Precipitation space downscaling method based on collaborative attention enhancement mechanism

Abstract

The invention belongs to the technical field of remote sensing image processing and deep learning, and particularly relates to a precipitation space downscaling method based on a collaborative attention enhancement mechanism, which comprises the following steps: firstly, performing global sensitivity analysis on candidate factors such as terrain and climate by using EFAST, quantifying factor contribution and screening significant sensitive variables to construct an optimal input data set; then, a CNN-ECGS downscaling model integrating ECAM and GLSAM is constructed, the ECAM utilizes an adaptive weight distribution mechanism to strengthen key channel features, and the GLSAM captures spatial heterogeneity by fusing global pooling information and local convolution features; in the model training process, a weighted mixed loss function combining a mean square error and a logarithmic error is adopted, and finally GDA is carried out in combination with actual measurement site data to carry out residual error correction on a prediction result. According to the method, the problems of feature extraction redundancy and high-frequency detail loss under complex terrains are effectively solved, and the spatial resolution and reconstruction precision of rainfall products in areas lacking data such as inland river mountainous areas are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and deep learning technology, and particularly relates to a spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism. Background Technology

[0002] Satellite Earth observation technology, leveraging its wide-area coverage, near-real-time monitoring advantages, and high spatial resolution, has gradually become an important means of acquiring global precipitation data. However, while existing satellite and reanalysis precipitation products have improved spatiotemporal coverage, their original resolution still falls short of the needs of refined meteorological and hydrological research in complex terrain regions. Especially in areas with scarce observation stations, existing precipitation products still exhibit significant errors. Therefore, conducting downscaling studies of precipitation products can not only capture the spatiotemporal variations of precipitation under different topographic conditions within mountainous watersheds more precisely, but also enhance the response capability of hydrological models to local precipitation levels. This is of great significance for improving the accuracy of watershed water resource management and eco-hydrological process simulation. The rapid development of remote sensing technology and large-scale reanalysis precipitation datasets provides an important data foundation for regional precipitation spatial downscaling research.

[0003] In the field of deep learning, recurrent neural networks (RNNs), due to their memory unit structure, are suitable for processing data with temporal relationships. However, for meteorological elements such as precipitation, which have strong spatial heterogeneity, convolutional neural networks (CNNs) are more advantageous. CNNs perform mathematical operations by sliding convolutional kernels across images, capturing spatial structure and local connectivity between pixels, thus effectively extracting local features and preserving spatial information. Simultaneously, pooling layers can reduce feature dimensionality and improve computational efficiency. Therefore, compared to RNNs, CNNs are more suitable for analyzing image-based datasets and grid-based spatiotemporal sequence data. Existing research shows that by constructing a CNN architecture and matching an appropriate loss function, the root mean square error of precipitation products after downscaling can be effectively reduced, and the ability to simulate extreme rainfall can be improved. Furthermore, recent research has begun to explore the use of more advanced deep learning techniques, such as generative adversarial networks (GANs) and attention-based networks, attempting to learn the underlying distribution of data in the absence of large-scale labeled data. Although various methods derived from GAN-based architectures can effectively improve the spatiotemporal resolution of precipitation data, existing methods still have limitations when handling multi-source data fusion in complex environments.

[0004] However, the aforementioned methods for downscaling precipitation products often overlook the fact that, under complex terrain conditions, a single network structure struggles to adaptively allocate computational resources, resulting in insufficient feature extraction capabilities for key terrain and climate factors. This leads to feature redundancy and loss of high-frequency details. Furthermore, due to the large topographic relief in mountainous areas, the variability of climate models, and the inherent systematic errors in precipitation products, traditional convolutional operations struggle to balance global background information with local detailed features. Therefore, constructing a multi-scale feature information extraction and fusion model with a collaborative attention enhancement mechanism is more conducive to improving the downscaling accuracy of precipitation products.

