Multi-source rainfall data fusion method based on hydrological model guidance and double transfer learning
By using hydrological model-guided and dual transfer learning methods, the problems of decreased accuracy and insufficient event identification in multi-source precipitation data fusion were solved, and high-precision precipitation estimation and event detection were achieved in sparse rain gauge areas.
Patent Information
- Application Number
- CN202511169142.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-20
- Publication Date
- 2025-12-05
AI Technical Summary
Existing multi-source precipitation data fusion methods show a significant decrease in accuracy in areas with sparse rain gauge distribution. Defects in the quality of the source data limit the fusion accuracy and result in insufficient ability to identify precipitation events.
We employ a hydrological model-guided and dual-transfer learning approach. By constructing an HBV hydrological model for data bias correction and combining it with a domain adversarial neural network for adversarial training, we build a multi-source precipitation data fusion model to achieve precipitation event classification and magnitude regression.
It improves the accuracy and reliability of precipitation estimation, enhances the ability to detect precipitation events and the accuracy of magnitude prediction, especially in adaptability and robustness in areas with sparse rain gauges.
Smart Images
Figure CN121071784A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of hydrology and meteorology, specifically relating to a multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning. Background Technology
[0002] Precipitation is a fundamental hydrometeorological variable and a core driver of the global water cycle, playing a decisive role in hydrological processes and serving as key input data for hydrological models. However, precipitation exhibits significant spatiotemporal heterogeneity, posing a persistent challenge to its accurate estimation. Up to 70–80% of the uncertainty in surface hydrological simulations stems from the impact of precipitation's spatiotemporal variability.
[0003] Currently, precipitation estimation mainly relies on four types of data sources: rain gauges, weather radar, satellite remote sensing products, and atmospheric reanalysis products. Each data source has its advantages and limitations: rain gauge and radar data have high accuracy, but their stations are usually sparse and unevenly distributed, making it difficult to fully characterize the spatiotemporal heterogeneity of precipitation, especially in high-altitude areas; satellite remote sensing and reanalysis products, with their high spatiotemporal resolution, provide strong support for obtaining continuous precipitation fields and have developed a variety of high-quality datasets (such as TRMM, GSMAP, IMERG, ERA5, etc.), which are widely used in hydrological simulation and water resource management at different scales. However, the accuracy of satellite and reanalysis products is still limited by various factors such as inversion algorithms, complex terrain, sensor resampling frequency, and uncertainties in assimilation techniques.
[0004] To improve the accuracy of precipitation estimation, the current main approach is to fuse multi-source grid precipitation products with rain gauge observation data. Existing fusion techniques can be summarized into two main types: (1) Precipitation correction methods. By establishing statistical relationships such as probability density function matching, quantile mapping, or multiplicative bias correction between grid precipitation products and ground observation data, the grid products are systematically corrected to generate deterministic hydrological input data; (2) Data fusion methods. Using algorithms such as weighted averaging or machine learning, multi-source data such as rain gauge observations, satellite remote sensing, and reanalysis products are systematically integrated, and the fusion results are usually better than those from a single data source.
[0005] Although the above-mentioned fusion methods have been proven to effectively improve the accuracy of precipitation estimation, they still have the following significant shortcomings in practical applications: (1) In areas with sparse distribution of rain gauges, the accuracy of the fusion algorithm decreases significantly due to insufficient ground observation data; (2) The quality defects of the source data itself are amplified through the fusion process, directly affecting the reliability of the final product; (3) Existing methods have inherent limitations in their ability to identify precipitation events, resulting in significant uncertainty in the fusion product's ability to identify precipitation events. These technical shortcomings severely restrict the practical application value of fused precipitation products in the field of hydrology and meteorology. Summary of the Invention
[0006] To address the aforementioned shortcomings in existing technologies, the multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning provided by this invention solves the problems of significantly reduced precipitation estimation accuracy in sparse rain gauge networks, limited precipitation fusion accuracy due to source data quality defects, and insufficient ability to identify precipitation events.
