A Multi-Dimensional Spatiotemporal Calibration-Based Intelligent Fusion Method and System for Land and Sea Multi-Source Real-Time Data
By employing a multi-dimensional spatiotemporal calibration method, the problems of uneven land-sea observation, systematic spatiotemporal inconsistency, and transitional region processing in the fusion of multi-source meteorological data were solved, generating a high-quality real-time meteorological element analysis dataset, achieving global coverage and improved data quality.
Patent Information
- Application Number
- CN202611319576.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively integrate multi-source meteorological data, resulting in uneven land and sea observation coverage, systematic spatiotemporal inconsistencies, failure to fully utilize the complementary advantages of station and grid data, a lack of processing strategies for land-sea transition areas, and insufficient quality of operational multi-source real-time fusion products.
A multi-dimensional spatiotemporal calibration method is adopted, which generates an enhanced site dataset and a fused grid dataset through data source quality assessment, systematic bias correction, time dimension calibration and spatial dimension calibration. A land-sea type matching factor is introduced and a closed-loop quality control mechanism is constructed by combining cross-validation.
It achieves effective coverage of global land and ocean areas, improves data quality and temporal continuity, enhances the fusion quality of land-sea transition areas, identifies and corrects data anomalies, and generates high-coverage, high-quality target meteorological element real-time analysis datasets.
Smart Images

Figure CN123116149A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological observation data processing technology, and in particular to a method and system for intelligent fusion of multi-source real-time data of target meteorological elements covering global land and sea areas. Background Technology
[0002] Meteorological data is the foundation for weather forecasting, climate monitoring, disaster early warning, and various meteorological services.
[0003] With the continuous improvement of the global meteorological observation system, the sources of data available for real-time analysis are becoming increasingly abundant, mainly including observation data from global meteorological telecommunications system exchange stations, observation data from ocean buoy stations, observation data from island stations, as well as various satellite inversion grid products and reanalysis grid products.
[0004] These data sources differ significantly in terms of spatial coverage, temporal resolution, observation accuracy, and quality characteristics. How to effectively integrate multi-source heterogeneous data to generate real-time analysis products with complete spatial coverage and reliable quality is a core technical requirement in meteorological operations and research.
[0005] The existing technology has the following main shortcomings: First, the uneven coverage of land and sea observations means that a single data source cannot meet global analysis needs.
[0006] GTS exchange stations and conventional ground stations are mainly distributed in land areas, while only sparse buoy stations and ship stations are used for observation in marine areas. Although island stations are located at the junction of land and sea, their number is limited.
[0007] While grid-based real-time products can provide global coverage, their accuracy drops significantly in areas with sparse sites, and there are also significant differences in accuracy, resolution, and systematic bias among different grid-based products.
[0008] Whether it's site observations or gridded products, a single data source is insufficient to independently meet the needs for high-quality real-time analysis across global land and sea areas.
[0009] Second, there is a systematic spatiotemporal inconsistency among the multi-source data.
[0010] The observation time, spatial representativeness, and data processing procedures of different data sources vary, leading to systematic biases among the data.
[0011] Site observations are point measurements, which are highly accurate but have limited spatial representativeness; grid products are area averages or pattern analysis values, which have good spatial coverage but have systematic biases; data from different sources show significant differences in temporal alignment and spatial matching.
[0012] Simply merging or averaging multi-source data can introduce inconsistencies and reduce the quality of analytical products.
[0013] Third, existing fusion methods do not fully utilize the complementary advantages of site and grid data.
[0014] Traditional real-time data fusion methods either use station data as the main source for spatial interpolation to generate gridded products, or use gridded products as background fields to assimilate station observations. The former has large interpolation errors in sparse station areas, while the latter relies excessively on the quality of the background field.
[0015] Existing methods have failed to establish a bidirectional complementary fusion mechanism between site data and grid data, making it difficult to generate high-quality site datasets and grid datasets simultaneously.
[0016] Fourth, there is a lack of effective processing strategies for data fusion in the land-sea transition zone.
[0017] The observation environment in coastal zones, islands, and other transitional areas between land and sea is complex, and the representativeness and accuracy of station observations and grid products in these areas are significantly different from those in purely land or purely marine areas.
[0018] Existing fusion methods typically do not address the specific needs of land-sea transition zones, resulting in poor quality fusion products in these critical areas.
[0019] Fifth, the existing integrated grid station solution has limitations in its technical approach.
[0020] The grid-station integrated solution that has emerged in recent years also involves two processing paths: station and grid. Its basic idea is to use numerical weather prediction products as the processing object and station observations as the true values to establish a correction model. The station correction forecast and the grid correction forecast are then integrated into a set of forecast products through optimal interpolation.
[0021] This scheme differs from the present invention in three aspects: In terms of the object of processing, the scheme is aimed at error correction of numerical weather prediction products, while the present invention is aimed at the fusion analysis of multi-source real-time data; In terms of product form, the scheme merges two paths into a single product, while the present invention independently produces two sets of complementary products at the station level and grid level and verifies each other; In terms of transition zone processing, the scheme does not perform special processing for the land-sea transition zone, while the present invention uses the land-sea type matching factor to segmentally modulate the interpolation weights.
[0022] Sixth, the existing business-oriented multi-source live fusion products have a relatively simple technical approach.
[0023] The meteorological operational departments have publicly launched multi-source fusion real-time analysis products, such as the 1km hourly multi-source fusion real-time analysis product for China and the 10km hourly precipitation multi-source fusion real-time analysis product for the world. Their technical approach aims to generate single-grid real-time products.
[0024] These products have the following limitations: in terms of product form, they only output a single grid dataset, and do not independently output an enhanced station dataset optimized through multi-source scoring; in terms of quality control, they do not set up a cross-verification between products of two independent paths and a closed-loop processing mechanism of verification-traceback-weight reduction or elimination-re-fusion-re-verification; in terms of transition zone processing, they do not take the sea-land transition zone as an independent partition to run through the whole process of data source quality scoring, deviation correction, interpolation weight modulation and boundary verification. Content of the Invention
[0025] Therefore, the technical problem to be solved by the present invention is to provide a sea-land multi-source live intelligent fusion method and system based on multi-dimensional spatiotemporal calibration, which intelligently fuses multi-source observation data to generate a target meteorological element live analysis dataset with high coverage and high quality.
[0026] In a first aspect, the present invention provides a sea-land multi-source live intelligent fusion method based on multi-dimensional spatiotemporal calibration, comprising the following steps: Step 1: Collect multi-source live data of target meteorological elements, and divide the surface types corresponding to the multi-source live data into land, ocean and sea-land transition zone; the multi-source live data includes station observation data and grid live product data. Step 2: Perform data source quality assessment, systematic deviation correction, time dimension calibration and space dimension calibration on the preprocessed multi-source live data. Step 3: Based on the processed multi-source live data, adopt a station fusion path to select the optimal value from the valid values of each data source at the station position for each station position, so as to generate an enhanced station dataset; at the same time, adopt a grid fusion path to interpolate the station observation data to the target grid, and perform weighted fusion with the grid live product data to generate a fused grid dataset. Step 4: Perform sea-land boundary gradient verification, smoothing processing and physical rationality check on the fused grid dataset. Step 5: Perform cross-verification on the enhanced station dataset and the fused grid dataset, and trigger feedback correction for abnormal data sources according to the verification result.
[0027] Further, in Step 1, the surface type partitioning includes: calculating the ocean area proportion p within a preset radius window around the target position. When p ≤ the first preset threshold, it is determined as land. When p ≥ the second preset threshold, it is determined as ocean. When the first preset threshold < p < the second preset threshold, it is determined as a sea-land transition zone.
[0028] Further, in Step 2, the data source quality assessment includes calculating a comprehensive quality score.
[0029] Among them, Q s (r) represents the overall quality score of data source s under land surface type r; C s (r) represents the coverage rate, which is the proportion of the number of valid observations of data source s within surface type r to the total number of observations to be made during the statistical period. T s For time continuity, it is the proportion of the number of valid observations of data source s within the statistical period to the total number of observations that should be made in that period; S s For spatial consistency, is the correlation coefficient between the observation time series of data source s and the reference observation series in the spatial neighborhood; B s Historical deviation is the average deviation of data source s relative to the reference benchmark within the sliding statistical window. B max This serves as a reference value for deviation normalization. w c w t w s w b These are the weighting coefficients for coverage, temporal continuity, spatial consistency, and historical bias, respectively.
