A GNSS-based optimization method and system for improving the precision of CMIP6 water vapor

Abstract

This invention proposes an optimization method and system for improving the accuracy of CMIP6 water vapor data based on GNSS, belonging to the field of water vapor accuracy prediction. The method includes: acquiring GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data; processing the acquired data: spatially, interpolating the GNSS observation data onto the station and grid; temporally, downsampling the GNSS observation data; using CMIP6 water vapor data, latitude, longitude, altitude, and annual day of birth of the location as input parameters, and GNSS observation data as the target parameter output; training the data using machine learning algorithms to construct a CMIP6 water vapor enhancement model based on GNSS-PWV; and inputting the measured data into the constructed model to obtain the enhanced CMIP6 water vapor data.

Description

Technical Field

[0001] This invention belongs to the field of water vapor accuracy prediction, and in particular relates to an optimization method and system for improving the water vapor accuracy of CMIP6 based on GNSS. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, numerous studies have extensively explored the historical reconstruction and future projection of climate change based on simulation results from the Coupled Model Intercomparison Project (CMIP6) Phase 6 model. Although the CMIP6 model has played a crucial role in climate change simulation and projection, its simulation results contain certain errors compared to observational references. For example, Tian's research indicates that simulated annual average precipitation is significantly overestimated in the southwestern river basins of China, while it is underestimated in the high-altitude regions of the northwest. A subset of the CMIP6 model can reproduce transient subsurface shortwave radiation well in some regions, but performs poorly in others. A series of studies have shown that the CMIP6 model plays a key role in global climate change simulation and prediction, but its simulation accuracy still exhibits certain regional and systematic biases.

[0004] Current research indicates that most studies utilizing CMIP6 to explore meteorological and climate evolution in study areas struggle to reduce errors. Furthermore, most studies merely integrate and average multiple global climate models to minimize CMIP6 model errors, without leveraging external constraints to reduce errors and improve accuracy. Especially driven by big data (parameterized schemes of models designed and designed by experts and institutions worldwide), there is a greater need to enhance and refine the accuracy of the CMIP6 model, breaking down the barriers of uncertainty in big data predictions and providing policymakers with more reliable information.

[0005] Global Navigation Satellite System (GNSS) meteorology is rapidly developing. GNSS has the advantages of high precision and high temporal resolution, but it is limited by the distribution of stations and spatial discontinuities; remote sensing has high spatial resolution and spatial continuity, but its revisit cycle is relatively poor. Based on these advantages and disadvantages, domestic and international research has integrated GNSS and remote sensing, resulting in many achievements that mutually improve water vapor accuracy between the two technologies.

[0006] In the context of the big data era, the spatiotemporal evolution of future water vapor and its high-precision simulation have become core challenges and key research areas in global climate change research. While traditional observation methods (such as GNSS and remote sensing) can provide high-precision instantaneous water vapor information, they are insufficient to meet the needs of long-term climate prediction. Global climate models (such as CMIP6) can simulate water vapor changes under future climate scenarios, but their spatiotemporal resolution is low and they suffer from regional simulation biases.

[0007] When using global climate models to perform numerical simulations, there are limitations such as regional simulation bias and insufficient observational constraints. In addition, there are problems such as insufficient spatiotemporal resolution and difficulty in accurately capturing water vapor changes. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, this invention provides an optimization method for improving the accuracy of CMIP6 water vapor simulation based on GNSS. This method constructs a water vapor model with high spatiotemporal resolution and high accuracy to improve the accuracy and robustness of water vapor simulation at the regional scale, thus providing a more reliable data foundation for climate change research and disaster prediction.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] Firstly, an optimization method for improving the water vapor accuracy of CMIP6 based on GNSS is disclosed, including:

[0011] Acquire GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data;

[0012] The acquired data is processed as follows: spatially, the target data is interpolated onto the station and grid; temporally, the target data is downsampled.

[0013] The CMIP6 water vapor data, the latitude, longitude, altitude, and annual day of CMIP6 water vapor are used as input parameters, and GNSS observation data is used as the output target parameter. The data are trained using machine learning algorithms to construct a CMIP6 water vapor enhancement model based on GNSS-PWV.

[0014] The measured data is input into the constructed model to obtain improved CMIP6 water vapor data.

[0015] Secondly, an optimization system for improving the water vapor accuracy of CMIP6 based on GNSS is provided, including:

[0016] The data acquisition module is configured to acquire GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data.

[0017] The data processing module is configured to process the acquired data: spatially, interpolate the target data onto the station and grid, and temporally, downsample the target data.