[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning, comprising the following steps: S100. Acquire and preprocess multi-source precipitation data and auxiliary meteorological data of the study area, and construct a standardized precipitation dataset, including precipitation correction input dataset, data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset. S200. Use the precipitation correction input dataset to calibrate the parameters of the HBV hydrological model, and use the calibrated HBV hydrological model parameter set to correct the bias of the data fusion input dataset. S300. Based on the corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset, divide the source domain and target domain, and construct the domain classification output dataset. S400: Construct a dual transfer learning architecture, use the domain classification output dataset to perform adversarial training on it, and obtain a multi-source precipitation data fusion model; S500: Using a multi-source precipitation data fusion model, precipitation data of the study area is retrieved to obtain a multi-source precipitation data fusion product.
[0008] Further, step S100 includes: Acquire multi-source precipitation data for the study area, including rain gauge observation data, hydrological station runoff observation data, auxiliary meteorological variable data, and gridded precipitation products; The acquired rain gauge observation data, gridded precipitation products, and auxiliary meteorological variable data were normalized. Based on the normalized data, precipitation, temperature and potential evapotranspiration grid data are extracted and their watershed average daily values are calculated to form a precipitation correction input dataset. For gridded precipitation products and auxiliary meteorological variable data, a system is constructed centered on each rain gauge station. Sub-mesh, extract Dimensional feature variables, combined The sliding window process generates a dimension of... High-dimensional input samples are used to form a data fusion input dataset; Rain gauge observation data are classified into rainy days and no rainy days using a set precipitation threshold as the discrimination criterion, forming a data fusion classification output dataset; Using actual precipitation as the regression label, the data from rain gauge observations are used to generate a data fusion regression output dataset. The data fusion input dataset, the data fusion classification output dataset, and the data fusion regression output dataset are used as the standard dataset.
[0009] Further, step S200 includes the following sub-steps: S201. Construct the HBV hydrological model and use the watershed water balance equation as the core physical constraint to simulate the runoff process. S202. Using the precipitation correction input dataset as the input data and the hydrological station runoff observation data as the output data, the parameters of the HBV hydrological model are calibrated to obtain the calibrated HBV hydrological model parameter set. S203. Based on the calibrated HBV hydrological model parameter set, the precipitation field in the data fusion input dataset is corrected for deviation, and the corrected data fusion input dataset is generated.
[0010] Further, in step S201, the watershed water balance equation is: In the formula, , , These represent precipitation, evapotranspiration, and runoff, respectively. , , , These represent snow cover, soil water, upper groundwater, and lower groundwater reserves, respectively. In step S203, the formula for deviation correction is: In the formula, , These are the rainfall correction factor and the snowfall correction factor, respectively. Indicates temperature, Indicates the temperature threshold. This indicates the precipitation before correction. This indicates the corrected precipitation.
[0011] Furthermore, in step S300: Rain gauge stations within the study area are taken as the target domain, and rain gauge stations with similar geographical features are taken as the source domain. They are binary classified and labeled. The corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset from the source and target domains are combined to construct a domain classification output dataset.
[0012] Further, step S400 includes: Constructing domain adversarial neural networks; An adversarial training mechanism for precipitation event classification is established. The data fusion input dataset is used as input, and the data fusion classification output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain a precipitation event classification model. An adversarial training mechanism for precipitation regression is established. The data fusion input dataset is used as input, and the data fusion regression output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain the precipitation regression model. Based on precipitation event classification models and precipitation regression models, a multi-source precipitation data fusion model based on a dual transfer learning architecture was constructed.
[0013] Furthermore, the domain adversarial neural network includes a feature extractor, a target predictor, and a domain discriminator; The feature extractor is used to extract spatiotemporal feature representations from the input data, the target predictor is used to generate precipitation event types or precipitation estimates for the target domain according to the purpose of the domain adversarial neural network, and the domain discriminator is used to distinguish whether the spatiotemporal feature representations come from the source domain or the target domain, so as to drive adversarial training.