[0030] Furthermore, in step two, the time dimension calibration includes: Define a uniform time matching window and assign time decay weights to data sources that have observations within the time window:
[0031] Where, Δt k Let k be the time difference between the kth observation and the analysis. The preset time decay scale parameter; For data sources with a time resolution lower than the analysis interval, they are not subject to the exclusion rules of the time matching window. The most recent valid time interval is selected for fusion, and the time decay weight is calculated based on the actual time difference.
[0032] Furthermore, in step two, the spatial dimension calibration includes: The station observations are corrected for vertical descent rate based on altitude differences; the grid point real-time products are statistically downscaled to match the variability characteristics of the station scale, and the statistical downscaling adopts variability scaling or quantile mapping.
[0033] Furthermore, in step three, the site fusion path includes: For each site location, all data sources with valid observations or grid values at that location are aggregated; for grid products, the grid values corresponding to each site location are extracted through spatial interpolation. For each data source at each site location, a multi-dimensional comprehensive score is calculated based on the comprehensive quality score, time decay weight, and spatial representativeness score of each data source. The spatial representativeness score is used to characterize the degree to which the data source represents the true value of the site location. The highest value is taken when the site is directly observed, and the value is determined according to the grid resolution when interpolating by grid points. For each site location, the value with the highest multi-dimensional comprehensive score from all available data sources is selected as the site's fusion value: The fused target meteorological element data is combined with the original station metadata to form an enhanced station dataset.
[0034] Furthermore, in step three, the grid fusion path includes: The calibrated station observation data is interpolated to the target grid, and a land-sea type matching factor is introduced into the interpolation weight to form station gridded data. The land-sea type matching factor is: when the land surface type of the station and the target grid is the same, the first weight value is taken; when they are different, the second weight value is taken, which is less than the first weight value. The gridded data of the sites is weighted and fused with the real-time product data of the grids to generate a fused grid dataset. The fusion weight of the gridded data of the sites is positively correlated with the number of valid sites around the target grid, and the fusion weight of the real-time product data of the grids is the comprehensive quality score.
[0035] Furthermore, the fusion weights of the gridded site data are:
[0036] In the formula: W sta,0 The baseline weight for site data; N sta (x g ) is the grid point x g Number of effective sites within the surrounding influence radius; N ref Number of sites for reference.
[0037] Furthermore, step five includes: The values at the site locations of the enhanced site dataset and the fused grid dataset are compared, and sites and their corresponding grid points with differences exceeding a preset threshold are marked as suspicious. By tracing back the processing records of each data source at this site, the source of the discrepancy can be identified. Data sources that cause the difference to exceed a preset threshold will be identified as abnormal data sources and will be downgraded or removed. Based on the remaining data sources, re-execute steps three and four to update the fusion values of the sites and corresponding grid points; If the re-inspection is passed, the marking will be removed. If the re-inspection still exceeds the threshold, it will be marked as low quality or missing.
[0038] Secondly, the present invention also provides a multi-dimensional spatiotemporal calibration-based intelligent fusion system for real-time data from multiple sources on land and sea, comprising: The data collection and preprocessing module is used to collect multi-source real-time data of target meteorological elements and divide the land surface type corresponding to the multi-source real-time data into land, ocean and land-sea transition zones; the multi-source real-time data includes station observation data and gridded real-time product data; The multi-dimensional spatiotemporal calibration module is used to perform data source quality assessment, systematic deviation correction, time dimension calibration, and spatial dimension calibration on the preprocessed multi-source real-time data. The dual-path intelligent fusion module is used to generate an enhanced site dataset by selecting the optimal value from the valid values of each data source at each site location based on the processed multi-source real-time data using a site fusion path; at the same time, it uses a grid fusion path to interpolate the site observation data to the target grid and weightedly fuse it with the grid real-time product data. The post-processing module is used to perform land-sea boundary gradient verification, smoothing, and physical rationality checks on the fused grid dataset. The cross-validation and feedback fusion module is used to perform cross-validation on the enhanced site dataset and the fused grid dataset, and trigger feedback correction on abnormal data sources based on the validation results.
[0039] Beneficial effects: First, by comprehensively integrating GTS exchange stations, buoy stations, island stations, and various grid point real-time products, effective coverage of the global land and sea areas has been achieved, and the data availability in sparsely populated areas has been significantly improved.
[0040] Second, two parallel processing chains, site fusion and grid fusion, were designed to independently produce two sets of products: enhanced site dataset and fused grid dataset. This balances the high-precision features of site observations with the spatial continuity of grid products, meeting the needs of different application scenarios.
[0041] Third, the multi-source data is comprehensively processed from four aspects: data source quality assessment, systematic bias correction, time dimension calibration, and spatial dimension calibration. This effectively eliminates systematic inconsistencies between different data sources and provides a reliable data foundation for subsequent fusion.
[0042] Fourth, a multi-dimensional comprehensive scoring optimization strategy is adopted in the site fusion path. The fusion value of each site always comes from the current best data source, which improves the temporal continuity and spatial coverage while ensuring data quality.
[0043] Fifth, by introducing a land-sea type matching factor during the interpolation and fusion process, combined with land-sea boundary consistency processing, the data discontinuity and unreasonable gradient problems in the land-sea transition area are effectively suppressed, and the fusion quality of key areas such as the coastal zone is improved.
[0044] Sixth, the products produced by the two independent paths are independent in terms of data organization and error sources. Cross-validation can detect data anomalies that are difficult to identify by a single path. After anomaly detection, a closed-loop process of backtracking, weight reduction / removal, re-fusion, and re-inspection is triggered, which transforms the anomaly identification capability into the system's self-correction capability, thus forming a complete closed-loop quality control mechanism. Attached Figure Description
[0045] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0046] Figure 1 This is a flowchart illustrating the overall process of the method in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the multi-dimensional spatiotemporal calibration process of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the dual-path intelligent fusion process in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the land-sea type matching factor in the site gridding interpolation of Embodiment 1 of the present invention. Detailed Implementation
[0047] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0048] The principles and features of the present invention are described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.
[0049] The examples given are for illustrative purposes only and are not intended to limit the scope of the invention.
[0050] Example 1: This embodiment presents a method for intelligent fusion of real-time data from multiple sources on land and sea based on multi-dimensional spatiotemporal calibration.
[0051] This invention provides independent fusion analysis methods for each target meteorological element.
[0052] The target meteorological element refers to the specific meteorological variable that needs to be subjected to real-time fusion analysis, that is, the physical quantity corresponding to the final analytical product of the method.
[0053] This method uses temperature and precipitation as examples.
[0054] I. Method for Fusion Analysis of Temperature Elements like Figure 1 As shown, the steps for fusing temperature elements in this method include: Step T1: Collection and Preprocessing of Multi-Source Temperature Data Collect a variety of real-time temperature monitoring data covering global land and ocean areas, specifically including: Data collection from site T1.1: (a) GTS Exchange Station Data: Collects ground weather station observation data exchanged by the Global Meteorological Telecommunication System, including 2m temperature (unit: °C), covering major land areas worldwide; (b) Buoy station data: Collect temperature observation data from ocean buoy stations around the world, covering major ocean areas; (c) Island station data: Collect temperature observation data from automatic weather stations on islands, covering the transition zone between land and sea and offshore islands; (d) Other available site observation data: including supplementary observation sources such as ship stations and automatic regional stations.
[0055] T1.2 grid point real-time data collection: Collect real-time products with a 10km resolution grid across the globe, including but not limited to: (a) Satellite-retrieved surface temperature products (such as MODISLST, converted into 2m temperature proxy values through statistical regression). (b) Further analyze gridded real-world products (such as ERA5, MERRA-2, etc.); (c) Other available global grid temperature products.
[0056] T1.3 Data Preprocessing: (a) Spatiotemporal matching: All data sources are uniformly matched to the standard analysis time (e.g., hourly or 6-hourly). For site data, the most recent valid observations of the analysis time are taken. For grid data, time interpolation is used to align to the standard time. (b) Spatial matching: Record the precise latitude and longitude coordinates of each station's data and interpolate each grid point product to a common reference grid (such as a 10km latitude and longitude grid). (c) Unit standardization: The temperature unit is standardized to degrees Celsius (°C); (d) Basic quality control: Perform extreme value checks (temperature: -70°C to +60°C) on each data source, mark missing values, and standardize data formats.
[0057] T1.4 Surface type zoning determination: High-resolution land-sea mask data (such as GSHHG global coastline data, or 1km resolution land-sea proportion field) is acquired in advance. For any station or grid point position x, the ocean area proportion p(x) (the ratio of the ocean area within the window to the total area of the window) in a circular window with a preset radius R0 (such as 5km) around it is calculated, and the surface type r(x) of this position is determined according to the following rules: p(x)≤0.2, land; 0.2<p(x)<0.8, land-sea transition zone; p(x)≥0.8, ocean.
