Coastal sea surface wind field correction method based on CFLGBM
By combining the Cascade Forest and LightGBM combined model (CFLGBM) with various ocean-atmosphere parameters, the problem of decreased accuracy in nearshore sea surface wind field inversion was solved, achieving efficient and accurate wind field correction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INFORMATION SCI & TECH UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-15
AI Technical Summary
In the inversion of wind fields on nearshore waters, existing technologies rely on complex physical inversion methods that require the removal of fuzzy wind direction solutions, while machine learning methods suffer from reduced inversion accuracy within 200 kilometers of the nearshore surface and cannot effectively correct for the impact of land pollution.
A combined model based on Cascade Forest and LightGBM (CFLGBM) is adopted. Through cascaded hierarchical information transfer and feature enhancement, combined with parameters such as sea surface temperature, significant wave height, ocean current and precipitation, a wind speed and wind direction correction model is constructed, which simplifies the parameter tuning process and improves computational efficiency and accuracy.
It significantly improves the observation accuracy and calculation speed of nearshore sea surface wind fields, effectively corrects nearshore wind field errors, and enhances the accuracy and reliability of wind field data.
Smart Images
Figure CN122046984A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine microwave remote sensing technology, and in particular to a nearshore sea surface wind field correction method based on CFLGBM. Background Technology
[0002] Ocean wind field observation is of great significance for meteorology, aviation, fisheries, marine transportation, and disaster early warning, providing crucial information and data support regarding momentum, heat, and mass exchange at the air-sea interface. Compared to observations from meteorological observation towers, wind towers, and buoys, satellite remote sensing of wind fields offers advantages in its broad global coverage and high spatiotemporal resolution, providing continuous global wind field monitoring. HY-2C is an advanced marine dynamic environment monitoring satellite in China, equipped with a radar altimeter, microwave scatterometer, and calibrated radiometer. Its microwave scatterometer utilizes microwave signals to detect sea surface roughness, effectively monitoring changes in sea surface wind fields with minimal impact from weather and cloud cover, providing critical data for marine monitoring. In recent years, due to the close relationship between coastal waters and human activities, nearshore wind fields have received increasing attention. Accurate nearshore wind field information is crucial for improving the accuracy of numerical weather prediction, deepening marine meteorological research, and strengthening disaster management and prevention. For example, nearshore wind fields play a key role in the prediction and monitoring of extreme weather events such as tsunamis and typhoons.
[0003] Currently, methods for inverting nearshore sea surface wind fields using microwave scatterometers mainly include physical inversion algorithms and empirical inversion algorithms. Among these, the physical inversion method requires special consideration of the impact of land pollution when processing microwave scatterometer observation data from coastal areas. To accurately obtain sea surface wind field information, it is essential to effectively reduce or eliminate land interference with the observation results. Therefore, two main techniques are typically used to address land pollution: the first is the Land Contribution Ratio (LCR) thresholding technique, and the second is the Empirical Land Mask (ELM) technique. After obtaining microwave scatterometer observation data from coastal areas, Bayesian maximum likelihood estimation (MLE) is used for sea surface wind field (SSW) inversion. However, the MLE method's point-by-point calculation of the objective function value is complex and does not consider the continuity of the SSW field. Furthermore, it requires removing wind direction ambiguities within wind vector cells based on the accurate background wind field, which increases the complexity of coastal SSW field inversion.