[0014] Furthermore, the domain adversarial neural network is trained adversarially through the first forward propagation and back propagation to obtain a precipitation event classification model; the domain adversarial neural network is trained through the second forward propagation and back propagation to obtain a precipitation regression model. During the first forward propagation process, the binary cross-entropy loss function is used to calculate the prediction bias of precipitation event classification and the difference in feature distribution between the source domain and the target domain. In the second forward propagation process, the mean square error loss function is used to calculate the prediction bias of precipitation regression, and the binary cross-entropy loss function is used to calculate the difference in characteristic distribution between the source domain and the target domain. Among them, the prediction bias of the precipitation event classification Represented as: In the formula, This indicates the number of samples in the source domain dataset. This represents the input data from the source domain. Indicates a precipitation event label. Indicates feature extractor, Represents the target predictor. This represents the model parameters of the feature extractor. This represents the model parameters of the target predictor; The prediction bias of the precipitation regression Represented as: In the formula, This represents rain gauge measurement data from the source region; The difference in feature distribution between the source domain and the target domain Represented as: In the formula, This indicates the number of samples from the source domain dataset and the target domain dataset. Indicates input data, Represents a field label.
[0015] Furthermore, during the backpropagation process, GRL is introduced to optimize the parameters of the domain adversarial neural network, thereby obtaining a precipitation event classification model / precipitation regression model; Introducing GRL to optimize the parameters of domain adversarial neural networks includes: Negative scaling is applied to the gradient of the domain adversarial neural network, and then the parameters are updated synchronously through a composite loss function; The composite loss function is: In the formula, This indicates an adjustable trade-off parameter.
[0016] Further, step S500 includes the following sub-steps: S501. Input the preprocessed data fusion and classification data within the grid cells of the study area into the trained precipitation event classification model to determine the rainy / non-rainy day status of the corresponding grid cells; S502. For grid cells classified as rainy days, the preprocessed data fusion regression data within the corresponding grid cell is input into the precipitation regression model to quantitatively estimate the precipitation value of the corresponding grid cell on a rainy day, thus obtaining a multi-source precipitation data fusion product.
[0017] The beneficial effects of this invention are as follows: (1) The present invention is based on a precipitation correction mechanism guided by a hydrological model. By establishing a dynamic filtering system through the response of the watershed hydrological system, the systematic deviation correction of grid precipitation products is realized, which effectively improves the accuracy and reliability of the source data.
[0018] (2) This invention introduces a transfer learning technology framework. By constructing a knowledge transfer channel between the source domain and the target domain, it enables sparse data regions to make full use of the learning experience of data-rich regions, thereby improving the robustness and regional adaptability of precipitation estimation.
[0019] (3) The present invention designs a dual transfer learning architecture, which improves the detection capability of precipitation events and the prediction accuracy of precipitation magnitude through a cascaded learning strategy of "precipitation event identification - precipitation magnitude regression". Attached Figure Description
[0020] Figure 1 The flowchart of the multi-source precipitation data fusion method based on hydrological model-guided dual transfer learning provided by this invention is shown in the figure.
[0021] Figure 2 The research area and site distribution map provided for this invention.
[0022] Figure 3 The flowchart for precipitation field correction based on the HBV hydrological model provided by this invention.
[0023] Figure 4 The domain adversarial neural network structure diagram provided by this invention.
[0024] Figure 5 This is a schematic diagram of the ConvLSTM layer structure provided by the present invention.
[0025] Figure 6 This is a schematic diagram of the multi-source precipitation data fusion process based on a dual transfer learning architecture provided by the present invention.
[0026] Figure 7 This is a scatter plot comparing the accuracy of MSPB with other precipitation products provided by this invention. Detailed Implementation
[0027] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0028] This invention provides a method for fusing multi-source precipitation data based on hydrological model guidance and dual transfer learning, such as... Figure 1 As shown, it includes the following steps: S100. Acquire and preprocess multi-source precipitation data and auxiliary meteorological data of the study area, and construct a standardized precipitation dataset, including precipitation correction input dataset, data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset. S200. Use the precipitation correction input dataset to calibrate the parameters of the HBV hydrological model, and use the calibrated HBV hydrological model parameter set to correct the bias of the data fusion input dataset. S300. Based on the corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset, divide the source domain and target domain, and construct the domain classification output dataset. S400: Construct a dual transfer learning architecture, use the domain classification output dataset to perform adversarial training on it, and obtain a multi-source precipitation data fusion model; S500: Using a multi-source precipitation data fusion model, precipitation data of the study area is retrieved to obtain a multi-source precipitation data fusion product.
