Medium and long term ionosphere TEC combined prediction method based on two-step loss fine tuning

Abstract

The embodiment of the invention discloses a medium and long term ionosphere TEC combined prediction method based on two-step loss fine tuning, and the method comprises the steps: (1) obtaining a deep learning training sample, and carrying out the preprocessing of data; (2) constructing a deep learning network model; (3) training a deep learning network model; (4) optimizing the model by a two-step loss fine tuning method and a combined prediction method; and (5) performing an SAR simulation experiment. After the model is trained, an input TEC is given, a target prediction TEC can be obtained through the trained model, and then a corresponding SAR ionosphere delay value is obtained through calculation. According to the model for predicting the TEC of the medium and long term ionosphere based on deep learning, a combined prediction and two-step loss fine adjustment method is introduced into the model, effective prediction of the TEC of the medium and long term ionosphere is achieved through optimal combination of the model and optimization of loss in the model training process, and the SAR real-time positioning precision is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of space science and remote sensing science, and specifically relates to a method for predicting medium- and long-term ionospheric total electron content (TEC) by using a deep learning technology combining prediction and two-step fine-tuning. Background Art

[0002] The ionosphere is usually defined as an ionized region in the Earth's atmosphere that contains charged particles that can affect the propagation of radio wave signals. The total electron content (TEC) is the total number of free electrons per unit area over the entire altitude range of the ionosphere. It can directly characterize the impact of the ionosphere on the propagation of spaceborne synthetic aperture radar (SAR) observation signals [1].

[0003] As countries vigorously develop low-frequency SAR satellites, the operating frequency of SAR satellites continues to decrease due to the inverse relationship between ionospheric delay and electromagnetic wave frequency, resulting in an increasing impact of ionospheric delay on slant range measurement accuracy [2]. Timely acquisition of the total electron content of the synthetic aperture radar satellite observation area can improve the real-time positioning accuracy of the synthetic aperture radar.

[0004] There are currently several problems with using ionospheric TEC to correct SAR ionospheric delay:

[0005] (1) The correction error of fixed TEC value is large: the existing correction methods mostly use fixed TEC value for correction, which fails to fully consider the temporal and spatial variation characteristics of the ionosphere, resulting in large errors and affecting the positioning accuracy.

[0006] (2) Existing TEC products have lags: The currently commonly used Global Ionospheric Map (GIM) TEC product has a time lag and is difficult to meet the real-time requirements of high-precision positioning correction.

[0007] (3) Short-term prediction models increase the onboard processing burden: Existing TEC prediction models are mostly short-term predictions and require frequent data updates, which places higher requirements on the real-time processing capabilities of the onboard system.

[0008] If TEC can be predicted over a longer period of time, the estimated ionospheric delay from these predictions can be applied to SAR data for correction, thereby reducing the impact of ionospheric delay effects. This will allow more accurate geographic location information to be obtained, which is crucial for applications that require accurate real-time data. The study of a universal medium- and long-term ionospheric TEC prediction method has important theoretical and practical significance. Summary of the invention

[0009] In order to solve the above scientific problems, the present invention provides a medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning. The method introduces combined prediction and two-step loss fine-tuning methods, and through the optimal combination of models and the optimization of losses during model training, it realizes effective prediction of medium- and long-term ionospheric TEC and further improves the SAR real-time positioning accuracy. The overall steps are:

[0010] S1: Obtain deep learning training samples and preprocess the data;

[0011] S2: Build a deep learning network model;

[0012] S3: Deep learning network model training;

[0013] S4: Two-step loss fine-tuning method and combined prediction method to optimize the model;

[0014] S5: SAR simulation test. After the model is trained, given the input TEC, the target predicted TEC can be obtained through the trained model, and then the corresponding SAR ionospheric delay value can be calculated.