[0018] The model building module is configured to take CMIP6 water vapor data, the latitude corresponding to CMIP6 water vapor, the longitude corresponding to CMIP6 water vapor, the altitude corresponding to CMIP6 water vapor, and the annual day corresponding to CMIP6 water vapor as input parameters, and output GNSS observation data as target parameters. It uses machine learning algorithms to train the data and builds a CMIP6 water vapor enhancement model based on GNSS-PWV.

[0019] The water vapor accuracy enhancement module is configured to input the data to be measured into the constructed model to obtain enhanced CMIP6 water vapor data.

[0020] The above one or more technical solutions have the following beneficial effects:

[0021] This invention incorporates GNSS water vapor data into the CMIP6 climate model and utilizes machine learning algorithms to improve the accuracy of CMIP6 water vapor data. Global Navigation Satellite System (GNSS) observations, with their high precision and high spatiotemporal resolution, provide reliable data support for PWV retrieval. By applying high-precision GNSS PWV data to correct the CMIP6 global climate model's water vapor data, the systematic bias of the model in different regions and under different climatic backgrounds can be effectively reduced, further guiding the optimization of model biases.

[0022] The technical solution of this invention proposes an improved method for improving the accuracy of CMIP6 water vapor observations using GNSS observations based on machine learning algorithms. This method not only enhances the model's ability to reconstruct historical data but also helps improve the reliability of CMIP6 in future weather forecasts and climate change predictions, providing a more solid scientific basis for risk assessment and adaptation strategy formulation for extreme weather and climate events.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a flowchart of the method according to an embodiment of the present invention;

[0026] Figure 2 The image shows a scatter plot of PWV values ​​for the training and test sets of the GNSS stations of this invention from 2020 to 2023, where (a) is the training set of the CNN algorithm; (b) is the test set of the CNN algorithm; (c) is the training set of the XGBoost algorithm; (d) is the test set of the XGBoost algorithm; (e) is the training set of the LSTM algorithm; and (f) is the test set of the LSTM algorithm.

[0027] Figure 3 The RMSE performance of CNN, XGBoost and LSTM models in improving PWV on training and test sets under different seasons;

[0028] Figure 4 The RMSE performance of CNN, XGBoost and LSTM models at different altitudes in terms of PWV improvement on training set, test set and unified station set;

[0029] Figure 5 Scatter plot of Improved_CMIP6-PWV and GNSS-PWV for validation set 1;

[0030] Figure 6 Scatter plot of Improved_CMIP6-PWV and ERA5-PWV for validation set 2;

[0031] Figure 7 Technical roadmap of sub-technical solutions in embodiments of the present invention. Detailed Implementation

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0034] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0035] Terminology Explanation:

[0036] PWV (Precipitable Water Vapor): Atmospheric precipitable water;

[0037] CMIP6 (Coupled Model Intercomparison Project Phase 6): The sixth international coupled model intercomparison project;

[0038] GNSS (Global Navigation Satellite System)

[0039] GCM (Global Climate Model)

[0040] CNN (Convolutional Neural Network): A neural network;

[0041] XGBoost (eXtreme Gradient Boosting): Extreme gradient boosting;

[0042] LSTM (Long Short-Term Memory): Long Short-Term Memory network;

[0043] ERA5 (European Centre for Medium-Range Weather Forecasts ReAnalysis v5): The fifth generation of reanalysis datasets from the European Centre for Medium-Range Weather Forecasts.

[0044] RMSE (Root Mean Square Error): Root mean square error.

[0045] The continuous improvement and robust development of the Global Navigation Satellite System (GNSS) has led to its widespread application in meteorology. One of its most typical applications is the acquisition of Precipitable Water Vapor (PWV) information, a crucial meteorological parameter that plays a vital role in weather forecasting, climate monitoring, and hydrological simulation. GNSS-PWV inversion is primarily based on the delay caused by water vapor as satellite signals pass through the atmosphere, particularly the Zenith Tropospheric Delay (ZTD), thereby achieving high temporal resolution and all-weather water vapor observation. Compared to traditional observation methods such as radiosondes and microwave radiometers, GNSS offers advantages such as high precision, high temporal resolution, low cost, continuous operation, and global coverage, making it particularly suitable for regions where conventional observations are scarce. In recent years, GNSS meteorology has been widely applied to the study of weather processes such as heavy precipitation, typhoons, monsoon systems, earthquakes, and atmospheric rivers, demonstrating its enormous potential in capturing rapid water vapor changes and revealing atmospheric water vapor transport mechanisms.

