Machine learning based near real-time satellite retrieval precipitation correction method and system

Abstract

The present application relates to the technical field of data processing, disclose a near real-time satellite inversion precipitation correction method and system based on machine learning, the method comprises: tensor remodeling is carried out to the inversion precipitation data of satellite inversion section, obtains the network adaptive feature primitive of satellite inversion section;The network adaptive feature primitive is carried out nonlinear transformation, obtains the precipitation correction field of satellite inversion section;Spectral fidelity evaluation is carried out to the inversion precipitation data, to obtain the quality mask matrix of satellite inversion section;The reliability of precipitation correction field is discriminated, and the spatial kriging interpolation is carried out to the reliability grid point set discriminated, obtains the compensation correction field of satellite inversion section;The edge smoothing fusion is carried out to the reliability grid point set and compensation correction field, obtains the seamless correction field of satellite inversion section;Data packaging is carried out to the seamless correction field, obtains the correction data stream of satellite inversion section;The present application can improve the efficiency of near real-time satellite inversion precipitation correction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a near real-time satellite inversion precipitation correction method and system based on machine learning. Background Technology

[0002] Existing satellite precipitation correction technologies lack a lightweight network architecture trained with structured pruning and quantitative sensing, resulting in high resource consumption and slow computation speed during the correction process. This makes it difficult to adapt to near real-time processing scenarios and respond quickly to the correction needs of large-scale precipitation data.

[0003] Existing technologies for precipitation data correction do not provide a comprehensive assessment of the spectral fidelity of the inverted data, lack a precise quality mask matrix to support credibility assessment, and lack effective spatiotemporal compensation interpolation and edge smoothing fusion mechanisms. This makes the corrected data prone to deviations, preventing the formation of seamless and accurate correction results and affecting the reliability of data applications. Summary of the Invention

[0004] This invention provides a near real-time satellite inversion precipitation correction method and system based on machine learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a near real-time satellite inversion precipitation correction method based on machine learning, comprising: S1. Based on a lightweight correction network, tensor reshaping is performed on the inverted precipitation data of the satellite inversion section to obtain the network-adaptive feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. S2. Perform a nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion section; S3. Evaluate the spectral fidelity of the retrieved precipitation data to obtain the mass mask matrix of the satellite retrieval section; S4. Based on the mass mask matrix, the credibility of the precipitation correction field is identified, and spatiotemporal kriging interpolation is performed on the identified credibility grid set to obtain the compensation correction field of the satellite inversion section. S5. Perform edge smoothing fusion on the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion section; S6. Encapsulate the seamless correction field to obtain the correction data stream of the satellite inversion section.

[0006] In a preferred embodiment, the tensor reshaping of the inverted precipitation data of the satellite inversion section based on the lightweight correction network yields network-adapted feature primitives for the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized sensing training, including: Acquire large-scale historical precipitation data in the satellite inversion section to train a high-capacity precipitation correction architecture; The high-capacity precipitation correction architecture is structurally pruned to obtain a sparse transition architecture for the satellite inversion section; The sparse transition architecture is quantized and retrained, and gradient backpropagation is performed on the trained architecture to obtain the deployable architecture of the satellite inversion segment. Edge nodes are deployed on the deployable architecture to obtain a lightweight correction network for the satellite inversion segment; Based on the lightweight correction network, the channel dimension of the inverted precipitation data of the satellite inversion section is reconstructed to obtain the network-adaptive feature primitives of the satellite inversion section.

[0007] In a preferred embodiment, the step of reconstructing the channel dimension of the retrieved precipitation data of the satellite inversion segment based on the lightweight correction network to obtain the network-adapted feature primitives of the satellite inversion segment includes: Based on the lightweight correction network, heterogeneous data fusion is performed on the retrieved precipitation data to obtain a multi-source feature data layer for the satellite retrieval section; The feature channels in the multi-source feature data layer are repositioned to obtain the rearranged data volume of the satellite inversion segment; The rearranged data volume is resampled using grids to obtain the data array of the satellite inversion segment; The relationship between the height, width and depth dimensions of the data array is encapsulated in three dimensions to obtain the network-adaptive feature primitives of the satellite inversion segment.

[0008] In a preferred embodiment, the step of performing a nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion section includes: By filtering the network-adaptive feature primitives layer by layer, condensed feature blocks of the satellite inversion segment are obtained; Neighborhood feature correlation analysis is performed on the condensed feature blocks to obtain the contextual correlation feature field of the satellite inversion segment; Based on the lightweight correction network, the abstract features in the context-related feature field are mapped to the initial intensity distribution field of the satellite inversion segment; The initial intensity distribution field is geospatial frame aligned to obtain the precipitation correction field for the satellite inversion section.

[0009] In a preferred embodiment, the step of evaluating the spectral fidelity of the retrieved precipitation data to obtain the mass mask matrix of the satellite inversion segment includes: The spectral smoothness of the retrieved precipitation data is tested to obtain the spectral continuity label of the satellite retrieval segment; Based on the preset typical ground cover spectral range, the spectral continuity label is used to determine spectral anomalies in order to obtain the distribution map of anomaly grid points in the satellite inversion section. Based on the spectral continuity label and the anomaly grid distribution map, the quality level of the retrieved precipitation data is determined, and the quality level label allocation result of the satellite inversion segment is obtained. Based on the spatial arrangement of the satellite inversion segments, the quality level identifier allocation results are organized and filled to obtain the quality mask matrix of the satellite inversion segments.

[0010] In a preferred embodiment, the step of performing confidence assessment on the precipitation correction field based on the mass mask matrix, and performing spatiotemporal kriging interpolation on the assessed confidence grid set to obtain the compensated correction field for the satellite inversion segment, includes: Based on the mass mask matrix, the credibility of the precipitation correction field is analyzed to obtain the credibility label of the satellite inversion segment; Based on the confidence labels, the precipitation correction field is divided into grid points to obtain the high confidence grid points and grid points to be compensated in the satellite inversion section; Based on the spatial location and value of the high-confidence grid points, a spatiotemporal neighborhood search is performed on the grid points to be compensated to obtain the interpolation reference grid point group of the satellite inversion section. Based on the interpolation reference grid group, the grid points to be compensated are weighted and evaluated to obtain the reconstructed value of the satellite inversion section; Spatial reconstruction is performed on the high-confidence grid points and the reconstructed values ​​to obtain the compensation and correction field of the satellite inversion section.

[0011] In a preferred embodiment, the formulas for calculating the contribution weight and reconstructed value in the satellite inversion segment are as follows: , ; In the formula, Reference grid points in the interpolation reference grid group For the grid points to be compensated among the grid points to be compensated Contribution weight, The reference grid point With the grid points to be compensated The Euclidean distance between them The reference grid point With the grid points to be compensated The absolute time difference between them For preset space-related length parameters, This is a preset time-related length parameter. The preset spatial distance attenuation index parameter, The preset time-distance decay index parameter, Reference grid points in the mass mask matrix The credibility coefficient For a preset small positive real number, The reference grid points in the retrieved precipitation data The precipitation intensity value, For the grid points to be compensated The reconstruction value, For the grid points to be compensated The corresponding reference grid group.

[0012] In a preferred embodiment, the step of smoothly fusing the confidence grid set and the compensation correction field to obtain the seamless correction field for the satellite inversion segment includes: The boundary between the confidence grid set and the compensation correction field is distinguished to obtain the boundary region of the satellite inversion segment; The outward expansion width of the boundary region is defined as the annular transition zone of the satellite inversion segment; Two-way distance fusion is performed on the grid points within the annular transition region to obtain the transition field of the satellite inversion section; The confidence grid set, the transition field, and the compensation correction field are spatially stitched together to obtain the seamless correction field of the satellite inversion section.

