Oceanic precipitation fusion method and device, system, storage medium
By eliminating false precipitation reports in the ocean region and correcting errors using a deep learning method with a hybrid Transformer architecture, combined with optimal weight calculation, the error and accuracy problems in ocean precipitation data fusion are solved, and high-precision ocean precipitation data acquisition is achieved.
Patent Information
- Application Number
- CN202610780376.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-02
- Publication Date
- 2026-08-25
AI Technical Summary
Existing technologies struggle to effectively integrate satellite and reanalysis precipitation products in marine areas lacking surface precipitation data, resulting in large errors and low accuracy. Furthermore, a single weighting strategy leads to weight uncertainty, impacting marine hydrology, meteorology, and climate research.
A categorical variable fusion algorithm is used to eliminate false precipitation reports, and a deep learning method based on a hybrid Transformer architecture is used for error correction. The optimal weights are calculated under the constraint of minimizing the root mean square error, thereby achieving high-precision ocean precipitation fusion.
By integrating error correction and multi-method optimal weighting, the accuracy of marine precipitation data was improved, the problems of weight uncertainty and disturbance of low-quality input data were solved, and high-precision marine precipitation fusion values were obtained.
Smart Images
Figure CN122634491A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of marine hydrology and meteorology, and in particular to a method, apparatus, system, and storage medium for marine precipitation fusion. Background Technology
[0002] Surface precipitation observation data for marine areas are scarce, and satellite and reanalysis precipitation products provide rich data support for research in marine hydrology, meteorology, and climate. However, satellite and reanalysis precipitation products have significant errors and uncertainties, which affect the depth and breadth of their application in marine areas and restrict research and development in marine hydrology, meteorology, and climate.
[0003] Multi-source precipitation data fusion is a powerful means to reduce precipitation errors and improve precipitation accuracy, and has become a research frontier and hot topic in the field of precipitation. Most existing methods use surface precipitation data as a reference value, correct for errors in satellite and reanalysis precipitation products, and fuse the data to produce high-precision fused precipitation estimates. The core of these methods relies on surface precipitation data from a high-density network of stations as a reference value, calculates the weights of different precipitation data, and obtains high-precision fused precipitation results. However, due to this crucial reliance, it is difficult to extend to marine areas lacking surface precipitation data.
[0004] To address the issue of precipitation data fusion in regions lacking surface data, scholars both domestically and internationally have focused on the applicability of the Triple Permutation Analysis (TC) method and the generalized triangular hat method for precipitation fusion in these regions. They have found that these two methods can effectively improve the accuracy of precipitation data in areas lacking surface data, providing a feasible technical solution for precipitation fusion in such areas.
[0005] However, using a single weighting strategy often results in significant uncertainty in the weights, making it difficult to achieve optimal weight allocation. Furthermore, the input data contains numerous false positives and large hit errors. Without false positive removal and hit error correction, precipitation data with large errors will mask or distort valid precipitation information, thereby weakening the overall effectiveness of the fusion method and hindering its ability to effectively improve the accuracy of ocean precipitation data. Summary of the Invention
[0006] This invention provides a method, apparatus, system, and storage medium for ocean precipitation fusion. It removes false precipitation reports from input data using a categorical variable fusion algorithm, then corrects the input data after removing false precipitation reports using a deep learning method based on a hybrid architecture of Swing Transformer, Vision Transformer, and Convolutional neural networks. Finally, based on the minimum root mean square error constraint, it calculates the optimal weights for different precipitation input data, thereby achieving the goal of obtaining high-precision ocean precipitation data.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for ocean precipitation fusion includes: S1. Obtain three independent precipitation products for the target ocean area; S2. In the absence of ground precipitation data as a reference value, eliminate false precipitation from the three independent precipitation products; S3. A deep learning method based on a hybrid Transformer architecture is used to perform error correction on the three independent precipitation products after removing false alarms, resulting in three error-corrected independent precipitation products. S4. Under the constraint of minimizing the root mean square error, obtain the theoretically optimal weights of the three independent precipitation products, and then perform weighted fusion to obtain a high-precision marine precipitation fusion value.
[0008] As a preferred approach, in step S3, the deep learning method based on the hybrid Transformer architecture is as follows: An architecture combining Swing Transformer, Vision Transformer, and Convolutional neural networks is constructed. The three stages of the Swing Transformer are used to capture precipitation features at convective, mesoscale, and synoptic scales, respectively. A global meteorological "token" is introduced at the network bottleneck layer. This "token" collects implicit deconstruction information about precipitation at different scales and facilitates information exchange between different "tokens," achieving multi-scale knowledge fusion from convective to synoptic scales. The Vision Transformer module is used to enhance the modeling ability of global spatial dependencies and extract global spatial features of precipitation. The Convolutional neural networks module is used to eliminate the "checkerboard" artifacts caused by the block processing of the Transformer, improving the spatial continuity and detail features of the results. Finally, the output features of the three branches are adaptively fused through an attention fusion module and then processed by a regression network to obtain three independent precipitation products after error correction.