[0004] Compared to physical algorithms, which offer clear physical meaning but involve complex inversion processes and require the removal of ambiguous solutions, machine learning inversion algorithms establish a fitting relationship between observed data and measured wind fields. This avoids the complexity of physical modeling and the need to remove ambiguous wind direction solutions, offering advantages such as high computational speed and timeliness. Therefore, machine learning methods, with their ability to autonomously learn feature information from data, are widely used in nearshore wind field inversion. However, due to various factors such as seabed topography, wind zone length, and strong tidal currents, the response of sea surface roughness to sea surface wind fields in coastal areas is relatively complex. Consequently, the inversion accuracy of both physical and machine learning algorithms decreases within 200 kilometers of the coast compared to open sea areas. Furthermore, the inversion error of the sea surface wind field increases with decreasing distance from the shore, necessitating further correction of the sea surface wind field in coastal areas to improve its observation accuracy. Summary of the Invention
[0005] The purpose of this invention is to propose a nearshore sea surface wind field correction method based on Cascade Forest-LightGBM (CFLGBM) for microwave scatterometers, aiming to address the accuracy degradation problem in nearshore sea surface wind field inversion using both machine learning and physical methods. Cascade Forest, as a tree-based ensemble learning method, effectively improves the predictive performance of the model through hierarchical information transfer and feature enhancement. However, it suffers from high computational cost and long training time when processing large-scale datasets, and the multi-layered cascaded structure complicates parameter tuning. Therefore, this invention uses LightGBM as the basic predictor of Cascade Forest, leveraging LightGBM's efficient histogram algorithm, leaf-based growth strategy, and automatic feature processing capabilities to significantly simplify parameter tuning and reduce computational complexity while improving training speed and reducing memory consumption. Through this structural adjustment, the powerful learning capabilities of the LightGBM model are introduced into the cascaded forest, thereby improving the overall performance of the model and enriching the input feature space of the next layer using the output features of LightGBM. This connection method retains the layer-by-layer optimization characteristics of cascaded forests while fully utilizing the modeling capabilities of LightGBM. The CFLGBM model fully combines the advantages of Cascade Forest and LightGBM. It not only retains the hierarchical structure advantage of Cascade Forest, but also uses the efficient LightGBM algorithm to quickly and accurately explore the correlation between sea surface wind field and multiple ocean-atmosphere parameters, thereby improving the observation accuracy of global nearshore sea surface wind field.
[0006] To achieve the above objectives, the present invention provides the following solution: A nearshore sea surface wind field correction method based on CFLGBM includes: A wind field dataset was constructed based on historical meteorological data; Wind speed correction model and wind direction correction model are constructed separately; both the wind speed correction model and the wind direction correction model are constructed based on the Cascade Forest model and the LightGBM model. Based on the wind field dataset, the wind speed correction model and the wind direction correction model were trained respectively. The wind field data of the nearshore sea surface were corrected using the trained wind speed correction model and wind direction correction model.
[0007] Optionally, the wind field dataset includes: HY-2C sea surface wind field data, ERA5 wind field data, ERA5 significant wave height data, GPM precipitation data, OSTIA sea surface temperature data, OSCAR sea surface current data, ASCAT-C Level 2 sea surface wind speed data, and NDBC buoy wind field data.
[0008] Optionally, the HY-2C sea surface wind field data includes: time, latitude and longitude, wind speed, wind direction, mass indicator, and land-sea indicator; The ERA5 wind field data includes: time, latitude and longitude, wind speed, and wind direction; The GPM precipitation data includes: time, latitude and longitude, and rainfall. The OSTIA sea surface temperature data includes: time, latitude and longitude, and sea surface temperature; OSCAR ocean surface current data includes: time, latitude and longitude, meridional current velocity, and zonal current velocity.
[0009] Optionally, the wind speed correction model and the wind direction correction model have the same model framework; both are built based on the CascadeForest model and the LightGBM model. The CFLGBM model includes: an input layer, a cascaded layer, and an output layer; The input features of the input layer include: latitude and longitude, wind speed of HY-2C, sea surface temperature, significant wave height, ocean currents, and rainfall; Each layer in the cascaded layer contains multiple LightGBM models. The LightGBM models are used to replace the base predictors in the cascaded forest. The cascaded layer is used to receive the input features and output results of the previous layer to train multiple LightGBM models in the current layer. The prediction results of multiple LightGBM models are then merged with the original input features to form the input features of the next layer. The output layer provides the ERA5 wind speed.
[0010] Optionally, before training the wind speed correction model and the wind direction correction model based on the wind field dataset, the process includes: preprocessing the wind field dataset; including: The pre-collected HY-2C data is matched with the coastline data to obtain nearshore HY-2C sea surface wind field data; Nearshore HY-2C sea surface wind field data are filtered using quality labels; Based on a preset spatiotemporal matching window, the filtered HY-2C sea surface wind field data is matched with pre-collected auxiliary parameters to obtain a matched dataset; wherein the auxiliary parameters include: GMP precipitation information, OSCAR ocean current information, OSTIA sea surface temperature data, and ERA5 significant wave height and wind field data. The matched dataset is divided into two preset time periods: the dataset from one time period is used as the training and validation set; the dataset from the other time period is matched with the pre-collected NDBC and ASCAT-C wind field data based on the preset spatiotemporal matching window to obtain the test set.