[0013] In a preferred embodiment, the step of encapsulating the seamless correction field to obtain the correction data stream of the satellite inversion segment includes: The seamless correction field is flattened to obtain a linear data array for the satellite inversion section; Metadata supplementation is performed on the linear data array to obtain the metadata-enhanced data sequence of the satellite inversion segment; The metadata-enhanced data sequence is encoded and converted to obtain the binary data of the satellite inversion segment; The binary data is encapsulated using a protocol to obtain the correction data stream for the satellite inversion segment.

[0014] To address the above problems, this invention also provides a near real-time satellite inversion precipitation correction system based on machine learning, the system comprising: The feature primitive reconstruction module is used to perform tensor reshaping on the inverted precipitation data of the satellite inversion section based on a lightweight correction network, so as to obtain the network-adapted feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. The nonlinear transformation module is used to perform nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion section. The mask matrix generation module is used to evaluate the spectral fidelity of the retrieved precipitation data in order to obtain the quality mask matrix of the satellite inversion section. The compensation correction interpolation module is used to perform credibility screening on the precipitation correction field based on the mass mask matrix, and to perform spatiotemporal kriging interpolation on the screened credibility grid set to obtain the compensation correction field of the satellite inversion section. A seamless correction fusion module is used to perform edge smoothing fusion of the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion segment; The calibration data encapsulation module is used to encapsulate the seamless calibration field to obtain the calibration data stream of the satellite inversion segment.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a lightweight correction network through structured pruning and quantitative perception training to perform tensor reshaping and channel dimension reconstruction on satellite-retrieved precipitation data. Simultaneously, it realizes heterogeneous data fusion, feature channel reorganization, and three-dimensional feature encapsulation, efficiently generating network-adaptive feature primitives, significantly improving the adaptability of data to subsequent processing procedures, accelerating the overall correction process, and ensuring the achievement of near real-time processing requirements.

[0016] 2. This invention generates an accurate quality mask matrix through spectral fidelity assessment, providing a reliable basis for the credibility identification of precipitation correction fields. It combines spatiotemporal neighborhood search and weighted assessment to complete the reconstruction of grid points to be compensated, and then achieves seamless splicing of the credibility grid point set and the compensation correction field through bidirectional distance fusion in the annular transition zone. The generated correction data stream has both integrity and consistency, which greatly improves the correction accuracy and practical application value of satellite-retrieved precipitation data. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a near real-time satellite inversion precipitation correction method based on machine learning provided in an embodiment of the present invention; Figure 2A functional block diagram of a near real-time satellite inversion precipitation correction system based on machine learning provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a near real-time satellite inversion precipitation correction method based on machine learning. The execution entity of the near real-time satellite inversion precipitation correction method based on machine learning includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the near real-time satellite inversion precipitation correction method based on machine learning can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a near-real-time satellite inversion precipitation correction method based on machine learning, according to an embodiment of the present invention. In this embodiment, the near-real-time satellite inversion precipitation correction method based on machine learning includes: S1. Based on a lightweight correction network, tensor reshaping is performed on the inverted precipitation data of the satellite inversion section to obtain the network-adaptive feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. In this embodiment of the invention, the tensor reshaping of the inverted precipitation data of the satellite inversion section based on the lightweight correction network to obtain the network-adapted feature primitives of the satellite inversion section, wherein the lightweight correction network is obtained through structured pruning and quantized sensing training, including: Acquire large-scale historical precipitation data in the satellite inversion section to train a high-capacity precipitation correction architecture; The high-capacity precipitation correction architecture is structurally pruned to obtain a sparse transition architecture for the satellite inversion section; The sparse transition architecture is quantized and retrained, and gradient backpropagation is performed on the trained architecture to obtain the deployable architecture of the satellite inversion segment. Edge nodes are deployed on the deployable architecture to obtain a lightweight correction network for the satellite inversion segment; Based on the lightweight correction network, the channel dimension of the inverted precipitation data of the satellite inversion section is reconstructed to obtain the network-adaptive feature primitives of the satellite inversion section.

[0021] The lightweight correction network is used to reconstruct the channel dimension of the inverted precipitation data in the satellite inversion section, resulting in network-adapted feature primitives for the satellite inversion section, including: Based on the lightweight correction network, heterogeneous data fusion is performed on the retrieved precipitation data to obtain a multi-source feature data layer for the satellite retrieval section; The feature channels in the multi-source feature data layer are repositioned to obtain the rearranged data volume of the satellite inversion segment; The rearranged data volume is resampled using grids to obtain the data array of the satellite inversion segment; The relationship between the height, width and depth dimensions of the data array is encapsulated in three dimensions to obtain the network-adaptive feature primitives of the satellite inversion segment.

[0022] Historical precipitation data was collected from satellite inversion sections within a fixed spatial range spanning 10°×10° latitude and longitude and a time span of nearly 5 years, with 4 daily observations. The data must include complete information such as precipitation intensity, observation time, and spatial coordinates. At the same time, satellite observation records and ground station measured data were integrated to ensure data integrity. The data was organized in a unified format, with the field order clearly defined as spatial coordinates, observation time, and precipitation intensity. The data precision was retained to three decimal places. After the data was organized, it was input into a preset network architecture and iterative training was carried out. The internal connection weights of the architecture were updated in each iteration. The mean absolute error between the architecture's predicted output and the actual precipitation data was calculated every 100 iterations. When the error fluctuation does not exceed 0.01 mm / h for 3 consecutive iterations, it is considered that the error is stable, and the training is stopped. Finally, a high-capacity precipitation correction architecture was obtained.

[0023] The neural connections of each layer in the high-capacity precipitation correction architecture are traversed, and the weight values ​​of each connection are extracted and recorded one by one. The absolute value of the weight is set to 0.01 as the pruning threshold. The absolute value of the weight of each connection is compared with the threshold one by one. Connection paths and corresponding nodes with absolute weight values ​​greater than the threshold are retained, and connection paths and nodes with absolute weight values ​​less than or equal to the threshold are removed. After pruning, the architecture is tested to confirm that the input and output dimensions are consistent with those before pruning and that the core computational logic is not affected. Under the premise of maintaining the input and output dimensions of the architecture unchanged, a sparse transition architecture that retains only the connections of key functions is formed.

[0024] All weight values ​​of the sparse transition architecture are proportionally converted to 8-bit integers between 0 and 255 according to the linear scaling rule, ensuring that the relative magnitude of the weights remains unchanged before and after mapping. Iterative training is carried out again using the sorted historical precipitation data as training samples, with a fixed iteration of 5000 times. Every 500 iterations, the error between the predicted output and the actual data is checked. If the error exceeds 0.2 mm / h, the gradient adjustment amplitude is accelerated. Error changes are recorded in real time during training. The weights of each layer are adjusted backward from the output layer to the input layer through the gradient backpropagation mechanism, so that the error continues to converge towards the target range, ensuring that the prediction error of the quantized architecture is always controlled within 0.2 mm / h. After training, a simplified and deployable architecture adapted to the operating environment is obtained.

[0025] Edge computing nodes with a 4-core CPU, 8GB of memory or higher, and a 64-bit Linux CentOS 7 or later operating environment were selected for satellite data receiving. The weight parameters of the deployable architecture, network structure configuration files, data reading interfaces, and computation scheduling scripts were packaged into a program package adapted to the operating system of the node. The program package was transmitted to the edge node via TCP / IP protocol. After the transmission was completed, the loading program in the node was started to load the program package into the computing unit. After loading, test data was input for verification to confirm that the architecture could output calculation results normally and the response time was controlled within 1 second. After deployment, a lightweight correction network capable of quickly processing satellite inversion precipitation data was formed.

