Precipitation space downscaling method and device based on physical constraint and deep learning

Abstract

The invention discloses a precipitation space downscaling method and device based on physical constraint and deep learning, and relates to the technical field of precipitation data processing. Comprising the following steps: forming training rainfall observation data by training low-spatial-resolution remote sensing rainfall data, training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data and training high-spatial-resolution soil moisture data; determining a target rainfall downscaling model constrained by atmospheric moisture and rainfall soil moisture based on a spatial cross validation method; and forming to-be-downscaled rainfall observation data by using the to-be-downscaled remote sensing rainfall data, the high-spatial-resolution meteorological data and the high-spatial-resolution soil moisture data, and inputting the to-be-downscaled rainfall observation data into the target rainfall downscaling model to output target downscaling remote sensing rainfall data. Through a rainfall downscaling model, atmospheric moisture and soil moisture are combined with physical constraint and spatial cross validation, and spatial generalization is ensured to improve the rainfall data spatial downscaling precision.

Description

Technical Field

[0001] The present invention relates to the technical field of precipitation data processing, and in particular, to a precipitation spatial downscaling method and device based on physical constraints and deep learning. Background Art

[0002] Precipitation, as an important part of the water cycle, plays a crucial role in the development and utilization of regional water resources and the construction of the ecological environment. At present, the main ways to obtain precipitation data include ground station observations, weather radar observations, and satellite remote sensing monitoring. With the development of satellite remote sensing technology, satellite precipitation products such as the Tropical Rainfall Measuring Mission (TRMM), CMORPH precipitation products, Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), Global Satellite Mapping of Precipitation (GSMaP), and Global Precipitation Measurement (GPM) have become important data sources for hydrological research in scarce areas because they can provide spatially covered data with large ranges and long time series, making up for the deficiencies of ground observations and radar data.

[0003] However, existing satellite precipitation products generally have the problem of low spatial resolution of precipitation data for precipitation observations, usually at scales of several kilometers to dozens of kilometers, which greatly limits their application in related research. For example, it affects the prediction accuracy of precipitation events, and there is an urgent need to improve the spatial resolution of precipitation data. Summary of the Invention

[0004] The present invention provides a precipitation spatial downscaling method and device based on physical constraints and deep learning, which solves the technical problem of low spatial resolution of existing precipitation data.

[0005] A precipitation spatial downscaling method based on physical constraints and deep learning provided by the first aspect of the present invention includes:

[0006] Obtain training low-spatial-resolution remote sensing precipitation data corresponding to multiple spatial regions, and corresponding training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data, and form multiple training precipitation observation data;

[0007] Use each of the training precipitation observation data to train and verify an initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint based on the spatial cross-validation method to determine a target precipitation downscaling model;

[0008] When receiving the to-be-downscaled remote sensing precipitation data, use the to-be-downscaled remote sensing precipitation data and the corresponding high-spatial-resolution meteorological data and high-spatial-resolution soil moisture data to form to-be-downscaled precipitation observation data, and input the to-be-downscaled precipitation observation data into the target precipitation downscaling model to output target downscaled remote sensing precipitation data.

[0009] Optionally, using each of the training precipitation observation data, based on the spatial cross-validation method, model training and verification are performed on the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint to determine the target precipitation downscaling model, including:

[0010] According to the spatial cross-validation method, each of the training precipitation observation data is composed into multiple groups of training sets - validation sets;

[0011] Based on the training precipitation observation data of any training set, input it into the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint for model training. After the loss function value converges, use the corresponding validation set for verification to determine the optimized precipitation downscaling model for each group of training sets - validation sets;

[0012] Integrate each of the optimized precipitation downscaling models to determine the target precipitation downscaling model.

[0013] Optionally, the precipitation downscaling model includes a vector quantization generative adversarial network and a temporal discriminator. The vector quantization generative adversarial network includes a generator and a spatial discriminator. The generator includes an encoder, a codebook module, an autoregressive transformer, and a decoder. Based on the training precipitation observation data of any training set, input it into the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint for model training. After the loss function value converges, use the corresponding validation set for verification to determine the optimized precipitation downscaling model for each group of training sets - validation sets, including:

[0014] Input the training precipitation observation data of any training set into the initial precipitation downscaling model, and encode the training precipitation observation data through the encoder to determine the encoded features;

[0015] Use the codebook module to vectorize the encoded features to obtain quantization features, and perform autoregressive processing on the quantization features based on the autoregressive transformer to generate transformed features;

[0016] Decode the transformed features through the decoder to output the predicted downscaled remote sensing precipitation data;

[0017] Input the training precipitation observation data and the predicted downscaled remote sensing precipitation data into the spatial discriminator to generate a spatial discrimination result;

[0018] According to the physical consistency scoring function of atmospheric moisture constraint and precipitation soil moisture constraint, calculate the first physical consistency score of the training precipitation observation data and the second physical consistency score of the predicted downscaled remote sensing precipitation data respectively;

[0019] After splicing the training precipitation observation data with the first physical consistency score and the predicted downscaled remote sensing precipitation data with the second physical consistency score, input them into the time discriminator to output the time discrimination result;

[0020] Determine the first-stage loss function value based on the spatial discrimination result and the time discrimination result. When the first-stage loss function value converges, determine the intermediate precipitation downscaling model;

[0021] Train the autoregressive transformer of the intermediate precipitation downscaling model based on the quantization features and determine the second-stage loss function value. After the second-stage loss function value converges, use the corresponding validation set for verification to determine the optimized precipitation downscaling model corresponding to each group of training set - validation set.