[0011] Optionally, the wind speed correction model and the wind direction correction model are trained using the same method, differing only in the setting of hyperparameters.
[0012] Optionally, the training methods for the wind speed correction model and the wind direction correction model include: Using Pearson correlation analysis, the input and output parameters of the model are selected to obtain the basic parameter combination; wherein, the basic parameter combination includes: input: latitude and longitude, wind speed of HY-2C, significant wave height, and output: wind speed of ERA5; The basic parameter combination is divided into several different input parameter combinations to train and validate the model, and obtain the final input parameters and final output parameters of the model; wherein, the final input is: latitude and longitude, wind speed of HY-2C, significant wave height, ocean current, precipitation, and sea surface temperature, and the final output is: wind speed of ERA5; The hyperparameters were optimized using grid search and five-fold cross-validation to determine the final hyperparameter values. The optimized hyperparameters included: the number of cascaded layers, the number of base forests in each layer, the size of each forest layer, and the number of decision trees in each forest layer. The wind speed correction model is trained based on the final input parameters, final output parameters, and final hyperparameter values.
[0013] Optionally, training the wind speed correction model and the wind direction correction model may also include: Regression indices are used to evaluate the wind speed and wind direction correction results of the wind speed correction model and the wind direction correction model; wherein, the regression indices include: mean deviation, root mean square error and correlation coefficient.
[0014] The beneficial effects of this invention are as follows: The advantages of the global nearshore wind field correction method based on CFLGBM in this invention are: (1) High computational efficiency: The CFLGBM model combines the advantages of Cascade Forest and LightGBM, adopts the Cascade Forest architecture, and can efficiently process large-scale data, improving computational speed; (2) High accuracy: Physical inversion methods and traditional machine learning inversion methods suffer from severe accuracy drops in nearshore areas, while the wind field correction method based on CFLGBM in this invention focuses on nearshore areas, effectively correcting errors in these areas and improving the accuracy of nearshore wind fields; (3) Comprehensive consideration of multiple influencing factors: Traditional physical inversion methods or machine learning inversion methods fail to fully consider other important ocean-atmosphere parameters that may affect the accuracy of nearshore wind fields. The global nearshore wind field correction method based on CFLGBM in this invention simultaneously considers parameters such as sea surface temperature, significant wave height, ocean currents, and precipitation. This multi-dimensional feature input can more accurately capture the complex relationships related to wind field changes, significantly improving the correction accuracy of nearshore wind fields. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is an architecture diagram of the CFLGBM-based wind field correction model according to an embodiment of the present invention; Figure 2 This is a scatter plot comparing the wind speed of the correction model in this embodiment of the invention with the wind speed of the NDBC buoy. Figure 3 This is a scatter plot comparing the wind direction of the correction model and the wind speed of the NDBC buoy in an embodiment of the present invention. Figure 4 This is a schematic diagram of a cascaded forest according to an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] This embodiment proposes a nearshore sea surface wind field correction method based on CFLGBM, and compares its results with other wind field products from the same period to illustrate that introducing multiple wind field-related parameters can positively improve the correction accuracy. The method steps of this embodiment include: A wind field dataset was constructed based on historical meteorological data; Wind speed correction model and wind direction correction model are constructed separately; both the wind speed correction model and the wind direction correction model are constructed based on the Cascade Forest model and the LightGBM model. Based on the wind field dataset, the wind speed correction model and the wind direction correction model were trained respectively. The wind field data of the nearshore sea surface were corrected using the trained wind speed correction model and wind direction correction model.
[0020] Specifically, this embodiment first collects the required data to construct a wind field dataset. The dataset includes HY-2C wind field data, ERA5 wind field data and significant wave height data, GPM precipitation data, OSTIA sea surface temperature data, OSCAR sea surface current data, ASCAT-C Level 2 sea surface wind speed data, and NDBC buoy wind field data. The specific steps are as follows: S11. A dataset covering the period from June 2020 to December 2020 was created.