[0026] After receiving the inverted precipitation data from the satellite inversion section, the lightweight correction network categorizes and splits the feature channels according to data attributes. Spatial height-related data is divided into 3 channels, spatial width-related data into 3 channels, and precipitation intensity and observation duration are each divided into 2 channels, for a total of 10 feature channels. These feature channels are arranged and recombined in the order of spatial height channel – spatial width channel – precipitation intensity channel – observation duration channel. The data of each channel after recombination is normalized by subtracting the minimum value from the maximum value of each channel to obtain the difference. Then, each data value is subtracted from the minimum value of that channel and divided by the difference to obtain normalized data in the range of 0-1. Finally, according to the definition of spatial vertical resolution corresponding to the data in the height dimension, spatial horizontal resolution corresponding to the width dimension, and the total number of feature channels after splitting and recombining in the channel number dimension, all channel data are integrated in the form of a three-dimensional array of height × width × number of channels. The product of the three is consistent with the total information content of the original data, and finally, the network-adaptive feature primitive is obtained.

[0027] After receiving the inverted precipitation data from the satellite inversion section, the lightweight correction network integrates three types of heterogeneous data: precipitation intensity data observed by the satellite, observation timestamp data, and spatial latitude and longitude data, as well as precipitation data measured by ground stations. First, all data are uniformly converted into binary format, spatial latitude and longitude data are aligned to a grid resolution of 0.1°×0.1°, and timestamps are uniformly calibrated to UTC minute level. Then, the data is divided into categories of "precipitation attribute - spatial information - temporal information", with each type of data as an independent feature channel. All channels are superimposed in sequence to form a multi-source feature data layer containing multi-dimensional information.

[0028] For the feature channels that have been divided in the multi-source feature data layer, the channel positions are readjusted according to the preset processing logic order. Specifically, the channels are arranged in the order of "longitude channel - latitude channel - observation time channel - satellite precipitation intensity channel - ground measured precipitation channel". The spatial arrangement order of the original grid data is kept unchanged within each channel, and the channels are continuously connected through data address offsets to form a rearranged data body with orderly channel positions and well-organized data.

[0029] The nearest neighbor interpolation method is used to resample the rearranged data volume. The preset target grid resolution is 0.05°×0.05°. For each target grid point, the spatial distance between it and the original grid points is calculated. The nearest original grid point data with a distance of less than 0.03° is directly assigned to the target grid point. If there are no original grid points within 0.03° of the target grid point, the average of the four nearest original grid points is taken as the value of the target grid point. After resampling, all grid point data are arranged in spatial latitude and longitude order to form channel data in the form of a two-dimensional matrix. The two-dimensional matrices of multiple channels are combined in sequence to obtain a three-dimensional data array.

[0030] The number of grid points in the latitudinal direction corresponding to the height dimension of the data array is clearly defined. Calculated according to the target resolution, a 10° latitude span corresponds to 200 grid points. The number of grid points in the longitudinal direction corresponding to the width dimension corresponds to 300 grid points in the longitude span. The total number of feature channels corresponds to 5. The data array is encapsulated according to a three-dimensional structure of "height × width × depth". The size information of each dimension is bound to the data content to ensure that the dimensions match when the data is stored and retrieved. During the encapsulation process, a data verification mechanism is used to confirm that no grid point data is lost or misaligned. Finally, a network-adaptive feature primitive that meets the input requirements of a lightweight correction network is formed.

[0031] The beneficial effects are as follows: by acquiring large-scale historical precipitation data from satellite inversion sections to train and form a high-capacity precipitation correction architecture, and after structured pruning to reduce redundant connections, quantitative perception retraining, and gradient backpropagation to optimize parameters, a lightweight correction network is deployed at the edge nodes. This effectively reduces resource consumption and improves operational efficiency during computer data processing. Subsequently, using this network to carry out channel dimension reconstruction operations such as heterogeneous data fusion of inverted precipitation data, feature channel location reorganization, grid resampling, and 3D feature encapsulation, it is possible to achieve efficient integration and normalization of multi-source precipitation-related information. The generated network-adaptive feature primitives have good network adaptability and information integrity, providing high-quality data support for subsequent satellite inversion precipitation correction processes, further improving the accuracy and coherence of overall data processing, and meeting the needs of efficient, standardized, and adaptable computer information processing.

[0032] S2. Perform a nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion section; In this embodiment of the invention, the step of performing a nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion segment includes: By filtering the network-adaptive feature primitives layer by layer, condensed feature blocks of the satellite inversion segment are obtained; Neighborhood feature correlation analysis is performed on the condensed feature blocks to obtain the contextual correlation feature field of the satellite inversion segment; Based on the lightweight correction network, the abstract features in the context-related feature field are mapped to the initial intensity distribution field of the satellite inversion segment; The initial intensity distribution field is geospatial frame aligned to obtain the precipitation correction field for the satellite inversion section.

[0033] The three-dimensional array of network-adaptive feature primitives is processed sequentially according to the hierarchical structure corresponding to the height, width, and depth dimensions. Each level of the three-dimensional array corresponds to a dimension of feature information. The threshold value of the feature response value for each level is explicitly set to 0.3. The response values ​​of all feature elements in the level are read one by one in row priority order. The relationship between the response value of each element and the threshold value is compared one by one. Elements with response values ​​higher than the threshold value are extracted. At the same time, the height index, width index, and depth index of each retained element in the three-dimensional array are accurately recorded to preserve the original spatial position information. Elements with response values ​​lower than or equal to the threshold value are directly removed. After the filtering operation of all levels is completed, the retained feature elements of each level are integrated according to their index positions according to the hierarchical order of the original three-dimensional array. During the integration process, the continuity and uniqueness of the element indexes are verified to ensure that no elements are omitted or repeated. Finally, a condensed feature block with simplified dimensions and concentrated core features is formed.

[0034] Taking each feature element in the condensed feature block as the center, a 3×3×3 neighborhood range is defined. Specifically, one neighboring element is selected above and below the center element in the height dimension, one neighboring element is selected to the left and right in the width dimension, and one neighboring element is selected before and after the depth dimension. This range completely covers the neighboring elements in the spatial height, width, and depth dimensions, totaling 26 neighboring elements. The feature value difference between the center element and each neighboring element is calculated one by one. The calculation method is to subtract the feature value of the center element from the feature value of the neighboring element and take the absolute value. When the difference value in the absolute value form is less than 0.1, it is determined that there is a correlation between the center element and the neighboring element. The index information of the neighboring element is associated with the difference value in the form of key-value pairs and attached to the attribute field of the center element. After all feature elements have completed the neighborhood association analysis, all feature elements carrying association information are integrated in strict accordance with the spatial coordinate arrangement order of the original condensed feature block to form a context association feature field containing complete association information between elements.

[0035] The lightweight correction network is pre-configured with feature mapping rules, which are established through training on large-scale historical precipitation data of satellite inversion sections. During training, various abstract features in the historical precipitation data are stored one-to-one with the corresponding measured precipitation intensity values ​​to form a fixed mapping table. The rule clarifies the unique correspondence between multi-dimensional abstract correlation information and precipitation intensity values ​​in the context-related feature field. The network traverses the spatial height, width, and depth dimensions of the context-related feature field in sequence, reads each abstract feature one by one, and searches for the unique precipitation intensity quantization value corresponding to the abstract feature in the pre-set mapping table. The abstract feature is then accurately converted into the corresponding precipitation intensity quantization value. All converted precipitation intensity quantization values ​​are arranged strictly according to the original spatial dimension distribution order of their corresponding abstract features in the context-related feature field to form an initial intensity distribution field that can completely reflect the regional precipitation intensity distribution.

