Tibet plateau global climate mode rainfall downscaling method based on double attention

Abstract

The invention discloses a Qinghai-Tibet Plateau global climate mode rainfall downscaling method based on double attention, and relates to the technical field of rainfall downscaling, and the method comprises the following steps: obtaining coarse resolution global climate mode CMIP6 rainfall data, high resolution reference rainfall data and multi-source auxiliary data, and carrying out the standardization preprocessing of all data; a generative adversarial network model is constructed, and a generator of the generative adversarial network model adopts a dual attention mechanism embedded with a channel attention module and a space attention module and is connected in series with a dense block to form a feature extraction main body; training the generative adversarial network model by adopting a two-stage sequential downscaling strategy; and inputting to-be-downscaled CMIP6 rainfall data into the final model, and outputting to obtain a high-resolution downscaled rainfall product. According to the method, the generative adversarial network fusing dense connection and a double attention mechanism is combined with the multi-scale discriminator, so that high-precision downscaling of the CMIP6 rainfall data is realized.

Description

Technical Field

[0001] This invention belongs to the field of precipitation downscaling technology, and in particular relates to a precipitation downscaling method based on dual attention for the Tibetan Plateau global climate model. Background Technology

[0002] Due to inherent coarse grid scales and limitations in physical parameterization, precipitation products from the China Climate Information Portal 6 (CMIP6) typically have a raw resolution of only tens to hundreds of kilometers. Using CMIP6 precipitation data for analysis leads to significant biases, specifically manifested as shifts in the location of orographic rainfall, an overestimation of the frequency of drizzle events, and underestimations or overestimations of extreme precipitation intensity. These limitations make CMIP6 precipitation data unsuitable for direct application to watershed hydrological simulations and engineering-scale analyses in the complex topographic region of the Tibetan Plateau.

[0003] To address the insufficient resolution of precipitation products from climate models, existing downscaling methods can be broadly categorized into four types: Statistical / Empirical methods: including multiple regression, geostatistics, weather type analogy, bias correction and spatial downscaling (BCSD), and quantile mapping (QDM). These methods are simple to implement and highly interpretable, but their effectiveness heavily relies on the assumption of statistical stationarity, making them inadequate for characterizing extreme events and non-stationary climate change, and prone to oversmoothing effects. Dynamical downscaling methods: such as those using regional climate models like WRF / RegCM. These methods better reflect the consistency of local topographic effects and physical processes. However, they are computationally expensive, highly sensitive to model parameterization schemes, and biases in upstream boundary conditions can propagate to downstream regions. Machine learning methods: such as Random Forest (RF), Support Vector Regression (SVR), and Gradient Boosting Tree (XGBoost). These methods excel at capturing the nonlinear relationships between precipitation and multi-source topographic / land surface factors. However, their performance is highly dependent on the quality and representativeness of the training samples, and their generalization ability and robustness still need improvement. Deep learning / super-resolution methods include Convolutional Neural Networks (CNNs), U-Nets, Generative Adversarial Networks (GANs), Transformers, and Graph Neural Networks. These methods demonstrate strong potential for learning multi-scale textures and reconstructing high-resolution details end-to-end. However, they may suffer from block artifacts, checkerboard effects, and physical non-conservation issues in applications, and they have extremely high requirements for high-quality, large-sample training data.

[0004] Given the unique geographical and climatic characteristics of the Tibetan Plateau, CMIP6 precipitation downscaling techniques for this region urgently need to strike an effective balance between high-resolution spatial structure restoration, accurate characterization of extreme events, and robust generalization capabilities across different time periods. Existing single statistical or learning strategies are insufficient to comprehensively address these requirements. Therefore, there is an urgent need to develop an interpretable scheme that can effectively incorporate prior knowledge of topography / land surface, and to combine uncertainty quantification techniques with multi-source observational data (such as site observations, satellite remote sensing, and reanalysis data) for collaborative calibration, in order to significantly improve the reliability and practicality of downscaling products. Summary of the Invention

