GNSS-R data processing method based on improved deep learning

Abstract

The invention discloses a GNSS-R data processing method based on improved deep learning, and belongs to the technical field of satellite remote sensing monitoring. The method comprises the specific steps that S1, original DDM data and ERA5 reanalysis data are acquired; s2, performing preprocessing and quality control on the original DDM data obtained in the S1; s3, based on the ERA5 reanalysis data obtained in the S1, performing balanced data set construction on the DDM data preprocessed in the S2; s4, designing a convolutional neural network model and training the convolutional neural network model; s5, verifying and deploying the designed convolutional neural network model; the technical bottlenecks that in existing sea ice monitoring, DDM data pollution is caused due to malformation time delay, the proportion of training set seawater and sea ice samples is unbalanced, and the convolutional neural network is not adaptive to a satellite-borne platform are solved, the sea ice detection precision and real-time performance are improved, and the method has remarkable scientific value and application prospects.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing monitoring technology, and in particular to a GNSS-R data processing method based on improved deep learning. Background Technology

[0002] As global warming intensifies, the rapid changes in polar sea ice are having a profound impact on the global climate system, the ecological environment, and human activities. Traditional sea ice monitoring methods have significant limitations: optical remote sensing (such as MODIS sensors) is constrained by cloud cover and polar night phenomena, with a data loss rate of up to 80% in key areas, making it impossible to meet the needs of continuous monitoring; microwave radiometers have coarse spatial resolution (about 25 km), making it difficult to identify thin ice and melt pool features that are crucial to shipping safety, and their signals are easily interfered with by liquid water on the sea surface; although synthetic aperture radar (SAR) can provide sub-meter high-resolution images, the equipment is expensive and the revisit cycle is long (several days to several weeks), making it difficult to achieve near real-time monitoring. In recent years, while deep learning models have shown potential in sea ice monitoring, existing technologies still suffer from two major common drawbacks: First, the ratio of seawater to sea ice time-delay Doppler maps (DDMs) in the training set is severely imbalanced (e.g., 7:1), leading to insufficient model generalization ability, manifested in thin ice identification root mean square errors exceeding 0.3 meters and melt pool detection accuracy generally below 65%; Second, there is a lack of a systematic process to filter out distorted DDM data affected by noise, systematic errors, or geometric effects, and noise contamination of input data significantly reduces model robustness. These technical bottlenecks severely restrict the development of high-precision, high-timeliness sea ice monitoring capabilities. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies, such as malformed DDM data pollution, severe imbalance in the ratio of seawater to sea ice samples in the training set, and the incompatibility of traditional convolutional neural network structures with spaceborne platforms. This invention proposes a GNSS-R data processing method based on improved deep learning. First, malformed DDMs are screened out by constructing an ADDM value with a threshold of less than 0.02. Then, a balanced training set construction strategy with a 1:1 ratio of seawater to sea ice is used to suppress decision bias. Finally, a suitable convolutional neural network architecture (including specific convolutional kernel configuration and Dropout regularization) is designed to overcome spaceborne computing power limitations, ultimately achieving a sea ice detection accuracy of up to 97.3%.

[0004] To achieve the above objectives, this invention provides a GNSS-R data processing method based on improved deep learning, comprising the following steps: S1. Obtain raw DDM data and ERA5 reanalysis data, including sea ice concentration, sea surface temperature and sea surface wind field. S2. Preprocess and quality control of the raw DDM data obtained in S1; S3. Based on the ERA5 reanalysis data obtained in S1, construct a balanced dataset from the preprocessed DDM data in S2; S4. Design a convolutional neural network model and train it; S5. Verify and deploy the designed convolutional neural network model.

[0005] Preferably, S2 specifically includes: Edge values, normalized moment of inertia, and related parameters are extracted from the original DDM data and its metadata file. The time information is then transformed and integrated into a NetCDF file. Finally, the aligned normalized DDM eigenvalues ​​(ADDM) are calculated, as shown in the following expression: ; In the formula, i is the Doppler cell index; j is the time delay cell index; This represents the normalized power value of the DDM at the i-th Doppler unit and the j-th time delay unit. Based on this feature value, a threshold of ADDM<0.02 was set to screen out malformed DDMs, and the seawater and sea ice categories of the remaining data were labeled using ERA5 reanalysis data to complete data quality control.

[0006] Preferably, the balanced dataset in S3 is constructed using DDM category labels determined based on ERA5 reanalysis data, and samples are randomly drawn at a seawater:sea ice ratio of 1:1 to construct the training set and test set.