[0036] The WGS84 coordinate system was explicitly set as the standard geospatial framework. The geodetic datum of this coordinate system is WGS84, with a semi-major axis of 6,378,137 meters and an flattening of 1 / 298.257223563. A uniform grid resolution of 0.05° × 0.05° was adopted. The latitude and longitude coordinates and intensity values ​​of each grid point in the initial intensity distribution field were read one by one. The original latitude and longitude coordinates were converted to plane coordinates in the WGS84 coordinate system through a coordinate transformation process. Six decimal places were retained during the transformation to ensure accuracy. Based on the transformed plane coordinates… The grid arrangement is adjusted from left to right and from top to bottom to fully align with the grid of the standard geospatial frame. At the same time, the intensity value of each grid point is checked one by one to see if it is within the reasonable precipitation intensity range of 0-100 mm / h. If the value is less than 0, it is directly corrected to 0; if the value is greater than 100 mm / h, it is corrected to 100 mm / h. After correction, all values ​​are checked again to confirm that they are all within the reasonable range. Finally, the coordinates of all grid points are checked to ensure that the deviation from the standard frame grid does not exceed 0.0001°, forming a precipitation correction field that is perfectly matched with the geospatial standard.

[0037] The beneficial effects include the ability to perform layer-by-layer screening of network-adaptive feature primitives, accurately condensing core features and eliminating redundant information. Subsequent analysis of neighborhood feature correlations can fully explore the contextual correlation information between features, forming a contextually correlated feature field rich in correlation logic. Relying on a lightweight correction network, abstract features are mapped to an initial intensity distribution field, realizing the transformation of abstract information into a concrete precipitation intensity distribution. Then, geospatial framework alignment ensures that the data is fully adapted to the standard spatial system. The entire process achieves precision in feature processing, completeness in correlation information, and standardization in data distribution. The generated precipitation correction field has the characteristics of concentrated information, close correlation, and spatial compliance, providing high-quality and highly adaptable basic data support for the subsequent correction process of satellite-retrieved precipitation data, meeting the needs of precision, standardization, and efficient adaptation in computer data processing.

[0038] S3. Evaluate the spectral fidelity of the retrieved precipitation data to obtain the mass mask matrix of the satellite retrieval section; In this embodiment of the invention, the step of evaluating the spectral fidelity of the retrieved precipitation data to obtain the mass mask matrix of the satellite inversion segment includes: The spectral smoothness of the retrieved precipitation data is tested to obtain the spectral continuity label of the satellite retrieval segment; Based on the preset typical ground cover spectral range, the spectral continuity label is used to determine spectral anomalies in order to obtain the distribution map of anomaly grid points in the satellite inversion section. Based on the spectral continuity label and the anomaly grid distribution map, the quality level of the retrieved precipitation data is determined, and the quality level label allocation result of the satellite inversion segment is obtained. Based on the spatial arrangement of the satellite inversion segments, the quality level identifier allocation results are organized and filled to obtain the quality mask matrix of the satellite inversion segments.

[0039] The retrieved precipitation data for the satellite inversion section is traversed grid by grid point in order of longitude from west to east and latitude from north to south. During the traversal, the grid points are sorted sequentially according to their latitude and longitude coordinates, ensuring that no valid grid point is missed. The spectral reflectance data of each grid point is extracted and retained to four decimal places. After data extraction, the corresponding latitude and longitude coordinates of the grid points are recorded in real time to establish a correlation. The difference in spectral reflectance between the current grid point and its right-hand adjacent grid point is calculated by subtracting the spectral reflectance data of the current grid point from the spectral reflectance data of the right-hand grid point, and then taking the absolute value of the difference. Simultaneously, the difference in spectral reflectance between the current grid point and its lower adjacent grid point is calculated by subtracting the spectral reflectance data of the current grid point from the spectral reflectance data of the lower grid point, and again taking the absolute value of the difference. A threshold of 0.02 for the absolute value of the difference was explicitly set. The absolute values ​​of the two differences were compared to this threshold. If both absolute values ​​were less than the threshold, the spectral data for that grid point was considered smooth and continuous, and a spectral continuity label "1" was assigned to it. This label was then bound to the grid point coordinates for storage. If either absolute value was greater than or equal to the threshold, the spectral data for that grid point was considered discontinuous, and a spectral continuity label "0" was assigned to it. The label was also bound to the coordinates. After all grid points were processed, all grid point data with bound labels were aggregated to form a set of spectral continuity labels that corresponded one-to-one with the grid point locations in the retrieved precipitation data, ensuring that each grid point had a unique spectral continuity label.

[0040] By classifying and statistically analyzing nearly five years of measured land cover data within the satellite inversion section, measured spectral reflectance data were collected for four core land cover types: vegetation, water bodies, bare land, and buildings. After removing extreme outliers from each data set, the minimum and maximum values ​​were used to establish a fixed spectral range for each land cover type. This range covers 99% of the measured spectral data for the corresponding land cover type. The spectral continuity label and corresponding spectral reflectance data for each grid point were read one by one. The land cover type corresponding to each grid point was determined by matching the grid point's latitude and longitude coordinates with the land cover type distribution map of the satellite inversion section. The spectral reflectance data of this grid point was precisely compared with the preset spectral range for the corresponding land cover type. If the spectral reflectance data was lower than the minimum or higher than the maximum value of the range, regardless of its spectral continuity label value, the grid point was determined to be an anomaly. Prepare a blank spatial distribution map with the same latitude and longitude range and grid resolution as the satellite inversion section. According to the spatial coordinate arrangement of the inverted precipitation data, mark the latitude and longitude coordinates of all anomalous grid points on the blank spatial distribution map with specific labels. When marking, ensure that the coordinate position deviation does not exceed 0.0001°. After all anomalous grid points are marked, the anomalous grid point distribution map of the satellite inversion section is formed.

[0041] The quality level is clearly divided into three levels, where "1" represents high confidence, "2" represents medium confidence, and "3" represents low confidence. Each level has a unique quality level description. The spectral continuity label of each grid point is retrieved one by one, and the location record of that grid point in the anomaly grid point distribution map is also checked for double verification. If the spectral continuity label of a grid point is "1" and there is no corresponding coordinate mark in the anomaly grid point distribution map, the grid point is determined to have an excellent quality level and is assigned a quality level label of "1". If the spectral continuity label of a grid point is "1" but there is a corresponding coordinate mark in the anomaly grid point distribution map, or if the spectral continuity label of a grid point is "0" but there is no corresponding coordinate mark in the anomaly grid point distribution map, the grid point is determined to have a medium quality level and is assigned a quality level label of "2". If the spectral continuity label of a grid point is "0" but there is a corresponding coordinate mark in the anomaly grid point distribution map, the grid point is determined to have a poor quality level and is assigned a quality level label of "3". The grade determination result of each grid point is bound to the grid point's latitude and longitude coordinates and stored. After all grid points have completed the grade determination, the grade label data of all bound coordinates are summarized to obtain the quality grade label allocation result of the satellite inversion section. After summarizing, the total number of grade labels is consistent with the total number of grid points to ensure that no determination is missed.