[0022] Optionally, the physical consistency scoring function includes:

[0023] ;

[0024] In the formula, is the physical consistency score, is the natural constant, is the first physical constraint weight, is the second physical constraint weight, is the atmospheric moisture conservation function, is the precipitation soil moisture loss function, is the specific humidity, is the time, is the east-west wind component at a height of 10 meters, is the coordinate axis in the east-west direction, is the coordinate axis in the north direction, is the north-south wind component at a height of 10 meters, is the east-west wind component at a height of 100 meters, is the north-south wind component at a height of 100 meters, is the derivative of the saturation vapor pressure with respect to temperature, is the meteorological constant, is the global radiation, is the latent heat of vaporization, is the change in soil moisture, is the number of grid cells.

[0025] Optionally, the process of determining the first-stage loss function value includes:

[0026] ;

[0027] In the formula, is the generation loss, is the reconstruction loss, For training precipitation observation data, For predicting downscaled remote sensing precipitation data, For the encoder, For the stop gradient operation, For the quantization feature, For the perceptual loss, For the expectation, For the adaptive weight, For the adversarial loss, For the spatial discriminator, For the gradient of the input to the final layer of the decoder, For the reconstruction loss, For a scalar for numerical stability, For the number of training precipitation observation data, For the th training precipitation observation data, For the temporal physical information loss, For the temporal discriminator, For the first physical consistency score, For the second physical consistency score.

[0028] Optionally, the process of determining the value of the second-stage loss function includes:

[0029] ;

[0030] In the formula, For the Transformer loss, For the expectation, For the probability distribution, For the th element of the index sequence, For the index sequence, For the sequence length.

[0031] A precipitation spatial downscaling device provided in the second aspect of the present invention includes:

[0032] A data acquisition module, configured to acquire training low-spatial-resolution remote sensing precipitation data corresponding to multiple spatial regions and corresponding training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data, and form multiple training precipitation observation data;

[0033] A model training module, configured to use each of the training precipitation observation data to perform model training and verification on an initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint based on the spatial cross-validation method, and determine a target precipitation downscaling model;

[0034] A remote sensing downscaling module, which is configured to, when receiving remotely sensed precipitation data to be downscaled, use the remotely sensed precipitation data to be downscaled, the corresponding high-spatial-resolution meteorological data, and the high-spatial-resolution soil moisture data to form precipitation observation data to be downscaled, and input the precipitation observation data into the target precipitation downscaling model to output the target downscaled remotely sensed precipitation data.

[0035] A computer device provided in the third aspect of the present invention includes a memory and a processor. A computer program is stored in the memory. When the computer program is executed by the processor, the processor is caused to execute the steps of the precipitation spatial downscaling method based on physical constraints and deep learning as described in any one of the above.

[0036] A computer-readable storage medium provided in the fourth aspect of the present invention has a computer program stored thereon. When the computer program is executed, it implements the precipitation spatial downscaling method based on physical constraints and deep learning as described in any one of the above.

[0037] A computer program product provided in the fifth aspect of the present invention includes a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the precipitation spatial downscaling method based on physical constraints and deep learning as described in any one of the above.

[0038] It can be seen from the above technical solutions that the present invention has the following advantages:

[0039] The above solution of the present invention provides a precipitation spatial downscaling method based on physical constraints and deep learning, including: obtaining training low-spatial-resolution remotely sensed precipitation data of multiple spatial regions and the corresponding training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data, and forming multiple training precipitation observation data; using each training precipitation observation data, based on the spatial cross-validation method, training and validating the initial precipitation downscaling model of the atmospheric moisture constraint and the precipitation soil moisture constraint to determine the target precipitation downscaling model; when receiving remotely sensed precipitation data to be downscaled, using the remotely sensed precipitation data to be downscaled, the corresponding high-spatial-resolution meteorological data, and the high-spatial-resolution soil moisture data to form precipitation observation data to be downscaled, and inputting the precipitation observation data into the target precipitation downscaling model to output the target downscaled remotely sensed precipitation data. Based on the above solution, through the designed precipitation downscaling model, the physical informed characteristic factors of atmospheric moisture and soil moisture closely related to precipitation are used to enhance the response sensitivity of precipitation events, and the combination of physical constraints and spatial cross-validation helps to enhance the generalization ability of the model, enabling it to adapt to precipitation downscaling tasks in different regions and different climate conditions, and effectively improving the spatial downscaling accuracy of precipitation data. Description of the Drawings

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0041] Figure 1 It is a flowchart of the steps of a precipitation spatial downscaling method based on physical constraints and deep learning provided by an embodiment of the present invention;

[0042] Figure 2 It is a schematic structural diagram of a PID-GAN precipitation downscaling model provided by an embodiment of the present invention;

[0043] Figure 3 It is a schematic comparison diagram of the spatial resolution of the original precipitation product and the precipitation downscaling product during the rainy season provided by an embodiment of the present invention;

[0044] Figure 4 It is a schematic comparison diagram of the accuracy of ground stations with the original precipitation product and the downscaling product provided by an embodiment of the present invention;

[0045] Figure 5 It is a structural block diagram of a precipitation spatial downscaling device based on physical constraints and deep learning provided by an embodiment of the present invention. Detailed implementation manners

[0046] The embodiments of the present invention provide a precipitation spatial downscaling method and device based on physical constraints and deep learning, which are used to solve the technical problem of the low spatial resolution of existing precipitation data.

