effectively reducing ERA5 precipitation data from 25km to 10km.

Description

Technical Field

[0001] This invention relates to the field of downscaling technology, and in particular to a deep learning-based method for downscaling ERA5 precipitation products. Background Technology

[0002] Precipitation estimation is fundamental to weather forecasting, aiming to accurately and timely predict precipitation intensity within a local area. This has always been a crucial issue, with typical precipitation estimation methods relying on numerical weather prediction and radar echo extrapolation. In recent years, with the development of artificial intelligence, methods focusing on learning potential precipitation representations within machine learning frameworks have emerged. However, these methods primarily concentrate on precipitation estimation within local areas, exhibiting relatively coarse spatial resolution, which fails to meet the needs of practical applications and cannot accurately reflect the spatiotemporal distribution of regional precipitation. High-resolution precipitation data helps characterize the spatiotemporal differentiation of precipitation and plays a vital role in the accurate simulation of watershed hydrological, meteorological, and ecological processes. Therefore, it is essential to spatially downscale low-resolution precipitation products to improve their resolution.

[0003] The key idea behind precipitation downscaling is to train a neural network using both low-resolution and high-resolution simulations, mapping the former to the latter. Currently, precipitation downscaling generally employs dynamic and statistical downscaling methods. Dynamic downscaling primarily simulates the physical environment in which precipitation occurs, generating high-resolution regional-scale precipitation data by embedding regional climate models into global climate models. However, this method requires significant computational resources and is subject to errors and uncertainties due to various influencing factors. Statistical downscaling achieves downscaling by establishing relationships between products of different resolutions. A common approach is to first interpolate low-resolution precipitation products onto a high-resolution grid, then correct the interpolation results point-by-point using historical high-resolution grid information. Common correction methods include linear functions, support vector machines, and artificial neural networks. However, traditional statistical downscaling methods often fail to learn this mapping relationship for regions lacking high-resolution observational data and tend to perform poorly for extreme events. Summary of the Invention

[0004] To address the above technical problems, this invention provides a deep learning-based method for downscaling ERA5 precipitation products, comprising the following steps:

[0005] S1. Acquire ERA5 reanalysis meteorological data, GPM satellite precipitation data, and ground rain gauge observation data;

[0006] S2, Data Preprocessing;

[0007] S3. Using the Pearson correlation coefficient method, calculate the correlation between ERA5 reanalysis meteorological data and ground rain gauge observation data, and select meteorological factors that affect precipitation as effective features.

[0008] S4. Clip each data point to the study area to obtain segmented data, and normalize the segmented data to construct a training dataset.

[0009] S5. Construct a precipitation downscaling model based on deep learning. The model includes a first branch, a second branch, and a residual attention network. The first branch takes the original ERA5 precipitation data as input, and the second branch takes meteorological factors that affect precipitation as input. The outputs of the first and second branches are fused and then input into the residual attention network for precipitation downscaling.

[0010] S6. Divide the training set and the test set, set the parameters and loss function of the precipitation downscaling model, and continuously train and adjust until the optimal parameter combination is obtained.

[0011] S7. Obtain the optimal precipitation downscaling model and use ground rain gauge observation data to analyze the accuracy of the downscaling results.

[0012] The technical solution further defined in this invention is:

[0013] Furthermore, in step S2, data preprocessing includes the following steps:

[0014] S2.1 Use the isnull() function in the Python pandas library to detect missing values;

[0015] S2.2 Detect outliers using box plot method;

[0016] S2.3. Based on spatial and temporal similarity, replace missing and outlier values ​​with the average value of the regions before, after, or adjacent to their locations.

[0017] In the aforementioned deep learning-based ERA5 precipitation product downscaling method, step S3 involves selecting meteorological factors that influence precipitation as effective features. The Pearson correlation coefficient method is used to calculate the correlation between each meteorological factor and ground rain gauge observation data. A preset threshold is established, and meteorological factors with a correlation greater than the preset threshold are considered to have an impact on precipitation. The Pearson correlation coefficient formula is as follows:

[0018]

[0019] Where Cov(X, Y) is the covariance of the eigenvalues ​​X and Y. and Let X and Y represent the variances of features X and Y, respectively.