[0042] Based on the grid arrangement of the WGS84 coordinate system used in the satellite inversion section, with WGS84 as the geodetic reference surface, a semi-major axis of the ellipsoid of 6,378,137 meters, and an flattening of 1 / 298.257223563, the number of rows in the quality mask matrix strictly corresponds to the number of grid points in the latitudinal direction, and the number of columns strictly corresponds to the number of grid points in the longitudinal direction. Each element of the matrix corresponds one-to-one with a single grid point. Following the order of longitude from west to east and latitude from north to south, consistent with the grid traversal of the inverted precipitation data, the quality grade label of each grid point in the quality grade label allocation result is filled into the corresponding row and column positions of the matrix. During the filling process, after every 100 grid points are filled with their level identifiers, a batch verification is performed. The corresponding matrix row and column numbers are calculated using the latitude and longitude coordinates of the grid points, and the calculation results are compared with the actual filling positions. After all filling is completed, an overall verification is performed to confirm that the number of rows and columns of the matrix completely matches the number of grid points in the latitude and longitude directions, and that there are no empty values, duplicate filling values, or misaligned filling values ​​in the matrix. Finally, a mass mask matrix is ​​formed that completely matches the spatial distribution of the satellite inversion section.

[0043] The beneficial effects include conducting spectral smoothness detection on inverted precipitation data, accurately capturing the spectral continuity features of the data and generating corresponding labels, and judging spectral anomalies by combining preset typical land cover spectral ranges. This can accurately identify anomalous grid points in the inverted data and form distribution maps. Based on the spectral continuity labels and the anomalous grid point distribution maps, quality level determination can be performed, enabling reasonable division and accurate labeling of the quality of inverted precipitation data. Then, the quality level labeling allocation results are organized and filled according to the spatial arrangement of satellite inversion segments. The resulting quality mask matrix is ​​highly compatible with the regional spatial distribution, providing reliable data support for the credibility identification of subsequent precipitation correction fields. This improves the accuracy and standardization of quality assessment in computer data processing, meeting the needs of precision and adaptability in computer information processing.

[0044] S4. Based on the mass mask matrix, the credibility of the precipitation correction field is identified, and spatiotemporal kriging interpolation is performed on the identified credibility grid set to obtain the compensation correction field of the satellite inversion section. In this embodiment of the invention, the step of performing confidence assessment on the precipitation correction field based on the mass mask matrix, and performing spatiotemporal kriging interpolation on the assessed confidence grid set to obtain the compensation correction field for the satellite inversion segment, includes: Based on the mass mask matrix, the credibility of the precipitation correction field is analyzed to obtain the credibility label of the satellite inversion segment; Based on the confidence labels, the precipitation correction field is divided into grid points to obtain the high confidence grid points and grid points to be compensated in the satellite inversion section; Based on the spatial location and value of the high-confidence grid points, a spatiotemporal neighborhood search is performed on the grid points to be compensated to obtain the interpolation reference grid point group of the satellite inversion section. Based on the interpolation reference grid group, the grid points to be compensated are weighted and evaluated to obtain the reconstructed value of the satellite inversion section; Spatial reconstruction is performed on the high-confidence grid points and the reconstructed values ​​to obtain the compensation and correction field of the satellite inversion section.

[0045] The formulas for calculating the contribution weight and reconstructed value in the satellite inversion segment are as follows: , ; In the formula, Reference grid points in the interpolation reference grid group For the grid points to be compensated among the grid points to be compensated Contribution weight, The reference grid point With the grid points to be compensated The Euclidean distance between them The reference grid point With the grid points to be compensated The absolute time difference between them For preset space-related length parameters, This is a preset time-related length parameter. The preset spatial distance attenuation index parameter, The preset time-distance decay index parameter, Reference grid points in the mass mask matrix The credibility coefficient For a preset small positive real number, The reference grid points in the retrieved precipitation data The precipitation intensity value, For the grid points to be compensated The reconstruction value, For the grid points to be compensated The corresponding reference grid group.

[0046] The mass mask matrix and precipitation correction field of the satellite inversion section are read, and the grid positions of the two are matched one by one in the order of latitude and longitude from west to east and latitude from north to south. The mass level label of each grid point in the mass mask matrix is ​​associated with the corresponding grid point in the precipitation correction field. The mass level label "1" is set to correspond to the confidence label "high", the label "2" to the confidence label "medium", and the label "3" to the confidence label "low". The mass level label of each grid point is converted into the corresponding confidence label one by one. During the conversion process, the consistency of grid point coordinates is checked to ensure that there are no mismatches. Finally, the confidence label that completely corresponds to the grid position of the precipitation correction field is obtained.

[0047] All grid points in the precipitation correction field are classified according to the type of confidence label. Grid points with a confidence label of "high" are defined as high-confidence grid points, and their original precipitation intensity values ​​and latitude and longitude coordinates are directly retained. Grid points with confidence labels of "medium" and "low" are uniformly defined as grid points to be compensated, and their latitude and longitude coordinates are recorded and marked as pending filling. During the classification process, the number of high-confidence grid points and grid points to be compensated are counted to ensure that the total number is consistent with the total number of grid points in the precipitation correction field. Finally, the high-confidence grid points and grid points to be compensated in the satellite inversion section are obtained.

[0048] Centered on the latitude and longitude coordinates of the grid point to be compensated, the spatial search range is set to a rectangular area of ​​0.1° × 0.1° around it, and the temporal search range is the time interval of 1 hour before and after the observation time of the grid point to be compensated. Grid points that meet both spatial and temporal range conditions are selected from the high-confidence grid points. During the selection, the latitude and longitude difference between the grid point to be compensated and the high-confidence grid points is calculated first to confirm that the spatial distance is within the set range, and then the absolute difference between the observation times of the two is calculated to confirm that the time difference is within the set time interval. All high-confidence grid points that meet the conditions are collected to form an interpolation reference grid point group exclusive to the grid point to be compensated. Each grid point to be compensated corresponds to a unique interpolation reference grid point group.

[0049] For each interpolation reference grid point group corresponding to a grid point to be compensated, the spatial straight-line distance between the reference grid point and the grid point to be compensated is first calculated. The closer the distance, the greater the spatial influence of the reference grid point. Then, the absolute difference in observation time between the two is calculated. The smaller the time difference, the greater the temporal influence of the reference grid point. At the same time, the quality level identifier corresponding to the reference grid point in the quality mask matrix is ​​extracted, and the confidence coefficient of each reference grid point is determined according to the standard that the identifier "1" corresponds to a confidence coefficient of 1.0. The spatial influence, temporal influence, and confidence coefficient are integrated in a product form to obtain the comprehensive contribution weight of each reference grid point to the grid point to be compensated. Then, the precipitation intensity value of each reference grid point is multiplied by its corresponding comprehensive contribution weight, the sum is divided by the sum of the comprehensive contribution weights of all reference grid points, and the reconstructed value of the grid point to be compensated is obtained. Each grid point to be compensated obtains a unique reconstructed value through this method.

[0050] Using the WGS84 coordinate system of the satellite inversion section as the spatial framework, a reconstruction benchmark is established with a grid resolution of 0.05°×0.05°. First, all high-confidence grid points are filled into their corresponding spatial locations according to their latitude and longitude coordinates, retaining their original precipitation intensity values. Then, the reconstructed values ​​of each grid point to be compensated are filled into their corresponding spatial locations according to their latitude and longitude coordinates. During filling, it is ensured that the units of the reconstructed values ​​are consistent with those of the high-confidence grid point values. After filling, the numerical range of all grid points is verified to ensure that they are within a reasonable range of 0-100 mm / h, ultimately forming a spatially continuous and numerically complete compensation and correction field.