[0047] To make the objectives, features, and advantages of the present invention more obvious and understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described below are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0048] Please refer to Figure 1 , Figure 1 It is a flowchart of the steps of a precipitation spatial downscaling method based on physical constraints and deep learning provided by an embodiment of the present invention.

[0049] A precipitation spatial downscaling method based on physical constraints and deep learning provided by the present invention includes:

[0050] Step 101: Obtain the training low-spatial-resolution remote sensing precipitation data, corresponding training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data for multiple spatial regions, and form multiple training precipitation observation data.

[0051] It should be noted that the training precipitation observation data includes the training low-spatial-resolution remote sensing precipitation data, training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data within a preset time range. The remote sensing precipitation data refers to the precipitation data monitored by remote sensing. The meteorological data includes meteorological observation values such as wind speed, wind direction, air temperature, dew point temperature, and global radiation. The soil moisture data includes soil moisture observation values such as soil moisture values. Different spatial regions can be understood as non-overlapping regions in geographical space.

[0052] Step 102: Use each training precipitation observation data to train and validate the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint based on the spatial cross-validation method to determine the target precipitation downscaling model.

[0053] Spatial cross-validation (Spatial CV) is a method for evaluating the performance of prediction models in geospatial applications, including geological hazard detection and prediction. In Spatial CV, the dataset is divided into multiple spatially non-overlapping subsets (folds). For each target fold, the model is trained on the remaining folds and validated on the target fold. By ensuring the spatial independence of the training and validation sets, Spatial CV can provide a more realistic assessment of the model performance.

[0054] It should be noted that in order to enable the model to exhibit good generalization ability in heterogeneous environments, the spatial cross-validation method is considered to train and validate the initial precipitation downscaling model for downscaling to determine the target precipitation downscaling model.

[0055] Step 102 includes the following sub-steps:

[0056] S1: Compose each training precipitation observation data into multiple training set-validation sets according to the spatial cross-validation method.

[0057] S2: Input the training precipitation observation data of any training set into the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint for model training. After the loss function value converges, use the corresponding validation set for validation to determine the optimized precipitation downscaling model for each training set-validation set.

[0058] S3: Determine the target precipitation downscaling model by synthesizing each optimized precipitation downscaling model.

[0059] It should be noted that according to the dataset partitioning strategy of the spatial cross-validation method, each set of training precipitation observation data is divided into multiple groups of training sets - validation sets. The training precipitation observation data of each group of training sets are respectively input into the initial precipitation downscaling model, and meteorological data and soil moisture data are used to assist in optimizing the remote sensing precipitation data for model training to determine the corresponding loss function values. The initial precipitation downscaling model is iteratively optimized according to the loss function values until the loss function values converge. Then, the corresponding validation sets are used for validation, and thus the corresponding optimized precipitation downscaling models are determined. On this basis, the performance evaluation results of the optimized precipitation downscaling models are comprehensively considered to determine the target precipitation downscaling model. In specific implementation, the entire spatial CV process can be set to repeat five times with different random seeds to adapt to and evaluate the uncertainty of data and models.

[0060] Furthermore, the precipitation downscaling model includes a vector quantization generative adversarial network and a temporal discriminator. The vector quantization generative adversarial network includes a generator and a spatial discriminator, and the generator includes an encoder, a codebook module, an autoregressive transformer, and a decoder.

[0061] It should be noted that in this embodiment, a precipitation downscaling model as shown in Figure 2 is designed and called the PID-GAN precipitation downscaling model. This model is a model combined based on a physics-informed neural network (PINN) and a generative adversarial network (GAN). And in the model, atmospheric moisture constraints and precipitation-soil moisture constraints are applied as physical constraints to enhance the sensitivity to precipitation time.

[0062] PINN (physics-informed neural network) is a machine learning model that combines deep learning and physics knowledge. This machine learning model integrates observational data and numerical model outputs into the training process. Essentially, it is a data assimilation problem, and a key feature is its simplicity in assimilating observational data. Time-varying observational data and / or spatial snapshots at various spatio-temporal scales can be easily incorporated into PINN training. It has some additional advantages compared to local-scale numerical models: First, PINN is a meshless method, which makes it an efficient and flexible downscaling tool. In practical applications, the downscaled solution of the region of interest can be obtained by constructing a PINN framework at a refined grid without relying on a complex grid generation process, which is particularly important in watershed models with complex terrains. Second, PINN can calculate the sub-grid solution of the coarse-scale model output by assimilating observational data without modifying the numerical algorithm or refining the grid resolution. Therefore, PINN can effectively handle the spatio-temporal variations of precipitation events, especially extreme precipitation events.