[0046] Water vapor plays a crucial role in numerical weather prediction, and as an important indicator of water vapor, precipitable water vapor (PWV) is closely related to actual precipitation. Therefore, improving the accuracy and spatiotemporal resolution of PWV will greatly benefit meteorological observation and forecasting, as well as climate evolution analysis. Given that the newly released CMIP6 continues to simulate and predict large-scale scenarios of more frequent and intense extreme heat and precipitation events under global warming, and that certain uncertainties still exist, improving the accuracy and precision of its simulation and prediction of spatiotemporal variations in atmospheric water vapor is particularly important.

[0047] This invention proposes an optimization method for improving the water vapor accuracy of the CMIP6 global climate model based on GNSS, combining machine learning algorithms, and verifies and analyzes it in Turkey. It falls within the interdisciplinary field of atmospheric water vapor research, specifically at the intersection of GNSS and atmospheric science.

[0048] In this embodiment, CMIP6 data can be understood as a large dataset, but its predictive accuracy for the future is lower than that of observational data. GNSS data, on the other hand, is observational data with high precision. Therefore, this embodiment introduces high-precision GNSS water vapor data to improve the accuracy of water vapor data from the CMIP6 large model.

[0049] In this implementation example, the use of GNSS to enhance CMIP6 water vapor is proposed for the first time, and experimental application verification is carried out.

[0050] Example 1

[0051] This invention falls within the interdisciplinary field of atmospheric water vapor research, specifically at the intersection of GNSS and atmospheric science. This embodiment utilizes water vapor data from the HadGEM3-GC31-HM_highres-future model, part of the High Resolution Model Comparison Project (HRP), a joint initiative by the Nevada Geodetic Laboratory (NGL) GNSS PWV, the Royal Netherlands Meteorological Service (Reindert J. Haarsma), and the UK Met Office (Malcolm J. Roberts). Based on different machine learning algorithms, and specifically for the evenly distributed monitoring stations in Turkey, a research method for improving CMIP6 water vapor accuracy based on GNSS fusion is proposed. The technical route is detailed in the appendix. Figure 7 As shown.

[0052] This embodiment specifically discloses an optimization method for improving the water vapor accuracy of CMIP6 based on GNSS, including:

[0053] First, data was acquired and spatiotemporal matching of GNSS, PWV, and CMIP6 water vapor data was performed.

[0054] Secondly, we used the training set data to construct CMIP6 water vapor enhancement models under different algorithms, used the test set to verify the accuracy at the spatial scale, and explored the enhancement effect of the models on CMIP6 water vapor under different seasons and altitudes.

[0055] Then, the model is validated on a time scale;

[0056] Finally, the accuracy of the model for grid-scale CMIP6 water vapor was verified using a new validation set.

[0057] Its detailed content and explanation are as follows: Figure 1 As shown, the present invention includes the following steps:

[0058] Step 1: Data Processing. The data required for PWV acquisition in this invention includes GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data.

[0059] It should be noted that GNSS technology has the advantages of high precision, all-weather operation, and low cost. The product data comes from the Nevada Geodetic Laboratory and is designated GNSS-PWV.

[0060] The CMIP6 water vapor data comes from the HadGEM3-GC31-HM_highres-future model in the CMIP6 high-resolution model comparison program, denoted as CMIP6-PWV.

[0061] The ERA5 reanalysis water vapor data comes from the total column water vapor data of the European Centre for Medium-Range Weather Forecasts (ECMWF), denoted as ERA5-PWV.

[0062] Because the spatiotemporal resolutions of the input CMIP6 water vapor data (CMIP6-PWV) and the target water vapor parameter (GNSS-PWV, i.e., GNSS observation data) are different, spatiotemporal matching is performed as follows: Spatially, bilinear interpolation is used to interpolate the target data onto the station and grid.

[0063] Specifically, CMIP6-PWV is 50 km resolution grid data, while GNSS-PWV is point data based on ground stations. The CMIP6 model data is interpolated onto the GNSS stations using bilinear interpolation to obtain the CMIP6-PWV of all stations participating in the experiment (station_CMIP6-PWV).

[0064] On the other hand, ERA5-PWV is used as the reference ground truth in the test set and validation set 1. ERA5-PWV is grid data with a resolution of 0.25°. Bilinear interpolation is used to interpolate the ERA5 data to the GNSS station to obtain the ERA5-PWV (station_ERA5-PWV) on the station as the validation data for the test set and validation set 1.

[0065] In addition, in validation set 2, the ERA5 data is bilinearly interpolated onto the original CMIP6 grid points to obtain ERA5-PWV data (Grid_ERA5-PWV) matching the CMIP6 grid points, which is used as validation data for validation set 2.