[0058] In the formula: r(x) is the determination result of the surface type at position x; p(x) is the ocean area proportion within the window of radius R0 around position x; 0.2 and 0.8 are the proportion determination thresholds for land / transition zone and transition zone / ocean, which are preferred values and can be adjusted according to the mask resolution and regional characteristics.
[0059] This zoning determination result is used throughout the subsequent steps: data source quality scores are counted by zone, systematic deviation correction is estimated by zone, the land-sea type matching factor is determined according to the zoning types of the station and the grid point, and the land-sea boundary gradient test is performed between grid points on both sides of the land-sea zoning boundary.
[0060] Step T2: Multi-dimensional spatiotemporal calibration of air temperature As Figure 2 shown, multi-dimensional spatiotemporal calibration is performed on the multi-source air temperature data collected in step T1 to eliminate systematic deviations and spatiotemporal inconsistencies between different data sources, which specifically includes: T2.1 Data source quality feature assessment: Assess the quality features of each data source s in different surface type regions respectively, including: (a) Data coverage Cs: the proportion of valid observation grid points or stations of this data source in the target region; (b) Temporal continuity Ts: the continuous observation capability of this data source in the time series; (c) Spatial consistency Ss: the spatial consistency degree between the observed values of this data source and the values of adjacent stations / grid points; (d) Historical deviation Bs: the historical average deviation of this data source relative to the reference benchmark.
[0061] The statistical caliber of each of the above sub-indicators is defined as follows: Data coverage Cs(r): within the statistical period, the ratio of the number of valid observations (number of valid stations or number of valid grid points) of data source s in region type r to the total number of observations that should be made in this region, with a value range of [0,1]; Time continuity Ts(r): The proportion of the number of valid observations of data source s to the total number of observations that should be made in that period within the statistical time period, with a value range of [0,1]. Spatial consistency Ss(r): The correlation coefficient between the observation time series of data source s and the reference observation series (corresponding observations of neighboring stations or other data sources) in its spatial neighborhood. The correlation coefficient is 0 when it is negative, and the value range is [0,1]. Historical bias Bs(r): The average deviation of the data source s relative to a reference baseline within a sliding statistical window, such as in reanalysis products or high-density site fusion fields; B max The reference value for deviation normalization is the maximum absolute value of the historical deviation of each candidate data source, or a preset physical upper limit. For temperature, it can be 5°C, and for precipitation, it can be the normalized value corresponding to 50% of the relative deviation.
[0062] Based on the above indicators, calculate the overall quality score Qs for each data source in each region:
[0063] In the formula: Q s (r) represents the overall quality score of data source s under region type r, with a value range of [0,1]; r represents the region type (land, sea, or land-sea transition zone); w c w t w s w b The weighting coefficients for coverage, temporal continuity, spatial consistency, and historical bias are respectively, satisfying the constraint w. c + w t + w s +w b =1.
[0064] T2.2 Systematic bias correction: By utilizing the overlapping regions and time periods of multi-source data, the systematic bias of each data source relative to the reference benchmark is estimated through least squares regression, and the original data is corrected for bias.
[0065] For the temperature element, additive deviation correction is used:
[0066] In the formula: The temperature value (°C) after deviation correction for data source s at location x and time t; The raw temperature observation (°C) of data source s at location x and time t; The additive bias (°C) of the data source s estimated by least squares regression is statistically analyzed by the land surface type region r(x) corresponding to the month and location x.
[0067] T2.3 Time Dimension Calibration: To address the differences in time representativeness across different data sources, time dimension calibration is performed. (a) Time window matching: Set a uniform time matching window Δt (e.g., ±30 minutes), and only include data that are observed within the time window into the fusion.
[0068] For data sources with a time resolution lower than the analysis interval, such as satellite products that pass over twice a day or daily products, the above window exclusion rules are not applied. Instead, the most recent valid time interval is selected for fusion, and the time decay weight is calculated based on its actual time difference. (b) Time decay weights: Time decay weights are assigned to observations participating in the fusion.
[0069] In the formula: This is the time decay weight for the k-th observation, with a value range of (0,1]. The time difference between the kth observation and analysis (in minutes); The time decay scale parameter determines how quickly the weight decays with time difference; temperature is taken as... =60min.
[0070] T2.4 Spatial Dimension Calibration: Spatial dimension calibration is performed to address the spatial representativeness differences between site data and grid data. (a) Site representativeness correction: For site observations, considering the difference between their altitude and the grid average altitude, an altitude lapse rate correction is performed:
[0071] In the formula: The temperature value (°C) is corrected for altitude. Γ represents the original temperature observation value (°C) at the station; Γ represents the vertical lapse rate of ambient temperature, taken as the ICAO standard atmospheric value of 0.0065°C / m. The average elevation (m) of the target grid points; This represents the actual altitude (m) of the station.
[0072] When the station's altitude is higher than the average altitude of the grid points, the correction value is negative, and vice versa.
[0073] (b) Grid representativeness correction: For grid products, the spatial smoothing effect is corrected by statistical downscaling to match the variability characteristics of the site scale.
[0074] Specifically, two methods can be used: variable rate scaling or quantile mapping. Under variable rate scaling, the variable rate adjustment coefficient is determined by the ratio of the standard deviation of the grid product time series to the standard deviation of the station observation time series within the overlapping area, and the anomaly value of the grid product is scaled. Under quantile mapping, the quantile correspondence is established based on the cumulative distribution function of the station observations and the grid product within the overlapping period, and the grid product value is mapped and corrected.
[0075] Step T3: Intelligent Fusion of Dual Temperature Paths like Figure 3 As shown, based on the multi-source temperature data calibrated in step T2, intelligent fusion is performed using both station fusion path and grid fusion path to generate two complementary real-time analysis datasets: Path 1: Generating a temperature station dataset based on multi-dimensional optimization of station locations T3.1.1 Multi-source site data aggregation: For each known site location xi, aggregate all data sources that have valid observations or grid values at that location.
[0076] For grid products, spatial interpolation is used to extract the grid values corresponding to each site location:
[0077] In the formula: To extract site locations from the grid field of data source s Temperature value (°C) at that location; For the gridded temperature field of data source s; Let be the latitude and longitude coordinates of the i-th station; Interp is the bilinear interpolation operator that spatially interpolates the grid field at the station location.
[0078] T3.1.2 Multi-dimensional comprehensive scoring and optimal selection: For each site location For each data source s, calculate a multi-dimensional comprehensive score:
[0079] In the formula: For data source s at site location The data source is evaluated across multiple dimensions, with higher scores indicating better quality on the site. The data source s calculated for step T2.1 is at the site Location type The overall quality score below; The time decay weight of data source s is , where This is the time difference between the observation time and the analysis time for this data source; For data sources s on the site Spatial representativeness score, when data comes from direct observations at the site. =1.0, when the data comes from gridded product interpolation. The value is set between 0.55 and 0.85, determined by the ratio of grid resolution to the representative scale of the site; the higher the resolution, the larger the value.
[0080] For each site location, the value with the highest overall score from all available data sources is selected as the site's merge value:
[0081] In the formula: For the site Temperature fusion value (°C) at the location; s * Number the data source with the highest overall score; Optimal data source Temperature at this station (°C); This indicates selecting from all available data sources. The data source index that has reached its maximum value.
[0082] T3.1.3 Enhanced Site Dataset Construction: The merged station temperature data is combined with the original station metadata (station number, station name, latitude and longitude, altitude, network to which it belongs, etc.) to form an enhanced station dataset.
[0083] This dataset has the following advantages compared to the original site data: (a) Higher coverage – grid products can supplement coverage when the original site is missing; (b) Higher quality – the value for each site comes from the current best data source; (c) Better temporal continuity – reducing data gaps through multi-source complementarity.
[0084] Path 2: Generation of temperature grid dataset based on grid interpolation fusion T3.2.1 Gridding of site data: The calibrated station observation data from step T2 are interpolated onto the target grid.
[0085] An adaptive interpolation method considering land-sea distribution is adopted: For the target grid point xg, search for valid site observations around it and calculate the interpolation weights:
[0086] In the formula: For station k, pair of target grid points Normalized interpolation weights; From station k to target grid point Geographical distance (km); It is a Gaussian distance decay function, where L is the radius of influence (for air temperature, L = 200 km). The land-sea type matching factor is defined as follows: = 1.0: The site and grid points have the same land and sea type; β: The land and sea types of the stations and grid points are different.