[0029] Step S100 of this embodiment of the invention includes: Acquire multi-source precipitation data for the study area, including rain gauge observation data, hydrological station runoff observation data, auxiliary meteorological variable data, and gridded precipitation products; The acquired rain gauge observation data, gridded precipitation products, and auxiliary meteorological variable data were normalized. Based on the normalized data, precipitation, temperature and potential evapotranspiration grid data are extracted and their watershed average daily values are calculated to form a precipitation correction input dataset. For gridded precipitation products and auxiliary meteorological variable data, a system is constructed centered on each rain gauge station. Sub-mesh, extract Dimensional feature variables, combined The sliding window process generates a dimension of... The high-dimensional input samples form a data fusion input dataset, including data on precipitation, temperature, humidity and air pressure for various precipitation products; Rain gauge observation data are classified into rainy days and no rainy days using a set precipitation threshold as the discrimination criterion, forming a data fusion classification output dataset; Using actual precipitation as the regression label, the data from rain gauge observations are used to generate a data fusion regression output dataset. The data fusion input dataset, the data fusion classification output dataset, and the data fusion regression output dataset are used as the standard dataset.
[0030] In this embodiment, as Figure 2Taking the Yellow River source region as an example, the multi-source precipitation data obtained includes rain gauge observation data from the National Surface Meteorological Station Daily Value Data Set provided by the China Meteorological Administration (CMA); hydrological station runoff observation data from the Yellow River Conservancy Commission Hydrological Bureau; auxiliary meteorological variables such as temperature, humidity, air pressure, and potential evapotranspiration, from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5-Land reanalysis dataset; and gridded precipitation products from NASA's IMERG satellite, JAXA's GSMAP satellite, and ECMWF ERA5-Land reanalysis precipitation products. The spatial resolution of the example is 0.1° × 0.1°.
[0031] In this embodiment, the formula for normalizing the rain gauge observation data, gridded precipitation products, and auxiliary meteorological variable data is as follows: In the formula, This represents the normalized data value. Indicates input data, and These represent the minimum and maximum values of the input dataset, respectively.
[0032] In this embodiment, when forming the data fusion input dataset, for the Yellow River source area, a 5×5 subgrid is constructed with each rain gauge station as the center, and 6-dimensional feature variables (including multi-source precipitation and auxiliary meteorological variables) are extracted. Combined with 3-day sliding window processing, a high-dimensional input sample with dimensions of 3×6×5×5 is generated to form the data fusion input dataset.
[0033] In this embodiment, when forming the data fusion classification output dataset, a binary classification label of rainy days / no rainy days is performed using a 0.1 mm precipitation threshold as the discrimination criterion (rainy day = 1, no rainy day = 0) to generate the data fusion classification output dataset, i.e., the precipitation event classification dataset.
[0034] like Figure 3 As shown, step S200 of this embodiment of the invention includes the following sub-steps: S201. Construct the HBV hydrological model and use the watershed water balance equation as the core physical constraint to simulate the runoff process. S202. Using the precipitation correction input dataset as the input data and the hydrological station runoff observation data as the output data, the parameters of the HBV hydrological model are calibrated to obtain the calibrated HBV hydrological model parameter set. S203. Based on the calibrated HBV hydrological model parameter set, the precipitation field in the data fusion input dataset is corrected for deviation, and the corrected data fusion input dataset is generated.
[0035] In step S201 of this embodiment, the HBV hydrological model adopts a modular structure design, including five functional units: noise reduction module, snow accumulation module, soil moisture module, response module, and river confluence module. The watershed water balance equation, which serves as the physical constraint, is as follows: In the formula, , , These represent precipitation, evapotranspiration, and runoff, respectively. , , , These represent snow cover, soil water, upper groundwater, and lower groundwater reserves, respectively. Furthermore, the parameters and calibration range of the HBV hydrological model are shown in Table 1; Table 1: Parameters and Calibration Range of the HBV Hydrological Model In step S202 of this embodiment, a genetic algorithm (GA) is used to calibrate the HBV hydrological model parameter set. The objective function is to minimize the mean square error of runoff simulation. The number of iterations is set to 200, and the number of individuals per generation is 100, ultimately obtaining the optimized HBV hydrological model parameter set. The objective function is: In the formula, Indicates the first Measured runoff at a time step Indicates the first Simulated runoff at a time step This represents the total number of time steps.
