Multi-modal feature and deep learning-based water vapor product resolution enhancement method and system

Abstract

This invention discloses a method and system for enhancing the resolution of water vapor products using multimodal features and deep learning. The invention first acquires a first atmospheric precipitable water vapor fusion field generated using a fusion method, along with multimodal feature data related to atmospheric water vapor. Then, it uses an empirical orthogonal function method to extract common modes from both, obtaining common variation patterns. Finally, it inputs the common variation patterns into a pre-trained AutoResNet model to generate a second atmospheric precipitable water vapor product with a higher spatiotemporal resolution than the first fusion field. The AutoResNet model is a neural network model based on an autoencoder and a residual network, which has established a nonlinear mapping relationship between multimodal features and atmospheric precipitable water vapor. This invention effectively breaks through the resolution limit of the original water vapor data source, significantly improving the temporal and spatial resolution of the fusion product, and can be applied to severe weather monitoring, regional water cycle research, and high-resolution water vapor inversion operations.

Description

Technical Field

[0001] This invention belongs to the fields of satellite remote sensing water vapor inversion, navigation meteorology, and multi-source heterogeneous data fusion technology, specifically involving a method and system for enhancing the resolution of water vapor products using multimodal features and deep learning. Background Technology

[0002] Satellite remote sensing can detect atmospheric water vapor by utilizing the absorption characteristics of water vapor to radiation of different wavelengths. However, due to the complexity of surface radiation, cloud influence, and limitations of sensor inversion conditions, single remote sensing water vapor products often suffer from insufficient accuracy or limited spatiotemporal resolution.

[0003] Navigation satellite inversion of water vapor data offers advantages such as high temporal resolution and good continuity, but its spatial sampling density depends on the distribution of the satellite network. Near-infrared satellite water vapor data possesses high spatial resolution, but its temporal resolution is limited and it is easily affected by clouds. Occultation and radiosonde data, on the other hand, have vertical structure information and high fidelity, respectively, but their spatial coverage and timeliness are limited. Therefore, there are significant differences in accuracy, sampling, and structure among water vapor data from different sources.

[0004] In existing technologies, multi-source water vapor fusion often employs physical models, traditional weighting methods, or shallow machine learning methods. While these methods can improve accuracy to some extent, they still have the following shortcomings: First, research on dynamic weighting mechanisms for the heterogeneity and regional differences of multi-source water vapor is insufficient. Second, traditional fusion results are usually limited by the maximum spatiotemporal resolution of the original data source, making it difficult to further improve the refinement of the product. Third, there is insufficient utilization of the complex nonlinear coupling relationships between water vapor and meteorological elements, topographic elements, surface parameters, and spatiotemporal trend factors. Summary of the Invention

[0005] To address the problems of inconsistent accuracy of multi-source water vapor data, insufficient spatiotemporal resolution of fusion results, and inadequate utilization of complex multimodal coupling relationships in existing technologies, this invention provides a method and system for enhancing the resolution of water vapor products using multimodal features and deep learning. By acquiring a first atmospheric precipitable water vapor fusion field generated using a fusion method and multimodal feature data related to atmospheric water vapor, the common modes of the two are extracted using an empirical orthogonal function method. The resulting common variation patterns are then input into a pre-trained AutoResNet model based on an autoencoder and residual network coupling to generate a second atmospheric precipitable water vapor product with a higher spatiotemporal resolution than the first fusion field, thereby improving the accuracy and enhancing the spatiotemporal resolution of regional atmospheric water vapor products.

[0006] According to one aspect of the present invention, a method for enhancing the resolution of water vapor products using multimodal features and deep learning is provided, comprising: Obtain the first atmospheric precipitable water fusion field generated using the fusion method; Acquire multimodal feature data related to atmospheric water vapor; The empirical orthogonal function method is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field to obtain common variation patterns; The common variation pattern is input into a pre-trained AutoResNet model to generate a second atmospheric precipitable water product. The spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field. The AutoResNet model is a neural network model based on an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established.

[0007] As a further technical solution, the fusion method includes: Using radiosonde atmospheric precipitable water as a reference, the relative weights of multi-source atmospheric precipitable water observation data are iteratively obtained by the Helmert variance component estimation method. The first atmospheric precipitable water field is obtained by weighting and fusing the multi-source atmospheric precipitable water observation data using the relative weights and spherical cap harmonic functions, solving for the spherical cap harmonic coefficients.