[0087] In the formula: The surface type of the location where station k is located; For target grid points The land surface type of the location; β is the cross-type attenuation coefficient, which is a positive number less than 1, preferably 0.2~0.5, used to reduce the influence weight of different land surface type stations on grid interpolation, and to avoid unreasonable propagation of land station values to marine grid points, and vice versa.
[0088] Figure 4 This is a schematic diagram of the land-sea type matching factor in a gridded interpolation of a site.
[0089] The gridded values for the site data are:
[0090] In the formula: Interpolate site observation data to target grid points Temperature value (°C) at that location; For station k, the normalized interpolation weights for the target grid points; The calibrated temperature observation (°C) for station k; Σ k Summing all valid sites within the search radius.
[0091] T3.2.2 Multi-source grid data fusion: The gridded data from the sites is then weighted and fused with the real-time data from each grid point collected in step T1:
[0092] In the formula: For target grid points Temperature fusion value (°C) at that location; After gridding the site data Temperature value (°C) at that location; For grid products s in Temperature value (°C) at that location; The fusion weight of the gridded data of a site is related to the number of valid sites around that grid point. Positive correlation:
[0093] In the formula: The baseline weight for station data (temperature is set to 3.0); For grid points Number of effective sites within the surrounding influence radius; N ref The number of reference sites can be set to 5 to control the rate of weight saturation.
[0094] When there are enough surrounding sites, the site data weights are close. When there are few sites, the weight automatically decreases, and the system relies more on grid products.
[0095] The fusion weight of each grid point product is determined by its quality score:
[0096] In the formula: For grid product s at the target grid The fusion weight at the location; For this data source in The overall quality score under the region type (calculated by step T2.1).
[0097] T3.2.3 Grid Dataset Construction: The fused grid data constitutes a complete global coverage grid dataset with a resolution equal to that of the target grid (e.g., 10 km).
[0098] This dataset combines the high precision of site observations with the wide coverage of gridded products. In densely populated areas, it prioritizes site data, while in sparsely populated areas (such as oceans and plateaus), it prioritizes gridded products, thus achieving adaptive weight allocation.
[0099] Step T4: Post-processing and quality control of temperature between land and sea zones Post-processing and rigorous quality control were performed on the dual-path fusion results, dividing them into land and sea zones. T4.1 Land-sea boundary consistency handling: For the gridded dataset, examine the temperature gradient between land and sea grid points near the coastal zone; if the gradient exceeds a physically reasonable threshold, perform smoothing. Temperature gradient threshold: ΔT / Δd ≤ 0.1°C / km Grid points exceeding the threshold are smoothed using a neighborhood-weighted average.
[0100] The above thresholds are preferred values and can be adjusted according to regional climate characteristics and seasonal configurations.
[0101] It should be noted that this gradient threshold is used to identify non-physical jumps caused by abnormal data sources, rather than to eliminate real land-sea thermal differences.
[0102] In practice, a joint judgment can be made by combining the consistency of adjacent time intervals and the consistency of neighboring grid points: smoothing is triggered only when the gradient exceeding the threshold occurs in isolation (without weather-scale continuity in both time and space); for real land-sea temperature differences with spatiotemporal continuity (such as periods of significant land-sea thermal differences in spring and autumn), smoothing is not triggered to avoid erasing the real land-sea boundary temperature gradient characteristics.
[0103] T4.2 Physical rationality check: Check whether the merged temperature meets the spatial continuity and vertical lapse rate constraints, and remove grid values that are abnormally high or low.
[0104] The removed grid points are marked as missing measurements and are filled by a weighted average of the valid grid point values in their neighborhood or by the value of the grid point product at that location.
[0105] T4.3 Consistency Cross-validation and Loop Closure for Two Datasets: Cross-compare the site dataset generated by path one with the grid dataset generated by path two, extract the values of the grid dataset at the site location, and compare them with the site fusion values.
[0106] If the difference exceeds the preset threshold (temperature: 3°C), the closed-loop processing procedure will be initiated: (a) Mark: Mark the site and its corresponding grid point as suspicious; (b) Retrospection: retrieve the quality scores, time decay weights, bias correction records and original observations of each data source at the site location to locate the data source that caused the discrepancy; (c) Deweighting / Removal: Data sources that are traced back and cause the difference to exceed the threshold are identified as abnormal data sources and are deweighted or removed. Auxiliary judgment conditions include: the frequency of recent abnormal events of the data source is higher than a preset value, or the deviation of its observed value from the consistency range of other data sources at the site location exceeds a preset range. (d) Re-fusion: Re-execute the dual-path fusion in step T3 based on the remaining data sources, and update the fusion value of the site and the corresponding grid point; (e) Re-inspection: After re-fusion, cross-comparison is performed again. If the difference still exceeds the threshold, the fusion value of the site / grid point is marked as low quality or missing, and the event log is recorded for subsequent statistical updates of the data source quality feature assessment (step T2.1).
[0107] The above-mentioned "inspection-backtracking-deweighting / removal-reintegration-re-inspection" process constitutes a closed-loop quality control, which utilizes the independence of the data organization methods and error sources of the two paths to identify and correct data anomalies that cannot be detected by a single path.
[0108] Step T5: Temperature Product Quality Assessment and Output T5.1 Product Quality Assessment: (a) Independent site verification: 10%-20% of sites are reserved not to participate in the integration process and are used as independent verification data.
[0109] By comparing the fused product with observations from independent sites, the bias, root mean square error (RMSE), and correlation coefficient (R) are calculated.
[0110] (b) Cross-data source validation: The site dataset of Path 1 and the grid dataset of Path 2 are cross-validated to evaluate the consistency of the two products.
[0111] (c) Coverage statistics: The improvement in spatial coverage and temporal continuity of statistical fusion products compared to single data sources.
[0112] T5.2 Product Output: Output two sets of standard format real-time temperature analysis products: (a) Enhanced site dataset: in CSV / BUFR format, containing fields such as station number, latitude and longitude, altitude, temperature fusion value, data source identifier, and quality score; (b) Fusion grid dataset: in NetCDF / GRIB2 format, containing a 10km resolution temperature grid analysis field, a mass label field, a data source contribution ratio field, etc.
[0113] II. Methods for Integrated Analysis of Precipitation Elements Step P1: Collection and Preprocessing of Multi-Source Precipitation Data Collect a variety of real-time precipitation monitoring data covering global land and ocean areas, specifically including: P1.1 Site observation data collection: (a) GTS Exchange Station Data: Collects precipitation observation data from ground meteorological stations exchanged by the Global Meteorological Telecommunication System, including hourly precipitation and cumulative precipitation (unit: mm), covering major land areas worldwide; (b) Buoy station data: Collect precipitation observation data from ocean buoy stations around the world (some buoys are equipped with rain gauges), covering major ocean areas worldwide; (c) Island station data: collect precipitation observation data from automatic weather stations on islands, covering the transition zone between land and sea and offshore islands; (d) Other available site observation data: including supplementary observation sources such as ship stations and automatic regional stations.
[0114] P1.2 Grid point real-time data collection: Collect real-time products with a 10km resolution grid across the globe, including but not limited to: (a) Satellite-retrieved gridded precipitation products; (b) Multi-source fused gridded precipitation products (such as GPMIMERG, CMORPH, etc.); (c) Further analyze gridded real-world products (such as ERA5); (d) Other available global grid precipitation products.
[0115] P1.3 Data Preprocessing: (a) Spatiotemporal matching: All data sources are uniformly matched to the standard analysis time (e.g., hourly or 6-hourly). For site data, the most recent valid observations of the analysis time are taken. For grid data, time interpolation is used to align to the standard time. (b) Spatial matching: Record the precise latitude and longitude coordinates of each station's data and interpolate each grid point product to a common reference grid (such as a 10km latitude and longitude grid). (c) Unit uniformity: The unit for precipitation is uniformly millimeters (mm); (d) Basic quality control: Perform extreme value checks on each data source (precipitation: 0 to 500 mm / h), mark missing values, and standardize data formats.
[0116] P1.4 Surface Type Zoning Determination: Similar to step T1.4, the proportion of ocean area within a preset radius window around each station and grid point is calculated based on high-resolution land-sea mask data. The land surface type is determined as land, land-sea transition zone, or ocean according to the proportion threshold. The determination result is used throughout the entire process of subsequent precipitation element quality scoring, deviation correction, land-sea type matching factor calculation, and land-sea boundary gradient verification.
[0117] Step P2: Multi-dimensional spatiotemporal calibration of precipitation Multi-dimensional spatiotemporal calibration is performed on the multi-source precipitation data collected in step P1 to eliminate systematic biases and spatiotemporal inconsistencies between different data sources. Specifically, this includes: P2.1 Data Source Quality Characteristics Assessment: For each data source s, evaluate its quality characteristics in different land surface types (land, ocean, and land-sea transition zone, as determined in step P1.4), including data coverage Cs, temporal continuity Ts, spatial consistency Ss, and historical bias Bs. The statistical definitions of each sub-indicator are the same as in step T2.1.