[0051] The spatial distance between the reference grid point and the grid point to be compensated is obtained by extracting their latitude and longitude coordinates and calculating them according to the rules of geographic spatial distance. This directly corresponds to the actual spatial position relationship of the grid points within the satellite inversion segment.

[0052] The absolute time difference between the reference grid point and the grid point to be compensated is taken from the timestamps recorded in the observation data of both, and is obtained by subtracting the time values ​​and taking the absolute value, reflecting the interval between the observation times of the two.

[0053] Spatial and temporal correlation length parameters are determined based on statistical analysis of large-scale historical precipitation data from satellite inversion sections. The statistical process covers precipitation distribution characteristics across different spatial ranges and time periods to ensure that the parameters conform to the spatiotemporal patterns of the region.

[0054] The spatial distance decay index parameter and the temporal distance decay index parameter were established through the training process of historical precipitation data. During training, the weight allocation effect under different indices was compared, and finally the value that minimizes the deviation between the interpolation result and the measured data was selected.

[0055] The confidence coefficient of the reference grid point is obtained by converting the quality level label corresponding to the reference grid point in the quality mask matrix. The higher the quality level label, the larger the confidence coefficient value, which is directly related to the grid point quality assessment result.

[0056] The precipitation intensity value of the reference grid point is the original precipitation intensity data recorded at that reference grid point in the inverted precipitation data, which is directly extracted and used after the previous data processing.

[0057] Small positive real numbers are preset fixed values. They are taken as extremely small values ​​to ensure that the denominator is not zero during the calculation process. They do not affect the reasonableness of the final result and are only used to avoid calculation errors.

[0058] The significance of the first formula is to determine the contribution weight of each reference grid point in the interpolation reference grid group to the corresponding grid point to be compensated. By comprehensively considering the spatial distance and time interval between the reference grid point and the grid point to be compensated, as well as the reliability of the reference grid point itself, the reference grid point with a closer spatial distance, shorter time interval, and higher reliability will have a greater contribution weight to the grid point to be compensated, thus ensuring that the weight allocation can truly reflect the influence of the reference grid point.

[0059] The significance of the second formula is to calculate the reconstructed value of the grid point to be compensated. By multiplying the precipitation intensity value of each reference grid point with its corresponding contribution weight and summing the results, and then dividing by the sum of the contribution weights of all reference grid points and the sum of small positive real numbers, a reasonable precipitation intensity value for the grid point to be compensated is obtained. This value takes into account both the information of the surrounding high-confidence grid points and the spatiotemporal correlation and confidence, ensuring that the reconstructed value can be integrated into the precipitation correction field to form a continuous and accurate compensation correction field.

[0060] The beneficial effects are as follows: Based on the mass mask matrix, the precipitation correction field is analyzed for reliability and gridded, accurately distinguishing between high-reliability grid points and grid points to be compensated, providing a clear target for subsequent interpolation compensation. Spatiotemporal neighborhood searches are conducted based on the spatial location and values ​​of high-reliability grid points, ensuring strong correlation and reliability of the interpolation reference grid point group. The reconstructed values ​​generated through weighted evaluation conform to the regional precipitation distribution patterns. Finally, spatial reconstruction integrates the high-reliability grid points and reconstructed values, forming a compensation correction field that is both complete and reasonable, effectively filling data gaps in grid points to be compensated, improving the accuracy and continuity of satellite-retrieved precipitation data processing, and meeting the targeting, correlation, and completeness requirements of computer data processing.

[0061] S5. Perform edge smoothing fusion on the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion section; In this embodiment of the invention, the step of performing edge smoothing fusion of the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion segment includes: The boundary between the confidence grid set and the compensation correction field is distinguished to obtain the boundary region of the satellite inversion segment; The outward expansion width of the boundary region is defined as the annular transition zone of the satellite inversion segment; Two-way distance fusion is performed on the grid points within the annular transition region to obtain the transition field of the satellite inversion section; The confidence grid set, the transition field, and the compensation correction field are spatially stitched together to obtain the seamless correction field of the satellite inversion section.

[0062] All grid data from the confidence grid set and the compensation correction field are read. Both are based on the WGS84 coordinate system and a grid resolution of 0.05°×0.05°. The grids are traversed one by one in the order of longitude from west to east and latitude from north to south. The classification of each grid point is determined by its latitude and longitude coordinates. The criteria for judgment are that grid points belonging to the confidence grid set are marked as "confidence class" and grid points belonging to the compensation correction field are marked as "compensation class". The classification of the four adjacent grid points above, below and to the left and right of each grid point is further checked. If a grid point belongs to a different classification than its adjacent grid points, the grid point and its adjacent grid points are marked as boundary grid points. All boundary grid points are summarized and arranged in order according to their latitude and longitude coordinates to form the boundary area of ​​the satellite inversion segment, ensuring that the boundary area completely covers the adjacent boundary of the two types of grid points.

[0063] Using the latitude and longitude coordinates of the outermost grid point in the boundary region as a reference, the width of the outward expansion is set to 0.02°. The expansion direction is simultaneously towards both the confidence grid point set and the compensation correction field. The resulting annular region is the annular transition zone of the satellite inversion section. The latitude and longitude range of the annular transition zone is determined by adding or subtracting 0.02° from the westernmost, easternmost, northernmost, and southernmost coordinates of the boundary region, respectively, to ensure that the transition zone completely encloses the boundary region. Furthermore, the grid point resolution within the transition zone remains at 0.05°×0.05°, consistent with the grid point density of the confidence grid point set and the compensation correction field, thus avoiding data dimension mismatch.

[0064] For each grid point within the annular transition zone, a two-way distance fusion calculation is performed individually. First, the spatial straight-line distance from the grid point to the nearest grid point on the boundary of the confidence grid point set is determined, and the actual geographic distance is calculated using the difference in latitude and longitude coordinates. Then, the spatial straight-line distance from the grid point to the nearest grid point on the boundary of the compensation and correction field is determined, and the calculation method is the same as the former. The sum of the two types of distances is used as the total distance base. The proportion of the distance from the grid point to the boundary of the confidence grid point set to the total distance base is used as the fusion weight of the compensation and correction field value, and the proportion of the distance to the boundary of the compensation and correction field to the total distance base is used as the fusion weight of the confidence grid point set value. The confidence grid point set reference value corresponding to the grid point is multiplied by its corresponding weight, and then the compensation and correction field reference value is multiplied by its corresponding weight. The sum of the two is used to obtain the fusion value of the grid point. After the fusion calculation of all grid points within the annular transition zone is completed, they are integrated in order of their latitude and longitude coordinates to form the transition field of the satellite inversion segment.

[0065] Using the latitude and longitude grid of the WGS84 coordinate system as the stitching reference, spatial stitching operations are carried out in the order of longitude from west to east and latitude from north to south. First, all grid points of the confidence grid set are accurately filled into the corresponding grid positions according to their latitude and longitude coordinates, retaining their original precipitation intensity values. Then, the fused values ​​of each grid point in the transition field are filled into the grid positions corresponding to the annular transition zone, ensuring that the transition zone values ​​are seamlessly connected with the confidence grid set values. Finally, all grid point values ​​of the compensation correction field are filled into the remaining grid positions. During the filling process, it is checked in real time whether each grid position is filled with only a unique value, without overlapping filling or gaps. At the same time, it is checked that the numerical difference between adjacent grid points does not exceed 0.5 mm / h, ensuring that the overall data is continuous and smooth, and finally forming a seamless correction field covering the entire range of the satellite inversion section with no abrupt changes in values.