[0063] In the PID-GAN precipitation downscaling model, it includes a vector quantization generative adversarial network (VQ-GAN) and a temporal discriminator. The vector quantization generative adversarial network includes a generator and a spatial discriminator. The generator (VQ-VAE) includes an encoder (Encoder, E), a codebook module, an autoregressive transformer, and a decoder (Decoder, G), and participates in calculating the physics consistency scores of real samples (Ground truth Samples), that is, the training precipitation observation data, and the physics consistency scores of generated samples (Generated Samples), that is, the predicted downscaled remote sensing precipitation data, with the constraints of physical parameters that take atmospheric moisture and precipitation soil moisture as important indicators of precipitation. By combining the structures of the generator and the spatial and temporal discriminators, a robust framework is ensured, which can efficiently learn the inherent spatial patterns and time series in the data to handle the complex dynamics in precipitation, helping to show significant effects in capturing the complex spatio-temporal relationships in the data. At the same time, the physical constraints of physical consistency are fed into the temporal discriminator as additional inputs, so as to correct the generation results by learning the latent distribution of marked points to ensure that the generated high-spatial-resolution data conforms to physical laws;

[0064] In specific implementation, the generator can be composed of a convolutional neural network (CNN) encoder (E) and a decoder (G). To improve the performance of the generator, deeper convolutional modules are introduced in the encoder, or the ResNet (residual network) structure is adopted to improve the feature extraction ability. At the same time, the decoder part can be combined with more deconvolution layers and skip connections to reconstruct higher-resolution and finer precipitation data;

[0065] The autoregressive transformer can effectively capture the dynamic relationships in the time series data and improve the coherence of the generated time series. To further optimize, an attention mechanism, that is, the self-attention module in the Transformer, is introduced to better capture the spatio-temporal relationships in the precipitation process;

[0066] The spatial discriminator is used to distinguish the spatial differences between the generated samples and the real samples, and a deep convolutional neural network (DCNN) can be used as the basis. To enhance the ability of the spatial discriminator, multiple convolutional layers are introduced and combined with the max pooling operation to extract high-level features of the image. In addition, an attention mechanism is added to enable the spatial discriminator to pay more attention to the key regions in the samples, further improving the discrimination effect. The temporal discriminator is responsible for analyzing the temporal consistency of the precipitation data sequence. The long short-term memory network (LSTM) can be used to model the long-term dependencies of the time series, and the combination of the self-attention mechanism (such as Transformer) can further improve the modeling ability of the temporal discriminator for time series data. At the same time, to improve the accuracy of the model, a multi-scale temporal feature extraction module can be considered, which can effectively capture the changes at different time scales during the precipitation process. Through the feedback mechanism of the temporal discriminator and the spatial discriminator for correction, the generated samples can be made closer to the actual observed data.

[0067] Following the PID-GAN framework, sub-step S2 includes:

[0068] Input the training precipitation observation data of any training set into the initial precipitation downscaling model, and encode the training precipitation observation data through the encoder to determine the encoded features.

[0069] Use the codebook module to vectorize the encoded features to obtain the quantization features, and perform autoregressive processing on the quantization features based on the autoregressive transformer to generate the transformed features.

[0070] Decode the transformed features through the decoder to output the predicted downscaled remote sensing precipitation data.

[0071] Input the training precipitation observation data and the predicted downscaled remote sensing precipitation data into the spatial discriminator to generate the spatial discrimination result.

[0072] According to the physical consistency scoring function of the atmospheric moisture constraint and the precipitation soil moisture constraint, calculate the first physical consistency score of the training precipitation observation data and the second physical consistency score of the predicted downscaled remote sensing precipitation data respectively.

[0073] After splicing the training precipitation observation data with the first physical consistency score and splicing the predicted downscaled remote sensing precipitation data with the second physical consistency score, input them into the temporal discriminator to output the temporal discrimination result.

[0074] Determine the value of the first-stage loss function based on the spatial discrimination result and the temporal discrimination result. When the value of the first-stage loss function converges, determine the intermediate precipitation downscaling model.

[0075] Train the autoregressive transformer of the intermediate precipitation downscaling model based on the quantization features, and determine the value of the second-stage loss function. After the value of the second-stage loss function converges, use the corresponding validation set for verification to determine the optimized precipitation downscaling model corresponding to each group of training set-validation set.

[0076] Furthermore, the physical consistency scoring function includes:

[0077] ;

[0078] where, is the physical consistency score, is the natural constant, is the first physical constraint weight, is the second physical constraint weight, is the atmospheric moisture conservation function, is the precipitation soil moisture loss function, is the specific humidity (g / g), is the time, is the zonal wind component at 10 m height, is the zonal axis, is the meridional axis, is the meridional wind component at 10 m height, is the zonal wind component at 100 m height, is the meridional wind component at 100 m height, is the derivative of the saturation vapor pressure with respect to temperature (Pa / °C), is the meteorological constant (J / m²), is the global radiation (Pa / °C), is the latent heat of vaporization (J / g), is the change in soil moisture, is the number of grid cells.

[0079] It can be understood that the relevant parameters for calculating the physical consistency score can be extracted from the corresponding meteorological data and soil moisture data, and the second physical consistency score for predicting the downscaled remote sensing precipitation data can be calculated using the training high-spatial-resolution meteorological data and training high-spatial-resolution soil moisture data.