[0118] Based on the above indicators, calculate the overall quality score Qs for each data source in each region:
[0119] The meanings of the variables in the formula are the same as in step T2.1: The overall quality score of data source s under region type r; For data coverage; For time continuity; For spatial consistency; This is due to historical bias. This is the reference value for deviation normalization; w c w t w s w b Weighting coefficients (w) c + w t + w s +w b =1).
[0120] P2.2 Systematic bias correction: By utilizing the overlapping regions and time periods of multi-source data, the systematic bias of each data source relative to the reference benchmark is estimated through least squares regression, and the original data is corrected for bias.
[0121] For precipitation elements, a multiplicative correction factor is used:
[0122] In the formula: The precipitation (mm) after correction for the deviation of the data source s at location x and time t; The original precipitation (mm) of data source s at location x and time t; The multiplicative correction factor (dimensionless) for the data source s estimated through the overlap period is used for statistical partitioning of the land surface type region r(x) corresponding to the month and location x.
[0123] Precipitation is corrected using a multiplicative method rather than an additive method because precipitation is a non-negative quantity and the deviation is usually proportional to the magnitude.
[0124] P2.3 Time dimension calibration: To address the differences in time representativeness across different data sources, time dimension calibration is performed. (a) Time window matching: Set a uniform time matching window Δt (e.g., ±15 minutes) and only include data that are observed within the time window into the fusion.
[0125] For data sources with a time resolution lower than the analysis interval (such as daily site interpolation products), the above window exclusion rules are not applied. Instead, the most recent valid time interval is selected for fusion, and the time decay weight is calculated based on its actual time difference. (b) Time decay weights: Time decay weights are assigned to observations participating in the fusion.
[0126] In the formula: This is the time decay weight for the k-th observation, with a value range of (0,1]. The time difference between the kth observation and analysis (in minutes); For the time decay scale parameter, precipitation is taken as... =30min (the time decay scale of precipitation is shorter than that of air temperature because precipitation is more instantaneous and localized).
[0127] P2.4 Spatial Dimension Calibration: Spatial dimension calibration is performed to address the spatial representativeness differences between site data and grid data. (a) Grid point representativeness correction: For grid point products, the spatial smoothing effect is corrected by statistical downscaling to match the precipitation variability characteristics at the station scale. The specific implementation method (variability scaling or quantile mapping) is the same as step T2.4(b). (b) Site location registration: For site observations, record their precise location information for subsequent gridded interpolation.
[0128] Precipitation is a cumulative amount, and its spatial distribution does not have a stable physical constraint similar to the vertical lapse rate of temperature. The influence of altitude on precipitation has significant regional and weather system dependence. Therefore, precipitation station observations are not corrected for altitude lapse rate, but only the precise location is recorded for gridded interpolation.
[0129] Step P3: Intelligent Fusion of Two Precipitation Pathways Based on the multi-source precipitation data calibrated in step P2, intelligent fusion was performed using both site fusion path and grid fusion path to generate two complementary real-time analysis datasets: Path 1: Generation of a precipitation station dataset based on multi-dimensional site optimization P3.1.1 Data aggregation from multiple sites: For each known site location xi, aggregate all data sources that have valid observations or grid values at that location.
[0130] For grid products, spatial interpolation is used to extract the grid values corresponding to each site location:
[0131] In the formula: To extract site locations from the grid field of data source s Rainfall at the location (mm); For the gridded precipitation field of data source s; Let be the latitude and longitude coordinates of the i-th station; Interp is the bilinear interpolation operator.
[0132] P3.1.2 Multi-dimensional comprehensive scoring and optimal selection: For each data source s at each site location xi, calculate the multi-dimensional comprehensive score:
[0133] In the formula: For data source s at site location A comprehensive score across multiple dimensions; The data source quality score calculated in step P2.1; For time decay weights, Δt_s is the time difference between the observation time and the analysis time of data source s; Spatial representativeness rating – direct observations at the site =1.0, and the value is taken from 0.55 to 0.85 when interpolating the grid points. It is determined according to the ratio of the grid resolution to the representative scale of the site. The higher the resolution, the larger the value.
[0134] For each site location, the value with the highest overall score from all available data sources is selected as the site's merge value:
[0135] In the formula: For the site Rainfall fusion value at location (mm); s * Number the data source with the highest overall score; For the optimal data source s * Rainfall (mm) at this station; This indicates selecting from all available data sources. The data source index that has reached its maximum value.
[0136] P3.1.3 Enhanced Site Dataset Construction: The fused station precipitation data is combined with the original station metadata (station number, station name, latitude and longitude, altitude, network, etc.) to form an enhanced station dataset.
[0137] This dataset has the following advantages compared to the original site data: (a) Higher coverage – grid products can supplement coverage when the original site is missing; (b) Higher quality – the value for each site comes from the current best data source; (c) Better temporal continuity – reducing data gaps through multi-source complementarity.
[0138] Path 2: Generation of precipitation grid dataset based on grid interpolation fusion P3.2.1 Gridding of site data: The calibrated station observation data from step P2 are interpolated onto the target grid.
[0139] An adaptive interpolation method considering land-sea distribution is adopted: For the target grid point Search for valid site observations in the surrounding area and calculate the interpolation weights:
[0140] In the formula: For station k, pair of target grid points Normalized interpolation weights; From station k to target grid point Geographical distance (km); It is a Gaussian distance decay function, where L is the radius of influence (for precipitation, L = 100 km, which is shorter than that for air temperature because precipitation has a greater spatial variability). Matching factor for land and sea types: = 1.0: The site and grid points have the same land and sea type; β: The land and sea types of the stations and grid points are different.
[0141] In the formula: The surface type of the location of station k (determined according to step P1.4); For target grid points The land surface type at the location; β is the cross-type attenuation coefficient, which is a positive number less than 1 (preferably 0.2~0.5 for precipitation), used to suppress cross-type propagation of values from different land surface types.
[0142] The gridded values for the site data are:
[0143] In the formula: Interpolate site observation data to target grid points Precipitation value (mm) at location; For station k, the normalized interpolation weights for the target grid points; Σ represents the calibrated precipitation observation value (mm) at station k; k Summing all valid sites within the search radius.
[0144] P3.2.2 Multi-source grid data fusion: The gridded data from the sites is then weighted and fused with the real-time data from each grid point collected in step P1:
[0145] In the formula: For target grid points Rainfall fusion value (mm) at the location; After gridding the site data Rainfall at the location (mm); For grid products s in Rainfall at the location (mm); The fusion weight of the gridded data of a site is related to the number of valid sites around that grid point. Positive correlation:
[0146] In the formula: The baseline weight for site data (2.5 for precipitation); For grid points Number of effective sites within the surrounding influence radius; N ref The number of reference sites can be set to 5 to control the rate of weight saturation.
[0147] The fusion weight of each grid point product is determined by its quality score:
[0148] In the formula: For grid product s at the target grid The fusion weight at the location; For this data source in The overall quality score under the region type (calculated by step P2.1).
[0149] P3.2.3 Grid Dataset Construction: The fused grid data constitutes a complete global coverage grid dataset with a resolution equal to that of the target grid (e.g., 10 km).
[0150] This dataset combines the high precision of site observations with the wide coverage of gridded products. In densely populated areas, it prioritizes site data, while in sparsely populated areas (such as oceans and plateaus), it prioritizes gridded products, thus achieving adaptive weight allocation.
[0151] Step P4: Post-processing and quality control of precipitation across land and sea zones Post-processing and rigorous quality control were performed on the dual-path fusion results, dividing them into land and sea zones. P4.1 Consistency handling of land-sea boundaries: For the gridded dataset, examine the precipitation gradient between land and sea grid points near the coastal zone; if the gradient exceeds a physically reasonable threshold, perform smoothing processing. Precipitation gradient threshold: ΔP / Δd ≤ 0.5 mm / km Grid points exceeding the threshold are smoothed using a neighborhood-weighted average.
[0152] The above thresholds are preferred values and can be adjusted according to regional climate characteristics and seasonal configurations.
[0153] Similar to step T4.1, this gradient threshold is used to identify isolated non-physical jumps caused by abnormal data sources. In practice, it combines the consistency of adjacent time intervals and the consistency of neighborhood grid points to jointly determine the gradient. Smoothing is only triggered for abnormal gradients that do not have spatiotemporal continuity, so as to avoid erasing the true spatial gradient characteristics of precipitation (such as coastal frontal precipitation).