[0066] In terms of time, the target data is downsampled and unified to eight timestamps per day: 01:30, 04:30, 07:30, 10:30, 13:30, 16:30, 19:30, and 22:30. CMIP6-PWV has a 3-hour resolution, corresponding to these eight timestamps per day; GNSS-PWV has a 300-second time resolution. Therefore, GNSS-PWV data needs to be downsampled, retaining only the eight timestamps per day from the GNSS-PWV time series itself. During matching, a conventional method of one-to-one location and one-to-one time correspondence is used. The PWV data information and uses after the spatiotemporal matching in step one are shown in Table 1 below.

[0067] Table 1. PWV data information and applications after spatiotemporal matching.

[0068]

[0069] The preprocessed data was used to train the model on the training set, test the model on the test set, and validate the model using validation sets 1 and 2. The Station_CMIP6-PWV data from the training set stations (2020-2023) was used as the input water vapor parameter for the training model, and the GNSS-PWV data from the same period was used as the output target water vapor data. This process was then used to train the model. The Station_CMIP6-PWV data from the test set stations (2020-2023) was used as the input water vapor parameter, and the GNSS-PWV and Station_ERA5-PWV data from the same test set stations were used as validation data for the test set. The Station_CMIP6-PWV data from validation set 1 stations (2024) was used as the input water vapor parameter, and the GNSS-PWV and Station_ERA5-PWV data from the same validation set 1 stations (2024) were used as validation data for validation set 1. The Grid_CMIP6-PWV data from the second validation set station in 2024 was used as the input water vapor parameter, and the Grid_ERA5-PWV data from the same station was used as the validation data. Spatial validation of the trained model was performed using the test set data, and temporal validation of the trained model was performed using the validation set data.

[0070] The advantages of this approach are as follows: First, it verifies the accuracy of the trained model on both spatial and temporal scales. Second, and this is the most innovative aspect of the invention, it is the first time that high-precision GNSS-PWV observation data has been used to train CMIP6-PWV data, which greatly reduces the error of the CMIP6 climate model data in future predictions and compensates for the low accuracy of climate models. Third, it verifies the model on both the station and grid scales to ensure the comprehensive application of both scales.

[0071] Step Two: Model Construction and Testing. This invention combines multi-source input features with different machine learning algorithms: Convolutional Neural Network (CNN), eXtreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). These are used to capture the spatial structure, nonlinear relationships, and temporal evolution patterns of the input features, respectively. The CNN model is designed as a lightweight convolutional structure with 13 layers and a batch size of 128, optimized with SDGM parameters. It is used to learn spatial distribution patterns and local neighborhood correlations, suitable for spatial error correction of large-scale grid data. The LSTM model is designed as a 5-layer LSTM network with a batch size of 64, optimized with Adam parameters. It is used to model the temporal fluctuation characteristics of diurnal water vapor, which is caused by the fluctuations in PWV content over time, including seasonal cycles, diurnal variations, abrupt weather changes, and trend changes. This effectively addresses the instability of CMIP6-PWV predictions in different seasonal periods and learns the intrinsic dynamic mechanism of PWV evolution over time. XGBoost modeling is based on a tree model structure, employing multi-feature nonlinear combinations to construct a boosting tree model that is sensitive to high-dimensional inputs and possesses strong generalization capabilities, especially demonstrating higher RMSE reduction performance in low- and mid-altitude regions. Building upon traditional meteorological factors, it integrates spatial geographic features and temporal dynamic features. Spatial geographic features, such as latitude and longitude, and temporal dynamic features, such as day-in-annual variation, range from 1 to 365, reflecting the changing patterns of seasonal climate cycles, forming a multi-dimensional input feature matrix. This multi-dimensional input feature matrix is ​​the number of samples multiplied by the number of input features, significantly enhancing the model's ability to identify regional topography and temporal periodic changes. In addition to water vapor data, geographic labels, including longitude, latitude, and altitude, are introduced as input features to enhance the model's adaptability to topographically sensitive areas. For topographically sensitive areas such as plateaus and mountains, time labels, such as day-in-annual variation, are introduced as input features to enhance the model's ability to capture temporal changes. The target output uses measured GNSS-PWV data as label values, representing the true water vapor value.

[0072] In this embodiment, time series fluctuation characteristics refer to the numerical fluctuation patterns exhibited by diurnal water vapor PWV over time, including but not limited to:

[0073] 1) Periodic fluctuations: A recurring pattern of change within similar time intervals, influenced by diurnal and seasonal climatic factors;

[0074] 2) Short-term fluctuations: Rapid rises and falls in PWV values ​​over a short period of time caused by weather systems (such as fronts, typhoons, and severe convection). LSTM networks can capture these fluctuation characteristics and learn the intrinsic dynamic mechanism of PWV evolution over time, thereby effectively reducing instability under different seasons and weather conditions in diurnal forecasting.