[0080] It should be noted that in the atmospheric moisture constraint, the moisture conservation equation of the atmospheric water balance physical equation describes the relationship between atmospheric moisture content, evaporation, and precipitation, and the formula is as follows:

[0081] ;

[0082] where, is the specific humidity (g / g), is the time, is the height east-west wind component (m / s), is the coordinate axis in the east-west direction, is the coordinate axis in the north-south direction, is the height north-south wind component (m / s), is the evapotranspiration rate (mm / hour), is the standardized precipitation value, is the vertical wind component (m / s), is the coordinate axis in the vertical direction; among them, the evapotranspiration rate can be estimated by the Makkink equation, which depends on temperature and solar radiation data and can provide relatively accurate results in cold and temperate humid climates;

[0083] Measuring vertical wind speed faces a series of challenges such as its low intensity, large spatial variability, the need for sensitive equipment, high requirements for meteorological stability, and the high cost and complexity of accurate measurement, and there is an easy lack of measurement data; Ignoring terms can simplify the process, but may ignore the moisture transport between the atmosphere layers, which may affect the three-dimensional dynamic balance of moisture in the atmosphere. To make up for the lack of vertical wind speed data, the focus is shifted to the horizontal wind components at different height levels and , and use the wind speed measurement at 100 meters height and (which can be provided by the ERA5 dataset) and the wind speed measurement at 10 meters height and (which can be provided by the AWS data); it can be understood that the interactions between the atmosphere layers have a significant impact on weather patterns, and these interactions continue up to the tropopause, at a height of about 10 kilometers. Therefore, focusing on the wind speed within the lower 100 meters, although it can provide some information, there is inevitably a certain degree of uncertainty because the atmosphere interactions beyond this range are often complex. By combining the horizontal wind components at 10 meters and 100 meters heights, the model can adapt to different height changes, thus better analyzing the moisture dynamics in the atmosphere; thereby, by simplifying the above moisture conservation equation, it is possible to approximately simulate the moisture transport between different heights without relying on vertical wind speed measurement, providing a more detailed perspective of the moisture dynamics conservation function in the atmosphere, that is, the atmospheric moisture conservation function in this embodiment;

[0084] In addition, given the existing clear positive correlation between precipitation and soil moisture, any deviation from this monotonic trend should be regarded as an anomaly. In particular, intense precipitation events identified through precipitation data should lead to a monotonic increase in soil moisture levels, and the change in soil moisture before and after precipitation should be positive. If a negative value appears, a precipitation loss factor should be introduced; a strong loss term is generated from soil moisture to form the model, which serves as an important reference for constraining the model results, thereby determining the precipitation soil moisture loss function.

[0085] Furthermore, the process of determining the value of the loss function in the first stage includes:

[0086] ;

[0087] where, is the generation loss, is the reconstruction loss, is the training precipitation observation data, is the predicted downscaled remote sensing precipitation data, is the encoder, is the stop gradient operation, is the quantization feature, is the perceptual loss, is the expectation, is the adaptive weight, is the adversarial loss, is the spatial discriminator, is the gradient of the input to the final layer of the decoder, is the reconstruction loss, is a scalar for numerical stability, is the number of training precipitation observation data, is the th training precipitation observation data, is the temporal physical information loss, is the temporal discriminator, is the first physical consistency score, is the second physical consistency score.

[0088] Furthermore, the process of determining the value of the loss function in the second stage includes:

[0089] ;

[0090] where, is the Transformer loss, is the expectation, is the probability distribution, is the th element of the index sequence, is the index sequence, is the sequence length.

[0091] It should be noted that the training and optimization process of VQ-GAN includes a first stage and a second stage; in the first stage, the whole is trained first, and at this time, the optimization objective of VQGAN is:

[0092] ;

[0093] In the second stage, the Transformer is trained for autoregressive generation of the prior, and the autoregressive transformer architecture aims to simulate the dynamic relationship between consecutive precipitation data; according to the aforementioned data processing process, the encoded features are quantized into quantization features , is vectorization, and the quantization features are represented as a sequence of indices which represents the index of the codebook in the VQ-GAN codebook module, represents the size of the codebook, represents the height, represents the width. These indices are converted into continuous vectors by an embedder and enhanced with positional embeddings to provide sequential information, and then input into the transformer for processing. The head module refines the output into logits representing the probability of using specific tokens. These logits can be used to calculate the cross-entropy loss to compare the predicted token probability distribution with the true token probability distribution; in this embodiment, for a given index , the transformer is trained to predict the distribution of subsequent indices . The autoregressive transformer uses a causal attention mechanism and only accesses the previously seen and current tokens when predicting the next token in the sequence, achieving efficient and context-sensitive output generation by maximizing the log-likelihood of the data representation.

[0094] Step 103: When receiving the remotely sensed precipitation data to be downscaled, the remotely sensed precipitation data to be downscaled is combined with the corresponding high-spatial-resolution meteorological data and high-spatial-resolution soil moisture data to form the precipitation observation data to be downscaled, and is input into the target precipitation downscaling model to output the target downscaled remotely sensed precipitation data.

[0095] It should be noted that when receiving the remotely sensed precipitation data to be downscaled, the remotely sensed precipitation data to be downscaled is combined with the corresponding high-spatial-resolution meteorological data and high-spatial-resolution soil moisture data to form the precipitation observation data to be downscaled, and the precipitation observation data to be downscaled is successively encoded, vectorized, autoregressively processed, and decoded by the generator of the target precipitation downscaling model to determine the target downscaled remotely sensed precipitation data.