[0154] P4.2 Physical rationality check: Ensure that the merged precipitation value is non-negative, and correct any minor negative values that may occur during the merging process to 0; check the rationality of the spatial distribution of precipitation, and remove isolated unreasonable precipitation extremes; mark the removed grid points as missing data, and fill them back with the weighted average of the effective grid point values in their neighborhood or the value of the grid point product at that location.
[0155] P4.3 Consistency cross-validation and loop closure for the two datasets: Cross-compare the site dataset generated by path one with the grid dataset generated by path two, extract the values of the grid dataset at the site location, and compare them with the site fusion values.
[0156] If the difference exceeds the preset threshold (precipitation: 5mm / h), the same closed-loop processing procedure as step T4.3 is initiated: suspicious sites / grids are marked, the quality scores and deviation correction records of each data source at the site location are traced back, and the data source located that causes the difference to exceed the threshold is determined as an abnormal data source (the auxiliary determination conditions are the same as in step T4.3(c)). It is downweighted or removed, and the dual-path fusion of step P3 is re-executed based on the remaining data sources, and a re-examination is performed; if the re-examination still exceeds the threshold, it is marked as low quality or missing and the event log is recorded for the data source quality feature statistics update in step P2.1.
[0157] Step P5: Precipitation Product Quality Assessment and Output P5.1 Product Quality Assessment: (a) Independent site verification: 10%-20% of sites are reserved not to participate in the integration process and are used as independent verification data.
[0158] By comparing the fusion product with independent site observations, the bias, root mean square error (RMSE), correlation coefficient (R), and fair skill score (ETS) were calculated.
[0159] (b) Cross-data source validation: The site dataset of Path 1 and the grid dataset of Path 2 are cross-validated to evaluate the consistency of the two products.
[0160] (c) Coverage statistics: The improvement in spatial coverage and temporal continuity of statistical fusion products compared to single data sources.
[0161] P5.2 Product Output: Output two sets of standard format precipitation data analysis products: (a) Enhanced site dataset: in CSV / BUFR format, containing fields such as station number, latitude and longitude, altitude, precipitation merge value, data source identifier, and quality score; (b) Fusion grid dataset: in NetCDF / GRIB2 format, containing a 10km resolution precipitation grid analysis field, quality label field, data source contribution ratio field, etc.
[0162] Here are two specific examples of temperature and precipitation: 1. Global hourly temperature fusion analysis This example uses a global range (90°S-90°N, 180°W-180°E) to illustrate the specific implementation process of the temperature element fusion analysis method of the present invention.
[0163] The target grid resolution is 0.1°×0.1° (approximately 10km), and the analysis is performed hourly.
[0164] 1.1 Collection and Preprocessing of Multi-Source Temperature Data (Step T1) 1.1.1 List of data sources for site observation:
[0165] 1.1.2 Grid Point Real-Time Product List:
[0166] 1.1.3 Quality control parameter settings: Temperature extreme value check threshold: -70°C to +60°C Missing test value code: uniformly -9999 1.1.4 Setting parameters for land surface type zoning determination:
[0167] Example: For a certain island station, the proportion of ocean area within a 5km window is p=0.55, which is determined to be a land-sea transition zone; for a certain GTS station in the North China Plain, p=0.0, which is determined to be land; for a certain grid point in the western Pacific Ocean, p=1.0, which is determined to be ocean.
[0168] 1.2 Multidimensional Spatiotemporal Calibration of Temperature (Step T2) 1.2.1 Data source quality scoring weighting coefficients:
[0169] In this example, the statistical scope for each sub-indicator is as follows: the statistical period is the most recent 30 days; data coverage C sThe proportion of the average daily number of valid stations (grid points) of each data source within the corresponding regional type during this period to the total number of observations; temporal continuity T s The proportion of valid observations within this period to the total number of hourly observations required; Spatial consistency S s The Pearson correlation coefficient between each data source and the nearest reference observation sequence (the mean sequence of the products of the 5 nearest other stations / grid points), with negative values taken as 0; historical bias B s B represents the average deviation relative to ERA5 within the recent 90-day sliding window. max Set the temperature to 5°C.
[0170] 1.2.2 Typical quality scores for each data source (temperature):
[0171] Note: "—" indicates that the data source has no data in this area.
[0172] 1.2.3 Systematic Deviation Correction Parameters (Ambient Temperature Addition Correction): Using ERA5 as a reference, statistics are divided by month and land surface type:
[0173] 1.2.4 Time decay scale parameter (air temperature):
[0174] 1.2.5 Spatial Dimension Calibration Parameters (Temperature): Vertical temperature lapse rate: Γ = 0.0065°C / m The grid representativeness correction adopts a scaling method with a scaling factor of 1.1 (determined by the ratio of the grid product in the overlapping area to the standard deviation of the site observations, and the scaling factor of the grid value is increased by 10% to match the site scale). 1.3 Intelligent Fusion of Dual Temperature Paths (Step T3) Path 1: Generating Temperature Station Datasets 1.3.1 Multi-dimensional scoring parameters: Spatial representativeness score D s (x i Value selection rules:
[0175] 1.3.2 Optimal Selection Implementation Example: Taking the temperature at a certain GTS station (station number 54511, Beijing, 39.8°N, 116.5°E) at 12 UTC on July 15, 2024 as an example:
[0176] Note: This timeframe is near the MODIS transit timeframe (the actual time difference between this MODIS timeframe and the analysis timeframe is approximately 40 minutes). MODISLST is a low temporal resolution product that is analyzed twice daily. According to the special case rule in step T2.3(a), it is not subject to the ±30-minute window exclusion rule. The most recent valid timeframe is used for fusion, and the time decay weight W is calculated based on the actual time difference. t =exp(-40 2 / (2×60 2 ))≈0.8.
[0177] The highest overall score was achieved by direct observation from GTSSYNOP (0.900), with the fused temperature at this station at this time being 34.2°C.
[0178] If a site has missing GTS data at a certain time (and that time is not during MODIS transit, MODIS will not participate in the fusion):
[0179] The ERA5 grid interpolation (0.638) is automatically selected, and the fused temperature is taken as 33.8°C.
[0180] This reflects the mechanism by which grid products automatically fill in the gaps when a site is missing data.
[0181] Path 2: Generation of Temperature Grid Dataset 1.3.3 Station gridded interpolation parameters (temperature):
[0182] When there are no valid stations within a range that meets the minimum number of stations requirement around a grid point, the grid point will not participate in the station gridging and will only be determined by the grid product fusion.
[0183] 1.3.4 Grid fusion weight parameters (temperature):
[0184] The site baseline weight is set relatively high (>1.0) to ensure that in densely populated areas, site data has a dominant influence on the grid fusion results.
[0185] 1.3.5 Example of grid fusion implementation (temperature): Example 1: Ocean grid (20.0°N, 130.0°E, Western Pacific) No observations were taken within 200 km of this grid point: N sta =0, W sta =0
[0186] Only one grid point temperature product, ERA5, is available for this ocean area, with a combined temperature of 28.5°C.
[0187] This example illustrates how gridded products determine the fusion value in marine areas without site coverage.
[0188] Example 2: Land grid (35.0°N, 115.0°E, North China Plain) There are 8 valid stations within 200km of this grid point: N sta =8 W sta =3.0×[1-exp(-8 / 5)]=3.0×0.798=2.39 The gridded temperature at the station is 33.5°C.
[0189] Fusion temperature = (2.39 × 33.5 + 0.75 × 32.8) / (2.39 + 0.75) =(80.07+24.60) / 3.14=33.34°C This example illustrates that in densely populated areas, site data dominates the fusion result (the site weight of 2.39 is much greater than the grid product weight of 0.75).
[0190] 1.4 Post-processing of temperature between land and sea zones (step T4) 1.4.1 Land-sea boundary gradient threshold (temperature):
[0191] 1.4.2 Cross-validation parameters (temperature):
[0192] 1.4.3 Example of Implementation of Cross-Check Closed-Loop Processing (Temperature): At a certain time, the fused temperature of path 1 at an island station (in the land-sea transition zone) was 26.5°C (selected from direct observations at the island station). The extracted value of path 2 grid points at that station location was 31.2°C, a difference of 4.7°C exceeding the 3°C threshold. Therefore, closed-loop processing was initiated. (a) Marking: The site and its corresponding grid point are marked as suspicious; (b) Retrospective analysis: By retrieving records from various data sources, it was found that the timestamps of the observations at the island station fell within the ±30-minute matching window due to system clock drift, while the actual observation time lagged behind the analysis time by about 55 minutes, resulting in the time decay weight being miscalculated as a high value. (c) De-weighting / removal: The data source of this island site has been traced back and located as the data source causing the difference, and is judged as an abnormal data source, and is removed in this instance; (d) Re-fusion: Path 1 is re-optimized, selecting ERA5 grid interpolation of 30.8°C; Path 2 re-executes gridding and weighted fusion, and the grid fusion value is updated to 30.9°C; (e) Re-inspection: The temperature difference between the two sets of products at this site is 0.1°C, which is below the threshold, so the suspicious label is removed; at the same time, the clock drift event of this island station is recorded in the log, and the time continuity T of this station in the subsequent statistical period is recorded. s The corresponding reduction.