[0075] In this embodiment, the time dynamic feature refers to a data attribute that changes with time, reflecting the periodicity, trend, and abrupt changes of water vapor over time. In this invention, the accumulated day of the year is used as the time dynamic feature. The value range of the accumulated day of the year is 1–365, used to reflect the changing pattern of seasonal climate cycles, and the input time dynamic feature is "accumulated day of the year".

[0076] In this embodiment, the feature matrix refers to a two-dimensional array composed of multiple feature variables arranged in a uniform format, used as input into the model for calculation. The "multi-dimensional input feature matrix" of this invention consists of "n samples × 5 input parameters." The input parameters include CMIP6-PWV, the latitude, longitude, altitude, and annual day of the corresponding CMIP6-PWV station / grid point. The output parameter is high-precision GNSS-PWV. Machine learning algorithms such as CNN, LSTM, and XGBoost are used to train the data and construct a CMIP6 water vapor enhancement model based on GNSS-PWV. The enhancement effect is evaluated on the test set using GNSS-PWV as the ground truth. The enhanced CMIP6-PWV result output by the model is denoted as Improved_CMIP6-PWV.

[0077] Step 3: Spatiotemporal Analysis: The model is divided into four seasons and five altitude categories. The water vapor enhancement indices (RMSE and percentage) of the model in these categories are analyzed. Further investigation is conducted to explore the performance of different models in improving CMIP6-PWV at different seasonal scales and geographical locations. The root mean square error (RMSE) and the percentage decrease in RMSE are used as indicators of model performance improvement. A smaller RMSE value indicates better model correction of CMIP6-PWV; conversely, a larger RMSE value indicates a less effective model correction.

[0078]

[0079]

[0080] In the seasonal analyses, the data were categorized by annual accumulated days (AR), divided into four seasons: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). For each season, the RMSE calculation formula was used to assess the error of the original CMIP6-PWV data and the model calibration results. Simultaneously, the percentage decrease in RMSE, i.e., the improvement rate, was calculated to reflect the degree of performance improvement of the model in each season, especially the response differences between wet seasons (e.g., summer) and dry seasons (e.g., winter). In the geographic spatial analyses, the GNSS stations in the study area were classified by elevation, using the natural discontinuity classification method to divide them into five elevation zones. The water vapor RMSE and improvement rate before and after model calibration were calculated for each of the five elevation zones to analyze the model's generalization ability and stability under different terrain conditions. This analysis was used to assess the model's performance in different seasons and geographic spaces.

[0081] Step 4: Accuracy Verification. Using GNSS-PWV as the ground truth, the improvement effect on the observation period of 2024 (validation set 1), which was not involved in the training, is verified: the accuracy of the 2024 station Improved_CMIP6-PWV output by the model is compared with the ground truth GNSS-PWV. Additionally, the original CMIP6 data is grid data, but the above part interpolates it to GNSS stations for accuracy improvement, which cannot reflect the spatial grid context characteristics of CMIP6. Therefore, this invention performs grid-scale accuracy verification to preserve the properties of the input product itself. Here, the ERA5-PWV, which was not involved in the modeling, is used as validation data for comparison, including both station and grid results. The new validation (i.e., validation set 2) compares the accuracy of the improved 2024 grid-scale CMIP6-PWV data (Improved_Grid_CMIP6-PWV) output by the model with the ground truth grid-scale ERA5 (Grid_ERA5-PWV). To ensure comparability, the Improved_CMIP6-PWV results of the station were also compared with the ERA5-PWV (station_ERA5-PWV) data of the station that was not involved in the modeling, using the accuracy of the test set and validation set 1.

[0082] Specifically, when conducting the new validation (validation set 2), it is necessary to extract grid point data with non-negative elevations. This is because the input characteristics of the grid points are affected by the equidistant latitude and longitude positions of the original data, and there are locations below sea level in the grid within the study area, whose elevations will inevitably have negative values. However, since the target parameter used in the model construction is GNSS-PWV observation data from the stations, and the elevations of the GNSS stations are all above 0 m, necessary data quality checks are required. This avoids the possibility of relatively long air columns at grid points with negative elevations, which could lead to overestimation of PWV results.