[0005] The purpose of this invention is to provide a method for downscaling precipitation data from the Tibetan Plateau global climate model based on dual attention. By integrating a generative adversarial network with dense connections and a dual attention mechanism, and combining a multi-scale discriminator with physical constraints, this method achieves high-precision downscaling of CMIP6 precipitation data, solving the problems of detail loss, obvious artifacts, and insufficient generalization ability of existing methods in complex terrain.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a precipitation downscaling method based on a dual-attention global climate model for the Tibetan Plateau. The downscaling method includes the following steps: Step S1, Data Preparation: Acquire coarse-resolution global climate model CMIP6 precipitation data, high-resolution reference precipitation data, and multi-source auxiliary data, and perform standardization preprocessing on all data; Step S2, Model Construction: Construct a generative adversarial network model. Its generator adopts a dual attention mechanism with embedded channel attention modules and spatial attention modules, and is connected in series with dense blocks to form the main body of feature extraction. Step S3, Model Training: The generative adversarial network model is trained using a two-stage sequential downscaling strategy, specifically as follows: In the first stage, the model is trained using the first coarse-resolution CMIP6 precipitation data and the multi-source auxiliary data as inputs, and the corresponding high-resolution reference precipitation data as the target, to obtain the model trained in the first stage. In the second stage, the weights of the model trained in the first stage are loaded for initialization. The model is fine-tuned by transfer learning using the second coarse-resolution CMIP6 precipitation data and the multi-source auxiliary data as inputs and the higher-resolution high-resolution reference precipitation data as the target, to obtain the final model. Step S4, Precipitation Downscaling: Input the CMIP6 precipitation data to be downscaled into the final model, and output a high-resolution downscaled precipitation product; Step S5, Evaluation: The quality of the downscaling precipitation product is evaluated using the Nash efficiency coefficient (NSE), correlation coefficient (CC), root mean square error (RMSE), and mean absolute error (MAE).

[0007] Furthermore, the construction of the generator specifically involves: The generator consists of multiple cascaded Dense Blocks, each containing multiple convolutional layers. The input to each layer is the concatenation of the channel dimensions of the feature maps output by all preceding layers within that Dense Block. The dual attention mechanism is embedded after each Dense Block. This mechanism includes a channel attention module and a spatial attention module. The channel attention module is used to generate channel weight vectors to emphasize hydrological feature channels related to precipitation. The spatial attention module is used to generate a spatial weight map to emphasize key spatial locations.

[0008] Furthermore, the channel attention module obtains the channel attention map Acam by calculating the correlation between the channels of the input feature map. The calculation process is as follows: In the formula, For the input feature map, For normalized exponential functions, This is a transpose operation.

[0009] Furthermore, the spatial attention module queries the matrix. and key matrix The interaction yields the spatial attention map Apam, whose computation process is represented as follows: In the formula, and These represent the query matrix and the key matrix, respectively. and From the input feature map Obtained through linear transformation, It is a normalized exponential function.

[0010] Furthermore, in the model training step, the loss function used is a composite loss function, which includes at least: pixel-level loss, used to constrain numerical consistency; perceptual loss, used to improve the ability to restore structure and texture; and adversarial loss, used to stabilize training and improve the realism of the results.

[0011] Furthermore, the adversarial loss uses Wasserstein distance as a metric and is coupled with a gradient penalty term.

[0012] Furthermore, the multi-source auxiliary data includes multiple variables from ERA5_Land, GPM_IMERGE, GLDAS-NOAH2.0 and SRTM data sources, with a preferred total of 37 variables, covering evapotranspiration, air temperature, runoff, albedo, soil moisture, digital elevation model (DEM), etc.

[0013] Furthermore, the two-stage sequential downscaling strategy is specifically as follows: The first phase involves downscaling CMIP6 precipitation data from 1° to 0.25°. The second phase will downscale the CMIP6 precipitation data from 0.25° to 0.1° or 0.05°.

[0014] The present invention has the following beneficial effects: This invention achieves high-precision downscaling of CMIP6 precipitation data by integrating a generative adversarial network with dense blocks and a dual attention mechanism, combined with a multi-scale discriminator. This method significantly improves the ability to restore spatial details under the complex topographic conditions of the Tibetan Plateau, suppresses artifacts and oversmoothing, enhances the accuracy of capturing extreme precipitation events, and improves the model's generalization ability under different climate scenarios. Compared with traditional statistical methods and existing deep learning models, this method shows significant advantages in preserving the spatial structure of precipitation, reasonably reflecting the influence of topography, and ensuring hydrological consistency, providing reliable technical support for high-precision climate and hydrological simulation.