[0021] S12. Download HY-2C product data from the National Satellite Meteorological Center website, including time, latitude and longitude, wind speed, wind direction, quality label, and land / sea label.
[0022] S13. Downloaded ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts, including time, latitude and longitude, wind speed, wind direction and significant wave height.
[0023] S14. I downloaded the GPM IMERG Level-3 product from the Global Precipitation Measurement Project website, which includes time, latitude and longitude, and rainfall.
[0024] S15. The OSTIA dataset, containing time, latitude and longitude, and sea surface temperature, was downloaded from the Copernicus Marine Environment Monitoring Service.
[0025] S16. I downloaded the OSCAR global ocean surface current field product from the NOAA website. It includes time, latitude and longitude, and meridional and zonal current velocities to indicate the direction and speed of ocean currents.
[0026] S17. Downloaded ASCAT-C Level 2 sea surface wind speed and direction products, including time, latitude and longitude, wind speed and wind direction.
[0027] S18. Downloaded NDBC buoy data to validate the results of the trained model.
[0028] In this embodiment, after constructing the wind field dataset, data preprocessing is performed. The obtained data are matched according to a predefined spatiotemporal window to obtain a matched dataset, which is then divided into training, validation, and test sets. The steps include the following: S21. First, the collected HY-2C data is matched with the coastline data to obtain the nearshore HY-2C sea surface wind field data.
[0029] S22. Then, the nearshore HY-2C sea surface wind field data is controlled by quality labeling to filter out data points with poor quality.
[0030] S23. Set the spatiotemporal matching windows to 30 minutes and 25 km respectively, and then use the quality-controlled wind field data to match with the collected auxiliary parameters, including GMP precipitation information, OSCAR ocean current information, OSTIA sea surface temperature data, ERA5 significant wave height and wind field data, to obtain a matched dataset.
[0031] S24. Divide the dataset into two time periods: June to November 2020 and December 2020. Select 75% of the data from the June to November 2020 period as the training set and name it Dataset 1, and use the remaining 25% as the validation set and name it Dataset 2, for training and validating the model.
[0032] S25. The December 2020 dataset was then matched with the wind field data from NDBC and ASCAT-C using the same spatiotemporal matching window to obtain dataset 3, which was used as a test set to evaluate the model's correction capability.
[0033] Furthermore, the wind speed correction model and the wind direction correction model share the same model framework, such as... Figure 1 As shown; all models are built based on the Cascade Forest model and the LightGBM model; The CFLGBM model includes: an input layer, a cascaded layer, and an output layer; The input features of the input layer include: latitude and longitude, wind speed of HY-2C, sea surface temperature, significant wave height, ocean currents, and rainfall; Each layer in the cascaded layers contains several LightGBM models to replace the base predictor in the cascaded forest. The main function of the cascaded layers is to receive the input features and output results of the previous layer, thereby training multiple LightGBM models in the current layer, and merging their prediction results with the original input features to form the input features of the next layer. The output layer provides the ERA5 wind speed.
[0034] Specifically, a cascaded forest is a framework, for example... Figure 4 As shown, each layer contains many predictors. Ordinary cascaded forests use random forests and fully random forests as predictors. Here, these two models are optimized into the LightGBM model, which is used as the new predictor. Then, this model is cascaded in this manner to obtain a new model.
[0035] Furthermore, the training methods for the wind speed correction model and the wind direction correction model are the same, only the hyperparameter settings are different.
[0036] The training methods for the wind speed correction model and the wind direction correction model include: Using Pearson correlation analysis, the parameters of the model input and output are selected to obtain the parameter combination; wherein, the basic parameter combination includes: input: latitude and longitude, wind speed of HY-2C, significant wave height, and output: wind speed of ERA5; The parameter combination is divided into several different input parameter combinations to train and validate the model, and to obtain the final input parameters and final output parameters of the model. The final inputs are: latitude and longitude, wind speed of HY-2C, significant wave height, ocean current, precipitation, and sea surface temperature; the final output is: wind speed of ERA5. The hyperparameters were optimized using grid search and five-fold cross-validation to determine the final hyperparameter values. The optimized hyperparameters included: the number of cascaded layers, the number of base forests in each layer, the size of each forest layer, and the number of decision trees in each forest layer. The wind speed correction model is trained based on the final input parameters, final output parameters, and final hyperparameter values.