[0008] As a further technical solution, the multi-source atmospheric precipitable water observation data includes atmospheric precipitable water retrieved from BeiDou navigation satellites, atmospheric precipitable water retrieved from Fengyun near-infrared remote sensing, atmospheric precipitable water retrieved from Fengyun 3C occultation, and atmospheric precipitable water retrieved from radiosonde.

[0009] As a further technical solution, the multimodal feature data includes meteorological elements, topographic elements, surface parameters, and spatiotemporal information; wherein, the meteorological elements include one or more of precipitation, surface air pressure, surface air temperature, wind speed, and specific humidity; the topographic elements include a digital elevation model; the surface parameters include an enhanced vegetation index; and the spatiotemporal information includes spatial location and time information.

[0010] As a further technical solution, the empirical orthogonal function method is used to extract the common mode from the multimodal feature data and the first atmospheric precipitable water fusion field, including: Anomaly processing is performed on the multimodal feature data and the first atmospheric precipitable water fusion field, respectively; Construct the covariance matrix between the anomaly-processed multimodal feature data and the first atmospheric precipitable water fusion field; The covariance matrix is ​​subjected to eigenvalue decomposition to extract the principal mode space function and its corresponding time coefficient, which are used as the common variation mode.

[0011] As a further technical solution, the AutoResNet model includes an input layer, an encoding layer, a hidden representation layer, a decoding layer, and an output layer; the encoding layer is used to progressively compress the dimensionality of the input features and enhance the training stability of the deep network through residual connections; the decoding layer is used to recover the high-level semantic representation and output the atmospheric precipitable water estimate at the target spatiotemporal resolution.

[0012] As a further technical solution, the AutoResNet model employs multiple convolutional layers, with the number of output channels increasing with the number of layers, and the loss function is weighted mean square error.

[0013] According to one aspect of the present invention, a water vapor product resolution enhancement system based on multimodal features and deep learning is provided, comprising: The data acquisition module is used to acquire the first atmospheric precipitable water fusion field generated by the fusion method, and to acquire multimodal feature data related to atmospheric precipitable water. The modality extraction module is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field using the empirical orthogonal function method, so as to obtain common variation patterns; The model acquisition module is used to acquire a pre-trained AutoResNet model. The AutoResNet model is a neural network model based on the coupling of an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established. The product generation module is used to input the common change pattern into the AutoResNet model to generate a second atmospheric precipitable water product, wherein the spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field.

[0014] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method.

[0015] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on the fusion of multiple water vapor sources, this invention introduces multimodal information such as meteorology, topography, surface and spatiotemporal trends, which breaks through the limitation of the resolution of the original water vapor data source itself and significantly improves the temporal and spatial resolution of the fusion product.

[0017] (2) This invention uses AutoResNet to learn complex nonlinear coupling relationships and uses an encoder-decoder structure to alleviate the local collinearity problem between input features, thereby achieving resolution enhancement of fused water vapor products. It is suitable for constructing enhancement models of high-dimensional, multimodal, and non-stationary water vapor fields.

[0018] (3) The present invention uses spherical cap harmonic functions to perform spectral expansion of the regional water vapor field, adapts to the modeling characteristics of the spherical cap domain, and reduces the boundary distortion problem in regional modeling of the traditional global spherical harmonic method.

[0019] (4) This invention utilizes Helmert variance component estimation to achieve dynamic weighting of multi-source data, which can adaptively adjust the weights according to the error level of each data source, thereby improving the robustness and reliability of the fusion results.

[0020] (5) The product obtained by the present invention can be applied to disaster weather monitoring, regional water cycle research, meteorological operational assimilation and high-resolution water vapor inversion scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the method for enhancing the resolution of water vapor products using multimodal features and deep learning, as provided in an embodiment of the present invention. Detailed Implementation

[0023] To facilitate understanding of this invention, the relevant terms are explained as follows: Precipitable water volume (PWV): refers to the height at which all the water vapor contained in a unit cross-sectional area of ​​a vertical air column would condense into liquid water. It is a standard physical quantity for measuring the total amount of water vapor in the atmosphere. In this invention, "water vapor," "water vapor content," and "precipitable water volume" all refer to the same physical quantity and can be used interchangeably.