[0096] To quantitatively evaluate the effect of precipitation spatial downscaling, indicators such as the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), relative bias (BIAS), and Pearson correlation coefficient (R), as well as nowcasting indicators such as the critical success index (CSI), false alarm rate (FAR), and probability of detection (POD) can be used:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] wherein, is the number of variables, represents the precipitation value at the ground observation station, represents the average value of the precipitation values at the ground observation stations; and respectively represent the precipitation value of the th grid point of the precipitation product and the average value of the corresponding values; H: is the number of precipitation events detected simultaneously by the precipitation product and the ground rain gauge station; M: is the number of rainfall events detected by the ground station, but not detected by the precipitation product; F: refers to the number of precipitation events detected by the precipitation product, while no rainfall event is detected by the ground rain gauge station;

[0104] To verify the effect of the precipitation downscaling model based on atmospheric moisture constraint and precipitation soil moisture constraint, exemplarily, the PID-GAN model with physical constraints is compared with the PID-GAN(P) model without physical constraints for downscaling, and the index evaluation results are shown in Table 1, where the precipitation thresholds for CSI and FAR are set to 1 and 6 mm respectively:

[0105] Table 1 Precision comparison of different types of precipitation

[0106]

[0107] As can be seen from Table 1, the PID-GAN model has made particularly significant progress in terms of root mean square error (RMSE) and Pearson correlation coefficient (PCC), indicating its strong correlation with actual precipitation events and high prediction accuracy. At the same time, it shows superior performance in the critical success index (CSI) under light (1 mm) and heavy (6 mm) precipitation thresholds. This performance indicates that the PID-GAN model can better detect and accurately predict extreme precipitation events. In addition, Table 1 also reveals an important insight: removing the physical constraints from the PID-GAN model (PID-GAN(P)) leads to a 45.8% and 76.9% decrease in CSI for light and heavy precipitation respectively. This decrease highlights the key role of physical constraints in the accuracy of extreme precipitation downscaling, further verifying the importance of integrating physical data into model design for performance improvement.

[0108] And according to Figure 3 the original precipitation spatial distribution measured by traditional precipitation products and the precipitation spatial distribution after downscaling in different regions of China during different time periods of the example. The gradient color from yellow to blue in the figure indicates the precipitation amount from low to high. It can be seen that the PID-GAN model shows better performance in the distribution from light rain to heavy rain after downscaling.

[0109] As Figure 4 shown, the relationship between the estimated precipitation of the traditional precipitation product GPM and the PID-GAN model of this embodiment and the precipitation based on ground observations such as rain gauges is compared in the form of a scatter plot. The red reference line in the figure represents the ideal situation where the estimated precipitation is equal to the ground observed precipitation, and the blue reference line represents the regression line obtained by data fitting to show the approximate relationship between the estimated value and the observed value. By comparing the two scatter distributions and statistical indicators, it can be seen that the PID-GAN model shows better capture ability in characterizing precipitation after downscaling.

[0110] In summary, the research results show that the PID-GAN model obtained by integrating physical data into the PINN framework shows good accuracy and reliability in precipitation spatial downscaling, especially in the identification and downscaling accuracy of extreme precipitation events. It can not only solve the limitations of traditional precipitation products or downscaling methods in spatial resolution and the handling of extreme precipitation events, but also provide higher-precision precipitation data for research such as hydrology and climate change.

[0111] Furthermore, constraints based on cloud physical processes can also be introduced, such as cloud top temperature, cloud water content, etc. For example, during the training process, the cloud water content can be constrained not to exceed a certain threshold under certain conditions, otherwise it may lead to excessive precipitation prediction. Furthermore, the relationship between local wind and precipitation can also be introduced. For example, the change of local wind affected by terrain uplift (such as the interaction between mountains and the ocean) will also affect the precipitation pattern.

[0112] In the embodiment of the present invention, based on the designed precipitation downscaling model, physical informed characteristic factors of atmospheric moisture and soil moisture closely related to precipitation are used to enhance the response sensitivity of precipitation events. Combining physical constraints with spatial cross-validation helps to enhance the generalization ability of the model, enabling it to adapt to precipitation downscaling tasks in different regions and different climate conditions. Especially in the application scenarios of extreme precipitation events, it can effectively improve the spatial downscaling accuracy of precipitation data.

[0113] Please refer to Figure 5 , Figure 5 which is a structural block diagram of a precipitation spatial downscaling device based on physical constraints and deep learning provided by the embodiment of the present invention.

[0114] A precipitation spatial downscaling device based on physical constraints and deep learning provided by the present invention includes:

[0115] A data acquisition module 501, configured to acquire training low-spatial-resolution remote sensing precipitation data and corresponding training low-spatial-resolution meteorological data, training low-spatial-resolution soil moisture data, training high-spatial-resolution meteorological data, and training high-spatial-resolution soil moisture data for multiple spatial regions, and form multiple training precipitation observation data;

[0116] A model training module 502, configured to use each training precipitation observation data to perform model training and verification on the initial precipitation downscaling model of atmospheric moisture constraint and precipitation soil moisture constraint based on the spatial cross-validation method, and determine the target precipitation downscaling model;

[0117] A remote sensing downscaling module 503, configured to when receiving the to-be-downscaled remote sensing precipitation data, form to-be-downscaled precipitation observation data by using the to-be-downscaled remote sensing precipitation data and corresponding high-spatial-resolution meteorological data and high-spatial-resolution soil moisture data, and input the to-be-downscaled precipitation observation data into the target precipitation downscaling model to output the target downscaled remote sensing precipitation data.

[0118] Optionally, the model training module 502 includes:

[0119] A data partitioning unit, configured to form multiple groups of training sets - validation sets from each training precipitation observation data according to the spatial cross-validation method;

[0120] A training and validation unit for inputting the training precipitation observation data of any training set into the initial precipitation downscaling model with atmospheric moisture constraint and precipitation soil moisture constraint for model training. After the loss function value converges, the corresponding validation set is used for validation to determine the optimized precipitation downscaling model for each group of training set - validation set;

[0121] A target model determination unit for comprehensively determining the target precipitation downscaling model from each optimized precipitation downscaling model.