[0037] Specifically, in this embodiment, a wind speed correction model based on CFLGBM is constructed; this includes the following steps: S31. First, Pearson correlation analysis was used to confirm the input and output parameters. The parameters with higher correlation weights were selected and their combination was set as follows: the input is latitude and longitude, wind speed of HY-2C, and significant wave height, and the output is wind speed of ERA5.
[0038] S32. To further determine the impact of precipitation, sea surface temperature, and ocean currents on the sea surface wind field correction effect, this study divided all parameters into five different input parameter combinations to train and validate the model. Ultimately, it was found that the model achieved optimal performance when all parameters were considered. The model inputs at this point were latitude and longitude, HY-2C wind speed, sea surface temperature, significant wave height, ocean currents, and precipitation; the output was ERA5 wind speed.
[0039] S33. Grid search and five-fold cross-validation were used to optimize hyperparameters. Ranges for each parameter were selected and labeled for testing. The number of cascaded layers, the number of base forests in each cascaded layer, the forest size in each cascaded layer, and the number of decision trees in each cascaded layer forest were selected from [2,3,4], [4,5,6], [6,7,8,9], and [100,200,300,500], respectively. After testing, these parameters were finally set to 3, 4, 8, and 300.
[0040] S34. Use the above parameter combination and hyperparameters to train the wind speed correction model using dataset 1, validate the wind speed correction model using dataset 2, and test the wind speed correction model using dataset 3.
[0041] In this embodiment, a wind direction correction model based on CFLGBM is constructed; specifically, it includes: S41. First, Pearson correlation analysis was used to confirm the input and output parameters. The parameters with higher correlation weights were selected and their combination was set as follows: the input is latitude and longitude, wind direction of HY-2C, and significant wave height, and the output is wind direction of ERA5.
[0042] S42. To further determine the impact of precipitation, sea surface temperature, and ocean currents on the sea surface wind field correction effect, this study divided all parameters into five different input parameter combinations to train and validate the model. Ultimately, it was found that the model achieved optimal performance when all parameters were considered. The model inputs at this point were latitude and longitude, HY-2C wind direction, sea surface temperature, significant wave height, ocean currents, and precipitation; the output was ERA5 wind direction.
[0043] S43. Grid search and five-fold cross-validation were used to optimize hyperparameters. Ranges for each parameter were selected and labeled for testing. The number of cascaded layers, the number of base forests in each cascaded layer, the forest size in each cascaded layer, and the number of decision trees in each cascaded layer forest were selected from [2,3,4], [4,5,6], [6,7,8,9], and [100,200,300,500], respectively. The final determined parameters were 3, 4, 7, and 200.
[0044] S34. Use the above parameter combination and hyperparameters to train the wind direction correction model using dataset 1, validate the wind direction correction model using dataset 2, and test the wind direction correction model using dataset 3.
[0045] Specifically, this embodiment trains and tests CFLGBM-based wind speed and direction correction models respectively, obtains the correction results of the two models, and evaluates the accuracy of the models in terms of wind speed and direction, including the following: S51. After training and saving the CFLGBM-based wind speed and direction correction models, use dataset 3 to test the results of the correction models. Compare the model outputs and the original data of HY-2C with those of ERA5, ASCAT-C, and NDBC buoys to evaluate the model's capabilities.
[0046] S52. Compare the inversion results of the two models respectively. Use regression indicators to evaluate the results of wind speed and direction correction, including mean bias (MB), root mean square error (RMSE), and correlation coefficient (R).
[0047] (1) (2) (3) Where X represents the corrected sea surface wind speed (SSWS) and sea surface wind direction (SSWD), Y represents the reference sea surface wind speed and sea surface wind direction from ERA5, and n represents the total number of samples. This represents the average value of the corrected sea surface wind speed and direction. This represents the average of the reference values.
[0048] S53. Wind speed comparison results are as follows: Figure 2 As shown, a comparison was made between the model output and NDBC data points. The MB value was 0.06 m / s, the RMSE value was 1.08 m / s, and the correlation coefficient was 0.91. The wind speed output by the model showed a high degree of consistency with the actual observed wind speed. The comparison results for wind direction are shown below. Figure 3 As shown, compared with the NDBC data, the MB value is -1.87°, the RMSE value is 22.85°, and the correlation coefficient is 0.90, which also shows a high degree of consistency.