[0024] The first atmospheric precipitable water distribution field refers to the spatially continuous and highly accurate initial atmospheric precipitable water distribution field obtained by weighted fusing multi-source atmospheric precipitable water observation data through fusion methods (such as multi-source data fusion based on spherical cap harmonic functions and Helmert variance component estimation). This fused field is the input object of the resolution enhancement method of this invention.

[0025] The second atmospheric precipitable water product refers to the atmospheric precipitable water distribution product with higher spatiotemporal resolution generated after inputting the aforementioned common variation pattern into a pre-trained AutoResNet model. This product is the final output of the resolution enhancement method of this invention, and its spatiotemporal resolution is higher than that of the first atmospheric precipitable water fusion field (i.e., the first atmospheric precipitable water product).

[0026] Relationship between fused field and product: In this invention, "fused field" emphasizes the spatial distribution field obtained through weighted fusion, and "product" emphasizes the data results available for subsequent applications. The first atmospheric precipitable water fused field is the first atmospheric precipitable water product, and the second atmospheric precipitable water product is the enhanced output product of this invention. Both are numerically representations of the spatial distribution of atmospheric precipitable water, but the second product has a higher spatiotemporal resolution.

[0027] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0029] This invention provides a method for enhancing the resolution of water vapor products using multimodal features and deep learning, such as... Figure 1 As shown, it includes the following steps:

[0030] Step 1: Acquisition and preprocessing of multi-source atmospheric precipitable water data.

[0031] In this embodiment, at least four types of atmospheric precipitable water observation data are acquired: atmospheric precipitable water retrieved from BeiDou navigation satellites, atmospheric precipitable water retrieved from Fengyun near-infrared remote sensing, atmospheric precipitable water retrieved from Fengyun 3C occultation, and atmospheric precipitable water retrieved from radiosonde.

[0032] The wet refractive index of the Fengyun-3C occultation data can be calculated based on the meteorological parameters layer by layer, and the zenith wet delay can be obtained by integration along the zenith direction, and then converted into atmospheric precipitable water content; the specific humidity of the radiosonde data can be calculated based on air pressure, temperature, relative humidity or dew point temperature, and the tropospheric water vapor content can be obtained by numerical integration.

[0033] Because different data sources differ in observational height benchmarks, spatial sampling methods, and temporal sampling frequencies, they need to be uniformly processed first. For discrete station observation data, spatial correspondence is achieved through address matching; for grid-type remote sensing water vapor data, adjacent grids around the point to be compared are extracted and bilinear interpolation is used to obtain the atmospheric precipitable water value at the target location; for data with vertical layer information, necessary vertical interpolation or height reduction is performed.

[0034] Specifically, the elevation benchmark unification includes reducing water vapor observations from different data sources to a preset reference height; the time unification includes matching according to the same observation time or a preset time window; and the spatial location unification includes bilinear interpolation of grid data, address matching of discrete station data, or horizontal / vertical interpolation.

[0035] Step 2, generation of the first atmospheric precipitable water fusion field.

[0036] In this embodiment, a first atmospheric precipitable water fusion field is first generated using a fusion method.

[0037] After completing the spatiotemporal matching, using radiosonde atmospheric precipitable water as a reference standard, the root mean square error, average deviation, and correlation between each data source and the reference value are calculated to assess the overall error level and regional distribution characteristics of various types of data.

[0038] Subsequently, based on the error levels of various water vapor data, the relative weights of multi-source atmospheric precipitable water observation data are iteratively obtained using the Helmert variance component estimation method. The basic process is as follows: first, it is assumed that various types of observation data have the same prior variance, and the initial fusion parameters are estimated using least squares; then, the unit weight variance is estimated based on the residuals of each observation group; if the unit weight variances of different observation groups are inconsistent, the relative weights of each data source are iteratively adjusted until the unit weight variance reaches the preset convergence threshold.

