CN-121980239-A - Dual-attention-based global climate mode water and scale reduction method for Qinghai-Tibet plateau

CN121980239ACN 121980239 ACN121980239 ACN 121980239ACN-121980239-A

Abstract

The invention discloses a global climate mode water-reducing and scale-reducing method for Qinghai-Tibet plateau based on dual attention, which relates to the technical field of water-reducing scale, and comprises the following steps of acquiring coarse resolution global climate mode CMIP6 water-reducing data, high resolution reference water-reducing data and multisource auxiliary data, and carrying out standardized pretreatment on all the data; the method comprises the steps of constructing a generated countermeasure network model, adopting a dual attention mechanism embedded with a channel attention module and a space attention module by a generator, connecting the dual attention mechanism with a Dense Block Dense Block in series to form a feature extraction main body, adopting a two-stage sequential downscaling strategy to train the generated countermeasure network model, inputting CMIP6 precipitation data to be downscaled into the final model, and outputting to obtain a downscaled precipitation product with high resolution. The invention combines the dense connection and the generation of the dual-attention mechanism with the multiscale discriminant through the countermeasure network, thereby realizing the high-precision downscaling of the CMIP6 precipitation data.

Inventors

SUN ZHANGLI
XIONG YI
Wei Xiangchao
LIAO YUXIN

Assignees

成都理工大学

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (8)

1. The global climate mode water-reducing and scale-reducing method for the Qinghai-Tibet plateau based on dual attention is characterized by comprising the following steps of: step S1, data preparation, namely acquiring coarse resolution global climate mode CMIP6 precipitation data, high resolution reference precipitation data and multisource auxiliary data, and carrying out standardized preprocessing on all the data; s2, constructing a model, namely constructing a generating countermeasure network model, wherein a generator of the model adopts a dual attention mechanism embedded with a channel attention module and a space attention module and is connected with a Dense Block Dense Block in series to form a feature extraction main body; Step S3, model training, namely training the generated countermeasure network model by adopting a two-stage sequential downscaling strategy, wherein the training comprises the following steps of: The first stage, taking first coarse resolution CMIP6 precipitation data and the multisource auxiliary data as input, taking high resolution reference precipitation data with corresponding resolution as a target, and training the model to obtain a model trained in the first stage; Loading the weight of the model trained in the first stage for initialization, taking second coarse resolution CMIP6 precipitation data and the multisource auxiliary data as input, and taking higher resolution high resolution reference precipitation data as a target, and carrying out migration fine adjustment on the model to obtain a final model; and S4, precipitation downscaling, namely inputting the CMIP6 precipitation data to be downscaling into the final model, and outputting to obtain a downscaling precipitation product with high resolution.
2. The dual-attention-based global climate pattern water and scale reduction method for Qinghai-Tibet plateau of claim 1, wherein the generator is constructed specifically as follows: the generator is composed of a plurality of cascaded Dense blocks, each Dense Block comprises a plurality of convolution layers, and the input of each layer is the channel dimension splicing of the output characteristic diagrams of all the previous layers in the Dense Block; The dual attention mechanism is embedded after each Dense Block, and comprises a channel attention module and a space attention module in sequence, wherein the channel attention module is used for generating a channel weight vector so as to emphasize a hydrologic characteristic channel related to precipitation, and the space attention module is used for generating a space weight graph so as to emphasize a key space position.
3. The global climate pattern water-down and scale-down method for the Qinghai-Tibet plateau based on dual attention according to claim 2, wherein the channel attention module obtains the channel attention map Acam by calculating the correlation between the channels of the input feature map, and the calculation process is expressed as: In the formula, In order to input the feature map, For the normalization of the exponential function, Is a transpose operation.
4. The dual-attention-based global climate pattern water-down and scale-down method for Qinghai-Tibet plateau of claim 2, wherein the spatial attention module is configured to query the matrix Key matrix Is a spatial attention diagram Apam, the calculation of which is expressed as: In the formula, And Representing the query matrix and the key matrix respectively, And From input feature maps Is obtained through the linear transformation of the materials, Is a normalized exponential function.
5. The global climate pattern water-reducing and scale-reducing method for Tibet plateau based on dual attention as claimed in claim 1, wherein the loss function adopted in the model training step is a composite loss function, and at least comprises pixel level loss for restricting numerical consistency, perception loss for improving structure and texture restoring capability, and countermeasures loss for stabilizing training and improving the authenticity of results.
6. The dual-attention-based global climate pattern water and scale reduction method for Qinghai-Tibet plateau of claim 5, wherein the countermeasures against loss using Wasserstein distance as a measure and matching gradient penalty terms.
7. The dual attention based global climate pattern water and scale reduction method for Qinghai-Tibet plateau of claim 1, wherein the multi-source assistance data comprises a plurality of variables from the group consisting of ERA5_Land, GPM_IMERGE, GLDAS-NOAH2.0 and SRTM data sources, preferably 37 total variables, covering vapor emission, air temperature, runoff, albedo, soil moisture, digital elevation model DEM.
8. The dual-attention-based global climate pattern water and scale reduction method for Qinghai-Tibet plateau of claim 1, wherein the two-stage sequential scale reduction strategy is specifically: the first stage downscales CMIP6 precipitation data from 1 ° to 0.25 °; the second phase downscales CMIP6 precipitation data from 0.25 ° to 0.1 ° or 0.05 °.