[0009] Preferably, step S1 further includes: unifying the spatiotemporal resolution of the precipitation products.
[0010] The present invention also provides a marine precipitation fusion device, comprising: The first processing module is used to acquire three independent precipitation products for the target ocean area; The second processing module is used to eliminate false precipitation from the three independent precipitation products without the need for ground precipitation data as a reference value. The third processing module is used to perform error correction on the three independent precipitation products after removing false alarms using a deep learning method based on a hybrid Transformer architecture, so as to obtain the three independent precipitation products after error correction. The fourth processing module is used to obtain the theoretically optimal weights of the three independent precipitation products under the constraint of minimizing the root mean square error, and to obtain a high-precision marine precipitation fusion value by weighted fusion.
[0011] As a preferred approach, the deep learning method based on a hybrid Transformer architecture is as follows: An architecture combining SwingTransformer, Vision Transformer, and Convolutional neural networks is constructed. The three stages of SwingTransformer are used to capture precipitation features at convective, mesoscale, and synoptic scales, respectively. A global meteorological "token" is introduced at the network bottleneck layer to collect implicit deconstruction information of precipitation at different scales and exchange information between different "tokens," achieving multi-scale knowledge fusion from convective to synoptic scales. A Vision Transformer module is employed to enhance the modeling ability of global spatial dependencies and extract global spatial features of precipitation. A Convolutional neural networks module is used to eliminate the "checkerboard" artifacts caused by the block processing of the Transformer, improving the spatial continuity and detail features of the results. Finally, an attention fusion module adaptively fuses the output features of the three branches, and through a regression network, three independent precipitation products are obtained after error correction.
[0012] Preferably, the first processing module is also used to unify the spatiotemporal resolution of the precipitation products.
[0013] The present invention also provides a marine precipitation fusion system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs a marine precipitation fusion method when executed by the processor.
[0014] The present invention also provides a storage medium storing a computer program that executes a marine precipitation fusion method when running.
[0015] Compared with the prior art, the technical solution disclosed in this invention has the following beneficial effects: By introducing a categorical variable fusion algorithm, false positive precipitation is removed from the three independent precipitation products without the need for ground reference values. Then, a deep learning method based on a hybrid Transformer architecture is used to correct the errors in the three independent precipitation products after removing false positives. Finally, by combining the advantages of multiple weighting methods, the theoretically optimal weights for the three precipitation products are derived under the constraint of minimizing the root mean square error. Thus, the method disclosed in this invention integrates error correction and optimal weighting using multiple methods, solving the problems of weight uncertainty caused by single weighting strategies and the disturbance of fusion performance by low-quality input data in existing methods, thereby obtaining high-precision ocean precipitation fusion values. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the marine precipitation fusion method according to an embodiment of the present invention; Figure 2 This is a technical roadmap for the marine precipitation fusion method according to an embodiment of the present invention. Detailed Implementation
[0018] 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, and 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.
[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] like Figure 1 , 2 As shown, this invention provides a method for ocean precipitation fusion. It removes false positives from the input data using a categorical variable fusion algorithm, and then corrects the input data error after removing false positives using a deep learning method based on a hybrid Transformer architecture. Under the constraint of minimizing the root mean square error, the theoretically optimal weights of each precipitation product are derived, and weighted fusion is performed to obtain high-precision ocean precipitation data. Specifically, it includes: S1: Obtain three independent precipitation products for the target ocean area and unify the spatiotemporal resolution of the precipitation products; that is, data preprocessing.
[0021] S2: Using a categorical variable fusion algorithm, false precipitation is eliminated from the three independent precipitation products without the need for surface precipitation data as a reference value; that is, false precipitation is eliminated by error correction.
[0022] S3: A deep learning method based on a hybrid Transformer architecture is used to perform error correction on the three independent precipitation products that have eliminated false precipitation; that is, error correction of the input data after eliminating false precipitation.
[0023] S4: Combining the advantages of multiple weighting methods, under the constraint of minimizing root mean square error, the theoretically optimal weights of the three independent precipitation products are obtained, and the weighted fusion is used to obtain a high-precision marine precipitation fusion value; that is, optimal weighting by multiple methods to obtain the optimal marine precipitation fusion value.