[0015] Furthermore, the step S1 includes the following methods and steps:

[0016] S11: Obtain ionospheric TEC data as training samples, including historical observation data and the Global Ionospheric Map (GIM);

[0017] S12: De-noise the data and fill in missing values ​​to ensure data quality and the validity of model input;

[0018] S13: Data normalization. To enhance the simulation fitting ability, the mean of ionospheric TEC data is normalized to between -1 and 1;

[0019] S14: Dataset division: Divide the sample data into three parts: training data, validation data, and test data in proportion.

[0020] The S2 is based on the ConvGRU model and designs a deep learning framework with the encoder-predictor architecture as the main body. The network consists of multi-channel time step input, normalization and denormalization layers, convolution and deconvolution layers, ConvGRU structure and output layer.

[0021] S21: Normalization and denormalization layer. Reduce the scale difference between different features and accelerate the convergence of the network;

[0022] S22: Convolution and deconvolution layers. Reduce the number of parameters required for subsequent ConvGRU unit training;

[0023] S23: ConvGRU structure. The hidden spatiotemporal features are learned by 3 layers of ConvGRU units.

[0024] Preferably, the tanh function is used as the activation function of the network in S2 to improve the convergence speed of the network. MAE, RMSE and R2 loss functions are selected as evaluation indicators of the model.

[0025] The deep learning network model training S3 includes the following sub-steps:

[0026] S31: Train the network model. Use the sample data in S1 and the network model built in S2 for training. Improve the prediction effect of network deep learning by optimizing parameters, including adjusting the learning rate and input batches. Judge the quality of model training by the change in the loss function value of the validation set;

[0027] S32: Test the network model. After the model is trained, the prediction error is calculated using the test set data to determine the prediction effect of the model. If the prediction accuracy is not ideal, return to step S31 and retrain.

[0028] The two-step loss fine-tuning method and combined prediction method optimization model S4 includes the following sub-steps:

[0029] S41: Train multiple models with different prediction scales. Adopt the model training method in S3, use the two-step loss fine-tuning strategy, and train five prediction models with prediction intervals of 1, 2, 3, 6, and 10 days respectively. These five models are used as the base models for combined prediction;

[0030] S42: Perform combined prediction. Perform combined prediction on the required prediction target, and the base model selection adopts a greedy algorithm, that is, selects the least base model to achieve the prediction target.

[0031] The S5 SAR simulation test is to apply the model trained in S42 to SAR systems of different frequency bands, compare it with the ionospheric delay value calculated by the measured GIM TEC, and calculate the corresponding ionospheric delay value error. It specifically includes the following sub-steps:

[0032] S51: Select SAR satellites of different frequency bands. Select SAR satellite sensor data of different frequency bands for simulation testing;

[0033] S52: Calculate the ionospheric delay error. According to the model predicted TEC and the measured GIM TEC, calculate the corresponding ionospheric delay, and then the ionospheric delay prediction value error.

[0034] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0035] (1) More accurate ionospheric delay correction: The present invention uses deep learning technology to accurately predict the ionospheric TEC and uses the predicted TEC value for delay correction. This dynamic adjustment method better reflects the actual changes in the ionosphere, is suitable for high-precision SAR satellite real-time positioning needs, and significantly improves positioning accuracy and stability, especially in complex ionospheric environments.

[0036] (2) Deep learning-driven prediction enhances real-time performance: Existing correction methods often rely on GIMTEC products, which have a certain delay and cannot meet the precise positioning requirements in real-time application scenarios. The present invention constructs an ionospheric TEC prediction model based on deep learning, which can predict the TEC changes in the ionosphere in advance and get rid of the reliance on the lagging GIMTEC data. This method can not only provide accurate TEC values ​​in real time, but also help enhance the real-time positioning capability of SAR satellites, especially when encountering ionospheric fluctuations under the satellite flight path, it can make corrections in time and improve positioning efficiency.