[0122] Optionally, the precipitation downscaling model includes a vector quantization generative adversarial network and a temporal discriminator. The vector quantization generative adversarial network includes a generator and a spatial discriminator, and the generator includes an encoder, a codebook module, an autoregressive transformer, and a decoder. The training and validation unit is specifically used for:

[0123] Input the training precipitation observation data of any training set into the initial precipitation downscaling model, and encode the training precipitation observation data through the encoder to determine the encoded features;

[0124] Use the codebook module to vectorize the encoded features to obtain quantization features, and perform autoregressive processing on the quantization features based on the autoregressive transformer to generate transformed features;

[0125] Decode the transformed features through the decoder to output the predicted downscaled remote sensing precipitation data;

[0126] Input the training precipitation observation data and the predicted downscaled remote sensing precipitation data into the spatial discriminator to generate a spatial discrimination result;

[0127] According to the physical consistency scoring function of the atmospheric moisture constraint and the precipitation soil moisture constraint, calculate the first physical consistency score of the training precipitation observation data and the second physical consistency score of the predicted downscaled remote sensing precipitation data respectively;

[0128] After splicing the training precipitation observation data with the first physical consistency score and the predicted downscaled remote sensing precipitation data with the second physical consistency score, input them into the temporal discriminator to output a temporal discrimination result;

[0129] Determine the first-stage loss function value based on the spatial discrimination result and the temporal discrimination result. When the first-stage loss function value converges, determine the intermediate precipitation downscaling model;

[0130] Train the autoregressive transformer of the intermediate precipitation downscaling model based on the quantization features and determine the second-stage loss function value. After the second-stage loss function value converges, use the corresponding validation set for validation to determine the optimized precipitation downscaling model corresponding to each group of training set - validation set.

[0131] Optionally, the physical consistency scoring function includes:

[0132] ;

[0133] In the formula, is the physical consistency score, is the natural constant, is the first physical constraint weight, is the second physical constraint weight, is the atmospheric moisture conservation function, is the precipitation soil moisture loss function, is the specific humidity, is the time, is the east-west wind component at 10 m height, is the coordinate axis in the east-west direction, is the coordinate axis in the north direction, is the north-south wind component at 10 m height, is the east-west wind component at 100 m height, is the north-south wind component at 100 m height, is the derivative of the saturation vapor pressure with respect to temperature, is the meteorological constant, is the global radiation, is the latent heat of vaporization, is the change in soil moisture, is the number of grid cells.

[0134] Optionally, the process of determining the value of the first-stage loss function includes:

[0135] ;

[0136] In the formula, is the generation loss, is the reconstruction loss, is the training precipitation observation data, is the predicted downscaled remote sensing precipitation data, is the encoder, is the stop gradient operation, is the quantization feature, is the perceptual loss, is the expectation, is the adaptive weight, is the adversarial loss, is the spatial discriminator, is the gradient of the input to the final layer of the decoder, is the reconstruction loss, is a scalar for numerical stability, is the number of training precipitation observation data, is the th training precipitation observation data, is the loss of time physical information, is the time discriminator, is the first physical consistency score, is the second physical consistency score.

[0137] Optionally, the process of determining the value of the loss function in the second stage includes:

[0138] ;

[0139] In the formula, is the Transformer loss, is the expectation, is the probability distribution, is the th element of the index sequence, is the index sequence, is the sequence length.

[0140] An embodiment of the present invention also provides a computer device, including a memory and a processor, and a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the precipitation spatial downscaling method based on physical constraints and deep learning in any of the above embodiments.

[0141] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program / instruction is stored, and when the computer program / instruction is executed by the processor, the steps of the precipitation spatial downscaling method based on physical constraints and deep learning in any of the above embodiments are implemented.

[0142] An embodiment of the present invention also provides a computer program product, including a computer program / instruction, and when the computer program / instruction is executed by the processor, the steps of the precipitation spatial downscaling method based on physical constraints and deep learning in any of the above embodiments are implemented.

[0143] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0144] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the devices or units can be in electrical, mechanical or other forms.

[0145] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0146] In addition, each functional unit in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0147] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical discs and other various media that can store program codes.

[0148] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. A precipitation spatial downscaling method based on physical constraints and deep learning, characterized in that: include: Acquire training low spatial resolution remote sensing precipitation data and corresponding training low spatial resolution meteorological data, training low spatial resolution soil moisture data, training high spatial resolution meteorological data and training high spatial resolution soil moisture data of multiple spatial regions, and form multiple training precipitation observation data; Using the training precipitation observation data, the initial precipitation downscaling model constrained by atmospheric moisture and precipitation soil moisture is trained and verified based on the spatial cross-validation method to determine the target precipitation downscaling model; When the remote sensing precipitation data to be downscaled is received, the remote sensing precipitation data to be downscaled is used together with the corresponding high spatial resolution meteorological data and high spatial resolution soil moisture data to form precipitation observation data to be downscaled, and is input into the target precipitation downscaling model to output the target downscaled remote sensing precipitation data.