[0005] This invention provides a spatial downscaling method and system for precipitation based on a collaborative enhanced attention mechanism. The method utilizes the collaborative enhanced attention mechanism to deeply extract multi-scale feature information from remote sensing images and precipitation products, and optimizes the global and local errors in the spatial downscaling task of precipitation through an improved hybrid loss function, thereby achieving a refined expression of multi-scale feature information and obtaining more accurate high-resolution precipitation products. Summary of the Invention

[0006] The purpose of this invention is to provide a spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism. By optimizing input parameters through global sensitivity analysis, a deep learning model integrating enhanced channel attention and global-local spatial attention is constructed. Combined with a hybrid loss function and geographic difference analysis, this method can significantly improve the reconstruction accuracy and spatial heterogeneity representation ability of high-resolution precipitation data.

[0007] The technical solution adopted in this invention is as follows: A spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism includes the following steps: S1: Acquire candidate predictors including topography, soil moisture, vegetation dynamics and climate drivers. Use the Extended Fourier Amplitude Sensitivity Test (EFAST) method to perform global sensitivity analysis on the candidate predictors. By calculating the main effect sensitivity index and the total sensitivity index, select the variables most sensitive to the spatial distribution of precipitation as the input parameter set. S2: The selected sensitive variables are processed for missing values ​​and standardized. Then, interpolation is used to resample the data to construct a low-resolution dataset of independent variables and a high-resolution dataset of the target variables. The dataset is then divided into a training set and a validation set. S3: Build a model based on a convolutional neural network framework. This model integrates the Enhanced Channel Attention Module (ECAM) and the Global-Local Spatial Attention Module (GLSAM), and uses a skip connection mechanism to fuse low-level spatial features and high-level semantic features. S4: Define a weighted mixed loss function, train the CNN-ECGS model using the training set, update the model parameters through the optimizer until the model converges or reaches the preset number of training iterations; S5: Input high-resolution independent variable factors into the trained model to obtain preliminary high-resolution precipitation prediction results; then use the Geographic Disparity Analysis (GDA) method, combined with measured station data, to perform residual correction on the prediction results, and obtain the final high-precision precipitation product downscaled data.

[0008] Preferably, in step S1, the Extended Fourier Amplitude Sensitivity Test (EFAST) method is used with a frequency domain sampling strategy to efficiently analyze the main effect sensitivity index in the high-dimensional parameter space (indicators of multi-factor coupling effects of precipitation, including topography, soil moisture, vegetation dynamics, and climate drivers). S i ) and the total effect sensitivity index (Total Effect, ST i The formulas are as follows: ; .

[0009] Among them, among them, V i For parameter 𝑥 i The variance; V ij For parameter 𝑥 i and 𝑥 j The variance of the contribution of interactions to the outcome variable; V ij…k The variance of the contribution of parameter interactions to the outcome variable across all studies; S i Reflects parameter 𝑥 i The direct contribution rate to the total variance of the results; ST i Reflects parameter 𝑥 i The sum of the contributions of the main effects and their interactions with other parameters to the total variance of the model results; S i A value >0.05 indicates that the factor makes a significant independent contribution and is considered a significant and sensitive parameter.

[0010] Preferably, the processing procedure of the Enhanced Channel Attention Module (ECAM) in step S3 is as follows: S31: Input feature map Global max pooling and global average pooling are performed separately to aggregate the spatial information of each channel of the input feature map, generating two channel vectors. f max ,f avg , The two vectors are fed into a shared multilayer perceptron (MLP) and then into an adaptive weighting module in parallel. The adaptive weighting module consists of two fully connected layers that generate the corresponding weights. w max , w avg : w max = FC m1 ( ReLU ( FC m0 ( f max ))); w avg = FC a1 ( ReLU ( FC a0 ( f avg ))); S32: Utilize the aforementioned weight integration feature to obtain the final channel weight vector. f cw And apply it to the input feature map: f cw = ReLU ( MLP ( f max )* w max +MLP ( f avg )* w avg ); Z j ’ = f cw * Z j ; in, Z j For the input feature map, Z j ’ This is for outputting feature maps.