[0037] (3) Medium- and long-term predictions reduce onboard processing pressure: Traditional ionospheric TEC prediction models mostly focus on short-term predictions and usually require frequent data updates, which places high real-time computing requirements on the onboard processing system and increases the onboard computing load. The deep learning model of the present invention has medium- and long-term ionospheric TEC prediction capabilities through optimized design, and can effectively predict the changing trend of ionospheric TEC on a longer time scale. By reducing the need for frequent data updates, the present invention significantly reduces the real-time computing pressure on the onboard processor while meeting high-precision positioning, thereby improving the overall operating efficiency and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall technical flow chart of the present invention.

[0039] Figure 2 Schematic diagram of the encoder-predictor model architecture.

[0040] Figure 3 Schematic diagram of two-step loss fine-tuning.

[0041] Figure 4 Schematic diagram of combined prediction. DETAILED DESCRIPTION

[0042] The features and exemplary embodiments of various aspects of the present invention will be described in detail below. In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and Examples. It should be understood that the specific embodiments described herein are only configured to explain the present invention and are not configured to limit the present invention. For those skilled in the art, the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by illustrating examples of the present invention.

[0043] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0044] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. Figure 1 The overall technical flow chart of the present invention is shown, which can be mainly divided into five steps: obtaining samples, building models, training models, optimizing models, and applying models. The implementation details of the present invention will be described below in conjunction with the following drawings.

[0045] S1: Obtain deep learning training samples and preprocess the data. The detailed steps are as follows:

[0046] (1) Obtain global ionospheric TEC data from 2015 to 2022 from the official website of NOAA;

[0047] (2) De-noising and missing value filling are performed on the data, and missing values ​​are filled according to the mean of adjacent elements;

[0048] (3) Normalize the data to the mean value between -1 and 1, that is:

[0049]

[0050] in, represents the normalized TEC value, X max , X min and X meanRepresents the maximum, minimum and mean values ​​of the original TEC;

[0051] (4) The sample data is divided into three parts: training set, validation set and test set in a ratio of 7:1:2. The training set is mainly responsible for model training. The validation set checks the model training status during model training, and the model stops early when the loss does not change for ten consecutive EPOCHs. The test set is used to test the final performance of the model.

[0052] S2: Construct a deep learning network model. The present invention uses the ConvGRU model as the base model and introduces the encoder-predictor network architecture. The network consists of multi-channel time step input, normalization and denormalization layers, convolution and deconvolution layers, ConvGRU structure and output layer. The overall network architecture diagram is attached. Figure 2 , as follows:

[0053] (1) Convolution and deconvolution layers. The convolution layer extracts local spatial features from the input data by using local receptive fields, and significantly reduces the number of parameters and computational overhead through parameter sharing. The deconvolution layer is used to restore the spatial information after convolution. Convolution and deconvolution compress the parameter space of the network, providing efficient and lightweight feature input for subsequent ConvGRU unit training;

[0054] (2) A deep network structure is constructed by three layers of ConvGRU units with convolution kernel sizes of (5×5), (3×3) and (1×1). ConvGRU is a recurrent neural network that combines convolution operations and can process temporal and spatial information simultaneously. The combination of three layers of convolution kernels of different sizes can learn the spatiotemporal characteristics of ionospheric TEC and improve the model's ability to capture complex spatiotemporal dependencies.

[0055] (3) The present invention preferably uses the tanh function as the activation function of the network. The range of the tanh function is [-1, 1], and the specific formula is as follows:

[0056]

[0057] (4) The present invention uses a variety of evaluation indicators, including mean absolute error (MAE), root mean square error (RMSE) and Pearson correlation coefficient (R 2 ). The details are as follows:

[0058]

[0059]

[0060]

[0061] in, is the measured TEC value, xpre is the model predicted TEC value, is the average TEC value.

[0062] S3: Deep learning network model training.