[0083] The experimental results achieved by this invention can be explained as follows:

[0084] Experimental Area: This invention utilizes data from 162 GNSS stations evenly distributed throughout Turkey, covering the study area from 2020 to 2024. Valid data were provided by 98 stations in 2020, 16 stations in 2021, 141 stations in 2022, 139 stations in 2023, and 71 stations in 2024.

[0085] Spatiotemporal matching of data: Spatially, bilinear interpolation was used to interpolate CMIP6 mode data onto GNSS stations to obtain CMIP6-PWV data from all participating stations. Temporally, CMIP6-PWV has a 3-hour resolution, with eight time points daily: 01:30, 04:30, 07:30, 10:30, 13:30, 16:30, 19:30, and 22:30. GNSS-PWV has a 300-second temporal resolution and requires downsampling, retaining only the eight time points daily from its own time series: 01:30, 04:30, 07:30, 10:30, 13:30, 16:30, 19:30, and 22:30.

[0086] On the other hand, the ERA5-PWV data was also spatiotemporally matched and interpolated onto the GNSS station and the original CMIP6 grid using bilinear interpolation.

[0087] Ultimately, spatiotemporally matched station CMIP6-PWV (station_CMIP6-PWV), GNSS-PWV, and station ERA5-PWV data (station_ERA5-PWV) for 2020-2024 were obtained, along with spatiotemporally matched grid-scale CMIP6-PWV (Grid_CMIP6-PWV) and grid-scale ERA5-PWV data (Grid_ERA5-PWV) for 2024.

[0088] Model Construction and Testing: Experimental results show that all models significantly improve the accuracy of CMIP6-PWV, demonstrating the effectiveness of GNSS integration, with varying degrees of improvement among different models. We denote the improved CMIP6 water vapor result output by the models as Improved_CMIP6-PWV. Using GNSS-PWV as the reference ground truth, we compare the accuracy of Improved_CMIP6-PWV with GNSS-PWV, as follows: Figure 2As shown in the figure, the RMSE of the training set using the CNN algorithm is 4.25 mm, and the RMSE of the test set is 4.57 mm, representing improvements of 38.49% and 36.17% respectively. The Person correlation coefficients are 0.79 and 0.76, respectively, demonstrating effective extraction of spatial features. XGBoost significantly reduces costs during the learning process, with RMSEs of 3.54 mm and 4.04 mm for the training and test sets, respectively, achieving improvements of 48.77% and 43.57%, with Person correlation coefficients of 0.86 and 0.82. LSTM performs excellently in time series modeling, with an RMSE of 4.85 mm for the training set (Improved_CMIP6-PWV) and 4.93 mm for the test set, representing improvements of 29.81% and 31.14% respectively. Overall, the XGBoost model clearly demonstrates substantial optimization effects on the CMIP6-PWV station.

[0089] Spatiotemporal analysis: To further investigate the performance of different models in improving CMIP6-PWV at different seasonal scales, the data were divided into spring (MAM), summer (JJA), autumn (SON), and winter (DJF). A smaller RMSE indicates a better model calibration of CMIP6-PWV; conversely, a larger RMSE indicates a less effective model calibration. Figure 3 The RMSE results of models built using CNN, XGBoost, and LSTM algorithms in different seasons on the training and test sets are shown. It is evident that the XGBoost model has a significantly lower RMSE and a significantly higher improvement rate than the CNN and LSTM models, respectively, demonstrating the best improvement effect. On the training set, the RMSEs for each season are 3.12 mm, 4.95 mm, 4.08 mm, and 2.53 mm, while on the test set, they are 3.60 mm, 5.62 mm, 4.58 mm, and 3.02 mm. Furthermore, summer has the highest RMSE among all seasons. This is because, although Turkey is dominated by a temperate continental climate with low rainfall throughout the year, summer is still a season of relatively active convection, with relatively high water vapor content and variability, leading to a larger simulation error of PWV in the CMIP6 model itself. Given this high baseline, summer also has a larger RMSE after improvement compared to other seasons. All models have an improvement rate of over 25%, with XGBoost achieving the highest improvement rate (over 40%). In winter, the water vapor content is low and stable, resulting in small temperature differences.