[0193] This example illustrates that Path 1 (point-by-point optimization) cannot detect "abnormal site observations that have been mistakenly assigned high time weights", and Path 2 (area weighting) also cannot locate the source of the anomaly independently. Only cross-validation of the two independent paths can trigger self-correction for this type of data source anomaly.
[0194] 1.5 Temperature Quality Assessment (Step T5) 1.5.1 Independent Validation Scheme: 15% of GTS sites and 20% of buoy sites were randomly selected as independent verification sites and did not participate in the fusion.
[0195] 1.5.2 Evaluation Index System (Temperature): Bias: Bias = (1 / N) × Σ(Ai - Oi) Root mean square error:
[0196] Correlation coefficient: R 1.5.3 Target accuracy indicators (temperature, design specifications): The table below shows the design target accuracy of the air temperature fusion product of this invention. The final accuracy will be based on the actual test results during operational trial.
[0197] 1.5.4 Coverage Improvement Design Target (Temperature):
[0198] 2. Fusion Analysis of Global Hourly Precipitation Data This example uses a global range (90°S-90°N, 180°W-180°E) to illustrate the specific implementation process of the precipitation element fusion analysis method of the present invention.
[0199] The target grid resolution is 0.1°×0.1° (approximately 10km), and the analysis is performed hourly.
[0200] 2.1 Collection and Preprocessing of Multi-Source Precipitation Data (Step P1) 2.1.1 List of data sources for site observation:
[0201] Note: Buoy precipitation observations are sparse; the actual number of available stations depends on data acquisition.
[0202] 2.1.2 Grid Point Real-Time Product List:
[0203] 2.1.3 Quality control parameter settings: Extreme precipitation threshold: 0 to 500 mm / h Missing test value code: uniformly -9999 2.1.4 Setting parameters for land surface type zoning determination: The study employed a GSHHG land-sea mask, a 5km decision window, and a 0.2 / 0.8 percentage threshold.
[0204] 2.2 Multidimensional Spatiotemporal Calibration of Precipitation (Step P2) 2.2.1 Weighting coefficients for data source quality scoring:
[0205] The statistical scope of each sub-indicator is the same as in 1.2.1, where the historical deviation Bs is calculated as the relative deviation (the ratio of the absolute value of the deviation to the mean of the reference benchmark). max Take the normalized value corresponding to a relative deviation of 50%.
[0206] 2.2.2 Typical quality scores (precipitation) for each data source:
[0207] Note: "—" indicates that the data source has no data in this area.
[0208] 2.2.3 Systematic bias correction parameters (precipitation multiplicative correction): Precipitation multiplication correction factor (statistics by month and land surface type):
[0209] 2.2.4 Time decay scale parameter (precipitation):
[0210] 2.2.5 Spatial Dimension Calibration Parameters (Precipitation): The grid representativeness correction adopts the variability scaling method, with a variability adjustment coefficient of 1.2 (scaling up the grid value variability by 20% to match the precipitation variability characteristics at the station scale). 2.3 Intelligent Fusion of Two Precipitation Pathways (Step P3) Path 1: Generation of Precipitation Station Datasets 2.3.1 Multi-dimensional scoring parameters: Spatial representativeness score Ds (x i Value selection rules:
[0211] 2.3.2 Optimal Selection Implementation Example: Taking the precipitation at a certain GTS station (station number 58362, Shanghai, 31.4°N, 121.5°E) at 12 UTC on July 15, 2024 as an example:
[0212] The highest overall score was achieved by direct observation at GTSSYNOP (0.900), with the fused precipitation rate at this station for this time period being 8.5 mm / h.
[0213] If the site has missing GTS data at a certain time:
[0214] The GPMIMERG grid interpolation (0.663) is automatically selected, and the fused precipitation is taken as 7.2 mm / h.
[0215] This reflects the mechanism by which grid products automatically fill in the gaps when a site is missing data.
[0216] Path 2: Generation of Precipitation Grid Dataset 2.3.3 Station gridded interpolation parameters (precipitation):
[0217] When there are no valid stations within a range that meets the minimum number of stations requirement around a grid point, the grid point will not participate in the station gridging and will only be determined by the grid product fusion.
[0218] 2.3.4 Grid fusion weight parameters (precipitation):
[0219] The site baseline weight is set relatively high (>1.0) to ensure that in densely populated areas, site data has a dominant influence on the grid fusion results.
[0220] 2.3.5 Example of grid fusion implementation (precipitation): Example 1: Ocean grid (20.0°N, 130.0°E, Western Pacific) No observations were taken within 100 km of this grid point: N sta =0, W sta =0
[0221] Combined precipitation = (0.80×2.5 + 0.72×3.1 + 0.70×2.8) / (0.80 + 0.72 + 0.70) =(2.00+2.23+1.96) / 2.22=2.79mm / h This example illustrates the quality-weighted fusion of gridded products in an ocean area without site coverage.
[0222] Example 2: Land grid (35.0°N, 115.0°E, North China Plain) There are 8 valid stations within 100km of this grid point: N sta =8 W sta =2.5×[1-exp(-8 / 5)]=2.5×0.798=2.00 Gridded precipitation at the monitoring station = 5.2 mm / h
[0223] Combined precipitation = (2.00×5.2+0.75×4.0+0.78×4.8) / (2.00+0.75+0.78) =(10.40+3.00+3.74) / 3.53=4.86mm / h This example illustrates that in densely populated areas, site data dominates the fusion result (the site weight of 2.00 is much greater than the grid product weight).
[0224] 2.4 Post-processing of precipitation in land-sea zones (step P4) 2.4.1 Land-sea boundary gradient threshold (precipitation):
[0225] 2.4.2 Physical rationality check: Ensure that the merged precipitation value is non-negative, and correct any minor negative values to 0; remove unreasonable extreme value grid points as missing data points, and fill them with the weighted average of the effective grid point values in the neighborhood or the value of the grid point product at that location.
[0226] 2.4.3 Cross-validation parameters (precipitation):
[0227] 2.5 Precipitation quality assessment (Step P5) 2.5.1 Independent Validation Scheme: 15% of GTS sites will be randomly selected as independent verification sites and will not participate in the integration.
[0228] 2.5.2 Evaluation Index System (Precipitation): Bias: Bias = (1 / N) × Σ(Ai - Oi) Root mean square error:
[0229] Correlation coefficient: R Fairness skill rating: ETS = (HH random ) / (H+M+FH random ) Where H is the hit count, M is the miss count, F is the null count, and H is the false alarm count. random =(H+M)(H+F) / N.
[0230] 2.5.3 Target accuracy indicators (precipitation, design indicators): The table below shows the design target accuracy of the precipitation fusion product of this invention. The final accuracy will be based on the actual measurement and verification results during operational trial runs:
[0231] 2.5.4 Coverage Improvement Design Target (Precipitation):
[0232] Example 2: This embodiment is a multi-dimensional spatiotemporal calibration-based intelligent fusion system for land and sea multi-source real-time data, used to implement the method of Embodiment 1.
[0233] The system includes: The data collection and preprocessing module is used to collect multi-source real-time data of target meteorological elements and divide the land surface type corresponding to the multi-source real-time data into land, ocean and land-sea transition zones; the multi-source real-time data includes station observation data and gridded real-time product data; The multi-dimensional spatiotemporal calibration module is used to perform data source quality assessment, systematic deviation correction, time dimension calibration, and spatial dimension calibration on the preprocessed multi-source real-time data. The dual-path intelligent fusion module is used to generate an enhanced site dataset by selecting the optimal value from the valid values of each data source at each site location based on the processed multi-source real-time data using a site fusion path; at the same time, it uses a grid fusion path to interpolate the site observation data to the target grid and weightedly fuse it with the grid real-time product data. The post-processing module is used to perform land-sea boundary gradient verification, smoothing, and physical rationality checks on the fused grid dataset. The cross-validation and feedback fusion module is used to perform cross-validation on the enhanced site dataset and the fused grid dataset, and trigger feedback correction on abnormal data sources based on the validation results.