[0011] Preferably, the processing procedure of the Global-Local Spatial Attention Module (GLSAM) in step S3 is as follows: S33: The input feature map is processed through a local feature extraction branch and a global feature extraction branch; the local feature extraction branch uses a 3×3 depthwise separable convolution to extract local feature responses. F local : F local = Conv 3×3 ( X ); S34: The global feature extraction branch uses global average pooling (GAP) and global max pooling (GMP) to extract global information. F global : ; S35: After extending the global features to the spatial dimension, a 1×1 convolutional layer is used to fuse the local and global information, and an attention map is generated using the Sigmoid activation function. ; in, This represents the fused feature information. It is the sigmoid activation function; S36: Use pixel-wise multiplication with an attention map on the original input. X Weighting is applied to enhance the features of key regions; the expression is as follows: ; in, This is the normalized attention map.

[0012] Preferably, the weighted mixture loss function described in step S4 L mix The calculation formula is as follows: ; in, N For the sample size, y i This represents the regional average precipitation. The precipitation predicted by the model. and These are the weight parameters of the hybrid loss function. Set to 0.7, Set to 0.3, To prevent abnormally small positive numbers in logarithmic operations, Set to 10 -6 .

[0013] Preferably, the specific process of the geographical difference analysis method in step S5 is as follows: S51: Extract the downscaled predicted raster data to the corresponding measured station location; S52: Calculate the residuals between downscaled raster data values ​​and measured precipitation values ​​at stations; S53: Spatial interpolation of the residuals is performed using the inverse distance weighting method (IDW) to generate an error distribution map with a resolution of 1 km; S54: Add and fuse the error distribution map with the high-resolution precipitation prediction results initially obtained in step S5 to obtain the calibrated precipitation data.

[0014] Preferably, in step S2, the spatial resolution of the low-resolution independent variable dataset is 0.1°, and the spatial resolution of the high-resolution target dataset and the output result of step S5 is 1km; the model training adopts the time cross-validation method.

[0015] The beneficial effects of this invention include: 1. The precipitation spatial downscaling method based on the collaborative enhanced attention mechanism provided in this invention quantifies the independent contribution and interaction of each factor to the spatial distribution of precipitation in a high-dimensional parameter space by introducing EFAST global sensitivity analysis. Compared with selecting input variables based on experience, this method eliminates redundant factors with low contribution to the model, which reduces the amount of computation and avoids interference from noisy data on model training, thus ensuring the rationality of the physical mechanism of the downscaling model from the source.

[0016] 2. This invention overcomes the shortcomings of traditional attention mechanisms that treat different pooling features indiscriminately by using an Enhanced Channel Attention Module (ECAM). Through adaptive weight allocation, the model can automatically identify and strengthen channel features that are more valuable for precipitation prediction, significantly improving the model's discriminative ability under complex features.

[0017] 3. To address the significant spatial differences in precipitation in mountainous areas, the designed Global-Local Spatial Attention Module (GLSAM) breaks through the limitations of traditional convolutional operations that are restricted to local receptive fields. By fusing global pooling information with local convolutional features, the model can perceive both large-scale climate background and accurately capture local topographic and rain details, effectively solving the texture loss problem in high-resolution reconstruction.

[0018] 4. The weighted hybrid loss function of MSE and LogError effectively solves the prediction bias caused by the imbalance in precipitation data distribution. MSE ensures the overall smoothness, while LogError improves the model's sensitivity to extreme precipitation (heavy precipitation) and trace precipitation, making the prediction results more numerically close to the actual precipitation distribution.

[0019] 5. Building upon the use of deep learning to uncover nonlinear relationships, Geographic Disparity Analysis (GDA) is incorporated as a post-processing step. Residual correction of the model output is performed using measured station data, effectively eliminating systematic biases caused by the satellite data itself or the model structure. This results in precipitation products with higher practical value and reliability in data-scarce regions.