[0049] This embodiment incorporates the LightGBM model into the Cascade Forest framework. By adding various physical parameters closely related to the wind field, features from these auxiliary parameters are extracted to correct wind field data that were originally more polluted near the coast, thereby improving the accuracy and reliability of the wind field.
[0050] The advantages of the global nearshore wind field correction method based on CFLGBM in this invention are: (1) High computational efficiency: The CFLGBM model combines the advantages of Cascade Forest and LightGBM, adopts the Cascade Forest architecture, and can efficiently process large-scale data, improving computational speed; (2) High accuracy: Physical inversion methods and traditional machine learning inversion methods suffer from severe accuracy drops in nearshore areas, while the wind field correction method based on CFLGBM in this invention focuses on nearshore areas, effectively correcting errors in these areas and improving the accuracy of nearshore wind fields; (3) Comprehensive consideration of multiple influencing factors: Traditional physical inversion methods or machine learning inversion methods fail to fully consider other important ocean-atmosphere parameters that may affect the accuracy of nearshore wind fields. The global nearshore wind field correction method based on CFLGBM in this invention simultaneously considers parameters such as sea surface temperature, significant wave height, ocean currents, and precipitation. This multi-dimensional feature input can more accurately capture the complex relationships related to wind field changes, significantly improving the correction accuracy of nearshore wind fields.
[0051] The following is a general explanation of the nearshore sea surface wind field correction method proposed in this embodiment based on Cascade Forest-LightGBM (CFLGBM). By analyzing the relationship between the nearshore sea surface wind field and multiple air-sea parameters, such as significant wave height, precipitation, ocean currents, and sea surface temperature, a global nearshore sea surface wind field correction model is established to improve the inversion accuracy of the scatterometer's sea surface wind field. The implementation steps of this embodiment are as follows: Step 1: Construct the dataset. First, download the HY-2C wind field product data from the National Satellite Meteorological Center website, covering the period from June to December 2020. The HY-2C wind field product data includes time, latitude and longitude, wind speed, wind direction, quality label, and land / sea label. Simultaneously, download the ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF), including time, latitude and longitude, wind speed, wind direction, and significant wave height. In addition, the GPM IMERG Level-3 product, which includes time, latitude and longitude, and precipitation, was downloaded from the Global Precipitation Measurement Project website; the OSTIA (Operational Sea Surface Temperature and Ice Analysis) dataset, which includes time, latitude and longitude, and sea surface temperature, was downloaded from the Copernicus Marine Environmental Monitoring Service site; and the OSCAR (Ocean Surface Current Analyses Real-time) global sea surface current field product, which includes time, latitude and longitude, and meridional (u) and zonal (v) current velocities, was downloaded from the NOAA website to represent the direction and speed of ocean currents. Finally, the second-level sea surface wind field product from MetOp-C ASCAT (Meteorological Operational Satellite Program - C Advanced Scatterometer) and buoy data from the National Data Buoy Center (NDBC) were downloaded to verify the model's actual performance.
[0052] Step Two: Data Preprocessing. Considering the differences in latitude, longitude, and time across different datasets, dataset matching is necessary. First, the collected HY-2C product data is matched with the coastline to obtain nearshore HY-2C sea surface wind field product data. Then, quality control indicators are used to filter the data, removing data with severe interference. Next, a 30-minute and 25km spatiotemporal window is used to match with auxiliary data, including GMP precipitation, OSCAR ocean current, OSTIA sea surface temperature, ERA5 significant wave height, and wind field data. Finally, the dataset is divided: 75% of the data from June to November 2020 is selected as the training set, and the remaining 25% as the validation set. The December 2020 dataset is matched with NDBC and ASCAT-C wind field data as the test set to evaluate the model's actual performance.