[0039] The relative weights and spherical cap harmonic functions are used to perform weighted fusion of multi-source atmospheric precipitable water observation data. The spherical cap harmonic function is an orthogonal and complete basis function within a spherical cap domain, composed of non-integer associated Legendre functions and trigonometric functions, which can perform spectral expansion of regional-scale water vapor fields. Its specific mathematical form can be found in existing techniques in this field.

[0040] An observation equation is constructed using multi-source atmospheric precipitable water observations. The spherical cap harmonic coefficients are solved using the relative weights (i.e., dynamic weights) obtained above, and the merged water vapor field is reconstructed to obtain the first atmospheric precipitable water field. This method can fully utilize the advantages of different data sources to achieve spatially continuous and highly accurate regional merged atmospheric precipitable water products.

[0041] Step 3: Acquisition of multimodal feature data.

[0042] To overcome the upper limit of the spatiotemporal resolution of the original water vapor data source, this embodiment further introduces multimodal feature data related to atmospheric precipitable water. The multimodal feature data includes meteorological elements, topographic elements, surface parameters, and spatiotemporal information.

[0043] The meteorological elements include one or more of precipitation, surface air pressure, surface air temperature, wind speed, and specific humidity; the topographic elements include a digital elevation model (DEM); the surface parameters include the enhanced vegetation index (EVI); and the spatiotemporal information includes spatial location (longitude, latitude) and temporal information (time, month, or season).

[0044] The selected features are spatiotemporally unified to ensure that they have the same spatial coverage and temporal sampling as the first atmospheric precipitable water fusion field.

[0045] Step 4: Extraction of common modes of empirical orthogonal functions.

[0046] In this embodiment, the Empirical Orthogonal Function (EOF) method is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field, thereby obtaining common variation patterns. The specific steps are as follows:

[0047] (1) Perform anomaly processing on the multimodal feature data and the first atmospheric precipitable water fusion field respectively;

[0048] (2) Construct the covariance matrix between the multimodal feature data after anomaly processing and the first atmospheric precipitable water fusion field.

[0049] (3) Perform eigenvalue decomposition on the covariance matrix to extract the principal mode space function and its corresponding time coefficient, which are used as the common variation pattern. This common variation pattern is used to characterize the cooperative variation structure between atmospheric precipitable water and multimodal features.

[0050] Step 5: AutoResNet model construction and resolution enhancement.

[0051] Considering the potential for local collinearity and complex nonlinear coupling among different modal features, this implementation pre-constructs and trains an AutoResNet model. The AutoResNet model is a neural network model based on the coupling of an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established.

[0052] The AutoResNet model comprises an input layer, an encoding layer, a hidden representation layer, a decoding layer, and an output layer. The encoding layer progressively compresses the dimensionality of the input features to reduce redundancy and collinearity, and enhances the training stability of the deep network through residual connections to reduce the impact of local collinearity issues among multimodal inputs on the modeling results. The decoding layer recovers high-level semantic representations and outputs an estimate of atmospheric precipitable water at the target spatiotemporal resolution.

[0053] In a specific example, the AutoResNet model has 8 convolutional layers with a kernel size of 3×3 and a stride of 1. The number of output channels is 64, 128, 256, and 512, respectively. During model training, the batch size is set to 64, the training epochs are 200, the learning rate is set to 0.0001, and the loss function is weighted mean squared error. The training samples consist of a fused field of first atmospheric precipitable water and multimodal features, and are randomly stratified and divided into training, validation, and test sets in a 6:2:2 ratio.

[0054] Those skilled in the art will understand that the data source types, threshold settings, network parameters, and feature combinations provided in this embodiment can be adjusted according to the actual data scale and computing resources, but using the above parameters can achieve better model performance.

[0055] After the model training is complete, the aforementioned common variation patterns are input into the pre-trained AutoResNet model to generate a second atmospheric precipitable water product. The spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field.

[0056] Step 6: Output and Verification of Results.

[0057] The generated second atmospheric precipitable water product can be used for severe weather monitoring, regional water cycle research, meteorological operational assimilation, and high-resolution atmospheric precipitable water retrieval scenarios. The output second atmospheric precipitable water product is superior to any single input data source in at least two of the following aspects: spatial resolution, temporal resolution, and retrieval accuracy.