[0020] 6. The proposed solution is significantly superior to existing mainstream technologies in both qualitative and quantitative evaluations, and can provide higher-precision, high-resolution precipitation reconstruction data for inland river and mountainous areas. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the spatial downscaling process of precipitation based on the collaborative enhanced attention mechanism in an embodiment of the present invention.

[0022] Figure 2 This is a framework diagram of the enhanced channel attention module based on the collaborative enhanced attention mechanism in an embodiment of the present invention.

[0023] Figure 3 This is a framework diagram of the global-local spatial attention module based on the collaborative enhanced attention mechanism in an embodiment of the present invention.

[0024] Figure 4 The following figures illustrate experimental results on a publicly available dataset for embodiments of the present invention. The CMFD portion represents the original 0.1° resolution precipitation product image, and the CNN-CEGS portion represents the experimental results of the present invention. Detailed Implementation

[0025] The following is in conjunction with the appendix Figures 1-4 The present invention will be further described in detail below: To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] See appendix Figure 1 As shown, a spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism includes the following steps: S1: Downscaling Factor Sensitivity Screening: Fifteen variables, including topography, soil moisture, vegetation dynamics, and climate drivers, were selected as candidate predictors. The Extended Fourier Amplitude Sensitivity Test (EFAST) method was used to map the input factors to radio frequency space and decompose the output signal using Fourier transform. Global sensitivity analysis was performed on the candidate predictors. By calculating the main effect sensitivity index and the total sensitivity index, the variables most sensitive to the spatial distribution of precipitation were selected as the input parameter set.

[0027] For the model 𝑦 = 𝑓(𝑥1, 𝑥2, … , 𝑥 𝑛 ), calculate each parameter 𝑥 i Main effect sensitivity index S i and total sensitivity index ST i Main effect sensitivity index S i Defined as: ; in, V i For parameters 𝑥 i The variance of Y is given by V(Y), where V(Y) is the total variance of the result Y. S i Reflects parameters 𝑥 i The direct contribution rate to the total variance of the results. In this embodiment, the threshold is set to 0.05, that is, when the main effect... S i When the value is greater than 0.05, the factor is considered to have made a significant independent contribution and is retained as a significant sensitive parameter to construct an optimized set of input parameters. This effectively eliminates redundant variables and reduces the complexity of model training.

[0028] S2: The selected sensitive variables are processed for missing values ​​and standardized. Interpolation is then used for resampling to construct a low-resolution dataset of independent variables and a high-resolution dataset of the target variables. The datasets are then divided into training and validation sets. The spatial resolution of the low-resolution dataset of independent variables is 0.1°, and the spatial resolution of the high-resolution dataset of the target variables and the output of step S5 is 1 km. Temporal cross-validation is used during model training.

[0029] Based on the sensitive variables selected in step S1, all data were first cropped to the study area and missing values ​​were imputed. Then, the variables were standardized. Inverse distance weighted interpolation (IDW) was used to resample the variables, constructing a low-resolution independent variable dataset (spatial resolution of 0.1°, consistent with the original CMFD precipitation data, used as input for model training) and a high-resolution target dataset (spatial resolution of 1km, used for model output and validation). Finally, according to the time series, for example, data from January 1, 2000 to December 31, 2018, were randomly divided into a training set (80%) and a validation set (20%).

[0030] S3: Constructing a CNN-ECGS Precipitation Spatial Downscaling Model: A model based on a Convolutional Neural Network (CNN) framework is built. Convolutional layers are used to capture the nonlinear relationships between precipitation data and other parameters. This model integrates an Enhanced Channel Attention Module (ECAM) and a Global-Local Spatial Attention Module (GLSAM) to refine the input features in two stages, thereby capturing key information in the features more comprehensively and improving the model's prediction results. Furthermore, to obtain pixel-by-pixel prediction results, an encoder-decoder structure similar to U-Net is adopted, and a skip connection mechanism is used to fuse features.