[0090] Turkish station elevations were classified into five categories according to the natural discontinuity classification method (as shown in Table 1). The improvement effects of different models' training sets, test sets, and unified station sets on CMIP6-PWV at different elevations were investigated. The results are as follows: Figure 4As shown, the XGBoost model achieved the best correction, with RMSEs for the five altitude classes across all station sets being 4.56, 3.28, 2.99, 2.89, and 2.42 mm, all representing improvements of over 40%. RMSE gradually decreased with increasing altitude. This is because as altitude increases, the air column shortens, reducing the magnitude of PWV content and naturally decreasing the RMSE. In the training, test, and unified station sets, the correction rate for the second category of low-to-medium altitude stations was the highest compared to the other four altitude classes, at 58.3%, 52.8%, and 56.5% respectively (XGBoost). Furthermore, the number of stations for each altitude class was unified to 16 station samples. XBGoost showed the best improvement in RMSE for CMIP6 PWV, at 4.49 mm, 3.28 mm, 2.97 mm, 3.11 mm, and 2.11 mm. In terms of improvement rate, the improvement effect was best in the Class II elevation region, with different models achieving 45.0%, 56.5%, and 34.8%, respectively. This was followed by the Class III elevation region, further strengthening the strong evidence that the model performs better in correcting for low and mid-elevation areas. The RMSE of water vapor in the unimproved CMIP6 model decreased significantly linearly. However, the RMSE of the Class II, III, and IV elevation regions after improvement using the XBGoost and CNN algorithms did not change significantly, indicating that the model learned the terrain features, and the correction in the mid-elevation region was comparable.

[0091] Table 2. Classification of effective stations in Türkiye by elevation zone

[0092]

[0093] Table 2 shows the elevation classification of effective stations in Turkey under the validation system built using GNSS-PWV. After all model improvements, the improved CMIP6-PWV for stations in the eastern and central regions still shows smaller errors than other stations. From a land-sea perspective, this illustrates a phenomenon: the further away from the ocean, the better the accuracy of CMIP6-PWV. Ocean water evaporation, high water vapor content, and drastic humidity changes make it difficult for CMIP6 under this parameterization scheme to accurately simulate this small-scale process. This strongly suggests that CMIP6's water vapor simulation has significant biases in the land-sea interface region. For applying CMIP6 to water vapor exploration, especially in coastal or land-sea interface areas, additional calibration or error correction in conjunction with GNSS is required. Comparing different models, XGBoost still shows the best improvement effect, and this improvement is useful for the entire Turkish region, with most improvements exceeding 50%.

[0094] Accuracy Validation: Time Validation: Data from the observation period of 2024, which was not involved in model training, was selected as validation set 1. The accuracy of the model's output station Improved_CMIP6-PWV was compared with the ground truth GNSS-PWV. Figure 5 As shown, the CNN model performed best, with an RMSE of 5.31 mm (an improvement of 29.20%) and a Person correlation coefficient of 0.73; followed by the LSTM model, with an RMSE of 5.61 mm (an improvement of 25.24%) and a Person correlation coefficient of 0.70; the XGBoost model had an RMSE of 5.94 mm (an improvement of 20.76%) and a Person correlation coefficient of 0.66.

[0095] The ERA5-PWV data not used in the modeling was compared with the ground truth. A new validation set (Validation Set 2) was the 2024 Turkish CMIP6-PWV grid data (Grid_CMIP6-PWV). Because the original data used a globally distributed, equally spaced latitude and longitude grid, and areas below sea level exist within the study region, grid point elevations were negative. However, the target parameter GNSS-PWV data used in this model construction all came from ground stations (elevation > 0 m). To ensure data consistency, strict quality control was necessary. Therefore, grid point data with negative elevations were removed, representing 0.40% of the total data. The accuracy of the model output grid-scale Improved_CMIP6-PWV was compared with the ground truth ERA5-PWV. Figure 6 As shown in the figure. The results show that CNN improved the CMIP6 water vapor data by 11.64% with an RMSE of 5.96 mm, XGBoost improved it by 17.71% with an RMSE of 5.55 mm, and LSTM improved it by 18.97% with an RMSE of 5.47 mm.

[0096] To ensure the comparability of the evaluation results, this invention uses ERA5-PWV data (station scale) that was not involved in modeling as a unified reference benchmark to systematically validate the improved_CMIP6-PWV results of the station: the RMSE of the test set under the three algorithms were 5.40 mm, 5.66 mm, and 5.41 mm, respectively, and the RMSE of validation set 1 were 5.36 mm, 6.03 mm, and 5.59 mm, respectively. Table 3 visually shows the evaluation results of all test and validation sets.

[0097] In summary, the evaluation results for all test and validation sets show a significant improvement in accuracy. This invention, through GNSS data co-correction, significantly improves the accuracy of CMIP6 water vapor products across different spatial scales (including the original grid scale and GNSS station locations), with an average improvement of approximately 20%.