2. The precipitation spatial downscaling method based on physical constraints and deep learning according to claim 1 is characterized in that: The training precipitation observation data are used to perform model training and verification on the initial precipitation downscaling model constrained by atmospheric moisture and precipitation soil moisture based on the spatial cross-validation method, and a target precipitation downscaling model is determined, including: According to the spatial cross-validation method, the training precipitation observation data are combined into multiple training set-validation sets; Based on the training precipitation observation data of any training set, the initial precipitation downscaling model with atmospheric moisture constraint and precipitation soil moisture constraint is input for model training until the loss function value converges, and then the corresponding validation set is used for validation to determine the optimized precipitation downscaling model of each training set-validation set; The target precipitation downscaling model is determined by comprehensively analyzing the optimized precipitation downscaling models.

3. The precipitation spatial downscaling method based on physical constraints and deep learning according to claim 2 is characterized in that: The precipitation downscaling model includes a vector quantization generative adversarial network and a time discriminator, the vector quantization generative adversarial network includes a generator and a spatial discriminator, the generator includes an encoder, a codebook module, an autoregressive transformer and a decoder; the training precipitation observation data based on any training set is input into the initial precipitation downscaling model with atmospheric moisture constraints and precipitation soil moisture constraints for model training, until the loss function value converges, and then the corresponding validation set is used for verification to determine the optimized precipitation downscaling model of each group of training set-validation set, including: Inputting the training precipitation observation data of any training set into the initial precipitation downscaling model, encoding the training precipitation observation data through an encoder, and determining encoding features; The coding feature is vectorized by using a codebook module to obtain a quantized feature, and the quantized feature is autoregressively processed based on an autoregressive transformer to generate a transformed feature; Decoding the transformation feature through a decoder to output predicted downscaled remote sensing precipitation data; Inputting the training precipitation observation data and the predicted downscaled remote sensing precipitation data into a spatial discriminator to generate a spatial discrimination result; According to the physical consistency scoring functions of the atmospheric moisture constraint and the precipitation soil moisture constraint, respectively calculating the first physical consistency score of the training precipitation observation data and the second physical consistency score of the predicted downscaled remote sensing precipitation data; After splicing the training precipitation observation data with the first physical consistency score and splicing the predicted downscaled remote sensing precipitation data with the second physical consistency score, the two are input into a time discriminator, and a time discrimination result is output; Determine the first-stage loss function value based on the spatial discrimination result and the temporal discrimination result, and determine the intermediate precipitation downscaling model when the first-stage loss function value converges; Based on the quantitative characteristics, the autoregressive transformer of the intermediate precipitation downscaling model is trained, and the second-stage loss function value is determined. When the second-stage loss function value converges, the corresponding validation set is used for verification to determine the optimized precipitation downscaling model corresponding to each training set-validation set.

4. The precipitation spatial downscaling method based on physical constraints and deep learning according to claim 3 is characterized in that: The physical consistency scoring functions include: ； In the formula, Score the physical consistency, is a natural constant, is the first physical constraint weight, is the second physical constraint weight, is the atmospheric moisture conservation function, is the precipitation soil moisture loss function, is the specific humidity, For time, is the east-west wind component at 10 meters height, is the east-west coordinate axis, is the north coordinate axis, is the north-south wind component at 10 meters height, is the east-west wind component at 100 m altitude, is the north-south wind component at 100 m altitude, is the derivative of saturated water vapor pressure with respect to temperature, is a meteorological constant, For global radiation, is the latent heat of vaporization, is the change in soil moisture, is the number of grids.

5. The precipitation spatial downscaling method based on physical constraints and deep learning according to claim 3 is characterized in that: The process of determining the loss function value in the first stage includes: ； In the formula, To generate the loss, To rebuild the losses, To train the precipitation observation data, To predict downscaled remote sensing precipitation data, For the encoder, To stop the gradient operation, To quantify the features, is the perceived loss, For expectations, is the adaptive weight, To combat losses, is the spatial discriminator, is the gradient of the decoder's final layer input, To rebuild the losses, is a numerically stable scalar, is the number of training precipitation observation data, For the training precipitation observation data, is the time-physical information loss, is the time discriminator, Score the first physical consistency, Score the second physical consistency.

6. The precipitation spatial downscaling method based on physical constraints and deep learning according to claim 3 is characterized in that: The process of determining the loss function value in the second stage includes: ； In the formula, is the Transformer loss, For expectations, is the probability distribution, The index sequence elements, is the index sequence, is the sequence length.

7. A precipitation spatial downscaling device based on physical constraints and deep learning, characterized in that: include: A data acquisition module is used to acquire training low spatial resolution remote sensing precipitation data and corresponding training low spatial resolution meteorological data, training low spatial resolution soil moisture data, training high spatial resolution meteorological data and training high spatial resolution soil moisture data of multiple spatial regions, and form multiple training precipitation observation data; A model training module is used to use the training precipitation observation data to perform model training and verification on the initial precipitation downscaling model constrained by atmospheric moisture and precipitation soil moisture based on a spatial cross-validation method, and determine a target precipitation downscaling model; The remote sensing downscaling module is used to, when receiving the remote sensing precipitation data to be downscaled, use the remote sensing precipitation data to be downscaled and the corresponding high spatial resolution meteorological data and high spatial resolution soil moisture data to form precipitation observation data to be downscaled, and input the target precipitation downscaling model to output the target downscaled remote sensing precipitation data.

8. A computer device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the precipitation spatial downscaling method based on physical constraints and deep learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the precipitation spatial downscaling method based on physical constraints and deep learning are implemented as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the precipitation spatial downscaling method based on physical constraints and deep learning are implemented as described in any one of claims 1 to 6.

Images