[0066] The beneficial effects are as follows: By distinguishing the boundaries between the confidence grid set and the compensation correction field, the boundary region between the two is accurately located, providing a clear target for subsequent smooth fusion. A ring-shaped transition zone, extending outward from the boundary region, establishes a buffer space connecting the two, avoiding data abrupt changes during the fusion process. Two-way distance fusion is performed on the grid points within the ring-shaped transition zone, fully considering the numerical characteristics of both the confidence grid set and the compensation correction field, ensuring a natural transition in the data within the transition zone. The generated transition field effectively connects the two types of data. Spatially stitching the confidence grid set, transition field, and compensation correction field achieves complete coverage of precipitation data in the satellite inversion section. The resulting seamless correction field features data continuity, smooth edges, and overall consistency, improving the integrity and usability of the satellite inversion precipitation correction data and providing high-quality foundational support for subsequent data applications.

[0067] S6. Encapsulate the seamless correction field to obtain the correction data stream of the satellite inversion section.

[0068] In this embodiment of the invention, the step of encapsulating the seamless correction field to obtain the correction data stream of the satellite inversion segment includes: The seamless correction field is flattened to obtain a linear data array for the satellite inversion section; Metadata supplementation is performed on the linear data array to obtain the metadata-enhanced data sequence of the satellite inversion segment; The metadata-enhanced data sequence is encoded and converted to obtain the binary data of the satellite inversion segment; The binary data is encapsulated using a protocol to obtain the correction data stream for the satellite inversion segment.

[0069] The seamless calibration field of the satellite inversion section is read. This data is based on the WGS84 coordinate system and a grid resolution of 0.05°×0.05°. It has a three-dimensional structure of height, width and depth. The data is traversed one by one in the order of "height dimension first, width dimension second, and depth dimension last". The precipitation intensity value of each grid point is extracted and three decimal places are retained. All values ​​are arranged in the order of traversal to form a one-dimensional continuous linear data array. During the arrangement process, the binding and verification between grid index and value are used to ensure that the element order of the linear data array corresponds one-to-one with the grid spatial position of the seamless calibration field, without misalignment or omission.

[0070] The core information contained in the metadata is clearly defined: the latitude and longitude range of the satellite inversion section, i.e., the specific values ​​of west longitude, east longitude, south latitude, and north latitude; the grid resolution; the data acquisition start and end times accurate to the second; the precipitation intensity value in mm / h; the data accuracy level is Level 1; and the data check code is a CRC32 check value calculated based on all values ​​of the linear data array. This metadata information is added to the beginning position of the linear data array in a fixed order of "latitude and longitude range - grid resolution - acquisition time - unit - accuracy level - check code". The metadata and the linear data array are separated by a fixed character "|", forming a metadata-enhanced data sequence containing the data body and auxiliary information, ensuring that the metadata and the data body are unambiguous in association.

[0071] The metadata-enhanced data sequence is encoded and converted using the UTF-8 encoding standard. Each element in the sequence is read one by one. Textual information in the metadata, including latitude and longitude range descriptions and units, is directly converted into corresponding byte data according to the UTF-8 encoding rules. Numerical information in the linear data array is first converted into decimal strings with three decimal places, and then converted into byte data according to the UTF-8 encoding rules. The encoding length of each element is recorded during the conversion process to ensure the integrity and consistency of the data before and after encoding. All converted byte data are concatenated continuously in the original sequence order to form uninterrupted binary data.

[0072] The binary data is encapsulated using the application layer encapsulation format of the TCP / IP protocol. The encapsulation structure is as follows: a 4-byte protocol identifier, which is the ASCII code byte corresponding to the fixed "PRCP"; a 4-byte data length field, which records the total number of bytes of binary data and stores it in big-endian byte order; the binary data body; and a 4-byte check field, which is the first 4 bytes of the MD5 checksum calculated based on the binary data body. After these fields are concatenated in order, a TCP / IP protocol header is added, which includes necessary fields such as source port, destination port, and sequence number, forming a corrected data stream that conforms to network transmission standards. After encapsulation, the total length of the verification data stream is consistent with the sum of the lengths of each field, ensuring that the data stream can be accurately identified, parsed, and verified by the receiving end.

[0073] The beneficial effects include data flattening of the seamless correction field, transforming three-dimensional structured data into a one-dimensional linear data array, simplifying data organization, and providing a convenient foundation for subsequent processing. The addition of key information such as latitude and longitude ranges, acquisition time, and data units during metadata supplementation provides complete auxiliary descriptions, enhancing data identifiability and traceability. The use of a unified encoding standard converts the data sequence into binary form, adapting to the underlying requirements of computer storage and transmission, ensuring the integrity and consistency of data conversion. Standardized protocols encapsulate binary data, giving it a structure conforming to network transmission specifications, ensuring stable and reliable transmission without loss or corruption. The resulting correction data stream combines structure, identifiability, and transmission adaptability, providing high-quality support for the subsequent application, sharing, and storage of satellite-derived precipitation correction data, meeting the standardization, efficiency, and practicality requirements of computer data processing.

[0074] like Figure 2 The diagram shown is a functional block diagram of a near real-time satellite inversion precipitation correction system based on machine learning provided in an embodiment of the present invention.

[0075] The near real-time satellite inversion precipitation correction system 100 based on machine learning described in this invention can be installed in an electronic device. Depending on the functions implemented, the near real-time satellite inversion precipitation correction system 100 may include a feature primitive reconstruction module 101, a correction field nonlinearization module 102, a mask matrix generation module 103, a compensation correction interpolation module 104, a seamless correction fusion module 105, and a correction data encapsulation module 106. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0076] In this embodiment, the functions of each module / unit are as follows: The feature primitive reconstruction module 101 is used to perform tensor reshaping on the inverted precipitation data of the satellite inversion section based on a lightweight correction network to obtain the network-adapted feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. The nonlinear transformation module 102 is used to perform nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field of the satellite inversion section. The mask matrix generation module 103 is used to evaluate the spectral fidelity of the retrieved precipitation data in order to obtain the quality mask matrix of the satellite inversion section. The compensation correction interpolation module 104 is used to perform credibility discrimination on the precipitation correction field based on the mass mask matrix, and to perform spatiotemporal kriging interpolation on the discriminated credibility grid set to obtain the compensation correction field of the satellite inversion section. The seamless correction fusion module 105 is used to perform edge smoothing fusion of the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion section; The correction data encapsulation module 106 is used to encapsulate the seamless correction field to obtain the correction data stream of the satellite inversion section.

[0077] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0078] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0080] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0081] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A near real-time satellite inversion precipitation correction method based on machine learning, characterized in that, The method includes: S1. Based on a lightweight correction network, tensor reshaping is performed on the inverted precipitation data of the satellite inversion section to obtain the network-adaptive feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. S2. Perform a nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field for the satellite inversion section, including: By filtering the network-adaptive feature primitives layer by layer, condensed feature blocks of the satellite inversion segment are obtained; Neighborhood feature correlation analysis is performed on the condensed feature blocks to obtain the contextual correlation feature field of the satellite inversion segment; Based on the lightweight correction network, the abstract features in the context-related feature field are mapped to the initial intensity distribution field of the satellite inversion segment; The initial intensity distribution field is geospatial frame aligned to obtain the precipitation correction field for the satellite inversion section; S3. Evaluate the spectral fidelity of the retrieved precipitation data to obtain the mass mask matrix of the satellite inversion section, including: The spectral smoothness of the retrieved precipitation data is tested to obtain the spectral continuity label of the satellite retrieval segment; Based on the preset typical ground cover spectral range, the spectral continuity label is used to determine spectral anomalies in order to obtain the distribution map of anomaly grid points in the satellite inversion section. Based on the spectral continuity label and the anomaly grid distribution map, the quality level of the retrieved precipitation data is determined, and the quality level label allocation result of the satellite inversion segment is obtained. Based on the spatial arrangement of the satellite inversion segments, the quality level identifier allocation results are organized and filled to obtain the quality mask matrix of the satellite inversion segments; S4. Based on the mass mask matrix, the credibility of the precipitation correction field is identified, and spatiotemporal kriging interpolation is performed on the identified credibility grid set to obtain the compensation correction field of the satellite inversion section. S5. Perform edge smoothing fusion on the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion section; S6. Encapsulate the seamless correction field to obtain the correction data stream of the satellite inversion section.

2. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 1, characterized in that, The lightweight correction network performs tensor reshaping on the inverted precipitation data of the satellite inversion section to obtain network-adapted feature primitives for the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized sensing training, including: Acquire large-scale historical precipitation data in the satellite inversion section to train a high-capacity precipitation correction architecture; The high-capacity precipitation correction architecture is structurally pruned to obtain a sparse transition architecture for the satellite inversion section; The sparse transition architecture is quantized and retrained, and gradient backpropagation is performed on the trained architecture to obtain the deployable architecture of the satellite inversion segment. Edge nodes are deployed on the deployable architecture to obtain a lightweight correction network for the satellite inversion segment; Based on the lightweight correction network, the channel dimension of the inverted precipitation data of the satellite inversion section is reconstructed to obtain the network-adaptive feature primitives of the satellite inversion section.

3. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 2, characterized in that, The lightweight correction network is used to reconstruct the channel dimension of the inverted precipitation data in the satellite inversion section, resulting in network-adapted feature primitives for the satellite inversion section, including: Based on the lightweight correction network, heterogeneous data fusion is performed on the retrieved precipitation data to obtain a multi-source feature data layer for the satellite retrieval section; The feature channels in the multi-source feature data layer are repositioned to obtain the rearranged data volume of the satellite inversion segment; The rearranged data volume is resampled using grids to obtain the data array of the satellite inversion segment; The relationship between the height, width and depth dimensions of the data array is encapsulated in three dimensions to obtain the network-adaptive feature primitives of the satellite inversion segment.

4. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 1, characterized in that, The process of determining the credibility of the precipitation correction field based on the mass mask matrix, and performing spatiotemporal kriging interpolation on the determined credibility grid set to obtain the compensation correction field for the satellite inversion segment includes: Based on the mass mask matrix, the credibility of the precipitation correction field is analyzed to obtain the credibility label of the satellite inversion segment; Based on the confidence labels, the precipitation correction field is divided into grid points to obtain the high confidence grid points and grid points to be compensated in the satellite inversion section; Based on the spatial location and value of the high-confidence grid points, a spatiotemporal neighborhood search is performed on the grid points to be compensated to obtain the interpolation reference grid point group of the satellite inversion section. Based on the interpolation reference grid group, the grid points to be compensated are weighted and evaluated to obtain the reconstructed value of the satellite inversion section; Spatial reconstruction is performed on the high-confidence grid points and the reconstructed values ​​to obtain the compensation and correction field of the satellite inversion section.

5. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 4, characterized in that, The formulas for calculating the contribution weight and reconstructed value in the satellite inversion segment are as follows: ， ； In the formula, Reference grid points in the interpolation reference grid group For the grid points to be compensated among the grid points to be compensated Contribution weight, The reference grid point With the grid points to be compensated The Euclidean distance between them The reference grid point With the grid points to be compensated The absolute time difference between them For preset space-related length parameters, This is a preset time-related length parameter. The preset spatial distance attenuation index parameter, The preset time-distance decay index parameter, Reference grid points in the mass mask matrix The credibility coefficient For a preset small positive real number, The reference grid points in the retrieved precipitation data The precipitation intensity value, For the grid points to be compensated The reconstruction value, For the grid points to be compensated The corresponding reference grid group.

6. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 1, characterized in that, The step of smoothly fusing the confidence grid set with the compensation correction field to obtain the seamless correction field for the satellite inversion segment includes: The boundary between the confidence grid set and the compensation correction field is distinguished to obtain the boundary region of the satellite inversion segment; The outward expansion width of the boundary region is defined as the annular transition zone of the satellite inversion segment; Two-way distance fusion is performed on the grid points within the annular transition region to obtain the transition field of the satellite inversion section; The confidence grid set, the transition field, and the compensation correction field are spatially stitched together to obtain the seamless correction field of the satellite inversion section.

7. The near real-time satellite inversion precipitation correction method based on machine learning as described in claim 1, characterized in that, The process of encapsulating the seamless correction field to obtain the correction data stream for the satellite inversion segment includes: The seamless correction field is flattened to obtain a linear data array for the satellite inversion section; Metadata supplementation is performed on the linear data array to obtain the metadata-enhanced data sequence of the satellite inversion segment; The metadata-enhanced data sequence is encoded and converted to obtain the binary data of the satellite inversion segment; The binary data is encapsulated using a protocol to obtain the correction data stream for the satellite inversion segment.

8. A near real-time satellite inversion precipitation correction system based on machine learning, characterized in that, The system for implementing the near real-time satellite inversion precipitation correction method based on machine learning as described in claim 1 includes: The feature primitive reconstruction module is used to perform tensor reshaping on the inverted precipitation data of the satellite inversion section based on a lightweight correction network, so as to obtain the network-adapted feature primitives of the satellite inversion section. The lightweight correction network is obtained through structured pruning and quantized perception training. The nonlinear transformation module for the correction field is used to perform nonlinear transformation on the network-adaptive feature primitives to obtain the precipitation correction field for the satellite inversion section, specifically for: By filtering the network-adaptive feature primitives layer by layer, condensed feature blocks of the satellite inversion segment are obtained; Neighborhood feature correlation analysis is performed on the condensed feature blocks to obtain the contextual correlation feature field of the satellite inversion segment; Based on the lightweight correction network, the abstract features in the context-related feature field are mapped to the initial intensity distribution field of the satellite inversion segment; The initial intensity distribution field is geospatial frame aligned to obtain the precipitation correction field for the satellite inversion section; The mask matrix generation module is used to evaluate the spectral fidelity of the retrieved precipitation data to obtain the quality mask matrix of the satellite inversion segment, specifically for: The spectral smoothness of the retrieved precipitation data is tested to obtain the spectral continuity label of the satellite retrieval segment; Based on the preset typical ground cover spectral range, the spectral continuity label is used to determine spectral anomalies in order to obtain the distribution map of anomaly grid points in the satellite inversion section. Based on the spectral continuity label and the anomaly grid distribution map, the quality level of the retrieved precipitation data is determined, and the quality level label allocation result of the satellite inversion segment is obtained. Based on the spatial arrangement of the satellite inversion segments, the quality level identifier allocation results are organized and filled to obtain the quality mask matrix of the satellite inversion segments; The compensation correction interpolation module is used to perform credibility screening on the precipitation correction field based on the mass mask matrix, and to perform spatiotemporal kriging interpolation on the screened credibility grid set to obtain the compensation correction field of the satellite inversion section. A seamless correction fusion module is used to perform edge smoothing fusion of the confidence grid set and the compensation correction field to obtain the seamless correction field of the satellite inversion segment; The calibration data encapsulation module is used to encapsulate the seamless calibration field to obtain the calibration data stream of the satellite inversion segment.

Images