Description

Dual-attention-based global climate mode water and scale reduction method for Qinghai-Tibet plateau Technical Field The invention belongs to the technical field of precipitation downscaling, and particularly relates to a global climate mode precipitation downscaling method of Qinghai-Tibet plateau based on dual attention. Background Precipitation products in global climate mode (CMIP 6) typically have an original resolution of only tens to hundreds of kilometers due to their inherent grid scale roughness and physical parameterization limitations. Analysis using CMIP6 precipitation data will result in significant deviations, particularly as problems with terrain rainfall position shifts, high frequencies of precipitation events, and under-or over-estimation of extreme precipitation intensities. These limitations make CMIP6 precipitation data difficult to apply directly to watershed hydrologic simulation and engineering scale analysis of the complex terrain area of the Qinghai-Tibet plateau. Aiming at the problem of insufficient resolution of the rainfall product in the climate mode, the existing downscaling method can be mainly divided into four major categories, namely a statistical/empirical method, including multiple regression, geostatistics, weather type analogy, deviation correction and spatial downscaling (BCSD), quantile mapping (QDM) and the like. Such methods have the advantage of being simple to implement and highly interpretative, but their effectiveness is highly dependent on statistical stationarity assumptions, performing inadequately in characterizing extreme events and non-stationary climate changes, and easily leading to excessive smoothing effects. Dynamic downscaling methods, such as using regional climate modes such as WRF/RegCM. Such methods can better exhibit local topographical effects and physical process consistency. However, it is computationally expensive, is highly sensitive to the mode parameterization scheme, and deviations in upstream boundary conditions can propagate to downstream regions. Machine learning methods such as Random Forest (RF), support Vector Regression (SVR), gradient-lifting tree (XGBoost), and the like. These methods are good at capturing the nonlinear relationship between precipitation and multi-source terrain/land factors. But its performance is strongly dependent on the quality and representativeness of the training samples and generalization ability and robustness remain to be improved. Deep learning/super resolution methods include Convolutional Neural Networks (CNN), U-Net, generating countermeasure networks (GAN), transformers, and graph neural networks. These methods exhibit a strong potential to learn multi-scale textures and reconstruct high resolution details end-to-end. However, they may suffer from blocking artifacts, checkerboard effects, and physical non-conservation in applications, and the need for high quality, large sample training data is extremely high. Given the unique geographic and climatic characteristics of the Qinghai-Tibet plateau, the CMIP6 precipitation downscaling technology for this region is highly desirable to find an effective balance between "high resolution spatial structure recovery, extreme event accurate characterization, and cross-period robust generalization capability". It is difficult for existing single statistical or learning strategies to fully address these needs. Therefore, there is an urgent need to develop an interpretable solution that can effectively introduce a priori knowledge of the terrain/land, and combine uncertainty quantization techniques with multi-source observation data (e.g., site observation, satellite telemetry, and analysis data) for collaborative calibration to significantly improve the reliability and practicality of downscaled products. Disclosure of Invention The invention aims to provide a global climate mode water-reducing and scale-reducing method for Tibet plateau based on dual attention, which combines a multiscale arbiter and physical constraint to realize high-precision scale-reducing of CMIP6 precipitation data by fusing dense connection and a generation countermeasure network of a dual attention mechanism, and solves the problems of detail loss, obvious artifact and insufficient generalization capability of the traditional method under complex terrains. In order to solve the technical problems, the invention is realized by the following technical scheme: The invention discloses a global climate mode water-reducing and scale-reducing method for Qinghai-Tibet plateau based on dual attention, which comprises the following steps: step S1, data preparation, namely acquiring coarse resolution global climate mode CMIP6 precipitation data, high resolution reference precipitation data and multisource auxiliary data, and carrying out standardized preprocessing on all the data; s2, constructing a model, namely constructing a generating countermeasure network model, wherein a generator of th