[0031] S4: Define a weighted mixed loss function, train the CNN-ECGS model using the training set, update the model parameters through the optimizer until the model converges or reaches the preset number of training iterations; S5: Input high-resolution independent variable factors into the trained model to obtain preliminary high-resolution precipitation prediction results; then use the Geographic Disparity Analysis (GDA) method, combined with measured station data, to perform residual correction on the prediction results, and obtain the final high-precision precipitation product downscaled data.

[0032] Preferably, in step S1, the Extended Fourier Amplitude Sensitivity Test (EFAST) method is used with a frequency domain sampling strategy to efficiently analyze the Main Effect (Si) and Total Effect (STi) sensitivity indices in the high-dimensional parameter space (indicators of the coupling effects of multiple precipitation factors, including topography, soil moisture, vegetation dynamics, and climate drivers). The formulas are as follows: ; .

[0033] Among them, among them, V i For parameter 𝑥 i The variance; V ij For parameter 𝑥 i and 𝑥j The variance of the contribution of interactions to the outcome variable; V ij…k The variance of the contribution of parameter interactions to the outcome variable across all studies; S i Reflects parameter 𝑥 i The direct contribution rate to the total variance of the results; ST i Reflects parameter 𝑥 i The sum of the contributions of the main effects and their interactions with other parameters to the total variance of the model results; S i A value >0.05 indicates that the factor makes a significant independent contribution and is considered a significant and sensitive parameter.

[0034] See Figure 2 This is a framework diagram of the enhanced channel attention module based on the collaborative enhanced attention mechanism in this embodiment. Traditional channel attention usually treats max pooling and average pooling features equally, while the enhanced channel attention module assigns different attention weights to the two through adaptive weight blocks.

[0035] The processing procedure of the Enhanced Channel Attention Module (ECAM) in step S3 is as follows: S31: Input feature map , The feature map is represented by its number of channels, height, and width. Global max pooling and global average pooling are performed respectively to aggregate the spatial information of each channel of the input feature map. The pooling operation produces two different channel vectors. f max , f avg , ; Two vectors f max , f avg The data are fed separately into a shared multilayer perceptron (MLP). Useless information is filtered out by first downsampling and then upsampling proportionally, while important information 𝑟 in the downscaling task is amplified. This process is performed in parallel. f max , f avg The weights are fed into an adaptive weighting module, which automatically adjusts the weights based on their importance to the final output. This adaptive weighting module consists of two fully connected layers that generate the corresponding weights. w max , w avg : w max = FC m1 ( ReLU ( FC m0 ( f max ))); w avg = FC a1 ( ReLU ( FC a0 ( f avg ))); S32: Max pooling and average pooling features are adaptively integrated to obtain the final channel weight vector by applying weights as follows. f cw And apply it to the input feature map: f cw = ReLU ( MLP ( f max )* w max +MLP ( f avg )* w avg ); Finally, the channel weights are applied to the input feature map as follows to obtain the refined feature map. Z j ’ ; Z j ’ = f cw * Z j ; in, Z j For the input feature map, Z j ’ This is for outputting feature maps.

[0036] This design enables the model to adaptively integrate features extracted by different pooling operations, and to more accurately respond to key channel information in precipitation prediction tasks.

[0037] See Figure 3 The diagram shows the framework of the Global-Local Spatial Attention (GLSAM) module based on the collaborative augmented attention mechanism. The processing procedure of the GLSAM module in step S3 is as follows: S33: To address the issue of traditional spatial attention being easily limited by local regions, the GLSAM module employs a dual-branch design. The GLSAM module first extracts local features from the input feature map using a 3×3 depthwise separable convolution. This extracts features to perceive the spatial relationships between adjacent pixels, allowing the attention mechanism to focus on local patterns (such as texture and edges). In other words, the input feature map is processed through a local feature extraction branch and a global feature extraction branch. The local feature extraction branch utilizes a 3×3 depthwise separable convolution to extract local feature responses. F local : F local = Conv 3×3 ( X ); in, X It is the input feature map. F local It is a local characteristic response S34: The global feature extraction branch uses global average pooling (GAP) and global max pooling (GMP) to extract global information. F global This is used to avoid feature loss due to noise interference and localization, allowing the attention mechanism to learn more comprehensive information. ; GAP refers to the input X Average pooling in the spatial dimension, GMP is a method for input... X Max pooling in the spatial dimension.