[0015] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the precipitation downscaling method of the Tibetan Plateau global climate model based on dual attention, as described in this invention. Figure 2 This is a schematic diagram of the generator structure of the present invention; Figure 3 This is a schematic diagram illustrating the discriminator and loss function of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-3 As shown, this invention is a precipitation downscaling method based on dual attention for the Tibetan Plateau global climate model, comprising the following steps: Step S1, Data Preparation: Acquire coarse-resolution global climate model CMIP6 precipitation data, high-resolution reference precipitation data, and multi-source auxiliary data, and perform standardization preprocessing on all data; Step S2, Model Construction: Construct a generative adversarial network model. Its generator adopts a dual attention mechanism with embedded channel attention modules and spatial attention modules, and is connected in series with dense blocks to form the main body of feature extraction. Step S3, Model Training: A two-stage sequential downscaling strategy is used to train the generative adversarial network model, specifically as follows: In the first stage, the model is trained using the first coarse-resolution CMIP6 precipitation data and multi-source auxiliary data as inputs and the corresponding high-resolution reference precipitation data as the target, resulting in the model trained in the first stage. In the second stage, the weights of the model trained in the first stage are loaded for initialization. The model is fine-tuned by transfer learning using the second coarse-resolution CMIP6 precipitation data and multi-source auxiliary data as inputs and the higher-resolution high-resolution reference precipitation data as the target, to obtain the final model. Step S4, Precipitation Downscaling: Input the CMIP6 precipitation data to be downscaled into the final model, and output a high-resolution downscaled precipitation product; Step S5, Evaluation: The quality of downscaled precipitation products is evaluated using the Nash efficiency coefficient (NSE), correlation coefficient (CC), root mean square error (RMSE), and mean absolute error (MAE).

[0020] The generator is constructed as follows: The generator consists of multiple cascaded Dense Blocks, each containing multiple convolutional layers. The input to each layer is the concatenation of the channel dimensions of the feature maps output by all the preceding layers in that Dense Block. A dual attention mechanism is embedded after each Dense Block, which consists of a channel attention module and a spatial attention module. The channel attention module is used to generate channel weight vectors to highlight hydrological feature channels related to precipitation. The spatial attention module is used to generate a spatial weight map to highlight key spatial locations.

[0021] The channel attention module obtains the channel attention map Acam by calculating the affinity between each channel of the input feature map. The calculation process is as follows: In the formula, For the input feature map, For normalized exponential functions, This is a transpose operation.

[0022] The spatial attention module queries the matrix. and key matrix The interaction yields the spatial attention map Apam, whose computation process is represented as follows: In the formula, and These represent the query matrix and the key matrix, respectively. and From the input feature map Obtained through linear transformation, It is a normalized exponential function.

[0023] In the model training process, the loss function used is a composite loss function, which includes at least: pixel-level loss, used to constrain numerical consistency; perceptual loss, used to improve the ability to restore structure and texture; and adversarial loss, used to stabilize training and improve the realism of the results.

[0024] The adversarial loss uses Wasserstein distance as a metric, along with a gradient penalty term.

[0025] The multi-source auxiliary data includes multiple variables from ERA5_Land, GPM_IMERGE, GLDAS-NOAH2.0 and SRTM data sources, with a preferred total of 37 variables, covering evapotranspiration, air temperature, runoff, albedo, soil moisture, digital elevation model (DEM), etc.

[0026] The two-stage sequential downscaling strategy is as follows: The first phase involves downscaling CMIP6 precipitation data from 1° to 0.25°. The second phase will downscale the CMIP6 precipitation data from 0.25° to 0.1° or 0.05°.

[0027] One specific application of this embodiment is: This method employs a coarse-to-fine sequential multi-level super-resolution strategy, progressively increasing the CMIP6 precipitation resolution from a 1° grid to 0.1°. The process consists of two stages: from 1° to 0.25°, model training is performed first to obtain robust training convergence and an objective validation foundation; subsequently, the final high-resolution results are obtained in the 0.25° to 0.1° stage. Both stages share the same GAN-DANet framework. The second stage uses the optimal weights from the first stage as initialization for transfer learning fine-tuning, thereby inheriting and refining the learned topographic-precipitation relationships and spatial structure. Each stage of the entire process uses 37 predictive factors (see Table 1) as conditional inputs, providing constraints and prior guidance for detailed reconstruction and physical consistency.