[0024] As one embodiment of the present invention, the spatiotemporal resolution is unified as 0.1° spatial resolution and 1-hour temporal resolution, which are obtained by interpolation and precipitation time accumulation methods, respectively, thereby unifying the spatiotemporal resolution of the three independent precipitation products.
[0025] As one embodiment of the present invention, S2 specifically includes: Based on the principle of categorical variable fusion algorithm, each precipitation product is converted into a binary time series of rainfall / no rainfall, with a rainfall occurrence flag incremented by 1 and a no rainfall occurrence flag decremented by 1. Three independent sets of rainfall / no rainfall categorical time series are defined as follows: , , Their covariance can be obtained using the following formula: (1) (2) (3) in, Q Covariance of binary time series with different rainfall / no rainfall conditions; P This represents the truth value of a binary time series with / without rainfall. A function of the statistical properties of P; To improve the accuracy of rainfall / no-rain classification for different precipitation products; Based on equations (1) to (3), the statistical measure of positive correlation between the accuracy of each rainfall / no rainfall classification can be calculated. v The specific calculation formula is as follows: (4) (5) (6) Through positive correlation statistics Weighted fusion yields high-precision time series classification of rainfall / no rainfall. The calculation expression is as follows: (7) (8) in, ( i =1,2,3) is the corresponding number i Weighting of individual precipitation products; Three independent time series classifications of rainfall / no rainfall , , High-precision rainfall / no-rainfall classification time series Grid-by-grid comparison over time: If there exists a grid point at time t... ,and If so, then the data is classified as hit precipitation data; ,and If so, the data is classified as false precipitation data; ,and If so, the data is classified as underreported precipitation data; Since the precipitation in the missed events is 0, the above method effectively identifies and removes false precipitation from the three independent precipitation products, while retaining the true precipitation data.
[0026] As one embodiment of the present invention, S3 specifically includes: Using buoy precipitation data as a reference, and taking three independent precipitation products after removing false alarms as data input, the errors of the three independent precipitation products after removing false alarms are corrected by a deep learning method based on a hybrid Transformer architecture.
[0027] As one embodiment of the present invention, the principle of the deep learning method based on the hybrid Transformer architecture is as follows: We construct a hybrid architecture of Swing Transformer, Vision Transformer, and Convolutional neural networks. We utilize the three stages of Swing Transformer to capture precipitation characteristics at convective, mesoscale, and synoptic scales, respectively. We introduce a global meteorological "token" at the network bottleneck layer. Through the "token," we collect implicit deconstruction information of precipitation at different scales and exchange information between different "tokens" to achieve multi-scale knowledge fusion from convective to synoptic scales. The Vision Transformer module is used to enhance the modeling capability of global spatial dependencies and extract global spatial features of precipitation. The Convolutional neural networks module is used to eliminate the "chessboard" artifacts that may be generated by the Transformer due to block processing, thereby improving the spatial continuity and detailed features of the results; Finally, the output features of the three branches are adaptively fused through the attention fusion module and then passed through a regression network to obtain three independent precipitation products after error correction.
[0028] As one embodiment of the present invention, S4 specifically includes: The various weighting methods include the Triple Permutation Analysis (TC) method, the Generalized Triangular Hat method, and the weighting framework of mathematical uncertainty analysis. Their weight calculation processes are listed below: The weight calculation process for the triple permutation analysis method TC is as follows: Assuming precipitation products that exclude false positives and error corrections with unknown truth value The relationship is: (9) In the formula, and These are the intercept and the slope, respectively. This represents observational error; in this case, the covariance between each pair of precipitation products after removing falsely reported precipitation and error correction is considered. The calculation formula can be expressed as: (10) Therefore, the variance of the error between the precipitation products after different false alarms and error corrections and the true value can be obtained. The calculation formula is as follows: (11) (12) (13) In the formula, For the firsti The root mean square error of precipitation products after eliminating false alarms and error correction is calculated using its reciprocal. As a weight, we have: (14) In the formula, For the triple permutation analysis method TC in the first i Weights of precipitation products after removing false alarms and error correction; The weight calculation process of the generalized triangular hat method is as follows: For the three precipitation products after removing false alarms and error correction, the time series of each product can be represented as follows: (15) In the formula, For the first i Time series of precipitation products For the true value of precipitation products, It is the first i Errors in precipitation products; If one precipitation product is chosen as the reference value, the difference between other precipitation products can be expressed as: (16) In the formula, For precipitation products that can be selected as a reference value, Error in the time series of precipitation products as a reference value; Constructing the difference matrix Y, we have: (17) In the formula, M represents the number of each type of precipitation product; The covariance matrix of the difference sequence matrix Y can then be expressed as: (18) In the formula, S For covariance matrix, For covariance operators, ,for and The variance (i=j) or covariance (i) between them j); Introduce a The covariance matrix R, where R is a symmetric matrix: (19) In the formula, ,yes and The variance (i=j) or covariance (i) between them j), diagonal elements It is an unknown quantity to be determined; At this point, we can construct an expression relating R and S: (20) In the formula, It is a submatrix of R with dimension 2. 