[0038] S35: After extending the global features to the spatial dimension, a 1×1 convolutional layer is used to fuse the local and global information. This allows the attention to perceive both local details and utilize global background information. An attention map is then generated using the Sigmoid activation function, as shown in the following expression: ; in, This represents the fused feature information. It is the sigmoid activation function; S36: Use pixel-wise multiplication with an attention map on the original input. X Weighting is applied to enhance the features of key regions; the expression is as follows: ; in, This is the normalized attention map.

[0039] This approach allows the model to perceive both the impact of local topography on precipitation and macro-climate background information.

[0040] Step S4: Input the processed 0.1° spatial resolution independent and dependent variables into the CNN-ECGS model. To optimize the model's global and local errors in the precipitation spatial downscaling task, this embodiment designs a weighted hybrid loss function. L mix The mixed loss function comprises two parts: mean squared error (MSE) and logarithmic error (LogError). This function fully considers the uneven distribution of precipitation. MSE primarily optimizes the global error, used for overall fitting of the precipitation distribution, while LogError focuses on the relative error of precipitation, better capturing the subtle differences between areas of heavy and light precipitation. The formula for this mixed loss function is as follows: The weighted mixture loss function L mix The calculation formula is as follows: ; in, N For the sample size, y i This represents the regional average precipitation. The precipitation predicted by the model. and These are the weight parameters of the hybrid loss function. Set to 0.7, Set to 0.3, To prevent abnormally small positive numbers in logarithmic operations, Set to 10 -6 .

[0041] The training parameters are set as follows: the input layer dimension is 13 (including CMFD precipitation and 12 selected sensitivity factors), the output layer dimension is 1 (daily precipitation), the Inception module consists of 1+1, 3+3 and 5+5 convolutional kernels, the learning rate is set to 0.0003, the batch size is set to 32, the Adam optimizer is used, the maximum epoch is set to 1000 and an early stopping mechanism is used.

[0042] After model training, independent variable factors with a 1km spatial resolution are input into the model to obtain preliminary 1km high-resolution precipitation prediction results. To further eliminate systematic bias, Geographic Disparity Analysis (GDA) is performed: First, downscaled precipitation data is extracted to the corresponding stations, and the interpolation between the downscaled raster data values ​​and the measured station precipitation is calculated; second, the stations are interpolated, and the inverse distance weighting (IDW) method is used to interpolate them to 1 km to obtain an error map; finally, the error map is fused with the downscaled results to obtain the final calibrated precipitation. The specific process of the Geographic Disparity Analysis (GDA) method in step S5 is as follows: S51: Extract the downscaled predicted raster data to the corresponding measured station location; S52: Calculate the residuals between downscaled raster data values ​​and measured precipitation values ​​at stations; S53: Spatial interpolation of the residuals is performed using the inverse distance weighting method (IDW) to generate an error distribution map with a resolution of 1 km; S54: Add and fuse the error distribution map with the high-resolution precipitation prediction results initially obtained in step S5 to obtain the calibrated precipitation data.

[0043] Reference Figure 4 The following figures illustrate experimental results on a publicly available dataset for embodiments of the present invention.

[0044] Figure 4 The top (CMFD) shows the original 0.1° resolution precipitation product image. It can be seen that the image is obviously mosaic-like and has low resolution, making it difficult to show the spatial details of precipitation, especially in areas with complex terrain, where the precipitation distribution appears too smooth.