[0028] Table 1 Data used Phase 1: 1° → 0.25°; Input and objective: Input (X): CMIP6 precipitation data (1°) + 37 predictors listed in Table 1; Target (Y): GPM IMERGE precipitation data (0.25°); Model and Training: Generator (G): GAN-DANet (with dual attention / channel and spatial features, used to highlight hydrological features); Discriminator (D): Discriminator (discriminates for 0.25° target domain texture / structure / perceptual consistency); Loss Design: Pixel-level loss (constraining numerical consistency); perceptual loss (improving structure / texture restoration); Wasserstein adversarial loss (stabilizing training); Optimization: Alternately update G / D and monitor the performance of the validation set to select the optimal weight.

[0029] Performance evaluation: Using Nash efficiency coefficient (NSE), correlation coefficient (CC), mean absolute error (MAE), and root mean square error (RMSE) as the main indicators, the 0.25° results are objectively tested to determine the optimal model weights for stage one. Phase Two: 0.25° → 0.1° Input and objective: Input (X): CMIP6 precipitation data (0.25°) + the same 37 predictors in Table I.

[0030] Initialization: Load the optimal generator / discriminator weights obtained in phase one and perform the transfer.

[0031] Target / Output (Ŷ): Precipitation product (final product) with a resolution of 0.1°.

[0032] Results and output: The final result of this embodiment is a 0.1° CMIP6 high-resolution product output. The quality evaluation was completed using the same metrics as in Phase 1, such as NSE, CC, MAE, and RMSE.

[0033] Dense Block: The generator employs a multi-level series of dense blocks as the main feature extraction unit. Its core idea is "layer-by-layer output and accumulation": each convolutional layer not only receives the output of the previous layer but also concatenates it with the feature maps of all previous layers using channel dimension concatenation before further convolution (commonly using "1×1 bottleneck + 3×3 convolution", with the growth rate k controlling the number of new channels per layer), thereby achieving efficient information reuse and short-path propagation of gradients. This design significantly alleviates information loss and gradient vanishing in deep networks, while also covering more semantic and scale cues with fewer parameters. Combined with transition layers controlling channel width and resolution, the network achieves a balance between enhancing details and controlling computational load. This approach is particularly effective for scenarios with strong topographic relief and significant spatial heterogeneity, such as the Qinghai-Tibet Plateau.

[0034] Within a dense block, the output of any layer is It is formed by connecting the feature maps of the previous layers. Therefore, it can be expressed as: The output feature map of the layer ; This indicates a convolution operation, which typically involves a convolutional layer, followed by batch normalization and a ReLU activation function. Indicates from level 0 to level 1 Within the block; this output They will also be connected and passed as input to all subsequent layers in the block, thereby further enhancing feature reuse; Dual attention mechanism (channel attention + spatial attention): A dual attention mechanism is embedded after each DenseBlock: first channel attention, then spatial attention. Channel attention refines the global response of each channel through global averaging / max pooling, and then shares the channel weight vector to dynamically "amplify" hydrological variables strongly correlated with precipitation (such as evapotranspiration, snow cover, and soil moisture) while suppressing redundant information. Spatial attention convolves the aggregated feature map to generate a location weight map, emphasizing location information. The combination of these two approaches is equivalent to first answering "which features are more important," and then answering "where are these features more important?" This allows for adaptive focusing on hydrologically sensitive areas of the Qinghai-Tibet Plateau without relying on external priors, thereby improving boundary continuity and the reliability of details.

[0035] and The matrices () and key matrices () represent the query matrix and key matrix, respectively, which are derived from the linear transformation of the input feature map. The dot product (·) operation generates an attention map, which, after passing through the softmax function, yields the spatial attention map. This attention map clearly highlights the most important spatial locations. Finally, the resulting attention map is applied to the input feature map to effectively refine and enhance the spatial features.