[0063] (1) Model training. Feed the training sample data divided in S1 into the network model built in S2. The quality of model training is judged by the changes in the loss function values ​​of the training set and the validation set. If the loss function values ​​of the two sets decrease synchronously and tend to be stable, the network training is completed, and the best trained model weights are saved for subsequent testing. At the same time, if the validation set decreases by no more than 0.0001 for ten consecutive cycles, it means that the model training is completed ahead of schedule and early stopping occurs;

[0064] (2) Model parameter optimization. The network training process is essentially a process of constantly updating weight parameters. With the help of a large number of samples and appropriate optimization methods, the weight parameters are continuously updated until the optimal weight parameters are found. This set of parameters is obtained through the optimization algorithm. The commonly used Admn is introduced in the present invention, and the expression is as follows:

[0065]

[0066] (3) Model testing. After training is completed, the model is verified using test set data that was not involved in the training, and the prediction error of the model is calculated. This process uses the test set for independent evaluation to ensure the generalization ability of the model, that is, its performance on new data;

[0067] (4) Error analysis. Calculate evaluation indicators such as MAE, RMSE, and R based on the prediction results on the test set and compare them with the expected accuracy requirements. If the prediction error of the model is small and the indicators meet the expected targets, it means that the model training is successful;

[0068] (5) Retraining (if necessary). If the test results are not ideal, for example, the prediction error is large or the evaluation index does not meet the requirements, return to step S31, readjust the model parameters and retrain. The prediction performance of the model can be gradually improved by changing the learning rate, optimizer or adjusting the network structure.

[0069] S4: Two-step loss fine-tuning method and combined prediction method to optimize the model;

[0070] (1) Initial model training. Using the model training method described in S3, five base models with different prediction time scales are constructed. The prediction intervals of the models are 1 day, 2 days, 3 days, 6 days, and 10 days. Each model is trained independently to enable it to predict ionospheric TEC for different time intervals. During the training process, the two-step loss fine-tuning strategy is adopted. Figure 3After training, these five base models will serve as the basis for subsequent combined predictions;

[0071] (2) Model evaluation. The five base models are tested using the validation set, and the prediction accuracy of each model is evaluated using three indicators. If some models do not achieve the expected results, they need to return to step S31 to retrain and adjust their loss functions and model parameters until all base models achieve the desired prediction performance;

[0072] (3) Design of combined prediction strategy. A combined prediction method is used to perform multi-scale combined prediction on the required prediction target. In order to ensure the efficiency and accuracy of combined prediction, a greedy algorithm is used to select the least number of base models for combination to minimize computational complexity and resource consumption. The greedy algorithm is used to select the optimal combination strategy:

[0073] Initial selection: First, select the model with the shortest prediction interval (1 day) for prediction to determine whether it can meet the prediction accuracy requirements.

[0074] Gradual expansion: If a single short-term model cannot achieve the required prediction accuracy, models with longer prediction intervals (such as 2 days, 3 days, etc.) are gradually introduced to improve the overall prediction accuracy by combining them with the short-term model.

[0075] Minimum combination principle: In the process of combined prediction, the minimum combination of base models is always given priority, that is, under the premise of meeting the prediction objectives, the number of models involved in the prediction is reduced as much as possible to reduce the computational cost and complexity.

[0076] S5: SAR simulation test. According to the test requirements, select SAR data of multiple frequency bands for simulation test, including P, L, S, C, X, and Ka bands, to ensure the application effect of the model in various actual scenarios.

[0077] (1) Calculate the model to predict the ionospheric delay. Based on the operating frequency band of the SAR satellite, the corresponding ionospheric delay value is calculated using the TEC value predicted by the model. The ionospheric delay is inversely proportional to the frequency of the TEC and SAR signals. Therefore, the impact of ionospheric delay on positioning accuracy varies in different frequency bands. The specific calculation formula is as follows:

[0078]

[0079] Where θ is the satellite incident angle, K = 40.28 [m^3 / s^2] is the reflection constant, and f is the satellite operating frequency.