[0020] The aforementioned deep learning-based ERA5 precipitation product downscaling method, in step S4, involves cropping the data to the study area, which includes the following steps.

[0021] S4.1.1 Convert ERA5 reanalysis meteorological data and GPM satellite precipitation data to Beijing time.

[0022] S4.1.2 Align the station coordinates with the grid data based on the latitude and longitude coordinates of each ground observation time;

[0023] S4.1.3. Divide the ERA5 reanalysis meteorological data and GPM satellite data into the study area and spread them outwards.

[0024] The aforementioned deep learning-based ERA5 precipitation product downscaling method, in step S4, the data normalization processing includes the following steps.

[0025] S4.2.1 Perform a Min-Max standardization operation on the split data. The calculation formula is as follows:

[0026]

[0027] Among them, y i x represents the standardized data. i Represents the original data, min 1≤j≤n {x j} represents the minimum value in the data, and max represents the maximum value. 1≤j≤n {x j} represents the maximum value in the data. This represents the normalized data;

[0028] S4.2.2. Apply Z-SCORE normalization to the standardized features to normalize the data features to the same dimension. The calculation formula is as follows:

[0029]

[0030] Among them, u i σ is the population mean of the i-th feature. i It is the overall standard deviation of the i-th feature.

[0031] The aforementioned deep learning-based ERA5 precipitation product downscaling method, in step S5, the first branch uses a convolutional neural network to extract the spatial features of the original ERA5 precipitation data, and the first branch includes 3 convolutional layers and one max pooling layer; the second branch uses a dense network structure, including 1 convolutional layer and 3 dense modules.

[0032] The dense module consists of three convolutional layers and three ReLU functions, with a ReLU function connected after each convolutional layer. The dense module inputs the features from each convolutional layer to all subsequent convolutional layers through skip connections.

[0033] In the aforementioned deep learning-based ERA5 precipitation product downscaling method, in step S5, the input in the residual attention network sequentially passes through a convolutional layer, a residual feature fusion module, a spatial and channel attention module, an upsampling layer, and another convolutional layer.

[0034] The residual feature fusion module includes three residual group modules. The tail output of each residual group module is made into a skip connection, and the element-wise AND operation is performed at the tail of the entire residual feature fusion module. The residual group module includes two 3*3 convolutional layers.

[0035] The spatial and channel attention modules include a spatial attention module and a channel attention module. In the spatial attention module, the input passes through parallel max pooling and average pooling layers. The pooled vectors are then fed into a fully connected layer for computation and summed. After passing through a sigmoid activation function, the sum is multiplied by the initial fusion feature to obtain the channel attention feature. In the channel attention module, max pooling and average pooling are performed spatially. The channels are concatenated and then convolved. Finally, the sum is multiplied by the channel attention map using a sigmoid activation function to obtain the final feature map.

[0036] In the aforementioned deep learning-based ERA5 precipitation product downscaling method, step S6 involves using a 10-fold cross-validation method to divide the training and test sets.

[0037] In the aforementioned deep learning-based ERA5 precipitation product downscaling method, step S6 includes various parameters of the precipitation downscaling model, such as adjustable learning rate, momentum parameter, total number of iterations, and weight decay parameter.

[0038] The adjustable learning rate is initially set to 10. -4 After 2×10 5 After the second backward iteration, the learning rate is halved, and the weight decay parameter is set to 10. -7 The momentum parameter was set to 0.5, and the precipitation downscaling model was trained using the Adam optimization method, where the first-order moment decay coefficient β1 = 0.9, the second-order moment decay coefficient β2 = 0.99, and the epsilon parameter ε in the Adam optimization method was 10. -8 The loss function for the precipitation downscaling model is set to L1 loss, as shown in the following formula:

[0039]

[0040] Where n represents the sample size, E iG represents the corresponding pixel value of the downscaled ERA5 precipitation data. i This indicates the pixel value corresponding to the GPM satellite precipitation data.

[0041] In the aforementioned deep learning-based ERA5 precipitation product downscaling method, step S7 involves analyzing the accuracy of the downscaling results using correlation coefficient, bias, and root mean square error, as shown in the following formulas.