[0045] Figure 4 The bottom (CNN-ECGS) shows the high-resolution (1km) precipitation results reconstructed using the method of this invention. Compared to the left, the results of this invention not only eliminate the grid effect of low-resolution data but also clearly present the spatial texture details of precipitation. Thanks to the introduction of ECAM and GLSAM modules, the model successfully recovers local precipitation features influenced by topography (such as precipitation enhancement areas on windward slopes) while maintaining a precipitation distribution trend consistent with the original data.

[0046] This invention compares with other methods for downscaling precipitation products, providing experimental data to demonstrate the effectiveness of this invention. The evaluation indicators of the comparative experimental results are shown in Table 1 below: Table 1 Evaluation indicators of experimental results Quantitative evaluation shows that the precipitation data generated by the method of this invention outperforms traditional interpolation methods and ordinary CNN downscaling methods in terms of correlation coefficient (CC) and root mean square error (RMSE), verifying the effectiveness and superiority of the collaborative enhanced attention mechanism in downscaling precipitation products in inland river and mountainous areas.

[0047] In summary, the precipitation spatial downscaling method and system proposed in this invention, based on a collaborative enhanced attention mechanism, firstly utilizes the Extended Fourier Amplitude Sensitivity Test (EFAST) to screen significant sensitive factors from multi-source data to construct the optimal input dataset. Then, low-resolution independent variable data is used as input to a deep network, and features are extracted using convolutional layers. An Enhanced Channel Attention Module (ECAM) adaptively allocates channel weights to amplify key information, while a Global-Local Spatial Attention Module (GLSAM) fuses local details with the global background, achieving efficient feature refinement and extraction. Simultaneously, a weighted mixed loss function is used during training to balance global and local errors, and a skip connection mechanism fuses features from shallow and deep layers. Finally, the preliminary prediction results output by the model are combined with station-measured data for Geographic Disparity Analysis (GDA) correction, resulting in high-precision precipitation data. The proposed scheme outperforms other precipitation spatial downscaling methods in both qualitative and quantitative evaluations, especially in complex terrain conditions such as inland river and mountainous areas, producing higher-precision high-resolution reconstruction results from remote sensing images.

[0048] Compared to existing methods, the advantages and positive effects of this invention are as follows: The CNN-ECGS downscaling model constructed in this invention not only ensures the physical correlation of input factors through EFAST screening, but also solves the problems of redundancy and loss of high-frequency details that occur when extracting feature information in complex terrain using existing methods through a collaborative enhanced attention mechanism. Specifically, the ECAM module overcomes the limitation of traditional channel attention treating different pooling features "equally," while the GLSAM module breaks through the bottleneck of convolution operations being limited by the local receptive field. Furthermore, by combining hybrid loss function optimization and geographical difference post-processing correction, this invention effectively eliminates systematic bias, enabling the acquisition of precipitation products with high spatial heterogeneity and high accuracy even in data-scarce areas, significantly enhancing the practical value of precipitation spatial downscaling technology in hydrological and ecological research.

[0049] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism, characterized in that, Includes the following steps: S1: Obtain candidate predictors including topography, soil moisture, vegetation dynamics and climate drivers, perform global sensitivity analysis on the candidate predictors, and select the variables most sensitive to the spatial distribution of precipitation as the input parameter set by calculating the main effect sensitivity index and the total sensitivity index. S2: The selected sensitive variables are processed for missing values ​​and standardized. Then, interpolation is used to resample the data to construct a low-resolution dataset of independent variables and a high-resolution dataset of the target variables. The dataset is then divided into a training set and a validation set. S3: Build a model based on a convolutional neural network framework. This model integrates an enhanced channel attention module and a global-local spatial attention module, and uses a skip connection mechanism to fuse low-level spatial features and high-level semantic features. S4: Define a weighted mixed loss function, train the CNN-ECGS model using the training set, update the model parameters through the optimizer until the model converges or reaches the preset number of training iterations; S5: Input high-resolution independent variable factors into the trained model to obtain preliminary high-resolution precipitation prediction results; then use the geographical difference analysis method, combined with the actual measurement station data, to perform residual correction on the prediction results, and obtain the final high-precision precipitation product downscaling data.