[0234] This invention can also be implemented in the following alternative ways, all of which fall within the protection scope of this invention: (1) Interpolation operator substitution: Bilinear interpolation operator can be replaced by Kriging interpolation, inverse distance weighted interpolation or natural neighborhood interpolation; Gaussian distance decay function in site gridding can be replaced by exponential decay or inverse distance squared decay function.
[0235] (2) Alternative method for determining scoring weights: Weighting coefficient w of data source quality scoring c w t w s w b In addition to manual setting, the weights can also be determined by training with machine learning methods (such as linear regression, random forest, gradient boosting tree) based on historical test samples, so that the weights can be adaptively adjusted according to the region and season.
[0236] (3) Generalization of land-sea type matching factor: land-sea type matching factor f type It can be generalized from a binary piecewise form to a continuous function: the attenuation coefficient is calculated based on the difference between the land-sea ratio (continuous land-sea mask value) of the station and the target grid point, and β is a monotonically decreasing function of the land-sea ratio difference, thus more finely depicting the spatial gradient characteristics of the land-sea transition zone.
[0237] Accordingly, the percentage threshold (0.2 / 0.8) and window radius R0 for determining the land surface type zoning can be adjusted according to the mask resolution and the complexity of the regional coastline.
[0238] (4) Reference benchmark replacement: The reference benchmark for systematic bias correction can be replaced by ERA5 with other reanalysis products (such as MERRA-2, CFSR) or high-density site fusion analysis field.
[0239] (5) Element expansion: The dual-path fusion framework of the present invention can be extended to the fusion analysis of other meteorological elements such as humidity, wind, and air pressure. Only the deviation correction method (addition / multiplication), time decay scale, influence radius and various threshold parameters need to be adjusted according to the physical characteristics of the elements.
[0240] (6) System implementation: Each sub-module of this system can be deployed on a single server or a distributed computing cluster, and the pipeline operation of data source access and product distribution is realized through message queue.
[0241] Obviously, the above embodiments are merely examples for clear illustration and are not intended to limit the implementation.
[0242] For those skilled in the art, other variations or modifications can be made based on the above explanation.
[0243] It is neither necessary nor possible to exhaustively list all possible implementation methods here.
[0244] However, any obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A method for intelligent fusion of multi-source real-time data from land and sea based on multi-dimensional spatiotemporal calibration, characterized in that, comprising the following steps: Step 1: collect multi-source real-time data of target meteorological elements, and divide the surface types corresponding to the multi-source real-time data into land, ocean and sea-land transition zone; the multi-source real-time data includes station observation data and grid real-time product data; Step 2: perform data source quality assessment, systematic deviation correction, time dimension calibration and space dimension calibration on the preprocessed multi-source real-time data; Step 3: based on the processed multi-source real-time data, adopt a station fusion path, for each station position, select the optimal value from the valid values of each data source at the station position to generate an enhanced station dataset; at the same time, adopt a grid fusion path, interpolate the station observation data to the target grid, and perform weighted fusion with the grid real-time product data to generate a fused grid dataset; Step 4: perform sea-land boundary gradient test, smoothing processing and physical rationality check on the fused grid dataset; Step 5: perform cross validation on the enhanced station dataset and the fused grid dataset, and trigger feedback correction for abnormal data sources according to the test result.
2. The method according to claim 1, characterized in that, In Step 1, the surface type division includes: calculating the proportion p of ocean area within a preset radius window around the target position; when p ≤ the first preset threshold, the position is determined as land; when p ≥ the second preset threshold, the position is determined as ocean; when the first preset threshold < p < the second preset threshold, the position is determined as a sea-land transition zone.
3. The method according to claim 1, characterized in that, In Step 2, the data source quality assessment includes calculating a comprehensive quality score: ; Among them, Q s (r) represents the overall quality score of data source s under land surface type r; C s (r) represents the coverage rate, which is the proportion of the number of valid observations of data source s within surface type r to the total number of observations to be made during the statistical period. T s For time continuity, it is the proportion of the number of valid observations of data source s within the statistical period to the total number of observations that should be made in that period; S s For spatial consistency, is the correlation coefficient between the observation time series of data source s and the reference observation series in the spatial neighborhood; B s Historical deviation is the average deviation of data source s relative to the reference benchmark within the sliding statistical window. B max This serves as a reference value for deviation normalization. w c w t w s w b These are the weighting coefficients for coverage, temporal continuity, spatial consistency, and historical bias, respectively.
4. The method according to claim 1, characterized in that, In Step 2, the time dimension calibration includes: setting a unified time matching window, and assigning a time decay weight to the data source that has observations within the time window: ; Where, Δt k Let k be the time difference between the kth observation and the analysis. The preset time decay scale parameter; for a data source whose time resolution is lower than the analysis time interval, it is not constrained by the exclusion rule of the time matching window, the valid time closest to the analysis time is taken to participate in fusion, and the time decay weight is calculated according to the actual time difference.
5. The method according to claim 1, characterized in that, In Step 2, the space dimension calibration includes: performing vertical lapse rate correction on station observations according to the altitude difference; performing statistical downscaling on the grid real-time products to match the variability characteristics of the station scale, wherein the statistical downscaling adopts variability scaling or quantile mapping.
6. The method according to claim 1, characterized in that, In Step 3, the station fusion path includes: for each station position, aggregate all data sources that have valid observations or grid values at the position; for grid products, extract the grid values corresponding to each station position through spatial interpolation; for each data source at each station position, calculate a multi-dimensional comprehensive score according to the comprehensive quality score, time decay weight and spatial representativeness score of each data source; the spatial representativeness score is used to characterize the degree to which the data source represents the true value of the station position, takes the highest value for direct station observation, and is determined according to the grid resolution for grid interpolation; for each station position, select the value of the data source with the highest multi-dimensional comprehensive score from all available data sources as the fusion value of the station: combining the fused target meteorological element data with the original station metadata to form an enhanced station dataset.
7. The method according to claim 1, characterized in that, In Step 3, the grid fusion path includes: The calibrated station observation data is interpolated to the target grid, and a land-sea type matching factor is introduced into the interpolation weight to form station gridded data. The land-sea type matching factor is: when the land surface type of the station and the target grid is the same, the first weight value is taken; when they are different, the second weight value is taken, which is less than the first weight value. The gridded data of the sites is weighted and fused with the real-time product data of the grids to generate a fused grid dataset. The fusion weight of the gridded data of the sites is positively correlated with the number of valid sites around the target grid, and the fusion weight of the real-time product data of the grids is the comprehensive quality score.
8. The method according to claim 7, characterized in that, The fusion weight of the gridded site data is: ; In the formula: W sta,0 The baseline weight for site data; N sta (x g ) is the grid point x g Number of effective sites within the surrounding influence radius; N ref Number of sites for reference.
9. The method according to claim 1, characterized in that, Step five includes: The values at the site locations of the enhanced site dataset and the fused grid dataset are compared, and sites and their corresponding grid points with differences exceeding a preset threshold are marked as suspicious. By tracing back the processing records of each data source at this site, the source of the discrepancy can be identified. Data sources that cause the difference to exceed a preset threshold will be identified as abnormal data sources and will be downgraded or removed. Based on the remaining data sources, re-execute steps three and four to update the fusion values of the sites and corresponding grid points; If the re-inspection is passed, the marking will be removed. If the re-inspection still exceeds the threshold, it will be marked as low quality or missing.
10. A multi-dimensional spatiotemporal calibration-based intelligent fusion system for real-time data from multiple land and sea sources, characterized in that: include: The data collection and preprocessing module is used to collect multi-source real-time data of target meteorological elements and divide the land surface type corresponding to the multi-source real-time data into land, ocean and land-sea transition zones; the multi-source real-time data includes station observation data and gridded real-time product data; The multi-dimensional spatiotemporal calibration module is used to perform data source quality assessment, systematic deviation correction, time dimension calibration, and spatial dimension calibration on the preprocessed multi-source real-time data. The dual-path intelligent fusion module is used to generate an enhanced site dataset by selecting the optimal value from the valid values of each data source at each site location based on the processed multi-source real-time data using a site fusion path; at the same time, it uses a grid fusion path to interpolate the site observation data to the target grid and weightedly fuse it with the grid real-time product data. The post-processing module is used to perform land-sea boundary gradient verification, smoothing, and physical rationality checks on the fused grid dataset. The cross-validation and feedback fusion module is used to perform cross-validation on the enhanced site dataset and the fused grid dataset, and trigger feedback correction on abnormal data sources based on the validation results.