[0036] This is a channel attention map, which quantifies the dependencies between various channels. During computation, the input feature map... First, it was reshaped into a Matrix, to make the dot product A C×C matrix is ​​generated to capture the pairwise correlations between channels. Then, a softmax function normalizes this matrix to ensure that the attention score for each channel has an appropriate proportion. The resulting attention map... By assigning higher attention scores to channels that demonstrate stronger relevance or contribute more significantly to a given task, the mechanism clearly identifies which channels are more important. This effectively guides the model to emphasize relevant channel features.

[0037] Compared with existing common deep learning model downscaling, our method has significant advantages in correlation coefficient (CC), Nash-Sutcliffe efficiency (NSE), root-mean-square error (RMSE), and mean absolute error (MAE), as shown in the table below.

[0038] Table 2 Comparison with existing common deep learning models In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0039] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for downscaling precipitation in a global climate model of the Tibetan Plateau based on dual attention, characterized in that... The downscaling method includes the following steps: Step S1, Data Preparation: Acquire coarse-resolution global climate model CMIP6 precipitation data, high-resolution reference precipitation data, and multi-source auxiliary data, and perform standardization preprocessing on all data; Step S2, Model Construction: Construct a generative adversarial network model. Its generator adopts a dual attention mechanism with embedded channel attention modules and spatial attention modules, and is connected in series with dense blocks to form the main body of feature extraction. Step S3, Model Training: The generative adversarial network model is trained using a two-stage sequential downscaling strategy, specifically as follows: In the first stage, the model is trained using the first coarse-resolution CMIP6 precipitation data and the multi-source auxiliary data as inputs, and the corresponding high-resolution reference precipitation data as the target, to obtain the model trained in the first stage. In the second stage, the weights of the model trained in the first stage are loaded for initialization. The model is fine-tuned by transfer learning using the second coarse-resolution CMIP6 precipitation data and the multi-source auxiliary data as inputs and the higher-resolution high-resolution reference precipitation data as the target, to obtain the final model. Step S4, Precipitation Downscaling: Input the CMIP6 precipitation data to be downscaled into the final model, and output a high-resolution downscaled precipitation product.

2. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 1, is characterized in that... The generator is constructed as follows: The generator consists of multiple cascaded Dense Blocks, each containing multiple convolutional layers. The input to each layer is the concatenation of the channel dimensions of the feature maps output by all preceding layers within that Dense Block. The dual attention mechanism is embedded after each Dense Block, which includes a channel attention module and a spatial attention module in sequence. The channel attention module is used to generate channel weight vectors to emphasize hydrological feature channels related to precipitation. The spatial attention module is used to generate a spatial weight map to emphasize key spatial locations.

3. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 2, is characterized in that... The channel attention module obtains the channel attention map Acam by calculating the correlation between the channels of the input feature map. The calculation process is as follows: In the formula, For the input feature map, For normalized exponential functions, This is a transpose operation.

4. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 2, is characterized in that... The spatial attention module queries the matrix. and key matrix The interaction yields the spatial attention map Apam, whose computation process is represented as follows: In the formula, and These represent the query matrix and the key matrix, respectively. and From the input feature map Obtained through linear transformation, It is a normalized exponential function.

5. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 1, is characterized in that... In the model training step, the loss function used is a composite loss function, which includes at least: pixel-level loss, used to constrain numerical consistency; perceptual loss, used to improve the ability to restore structure and texture; and adversarial loss, used to stabilize training and improve the realism of the results.

6. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 5, is characterized in that... The adversarial loss uses Wasserstein distance as a metric and is coupled with a gradient penalty term.

7. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 1, is characterized in that... The multi-source auxiliary data includes multiple variables from ERA5_Land, GPM_IMERGE, GLDAS-NOAH2.0 and SRTM data sources, with a preferred total of 37 variables covering evapotranspiration, air temperature, runoff, albedo, soil moisture, and digital elevation model (DEM).

8. The method for downscaling precipitation in the Tibetan Plateau global climate model based on dual attention, as described in claim 1, is characterized in that... The two-stage sequential downscaling strategy is as follows: The first phase involves downscaling CMIP6 precipitation data from 1° to 0.25°. The second phase will downscale the CMIP6 precipitation data from 0.25° to 0.1° or 0.05°.

Images