2. The spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 1, characterized in that, In step S1, the extended Fourier amplitude sensitivity test method is used to perform global sensitivity analysis on the candidate predictors. Through a frequency domain sampling strategy, the main effect sensitivity index in the high-dimensional parameter space is efficiently analyzed. S i and total effect sensitivity index ST i The formulas are as follows: ； ； Among them, among them, V i For parameter 𝑥 i The variance; V ij For parameter 𝑥 i and 𝑥 j The variance of the contribution of interactions to the outcome variable; V ij…k The variance of the contribution of parameter interactions to the outcome variable across all studies; S i Reflects parameter 𝑥 i The direct contribution rate to the total variance of the results; ST i Reflects parameter 𝑥 i The sum of the contributions of the main effects and their interactions with other parameters to the total variance of the model results; S i A value >0.05 indicates that the factor makes a significant independent contribution and is considered a significant and sensitive parameter.

3. The spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 1, characterized in that, The processing procedure of the enhanced channel attention module in step S3 is as follows: S31: Input feature map Global max pooling and global average pooling are performed separately to aggregate the spatial information of each channel of the input feature map, generating two channel vectors. f max , f avg , The two vectors are fed into a shared multilayer perceptron (MLP) and then fed into the adaptive weight module in parallel. The adaptive weight module consists of two fully connected layers, which generate the corresponding weights. w max , w avg : w max = FC m1 ( ReLU ( FC m0 ( f max )))； w avg = FC a1 ( ReLU ( FC a0 ( f avg )))； S32: Utilize the weight integration feature to obtain the final channel weight vector. f cw And apply it to the input feature map: f cw = ReLU ( MLP ( f max )* w max + MLP ( f avg )* w avg )； Z j ’ = f cw * Z j ； in, Z j For the input feature map, Z j ’ This is for outputting feature maps.

4. The spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 3, characterized in that, The processing procedure of the global-local spatial attention module in step S3 is as follows: S33: Process the input feature map through local feature extraction branch and global feature extraction branch; The local feature extraction branch uses a 3×3 depthwise separable convolution to extract local feature responses. F local : F local = Conv 3×3 ( X )； S34: The global feature extraction branch uses global average pooling and global max pooling to extract global information. F global : ； S35: After extending the global features to the spatial dimension, a 1×1 convolutional layer is used to fuse the local and global information, and an attention map is generated using the Sigmoid activation function. ； in, This represents the fused feature information. It is the sigmoid activation function; S36: Use pixel-wise multiplication with an attention map on the original input. X Weighting is applied to enhance the features of key regions; the expression is as follows: ； in, This is the normalized attention map.

5. The spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 1, characterized in that, The weighted mixture loss function described in step S4 L mix The calculation formula is as follows: ； in, N For the sample size, y i This represents the regional average precipitation. The precipitation predicted by the model. and These are the weight parameters of the hybrid loss function. To prevent abnormally small positive numbers in logarithmic operations.

6. The spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 1, characterized in that, The specific process of the geographical difference analysis method in step S5 is as follows: S51: Extract the downscaled predicted raster data to the corresponding measured station location; S52: Calculate the residuals between downscaled raster data values ​​and measured precipitation values ​​at stations; S53: Spatial interpolation of the residuals is performed using the inverse distance weighting method to generate an error distribution map with a resolution of 1km; S54: Add and fuse the error distribution map with the high-resolution precipitation prediction results initially obtained in step S5 to obtain the calibrated precipitation data.

7. A spatial downscaling method for precipitation based on a collaborative enhanced attention mechanism according to claim 6, characterized in that, In step S2, the spatial resolution of the low-resolution independent variable dataset is 0.1°, and the spatial resolution of the high-resolution target dataset and the output result of step S5 is 1km. The model is trained using time-based cross-validation.

Images