[0058] The results were validated using independent external observation data (such as radiosonde data not involved in the fusion). Validation included overall accuracy (root mean square error, mean bias), spatial detail recovery capability of atmospheric precipitable water, and temporal continuity enhancement. Practical applications show that the augmented product obtained using this method significantly outperforms single-source data sources or traditional fusion products in both spatial and temporal resolution.

[0059] The overall accuracy statistics are shown in Table 1. This invention used the constructed model to calculate the fitting model, ten-fold cross-validation accuracy, and correlation between the model's water vapor estimate and GNSS water vapor value samples for 238 GNSS station locations in a certain area. The results are shown in Table 1. Table 1 shows that the root mean square error (RMSE) and bias of the multimodal augmentation model are 0.89 mm and 0.02 mm, respectively, and the correlation coefficient (R) between GNSS and model water vapor is 0.995. The ten-fold cross-validation results show that the model's RMSE and bias are 1.35 mm and 0 mm, respectively, and the cross-validation correlation is 0.989. This result indicates that in the first-stage reconstruction, the MLR model performs the worst, while the GRNN model performs the best.

[0060] Table 1. Fitting and cross-validation results of the multimodal enhancement model (mm) .

[0061] Corresponding to the above method embodiments, this embodiment also provides a multimodal feature and deep learning-based water vapor product resolution enhancement system, including:

[0062] The data acquisition module is used to acquire the first atmospheric precipitable water fusion field generated by the fusion method, and to acquire multimodal feature data related to atmospheric precipitable water.

[0063] The modality extraction module is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field using the empirical orthogonal function method, so as to obtain common variation patterns;

[0064] The model acquisition module is used to acquire a pre-trained AutoResNet model. The AutoResNet model is a neural network model based on the coupling of an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established.

[0065] The product generation module is used to input the common change pattern into the AutoResNet model to generate a second atmospheric precipitable water product, wherein the spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field.

[0066] The specific implementation methods of each of the above modules correspond one-to-one with the steps described above, and will not be repeated here.

[0067] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above method embodiments.

[0068] Specifically, the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), or an application-specific integrated circuit (ASIC), etc. The memory may be non-volatile memory (such as a hard disk or solid-state drive) or volatile memory (such as random access memory). When the processor executes a computer program in the memory, it performs the following operations:

[0069] A first atmospheric precipitable water content fusion field generated using a fusion method is obtained; multimodal feature data related to atmospheric water vapor is obtained; common modes are extracted from the multimodal feature data and the first atmospheric precipitable water content fusion field using an empirical orthogonal function method to obtain common variation patterns; the common variation patterns are input into a pre-trained AutoResNet model to generate a second atmospheric precipitable water content product.

[0070] The electronic device can be a server, a personal computer, an embedded device, or a mobile terminal; the specific form is not limited, as long as it can execute the above program.

[0071] This embodiment also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the steps in any of the above method embodiments.

[0072] The computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), hard disk, optical disk (CD-ROM, DVD), flash memory, magnetic tape, magnetic disk or other magnetic storage device, or any other medium capable of carrying or storing program code.

[0073] When the computer program runs on the processor, it can implement the atmospheric precipitable water product resolution enhancement method of the present invention, including: acquiring a first atmospheric precipitable water fusion field and multimodal feature data, performing EOF common mode extraction, and generating a second atmospheric precipitable water product with higher spatiotemporal resolution using a pre-trained AutoResNet model.

[0074] In summary, the method of this invention first acquires multi-source water vapor data, including water vapor retrieved from BeiDou, near-infrared water vapor from Fengyun 3C occultation, and radiosonde water vapor. It then performs unified coordinate representation, time matching, elevation reduction, and necessary horizontal / vertical interpolation on various types of water vapor data. Subsequently, using radiosonde data as a reference, it calculates error indices for various types of water vapor data and establishes heterogeneity error characteristics. Based on this, it dynamically calculates the relative weights of each data source using the Helmert variance component estimation method, and uses the spherical cap harmonic function as the unified basis function for representing the regional water vapor field, achieving weighted fusion of multi-source water vapor. Furthermore, it introduces multi-modal features such as meteorological elements, topographic surface parameters, and spatiotemporal information, and uses empirical orthogonal functions to extract common variation modes. Finally, it establishes a nonlinear mapping between multi-modal features and the first atmospheric precipitable water vapor fusion field using the AutoResNet model, thereby generating a high-precision, high spatiotemporal resolution fused water vapor product. This invention takes into account the differences in the accuracy of multi-source observations, the modeling characteristics of regional spherical canopy domains, and the multimodal nonlinear coupling features. It can output high-precision, high spatiotemporal resolution regional atmospheric water vapor products, which are suitable for severe weather monitoring, regional water cycle analysis, and high-resolution water vapor inversion operations.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for enhancing the resolution of water vapor products using multimodal features and deep learning, characterized in that, include: Obtain the first atmospheric precipitable water fusion field generated using the fusion method; Acquire multimodal feature data related to atmospheric water vapor; The empirical orthogonal function method is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field to obtain common variation patterns; The common variation pattern is input into a pre-trained AutoResNet model to generate a second atmospheric precipitable water product. The spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field. The AutoResNet model is a neural network model based on an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established.

2. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 1, characterized in that, The fusion method includes: Using radiosonde atmospheric precipitable water as a reference, the relative weights of multi-source atmospheric precipitable water observation data are iteratively obtained by the Helmert variance component estimation method. The first atmospheric precipitable water field is obtained by weighting and fusing the multi-source atmospheric precipitable water observation data using the relative weights and spherical cap harmonic functions, solving for the spherical cap harmonic coefficients.

3. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 2, characterized in that, The multi-source atmospheric precipitable water observation data includes atmospheric precipitable water retrieved from BeiDou navigation satellites, atmospheric precipitable water retrieved from Fengyun near-infrared remote sensing, atmospheric precipitable water retrieved from Fengyun 3C occultation, and atmospheric precipitable water retrieved from radiosondes.

4. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 1, characterized in that, The multimodal feature data includes meteorological elements, topographic elements, surface parameters, and spatiotemporal information; wherein, the meteorological elements include one or more of precipitation, surface air pressure, surface air temperature, wind speed, and specific humidity; the topographic elements include a digital elevation model; the surface parameters include an enhanced vegetation index; and the spatiotemporal information includes spatial location and time information.

5. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 1, characterized in that, The empirical orthogonal function method is used to extract the common mode from the multimodal feature data and the first atmospheric precipitable water fusion field, including: Anomaly processing is performed on the multimodal feature data and the first atmospheric precipitable water fusion field, respectively; Construct the covariance matrix between the anomaly-processed multimodal feature data and the first atmospheric precipitable water fusion field; The covariance matrix is ​​subjected to eigenvalue decomposition to extract the principal mode space function and its corresponding time coefficient, which are used as the common variation mode.

6. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 1, characterized in that, The AutoResNet model includes an input layer, an encoding layer, a hidden representation layer, a decoding layer, and an output layer. The encoding layer is used to progressively compress the dimensionality of the input features and enhance the training stability of the deep network through residual connections. The decoding layer is used to recover the high-level semantic representation and output the atmospheric precipitable water estimate at the target spatiotemporal resolution.

7. The method for enhancing the resolution of water vapor products using multimodal features and deep learning according to claim 6, characterized in that, The AutoResNet model employs multiple convolutional layers, with the number of output channels increasing with the number of layers, and the loss function is weighted mean square error.

8. A multimodal feature and deep learning-based water vapor product resolution enhancement system, characterized in that, include: The data acquisition module is used to acquire the first atmospheric precipitable water fusion field generated by the fusion method, and to acquire multimodal feature data related to atmospheric precipitable water. The modality extraction module is used to extract common modes from the multimodal feature data and the first atmospheric precipitable water fusion field using the empirical orthogonal function method, so as to obtain common variation patterns; The model acquisition module is used to acquire a pre-trained AutoResNet model. The AutoResNet model is a neural network model based on the coupling of an autoencoder and a residual network, and a nonlinear mapping relationship between multimodal features and atmospheric precipitable water has been established. The product generation module is used to input the common change pattern into the AutoResNet model to generate a second atmospheric precipitable water product, wherein the spatiotemporal resolution of the second atmospheric precipitable water product is higher than that of the first atmospheric precipitable water fusion field.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the water vapor product resolution enhancement method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the water vapor product resolution enhancement method of any one of claims 1 to 7 using multimodal features and deep learning.

Images