[0042]

[0043] Among them, X i and Y i These represent the high-resolution precipitation data obtained after downscaling and the observation data from surface rain gauges, respectively. and These represent the average values ​​of the high-resolution precipitation data obtained after downscaling and the average values ​​of the observation data from ground rain gauges, respectively.

[0044] The beneficial effects of this invention are:

[0045] (1) In this invention, the elemental and spatial correlation of precipitation processes are fully considered. By selecting meteorological factors that have potential impact on precipitation, a precipitation downscaling model is constructed using a dual-branch network structure. GPM high-resolution precipitation data is used as labels. The model is continuously trained and optimized to obtain the optimal network parameters, thereby obtaining the best deep learning-based ERA5 precipitation product downscaling model.

[0046] (2) In this invention, a dual-branch network structure is used. One branch extracts the spatial precipitation features of the original coarse-resolution precipitation through a convolutional neural network (CNN). In order to extract powerful auxiliary guidance for downscaling precipitation and obtain a more accurate precipitation feature map, the other branch builds a dense network based on CNN, taking meteorological factors related to precipitation as input. All auxiliary meteorological elements share a common structure with shared parameters. This strategy enables the downscaling network to learn more specific features with meteorological relevance.

[0047] (3) In this invention, the features of the two branches in the dual-branch network structure are fused and input into the residual attention network to complete the precipitation downscaling. The residual attention network integrates the residual feature fusion mechanism and the spatial and channel attention mechanism. This network fuses the trained features in each residual block through skip connections to form a new set of features. This set of precipitation features contains richer and more complete information, and the spatial and channel attention module can assign weights to the features without missing a large number of features. The results show that the downscaling result of ERA5 has a more refined spatial structure than ERA5 and can reproduce the precipitation spatial pattern revealed by high-resolution simulation, showing a great improvement in precipitation spatial features. The downscaled ERA5 retains the temporal characteristics of the original ERA5 and is more consistent with the ground rain gauge data compared with the high-resolution simulation. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of the deep learning-based precipitation downscaling model in this invention;

[0050] Figure 3 This is a schematic diagram of the structure of the dense module in this invention;

[0051] Figure 4 This is a schematic diagram of the residual attention network in this invention;

[0052] Figure 5 This is a schematic diagram of the residual group module in this invention;

[0053] Figure 6 This is a schematic diagram of the spatial and channel attention module in this invention. Detailed Implementation

[0054] This embodiment provides a deep learning-based method for downscaling ERA5 precipitation products, such as... Figure 1 As shown, it includes the following steps

[0055] S1. Acquire ERA5 reanalysis meteorological data, GPM satellite precipitation data, and ground rain gauge observation data;

[0056] The study area in this embodiment is the southeastern region of China. The dataset mainly uses ERA5 reanalysis meteorological data for precipitation downscaling. ERA5 is the fifth generation atmospheric reanalysis dataset of ECMWF for global climate, which provides hourly estimates of a large number of atmospheric, terrestrial and oceanic climate variables with a spatial resolution of 25 km and a temporal resolution of 1 h.

[0057] The multi-satellite retrieval GPM_IMERG(Final) diurnal dataset was selected as the label. This data source has been widely proven to be relatively accurate among publicly available satellite precipitation datasets, with a spatial resolution of 10 km and a temporal resolution of 24 hours. This invention uses daily precipitation data from 858 meteorological stations in the study area provided by the China Meteorological Data Service Center to evaluate and analyze the downscaling results.

[0058] S2. Based on the acquired ERA5 reanalysis meteorological data, GPM satellite precipitation data, and ground rain gauge observation data, preprocessing is performed before inputting into the neural network model. The specific process is as follows: First, the isnull() function in the Python pandas library is used to detect missing values, and the box plot method is used to detect outliers. Based on the similarity in spatial and temporal dimensions, the average value of the area before and after or in the vicinity of the value is used for replacement.

[0059] S3. Using the Pearson correlation coefficient method, calculate the correlation between ERA5 reanalysis meteorological data and surface rain gauge observation data. A preset threshold of 0.2 is used. All features exceeding this threshold are considered valid features, i.e., meteorological factors affecting precipitation (ERA5 reanalysis meteorological data includes precipitation, wind, temperature, and many other meteorological data points; meteorological factors refer to meteorological elements other than precipitation in the ERA5 reanalysis meteorological data). This removes any unnecessary noise input that may negatively impact performance. Temperature, relative humidity, and wind are selected as meteorological factors affecting precipitation. The Pearson correlation coefficient formula is:

[0060]

[0061] Where Cov(X, Y) is the covariance of the eigenvalues ​​X and Y. and Let X and Y represent the variances of features X and Y, respectively.