2; A vector of dimension 2; yes The variance; , I is 2 A unit array of 2; A vector of dimension 2; To obtain the root mean square error, we need to solve for the covariance matrix R; R has 3... (3+1) / 2 = 6 unknowns, but only 3 are known. (3-1) / 2 = 3 equations (the number of distinct elements in S), making it impossible to calculate the solution to formula (14); the remaining 3 free parameters need to be uniquely solved by defining an objective function that minimizes it. The defined objective function is expressed as: (twenty one) In the formula, ; To satisfy the condition det(R)>0, the constraint conditions of the objective function (15) can be expressed as: (twenty two) To ensure that the initial value is within the constraints, the initial value for iteration is set as follows: (twenty three) (twenty four) Minimize the objective function (21) under constraint (22) to obtain the three free parameters. The unique solution to ) is given; the other unknowns can be calculated using the following formula: (25) Calculate the root mean square error for each precipitation product point by point: (26) In the formula, For the first i The root mean square error of precipitation products after eliminating false alarms and error correction is calculated using its reciprocal. As a weight; (27) In the formula, For the generalized triangular cap method in the first i Weights of precipitation products after removing false alarms and error correction; The weight calculation process of the weighting framework in mathematical uncertainty analysis is as follows: By using the error independence between TC and precipitation products through triple permutation analysis, a QC based on quadruple permutation analysis is constructed. The QC method was used to calculate the error of precipitation products after removing false alarms and error correction, as well as the cross-correlation of errors among the products, resulting in the error matrix of precipitation products after removing false alarms and error correction. ; (28) In the formula, For covariance, For the first i ( j After removing false positives and correcting for errors, the variance of the precipitation product is removed. Based on the constructed error matrix of precipitation products after removing false alarms and error correction (Equation (28)), the weights of each precipitation product after removing false alarms and error correction are calculated: (29) In the formula, The weighting framework for mathematical uncertainty analysis in the first i Weights of precipitation products after removing false alarms and error correction; To ensure the accuracy and fairness of the results, the three weighted outcomes were considered. , and Standardize them separately so that they are all dimensionless values in the 0-1 range; Using these three different weighting methods, the initial weights of each precipitation product after removing false alarms and error correction are calculated respectively, and denoted as the first weight. j The first method i The weight of each product is (i,j=1,2,3), satisfying the weight normalization constraint: , j=1,2,3(30) To integrate the advantages of different weighting methods, a method fusion weight is introduced. , , ,satisfy: (31) This leads to the final weighting of each precipitation product. : (32) Ultimately, the weights of each precipitation product satisfy the normalization constraint: (33) Based on this, construct the objective function that minimizes the root mean square error: (34) In the formula, For the first i Precipitation data after removing false positives and error correction. T This is a reference value for surface precipitation; The objective function is a constrained convex optimization problem, with the following constraints: According to convex optimization theory, this problem has a unique global optimal solution, which can be obtained by nonnegative least squares method or sequential quadratic programming algorithm, yielding the optimal weight solution for each weighting method. The optimal fusion weights of each precipitation product are calculated using equation (32). Finally, high-precision ocean precipitation fusion values are generated. : (35) Example 2 The present invention also provides a marine precipitation fusion device, comprising: The first processing module is used to acquire three independent precipitation products for the target ocean area; The second processing module is used to eliminate false precipitation from the three independent precipitation products without the need for ground precipitation data as a reference value. The third processing module is used to perform error correction on the three independent precipitation products after removing false alarms using a deep learning method based on a hybrid Transformer architecture, so as to obtain the three independent precipitation products after error correction. The fourth processing module is used to obtain the theoretically optimal weights of the three independent precipitation products under the constraint of minimizing the root mean square error, and to obtain a high-precision marine precipitation fusion value by weighted fusion.