[0053] Step 3: Construct a wind speed correction model based on Cascade Forest - LightGBM. CFLGBM inherits the multi-layered cascaded structure of Cascade Forest, with each layer containing multiple base learners (i.e., LightGBM models). The output of the previous layer serves as the input to the next layer. Through this hierarchical structure, the model can progressively extract features, thereby improving prediction accuracy. Specifically, the architecture of CFLGBM is as follows: The input features of the input layer include latitude and longitude, HY-2C wind speed, sea surface temperature, significant wave height, ocean currents, and precipitation. Each cascaded layer contains multiple LightGBM models, with the output features of the previous layer and the original input features serving as the input to the next layer. Finally, the output layer outputs the ERA5 wind speed. To optimize model performance, random search and cross-validation are used to optimize the hyperparameters. The optimized hyperparameters include the number of cascaded layers, the number of base forests in each layer, the size of each forest layer, and the number of decision trees in each forest layer. After optimization, the final determined hyperparameter values are: 3 (number of cascaded layers), 4 (number of basic forests per layer), 8 (size of forests per layer), and 300 (number of decision trees per layer).
[0054] Assuming the input feature vector X Include d One characteristic: (4) in x 1 , x 2 , …, x d These represent the input parameters: latitude and longitude, wind speed of HY-2C, sea surface temperature, significant wave height, ocean current, and precipitation, respectively.
[0055] For the l Layer, assuming the layer contains N There are 1 LightGBM model, and the prediction output of each model is: f l,i (X) Then the first l The total output of the layer is: (5) in f l,i (X) Indicates the first l Layer i Each LightGBM model for input features X The predicted value.
[0056] No. l+1 The input of the layer is the first lThe concatenation of the layer's output with the original input features: (6) This means the first l+1 The input feature vector of the layer X+1 Includes original features and the first l All predicted outputs of the layer.
[0057] go through L After layer concatenation, the final output is the prediction result of the last layer: (7) in f L (X) Indicates the first L Total predicted value for the layer.
[0058] Step 4: Construct a CFLGBM-based wind direction correction model. The model framework is the same as the wind speed correction model in Step 3, but the hyperparameter settings have been further adjusted based on the data characteristics. Specifically, the number of cascaded layers is set to 3, the number of basic forests in each cascaded layer is set to 4, the forest size in each cascaded layer is set to 7, and the number of decision trees in each forest in each cascaded layer is set to 200. The model's input features include latitude and longitude, HY-2C wind direction, sea surface temperature, significant wave height, ocean currents, and precipitation; the output is the ERA5 wind direction.
[0059] Step 5: Train and test the CFLGBM-based wind speed and direction correction models separately, obtain the correction results of the two models, and evaluate the accuracy of the models in wind speed and direction. Evaluation metrics include mean bias (MB), root mean square error (RMSE), and correlation coefficient (R).
[0060] This invention aims to propose a nearshore sea surface wind field correction method based on Cascade Forest-LightGBM (CFLGBM) for microwave scatterometers, addressing the accuracy degradation issue in nearshore sea surface wind field inversion using both machine learning and physical methods. Cascade Forest, as a tree-based ensemble learning method, effectively improves model prediction performance through hierarchical information transfer and feature enhancement. However, it suffers from high computational cost and long training time when processing large-scale datasets, and its multi-layered cascaded structure complicates parameter tuning. Therefore, this invention employs LightGBM as the base predictor for Cascade Forest, leveraging its efficient histogram algorithm, leaf-based growth strategy, and automatic feature processing capabilities to significantly simplify parameter tuning and reduce computational complexity while improving training speed and reducing memory consumption. The CFLGBM model combines the advantages of Cascade Forest and LightGBM, retaining the hierarchical architecture of Cascade Forest while using the efficient LightGBM algorithm to quickly and accurately uncover the correlation between sea surface wind fields and multiple air-sea parameters, thereby improving the observation accuracy of global nearshore sea surface wind fields.
[0061] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A nearshore sea surface wind field correction method based on CFLGBM, characterized in that, include: A wind field dataset was constructed based on historical meteorological data; Wind speed correction model and wind direction correction model are constructed separately; both the wind speed correction model and the wind direction correction model are constructed based on the Cascade Forest model and the LightGBM model. Based on the wind field dataset, the wind speed correction model and the wind direction correction model were trained respectively. The wind field data of the nearshore sea surface were corrected using the trained wind speed correction model and wind direction correction model.