[0062] S4. Clip the data to the study area, which includes contextual information about the spatial characteristics of the study area. This improves spatial relevance and allows for better extraction of spatial information distribution. Divide the ERA5 reanalysis meteorological data and GPM satellite data into the study area and spread them outwards to consider the impact of the surrounding environment on precipitation in the target area. Accumulate the collected ERA5 reanalysis meteorological data to a daily precipitation scale to maintain consistent temporal resolution. Add 8 hours to the ERA5 reanalysis meteorological data and GPM satellite data to convert them to Beijing time.

[0063] Considering the different dimensions of various meteorological features in GPM satellite precipitation data and ERA5 reanalysis meteorological data, which may affect network training, data normalization is necessary. To avoid introducing large outliers, the segmented data is first subjected to a Min-Max standardization operation, calculated using the following formula:

[0064]

[0065] Among them, y i x represents the standardized data. i Represents the original data, min 1≤j≤n {x j} represents the minimum value in the data, and max represents the maximum value. 1≤j≤n {x j} represents the maximum value in the data. This represents the normalized data;

[0066] Then, the standardized features are normalized using the Z-SCORE normalization method, calculated as follows:

[0067]

[0068] Data features are normalized to the same dimension to eliminate differences in feature scale, where u i σ is the population mean of the i-th feature. i It is the overall standard deviation of the i-th feature.

[0069] S5. Construct a precipitation downscaling model based on the deep learning PyTorch framework. The model structure is as follows: Figure 2 As shown, the model structure is divided into two branches. The first branch extracts the spatial features of the original coarse-resolution ERA5 precipitation data based on the convolutional neural network. This branch consists of a 5*5 convolutional layer conv1, a 3*3 convolutional layer conv2, a 2*2 convolutional layer, and a 2*2 max pooling layer.

[0070] Another branch adopts a dense network structure, which uses meteorological factors (relative humidity, temperature, wind) selected in step S3 that are highly correlated with precipitation as input to learn auxiliary guidance for accurate downscaling, so that the downscaling network can learn more specific features with meteorological correlation. This branch mainly consists of a 5*5 convolutional layer and three dense modules.

[0071] The structure of dense modules is as follows Figure 3As shown, each dense module consists of three convolutional layers and three ReLU functions. One convolutional layer and one ReLU function are sequentially connected to form a group. The dense network inputs the features of each convolutional layer to all subsequent convolutional layers through skip connections. This can reduce the computational complexity of the overall convolutional neural network parameters and compensate for the loss of meteorological factor features due to the establishment of deep convolutional neural network models.

[0072] The output features of the two branches are fused and used as input to the residual attention network, such as... Figure 4 As shown, the residual attention network includes a convolutional layer, a residual feature fusion module (RFAM), a spatial and channel attention module (CBAM), an upsampling layer, and a convolutional layer, which fully extracts the features of the input data. This residual cascade structure can build a very deep feature mapping network to improve the performance of super-resolution in order to achieve water reduction and downscaling.

[0073] The residual attention network can be divided into three stages: shallow feature extraction, deep feature mapping, and reconstruction. In the shallow feature extraction stage, there is only one convolutional layer with a kernel size of 3*3. The shallow features obtained are used as the input for the deep features.

[0074] The deep feature mapping module mainly consists of a residual feature fusion module and a spatial and channel attention module. The residual feature fusion module mainly includes three residual group modules (RBs), such as... Figure 5 As shown, the residual group module consists of two 3*3 convolutional layers. In order to ensure that the spatial and channel attention modules integrate and assign weights to the features without omission, a skip connection is made at the tail output of each residual group module. The element-wise sum operation is performed at the tail of the entire residual feature fusion module, and the resulting precipitation features have a fusion of low-frequency and high-frequency information.