[0080] (2) Comparison of measured GIM TEC. Using the measured TEC data provided by GIM, the corresponding ionospheric delay value is calculated by the same formula. GIM TEC has a certain hysteresis, but it can be used as a reference value for comparison and analysis with the TEC predicted by the model;

[0081] (3) Calculation of ionospheric delay error. Compare the ionospheric delay value predicted by the model with the delay value calculated by the measured GIM TEC to calculate the prediction error. Error analysis can be performed using the following formula:

[0082] Error = |Δs 预测 -Δs 实测 |

[0083] The accuracy of the model is evaluated by statistically analyzing the error sizes in different frequency bands.

[0084] References

[0085] [1]M.Jehle, O.Frey, D.Small, and E.Meier, "Measurement of Ionospheric TECin Spaceborne SAR Data," IEEE Trans.Geosci.Remote Sens., vol.48, no.6, pp.2460-2468, Jun.2010.

[0086] [2] M.Jehle, "Estimation of path delays, TEC and Faraday rotation from SAR data," PhD diss., University of Zurich, 2009.

Claims

1. A medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning, characterized in that: The method comprises the following steps: S1: Obtain deep learning training samples and preprocess the data; S2: Build a deep learning network model; S3: Deep learning network model training; S4: Two-step loss fine-tuning method and combined prediction method to optimize the model; S5: SAR simulation test. After the model is trained, given the input TEC, the target predicted TEC can be obtained through the trained model, and then the corresponding SAR ionospheric delay value can be calculated.

2. The medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning according to claim 1 is characterized in that: The step S1 includes the following methods and steps: S11: Obtain ionospheric TEC data as training samples, including historical observation data and the Global Ionospheric Map (GIM); S12: De-noise the data and fill in missing values ​​to ensure data quality and the validity of model input; S13: Data normalization. To enhance the simulation fitting ability, the mean of ionospheric TEC data is normalized to between -1 and 1; S14: Dataset division: Divide the sample data into three parts: training data, validation data, and test data in proportion.

3. The medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning according to claim 1 is characterized in that: The step S2 includes the following sub-steps: S21: Normalization and denormalization layer. Reduce the scale difference between different features and accelerate the convergence of the network; S22: Convolution and deconvolution layers. Reduce the number of parameters required for subsequent ConvGRU unit training; S23: ConvGRU structure. The hidden spatiotemporal features are learned by 3 layers of ConvGRU units; Preferably, the tanh function is used as the activation function of the network in S2 to improve the convergence speed of the network. MAE, RMSE and R2 loss functions are selected as evaluation indicators of the model.

4. The medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning according to claim 1 is characterized in that: The step S3 includes the following sub-steps: S31: Train the network model. Use the sample data in S1 and the network model built in S2 for training. Improve the prediction effect of network deep learning by optimizing parameters, including adjusting the learning rate and input batches. Judge the quality of model training by the change in the loss function value of the validation set; S32: Test the network model. After the model is trained, the prediction error is calculated using the test set data to determine the prediction effect of the model. If the prediction accuracy is not ideal, return to step S31 and retrain.

5. The medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning according to claim 1 is characterized in that: The step S4 includes the following sub-steps: S41: Train multiple models with different prediction scales. Adopt the model training method in S3, use the two-step loss fine-tuning strategy, and train five prediction models with prediction intervals of 1, 2, 3, 6, and 10 days respectively. These five models are used as the base models for combined prediction; S42: Perform combined prediction. Perform combined prediction on the required prediction target, and the base model selection adopts a greedy algorithm, that is, selects the least base model to achieve the prediction target.

6. The medium- and long-term ionospheric TEC combined prediction method based on two-step loss fine-tuning according to claim 1 is characterized in that: The step S5 SAR simulation test specifically includes the following sub-steps: S51: Select SAR satellites of different frequency bands. Select SAR satellite sensor data of different frequency bands for simulation testing; S52: Calculate the ionospheric delay error. According to the model predicted TEC and the measured GIM TEC, calculate the corresponding ionospheric delay, and then obtain the ionospheric delay prediction value error.

Images