2. The nearshore sea surface wind field correction method based on CFLGBM according to claim 1, characterized in that, The wind field dataset includes: HY-2C sea surface wind field data, ERA5 wind field data, ERA5 significant wave height data, GPM precipitation data, OSTIA sea surface temperature data, OSCAR sea surface current data, ASCAT-C Level 2 sea surface wind speed data, and NDBC buoy wind field data.
3. The nearshore sea surface wind field correction method based on CFLGBM according to claim 2, characterized in that, The HY-2C sea surface wind field data includes: time, latitude and longitude, wind speed, wind direction, mass indicator, and land-sea indicator; The ERA5 wind field data includes: time, latitude and longitude, wind speed, and wind direction; The GPM precipitation data includes: time, latitude and longitude, and rainfall. The OSTIA sea surface temperature data includes: time, latitude and longitude, and sea surface temperature; OSCAR ocean surface current data includes: time, latitude and longitude, meridional current velocity, and zonal current velocity.
4. The nearshore sea surface wind field correction method based on CFLGBM according to claim 1, characterized in that, The wind speed correction model and the wind direction correction model have the same model framework; both are built based on the Cascade Forest model and the LightGBM model. The CFLGBM model includes: an input layer, a cascaded layer, and an output layer; The input features of the input layer include: latitude and longitude, wind speed of HY-2C, sea surface temperature, significant wave height, ocean currents, and rainfall; Each layer in the cascaded layer contains multiple LightGBM models. The LightGBM models are used to replace the base predictors in the cascaded forest. The cascaded layer is used to receive the input features and output results of the previous layer to train multiple LightGBM models in the current layer. The prediction results of multiple LightGBM models are then merged with the original input features to form the input features of the next layer. The output layer provides the ERA5 wind speed.
5. The nearshore sea surface wind field correction method based on CFLGBM according to claim 1, characterized in that, Before training the wind speed correction model and the wind direction correction model based on the wind field dataset, the process includes: preprocessing the wind field dataset; including: The pre-collected HY-2C data is matched with the coastline data to obtain nearshore HY-2C sea surface wind field data; Nearshore HY-2C sea surface wind field data are filtered using quality labels; Based on a preset spatiotemporal matching window, the filtered HY-2C sea surface wind field data is matched with pre-collected auxiliary parameters to obtain a matched dataset; wherein the auxiliary parameters include: GMP precipitation information, OSCAR ocean current information, OSTIA sea surface temperature data, and ERA5 significant wave height and wind field data. The matched dataset is divided into two preset time periods: the dataset from one time period is used as the training and validation set; the dataset from the other time period is matched with the pre-collected NDBC and ASCAT-C wind field data based on the preset spatiotemporal matching window to obtain the test set.
6. The nearshore sea surface wind field correction method based on CFLGBM according to claim 1, characterized in that, The wind speed correction model and the wind direction correction model are trained using the same method, differing only in the settings of hyperparameters.
7. The nearshore sea surface wind field correction method based on CFLGBM according to claim 6, characterized in that, The training methods for the wind speed correction model and the wind direction correction model include: Using Pearson correlation analysis, the input and output parameters of the model are selected to obtain the basic parameter combination; wherein, the basic parameter combination includes: input: latitude and longitude, wind speed of HY-2C, significant wave height, and output: wind speed of ERA5; The basic parameter combination is divided into several different input parameter combinations to train and validate the model, and obtain the final input parameters and final output parameters of the model; wherein, the final input is: latitude and longitude, wind speed of HY-2C, significant wave height, ocean current, precipitation, and sea surface temperature, and the final output is: wind speed of ERA5; The hyperparameters were optimized using grid search and five-fold cross-validation to determine the final hyperparameter values. The optimized hyperparameters included: the number of cascaded layers, the number of base forests in each layer, the size of each forest layer, and the number of decision trees in each forest layer. The wind speed correction model is trained based on the final input parameters, final output parameters, and final hyperparameter values.
8. The nearshore sea surface wind field correction method based on CFLGBM according to claim 1, characterized in that, Training the wind speed correction model and the wind direction correction model also includes: Regression indices are used to evaluate the wind speed and wind direction correction results of the wind speed correction model and the wind direction correction model; wherein, the regression indices include: mean deviation, root mean square error and correlation coefficient.