[0075] Following this residual feature fusion module, a dual spatial and channel attention module is further added to filter out higher-level features from the trained precipitation features. The results of the spatial and channel attention module are as follows: Figure 6 As shown, the channel attention feature is first passed through a channel attention module, then through parallel max pooling and average pooling layers. The pooled vectors are fed into a fully connected layer for computation and then added together. After passing through the Sigmoid activation function, the vectors are multiplied by the initial fusion feature to obtain the channel attention feature.

[0076] Then, after passing through a spatial attention module, max pooling and average pooling are performed spatially. The channels are concatenated and then convolutional. Finally, the Sigmoid activation function is used to multiply the channel attention map to obtain the final feature map.

[0077] Finally, after upsampling and convolution in the reconstruction part, high-resolution precipitation data is obtained.

[0078] The model includes a first branch, a second branch, and a residual attention network. The first branch takes the original ERA5 precipitation data as input, and the second branch takes meteorological factors that affect precipitation as input. The outputs of the first and second branches are fused and then input into the residual attention network for precipitation downscaling.

[0079] S6. Use 10-fold cross-validation to divide the training set and test set samples, set the parameters and loss function of the precipitation downscaling model, update the model parameters according to the backpropagation algorithm, and train continuously until the loss error reaches the minimum; the parameters of the precipitation downscaling model include adjustable learning rate, momentum parameter, total number of iterations and weight decay parameter.

[0080] After repeated experiments and model training, the specific parameters of the model are set as follows: the adjustable learning rate is initially set to 10. -4 After 2×10 5 After the second backward iteration, the learning rate is halved, and the weight decay parameter is set to 10. -7 The momentum parameter was set to 0.5, and the precipitation downscaling model was trained using the Adam optimization method, where the first-order moment decay coefficient β1 = 0.9, the second-order moment decay coefficient β2 = 0.99, and the epsilon parameter ε in the Adam optimization method was 10. -8 The loss function for the precipitation downscaling model is set to L1 loss, as shown in the following formula:

[0081]

[0082] Where n represents the sample size, E i This represents the corresponding pixel value of the downscaled ERA5 precipitation data (high resolution), G i This indicates the pixel value corresponding to the GPM satellite precipitation data.

[0083] S7. Obtain the optimal precipitation downscaling model. Based on ground rain gauge observation data, evaluate and analyze the accuracy of the downscaling results to demonstrate the superiority of the downscaling method. The accuracy evaluation and analysis indicators mainly use the correlation coefficient (R), root mean square error (RMSE), and bias (BIAS). The formulas for the correlation coefficient, bias, and root mean square error are shown below.

[0084]

[0085] Among them, X i and Y i These represent the high-resolution precipitation data obtained after downscaling and the observation data from ground rain gauges, respectively. and These represent the average values ​​of the high-resolution precipitation data obtained after downscaling and the average values ​​of the observation data from ground rain gauges, respectively.

[0086] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A deep learning-based method for downscaling ERA5 precipitation products, characterized in that: Includes the following steps S1. Acquire ERA5 reanalysis meteorological data, GPM satellite precipitation data, and ground rain gauge observation data; S2, Data Preprocessing; S3. Using the Pearson correlation coefficient method, calculate the correlation between ERA5 reanalysis meteorological data and ground rain gauge observation data, and select meteorological factors that affect precipitation as effective features. S4. Clip each data point to the study area to obtain segmented data, and normalize the segmented data to construct a training dataset. S5. Construct a deep learning-based precipitation downscaling model. This model includes a first branch, a second branch, and a residual attention network. The first branch takes the original ERA5 precipitation data as input, and the second branch takes meteorological factors that affect precipitation as input. The outputs of the first and second branches are fused and then input into the residual attention network for precipitation downscaling. The first branch uses a convolutional neural network to extract the spatial features of the original ERA5 precipitation data. The first branch includes three convolutional layers and one max pooling layer. The second branch adopts a dense network structure, including one convolutional layer and three dense modules. The dense module consists of 3 convolutional layers and 3 ReLU functions, with a ReLU function connected after each convolutional layer. The dense module inputs the features from each convolutional layer to all subsequent convolutional layers through skip connections. In the residual attention network, the input passes through a convolutional layer, a residual feature fusion module, a spatial and channel attention module, an upsampling layer, and another convolutional layer in sequence. The residual feature fusion module includes three residual group modules. The tail output of each residual group module is made into a skip connection, and the element-wise AND operation is performed at the tail of the entire residual feature fusion module. The residual group module includes two 3*3 convolutional layers. The spatial and channel attention modules include a spatial attention module and a channel attention module. In the spatial attention module, the input passes through parallel max pooling and average pooling layers. The pooled vectors are fed into a fully connected layer for computation and then added together. After passing through the Sigmoid activation function, the vectors are multiplied by the initial fusion features to obtain the channel attention features. In the channel attention module, max pooling and average pooling are performed spatially. After concatenating the channels, a convolution operation is performed, and then the Sigmoid activation function is used to multiply it with the channel attention map to form the final feature map. S6. Divide the training set and the test set, set the parameters and loss function of the precipitation downscaling model, and continuously train and adjust until the optimal parameter combination is obtained. S7. Obtain the optimal precipitation downscaling model and use ground rain gauge observation data to analyze the accuracy of the downscaling results.

2. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S2, data preprocessing includes the following steps: S2.1 Use the isnull() function in the Python pandas library to detect missing values; S2.2 Detect outliers using box plot method; S2.

3. Based on spatial and temporal similarity, replace missing and outlier values ​​with the average value of the regions before, after, or adjacent to their locations.

3. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S3, meteorological factors that influence precipitation are selected as effective features. The correlation between each meteorological factor and ground rain gauge observation data is calculated using the Pearson correlation coefficient method. A preset threshold is established, and meteorological factors with a correlation greater than the preset threshold are considered to have an impact on precipitation. The Pearson correlation coefficient formula is as follows: Where Cov(X,Y) is the covariance of the eigenvalues ​​X and Y. and Let X and Y represent the variances of features X and Y, respectively.

4. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S4, cropping the data to the study area includes the following steps: S4.1.1 Convert ERA5 reanalysis meteorological data and GPM satellite precipitation data to Beijing time. S4.1.2 Align the station coordinates with the grid data based on the latitude and longitude coordinates of each ground observation time; S4.1.

3. Divide the ERA5 reanalysis meteorological data and GPM satellite data into the study area and spread them outwards.

5. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S4, the data normalization process includes the following steps: S4.2.1 Perform a Min-Max standardization operation on the split data. The calculation formula is as follows: Among them, y i The data represents the standardized data, xi represents the original data, and min represents the normalized data. 1≤j≤n {x j } represents the minimum value in the data, and max represents the maximum value. 1≤j≤n {x j } represents the maximum value in the data. This represents the normalized data; S4.2.

2. Apply Z-SCORE normalization to the standardized features to normalize the data features to the same dimension. The calculation formula is as follows: Among them, u i σ is the population mean of the i-th feature. i It is the overall standard deviation of the i-th feature.

6. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S6, the training set and the test set are divided using the 10-fold cross-validation method.

7. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S6, the parameters of the precipitation downscaling model include the adjustable learning rate, momentum parameter, total number of iterations, and weight decay parameter. The adjustable learning rate is initially set to 10. -4 After 2×10 5 After the second backward iteration, the learning rate is halved, and the weight decay parameter is set to 10. -7 The momentum parameter was set to 0.5, and the precipitation downscaling model was trained using the Adam optimization method, where the first-order moment decay coefficient β1 = 0.9, the second-order moment decay coefficient β2 = 0.99, and the epsilon parameter ε in the Adam optimization method was 10. -8 The loss function for the precipitation downscaling model is set to L1 loss, as shown in the following formula: Where n represents the sample size, E i G represents the corresponding pixel value of the downscaled ERA5 precipitation data. i This indicates the pixel value corresponding to the GPM satellite precipitation data.

8. The method for downscaling ERA5 precipitation products based on deep learning according to claim 1, characterized in that: In step S7, the accuracy of the downscaling results is analyzed using the correlation coefficient, bias, and root mean square error, respectively, as shown in the following formulas. Among them, X i and Y i These represent the high-resolution precipitation data obtained after downscaling and the observation data from surface rain gauges, respectively. and These represent the average values ​​of the high-resolution precipitation data obtained after downscaling and the average values ​​of the observation data from ground rain gauges, respectively.

Images