[0036] In step S203 of this embodiment, the formula for deviation correction is: In the formula, , These are the rainfall correction factor and the snowfall correction factor, respectively. Indicates temperature, Indicates the temperature threshold. This indicates the precipitation before correction. This indicates the corrected precipitation.
[0037] In step S300 of this embodiment of the invention: Rain gauge stations within the study area are taken as the target domain, and rain gauge stations with similar geographical features are taken as the source domain. They are binary classified and labeled. The corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset from the source and target domains are combined to construct a domain classification output dataset.
[0038] In this embodiment, for the Yellow River region, 18 rain gauge stations within the study area are used as the target domain. Based on the Euclidean distance metric of spatial characteristic variables such as longitude, latitude, and elevation, 20 rain gauge stations with similar geographical characteristics to the target domain are selected as the source domain.
[0039] Step S400 of this embodiment of the invention includes: Constructing domain adversarial neural networks; An adversarial training mechanism for precipitation event classification is established. The data fusion input dataset is used as input, and the data fusion classification output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain a precipitation event classification model. An adversarial training mechanism for precipitation regression is established. The data fusion input dataset is used as input, and the data fusion regression output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain the precipitation regression model. Based on precipitation event classification models and precipitation regression models, a multi-source precipitation data fusion model based on a dual transfer learning architecture was constructed.
[0040] In embodiments of the present invention, such as Figure 4 As shown, a domain adversarial neural network includes a feature extractor. Target predictor and domain discriminator ; Specifically, the feature extractor is used to extract spatiotemporal feature representations from the input data, the target predictor is used to generate precipitation event types or precipitation estimates for the target domain according to the purpose of the domain adversarial neural network, and the domain discriminator is used to distinguish whether the spatiotemporal feature representations come from the source domain or the target domain, so as to drive adversarial training.
[0041] In this embodiment, the feature extractor employs a spatiotemporal feature extraction network constructed using stacked ConvLSTM hidden layers, such as... Figure 5 As shown, the ConvLSTM hidden layer includes an input gate ( Forgotten Gate ( ) and output gate ( ),in, and These represent the input and output vectors of the ConvLSTM hidden layer, respectively. As a memory unit, it is used to store the temporal characteristics of historical information. Information updating and transmission are achieved through the following gating mechanism: In the formula, This represents the weight matrix of each gate. This represents the corresponding bias vector. It is the sigmoid activation function. The hyperbolic tangent activation function maps the input to the interval [–1, 1].
[0042] In this embodiment, the domain adversarial neural network is trained through first forward propagation and back propagation to obtain a precipitation event classification model; the domain adversarial neural network is trained through second forward propagation and back propagation to obtain a precipitation regression model; the specific structures of the corresponding precipitation event classification model and precipitation regression simulation are shown in Tables 2 and 3.
[0043] In this embodiment, during the establishment of the adversarial training mechanism for precipitation event classification and precipitation regression, an early stopping mechanism is used to avoid overfitting. A precipitation event classification model / precipitation regression model based on a domain adversarial neural network is constructed, specifically configured as follows: one ConvLSTM layer, 64 hidden neurons, a learning rate of 0.001, 100 training epochs, and an early stopping patience value of 10. The parameters are then balanced. =1, the optimizer uses the Adam algorithm.
[0044] Specifically, during the first forward propagation process, the binary cross-entropy loss function is used to calculate the prediction bias of precipitation event classification and the difference in feature distribution between the source domain and the target domain. In the second forward propagation process, the mean square error loss function is used to calculate the prediction bias of precipitation regression, and the binary cross-entropy loss function is used to calculate the difference in characteristic distribution between the source domain and the target domain. Among them, the prediction bias of the precipitation event classification Represented as: In the formula, This indicates the number of samples in the source domain dataset. This represents the input data from the source domain. Indicates a precipitation event label. Indicates feature extractor, Represents the target predictor. This represents the model parameters of the feature extractor. This represents the model parameters of the target predictor; Prediction bias of precipitation regression Represented as: In the formula, This represents rain gauge measurement data from the source region; Differences in feature distribution between source and target domains Represented as: In the formula, This indicates the number of samples from the source domain dataset and the target domain dataset. Indicates input data, Represents a field label.