[0098] Table 3. Accuracy verification of Improved_CMIP6-PWV compared to ERA5-PWV at station and grid scales

[0099]

[0100] The best implementation results can be achieved using ArcGIS software and MATLAB. This invention is based on the improvement and optimization of water vapor accuracy in CMIP6 using high-precision GNSS, and conducts a series of accuracy verifications. In terms of implementation, GNSS is integrated, and different models are constructed using machine learning algorithms to improve the water vapor accuracy of the CMIP6 high-resolution model comparison program, and accuracy verification is performed at multiple scales.

[0101] This invention takes Turkey as the study area and uses GNSS PWV to construct the CMIP6 water vapor lifting model, and conducts spatiotemporal tests, verifications and analyses at both the station scale and the grid scale.

[0102] Example 2

[0103] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0104] Example 3

[0105] The purpose of this embodiment is to provide a computer-readable storage medium.

[0106] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0107] Example 4

[0108] The purpose of this embodiment is to provide an optimization system for improving the water vapor accuracy of CMIP6 based on GNSS, including:

[0109] The data acquisition module is configured to acquire GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data.

[0110] The data processing module is configured to process the acquired data: spatially, interpolate the target data onto the station and grid, and temporally, downsample the target data.

[0111] The model building module is configured to take CMIP6 water vapor data, the latitude corresponding to CMIP6 water vapor, the longitude corresponding to CMIP6 water vapor, the altitude corresponding to CMIP6 water vapor, and the annual day corresponding to CMIP6 water vapor as input parameters, and output GNSS observation data as target parameters. It uses machine learning algorithms to train the data and builds a CMIP6 water vapor enhancement model based on GNSS-PWV.

[0112] The water vapor accuracy enhancement module is configured to input the data to be measured into the constructed model to obtain enhanced CMIP6 water vapor data.

[0113] Example 5

[0114] The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments.

[0115] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0116] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0117] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An optimization method for improving the water vapor accuracy of CMIP6 based on GNSS, characterized in that, include: Acquire GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data; The GNSS observation data is based on point data from ground stations. Bilinear interpolation is used to interpolate the CMIP6 mode data onto the GNSS stations to obtain data from all stations participating in the experiment. The acquired data is processed as follows: spatially, the target data is interpolated onto the station and grid; temporally, the target data is downsampled. The CMIP6 water vapor data, the latitude, longitude, altitude, and annual day of CMIP6 water vapor are used as input parameters, and GNSS observation data is used as the output target parameter. The data are trained using machine learning algorithms to construct a CMIP6 water vapor enhancement model based on GNSS-PWV. The measured data is input into the constructed model to obtain improved CMIP6 water vapor data; In the process of constructing the CMIP6 water vapor enhancement model based on GNSS-PWV, ERA5 reanalysis water vapor data is used as the reference truth in both the test and validation sets. ERA5 reanalysis water vapor data is grid data with a set resolution. Bilinear interpolation is used to interpolate the ERA5 reanalysis water vapor data to the GNSS station to obtain the ERA5 reanalysis water vapor data at the station, which is then used as the validation data for both the test and validation sets.

2. The optimization method for improving CMIP6 water vapor accuracy based on GNSS as described in claim 1, characterized in that, Spatially, bilinear interpolation is used to interpolate the target data onto the station and grid.

3. The optimization method for improving CMIP6 water vapor accuracy based on GNSS as described in claim 1, characterized in that, The CMIP6 water vapor data is grid data with a set resolution.

4. The optimization method for improving CMIP6 water vapor accuracy based on GNSS as described in claim 1, characterized in that, In another validation set, the ERA5 reanalysis water vapor data were bilinearly interpolated onto the original CMIP6 grid points to obtain ERA5 reanalysis water vapor data matching the CMIP6 grid points, which served as validation data for another validation set.

5. An optimization system for improving CMIP6 water vapor accuracy based on GNSS, employing the optimization method for improving CMIP6 water vapor accuracy based on GNSS as described in any one of claims 1-4, characterized in that, include: The data acquisition module is configured to acquire GNSS observation data, CMIP6 water vapor data, and ERA5 reanalysis water vapor data. The data processing module is configured to process the acquired data: spatially, interpolate the target data onto the station and grid, and temporally, downsample the target data. The model building module is configured to take CMIP6 water vapor data, the latitude corresponding to CMIP6 water vapor, the longitude corresponding to CMIP6 water vapor, the altitude corresponding to CMIP6 water vapor, and the annual day corresponding to CMIP6 water vapor as input parameters, and output GNSS observation data as target parameters. It uses machine learning algorithms to train the data and builds a CMIP6 water vapor enhancement model based on GNSS-PWV. The water vapor accuracy enhancement module is configured to input the data to be measured into the constructed model to obtain enhanced CMIP6 water vapor data.

6. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-4 above.

Images