[0029] As one embodiment of the present invention, the deep learning method based on the hybrid Transformer architecture is as follows: A hybrid architecture of Swing Transformer, Vision Transformer, and Convolutional neural networks is constructed. The three stages of the Swing Transformer are used to capture precipitation features at convective, mesoscale, and synoptic scales, respectively. A global meteorological "token" is introduced at the network bottleneck layer. This "token" collects implicit deconstruction information of precipitation at different scales and facilitates information exchange between different "tokens," achieving multi-scale knowledge fusion from convective to synoptic scales. The Vision Transformer module is used to enhance the modeling ability of global spatial dependencies and extract global spatial features of precipitation. The Convolutional neural networks module is used to eliminate the "checkerboard" artifacts caused by the block processing of the Transformer. An attention fusion module adaptively fuses the output features of the three branches, and through a regression network, three independent precipitation products are obtained after error correction.
[0030] As one embodiment of the present invention, the first processing module is also used to unify the spatiotemporal resolution of the precipitation products.
[0031] Example 3 The present invention also provides a marine precipitation fusion system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs a marine precipitation fusion method when executed by the processor.
[0032] Example 4 The present invention also provides a storage medium storing a computer program that executes a marine precipitation fusion method when running.
[0033] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for merging marine precipitation, characterized in that, include: S1. Obtain three independent precipitation products for the target ocean area; S2. In the absence of ground precipitation data as a reference value, eliminate false precipitation from the three independent precipitation products; S3. A deep learning method based on a hybrid Transformer architecture is used to perform error correction on the three independent precipitation products after removing false alarms, resulting in three error-corrected independent precipitation products. S4. Under the constraint of minimizing the root mean square error, obtain the theoretically optimal weights of the three independent precipitation products, and then perform weighted fusion to obtain a high-precision marine precipitation fusion value.
2. The marine precipitation fusion method as described in claim 1, characterized in that, In step S3, the deep learning method based on the hybrid Transformer architecture is as follows: An architecture combining Swing Transformer, Vision Transformer, and Convolutional neural networks is constructed. The three stages of the Swing Transformer are used to capture precipitation characteristics at convective, mesoscale, and synoptic scales, respectively. A global meteorological "token" is introduced at the network bottleneck layer. This "token" collects implicit deconstruction information about precipitation at different scales and facilitates information exchange between different "tokens," achieving multi-scale knowledge fusion from convective to synoptic scales. The Vision Transformer module is employed to enhance the modeling ability of global spatial dependencies and extract global spatial features of precipitation. By utilizing the Convolutional neural networks module, the "checkerboard" artifacts that may be generated by the Transformer due to block processing are eliminated, improving the spatial continuity and detailed features of the results. Finally, through the attention fusion module, the output features of the three branches are adaptively fused and passed through the regression network to obtain three independent precipitation products after error correction.
3. The marine precipitation fusion method as described in claim 2, characterized in that, Step S1 also includes: unifying the spatiotemporal resolution of the precipitation products.
4. A marine precipitation fusion device, characterized in that, include: The first processing module is used to acquire three independent precipitation products for the target ocean area; The second processing module is used to eliminate false precipitation from the three independent precipitation products without the need for ground precipitation data as a reference value. The third processing module is used to perform error correction on the three independent precipitation products after removing false alarms using a deep learning method based on a hybrid Transformer architecture, so as to obtain the three independent precipitation products after error correction. The fourth processing module is used to obtain the theoretically optimal weights of the three independent precipitation products under the constraint of minimizing the root mean square error, and to obtain a high-precision marine precipitation fusion value by weighted fusion.
5. The marine precipitation fusion device as described in claim 4, characterized in that, The deep learning method based on the hybrid Transformer architecture is as follows: An architecture combining Swing Transformer, Vision Transformer, and Convolutional neural networks is constructed. The three stages of the Swing Transformer are used to capture precipitation characteristics at convective, mesoscale, and synoptic scales, respectively. A global meteorological "token" is introduced at the network bottleneck layer. This "token" collects implicit deconstruction information about precipitation at different scales and facilitates information exchange between different "tokens," achieving multi-scale knowledge fusion from convective to synoptic scales. A Vision Transformer module is employed to enhance the modeling ability of global spatial dependencies and extract global spatial features of precipitation. The Convolutional neural networks module is used to eliminate the "checkerboard" artifacts caused by the block processing of the Transformer; the output features of the three branches are adaptively fused through the attention fusion module, and then passed through the regression network to obtain three independent precipitation products after error correction.
6. The marine precipitation fusion device as described in claim 5, characterized in that, The first processing module is also used to unify the spatiotemporal resolution of the precipitation products.
7. A marine precipitation fusion system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the ocean precipitation fusion method as described in any one of claims 1-3 when executed by the processor.
8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the marine precipitation fusion method as described in any one of claims 1-3.