[0045] In this embodiment, during the backpropagation process, GRL is introduced to optimize the parameters of the domain adversarial neural network, thereby obtaining a precipitation event classification model / precipitation regression model. Introducing GRL to optimize the parameters of domain adversarial neural networks includes: Negative scaling is applied to the gradient of the domain adversarial neural network, and then the parameters are updated synchronously through a composite loss function; The composite loss function is: In the formula, This indicates an adjustable trade-off parameter.
[0046] In this embodiment of the invention, a multi-source precipitation data fusion model based on the aforementioned dual transfer learning architecture is used, as follows: Figure 6 As shown, step S500 includes the following sub-steps: S501. Input the preprocessed data fusion and classification data within the grid cells of the study area into the trained precipitation event classification model to determine the rainy / non-rainy day status of the corresponding grid cells; S502. For grid cells classified as rainy days, the preprocessed data fusion regression data within the corresponding grid cell is input into the precipitation regression model to quantitatively estimate the precipitation value of the corresponding grid cell on a rainy day, thus obtaining a multi-source precipitation data fusion product.
[0047] In this embodiment of the invention, for the purpose of multi-source precipitation data fusion in the Yellow River region, a fully trained multi-source precipitation data fusion model is used to generate precipitation fusion products. Based on the classification and regression evaluation index system listed in Table 4, the performance of this method in both precipitation event identification and precipitation value prediction is systematically evaluated. Through comparative analysis with the original precipitation products and HBV hydrological model correction products (e.g., ... Figure 7The scatter plot of the predicted results for each station shown in this paper verifies that the multi-source precipitation data fusion method proposed in this invention has a significant accuracy improvement effect.
[0048] Note: The number of samples; and They represent the first Observed and predicted values for the day; and These represent the average values of the observed and predicted values, respectively. This indicates the number of precipitation events detected simultaneously by observation and forecasting; and These represent the number of precipitation events detected only in observed data or predicted data, respectively. , and These represent the Pearson correlation coefficient, standard deviation, and correlation index, respectively.
[0049] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0050] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning, characterized in that, Includes the following steps: S100. Acquire and preprocess multi-source precipitation data and auxiliary meteorological data of the study area, and construct a standardized precipitation dataset, including precipitation correction input dataset, data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset. S200. Use the precipitation correction input dataset to calibrate the parameters of the HBV hydrological model, and use the calibrated HBV hydrological model parameter set to correct the bias of the data fusion input dataset. S300. Based on the corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset, divide the source domain and target domain, and construct the domain classification output dataset. S400: Construct a dual transfer learning architecture, use the domain classification output dataset to perform adversarial training on it, and obtain a multi-source precipitation data fusion model; S500: Using a multi-source precipitation data fusion model, precipitation data of the study area is retrieved to obtain a multi-source precipitation data fusion product.
2. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 1, characterized in that, Step S100 includes: Acquire multi-source precipitation data for the study area, including rain gauge observation data, hydrological station runoff observation data, auxiliary meteorological variable data, and gridded precipitation products; The acquired rain gauge observation data, gridded precipitation products, and auxiliary meteorological variable data were normalized. Based on the normalized data, precipitation, temperature and potential evapotranspiration grid data are extracted and their watershed average daily values are calculated to form a precipitation correction input dataset. For gridded precipitation products and auxiliary meteorological variable data, a system is constructed centered on each rain gauge station. Sub-mesh, extract Dimensional feature variables, combined The sliding window process generates a dimension of... High-dimensional input samples are used to form a data fusion input dataset; Rain gauge observation data are classified into rainy days and no rainy days using a set precipitation threshold as the discrimination criterion, forming a data fusion classification output dataset; Using actual precipitation as the regression label, the data from rain gauge observations are used to generate a data fusion regression output dataset. The data fusion input dataset, the data fusion classification output dataset, and the data fusion regression output dataset are used as the standard dataset.
3. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 1, characterized in that, Step S200 includes the following sub-steps: S201. Construct the HBV hydrological model and use the watershed water balance equation as the core physical constraint to simulate the runoff process. S202. Using the precipitation correction input dataset as the input data and the hydrological station runoff observation data as the output data, the parameters of the HBV hydrological model are calibrated to obtain the calibrated HBV hydrological model parameter set. S203. Based on the calibrated HBV hydrological model parameter set, the precipitation field in the data fusion input dataset is corrected for deviation, and the corrected data fusion input dataset is generated.
4. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 3, characterized in that, In step S201, the watershed water balance equation is: In the formula, , , These represent precipitation, evapotranspiration, and runoff, respectively. , , , These represent snow cover, soil water, upper groundwater, and lower groundwater reserves, respectively. In step S203, the formula for deviation correction is: In the formula, , These are the rainfall correction factor and the snowfall correction factor, respectively. Indicates temperature, Indicates the temperature threshold. This indicates the precipitation before correction. This indicates the corrected precipitation.
5. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 1, characterized in that, In step S300: Rain gauge stations within the study area are taken as the target domain, and rain gauge stations with similar geographical features are taken as the source domain. They are binary classified and labeled. The corrected data fusion input dataset, data fusion classification output dataset, and data fusion regression output dataset from the source and target domains are combined to construct a domain classification output dataset.
6. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 2, characterized in that, Step S400 includes: Constructing domain adversarial neural networks; An adversarial training mechanism for precipitation event classification is established. The data fusion input dataset is used as input, and the data fusion classification output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain a precipitation event classification model. An adversarial training mechanism for precipitation regression is established. The data fusion input dataset is used as input, and the data fusion regression output dataset and the domain classification output dataset are used as adversarial outputs. The domain adversarial neural network is trained adversarially to obtain the precipitation regression model. Based on precipitation event classification models and precipitation regression models, a multi-source precipitation data fusion model based on a dual transfer learning architecture was constructed.
7. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 6, characterized in that, The domain adversarial neural network includes a feature extractor, a target predictor, and a domain discriminator; The feature extractor is used to extract spatiotemporal feature representations from the input data, the target predictor is used to generate precipitation event types or precipitation estimates for the target domain according to the purpose of the domain adversarial neural network, and the domain discriminator is used to distinguish whether the spatiotemporal feature representations come from the source domain or the target domain, so as to drive adversarial training.
8. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 6, characterized in that, A precipitation event classification model is obtained by training the domain adversarial neural network through the first forward and back propagation; a precipitation regression model is obtained by training the domain adversarial neural network through the second forward and back propagation. During the first forward propagation process, the binary cross-entropy loss function is used to calculate the prediction bias of precipitation event classification and the difference in feature distribution between the source domain and the target domain. In the second forward propagation process, the mean square error loss function is used to calculate the prediction bias of precipitation regression, and the binary cross-entropy loss function is used to calculate the difference in characteristic distribution between the source domain and the target domain. Among them, the prediction bias of the precipitation event classification Represented as: In the formula, This indicates the number of samples in the source domain dataset. This represents the input data from the source domain. Indicates a precipitation event label. Indicates feature extractor, Represents the target predictor. This represents the model parameters of the feature extractor. This represents the model parameters of the target predictor; The prediction bias of the precipitation regression Represented as: In the formula, This represents rain gauge measurement data from the source region; The difference in feature distribution between the source domain and the target domain Represented as: In the formula, This indicates the number of samples from the source domain dataset and the target domain dataset. Indicates input data, Represents a field label.
9. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 8, characterized in that, During backpropagation, GRL is introduced to optimize the parameters of the domain adversarial neural network, thereby obtaining a precipitation event classification model / precipitation regression model; Introducing GRL to optimize the parameters of domain adversarial neural networks includes: Negative scaling is applied to the gradient of the domain adversarial neural network, and then the parameters are updated synchronously through a composite loss function; The composite loss function is: In the formula, This indicates an adjustable trade-off parameter.
10. The multi-source precipitation data fusion method based on hydrological model guidance and dual transfer learning according to claim 6, characterized in that, Step S500 includes the following sub-steps: S501. Input the preprocessed data fusion and classification data within the grid cells of the study area into the trained precipitation event classification model to determine the rainy / non-rainy day status of the corresponding grid cells; S502. For grid cells classified as rainy days, the preprocessed data fusion regression data within the corresponding grid cell is input into the precipitation regression model to quantitatively estimate the precipitation value of the corresponding grid cell on a rainy day, thus obtaining a multi-source precipitation data fusion product.