[0007] Preferably, S4 specifically includes: S401. Design the network structure of a convolutional neural network model; S402. Initialize the parameters of the convolutional neural network model; S403, Configure the optimizer for the convolutional neural network model; S404. Construct the loss function; S405. Use regularization strategies to suppress overfitting in convolutional neural network models.

[0008] Preferably, the network structure of the convolutional neural network model in S401 adopts a three-layer convolutional neural network structure, including an input layer, convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, convolutional layer 3, fully connected layer 1, dropout layer, and output layer.

[0009] Preferably, the specific content of initializing the parameters of the convolutional neural network model in S402 is as follows: The Kaiming initialization method is used to set initial values ​​for the weights, ensuring stable gradient propagation during training and avoiding gradient vanishing or exploding problems. For the ReLU activation function, the standard deviation expression for initializing the weights W is as follows: ; In the formula, The number of input neurons in the current layer; Preferably, the optimizer configured in S403 for the convolutional neural network model contains the following details: The Adam optimizer dynamically adjusts the learning rate by calculating the exponential moving average of the first and second moments of the gradient, as shown in the following expression: The first-order moment estimate is expressed as follows: ; In the formula, Let the first-order momentum be at the current moment; The first-order momentum at the previous moment; The gradient at the current time step; The decay coefficient of first-order momentum; The second-order moment estimate is expressed as follows: ; In the formula, Let be the second momentum at the current moment; This is the second momentum of the previous moment; The element-wise square of the current gradient; The decay coefficient of the second momentum; The first-order moment correction is expressed as follows: ; In the formula, Let the first-order momentum be at the current moment; The first-order momentum corrected at the current moment; The decay coefficient of the first-order momentum at the current moment; The second-order moment correction is expressed as follows: ; In the formula, Let be the second momentum at the current moment; The corrected second-order momentum at the current moment; The decay coefficient of the second momentum at the current moment; The parameter is updated using the following expression: ; In the formula, These are the parameters of the convolutional neural network model at the current moment; These are the updated parameters for the convolutional neural network model. The learning rate; It is a small constant that prevents division by zero.

[0010] Preferably, the specific content of constructing the loss function in S404 is as follows: The cross-entropy loss function is used as the optimization objective to measure the difference between the predicted probability distribution and the true label distribution of the convolutional neural network model. The expression is as follows: ; In the formula, Number of categories; This is a real label; To predict the probability of a convolutional neural network model belonging to the i-th class, This represents the loss function.

[0011] Preferably, the specific details of the regularization strategy used in S405 to suppress overfitting in the convolutional neural network are as follows: This approach combines L2 regularization and Dropout to suppress overfitting of the convolutional neural network (CNN) model to the training data by randomly dropping neurons with a set probability during training. A Dropout layer is introduced into the CNN model, setting a dropout probability after the first two fully connected layers. This weakens the co-adaptation relationship between neurons; a dropout probability is set after the last fully connected layer near the output layer. Maintain a certain regularization strength while retaining effective information for the final classification of seawater and sea ice.

[0012] Preferably, in S5, the convolutional neural network model designed in S4 is validated and deployed using four metrics: accuracy, precision, recall, and F1 score.

[0013] Therefore, the present invention employs the above-described method, which has the following beneficial effects: (1) By using an ADDM threshold of less than 0.02 to automatically screen out malformed DDM data, the screening rate is close to 100% and the false screening rate is less than 0.5%, which significantly improves the quality of input data and increases the thin ice recognition rate by 15 percentage points, effectively solving the feature space offset problem caused by noise pollution. (2) Innovatively adopting the strategy of constructing a 1:1 balanced training set of seawater and sea ice, and eliminating class bias through hierarchical random sampling technology, the model's cross-scene accuracy fluctuation is less than 0.8%, which greatly enhances the generalization ability in complex scenarios such as polar night and different seasons; (3) Design a three-layer convolutional neural network architecture (with 4-8-16 convolutional kernels and Dropout 0.2 regularization) to meet the real-time processing requirements of the satellite. Through triple innovations in data quality control, balanced training strategies, and convolutional neural network model design, synergistic optimization is achieved in three dimensions: accuracy, speed, and power consumption. This significantly improves the reliability, real-time performance, and practicality of sea ice monitoring, providing a complete technical solution for building an operational polar environment monitoring system. It has significant scientific value and broad application prospects.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a flowchart of a GNSS-R data processing method based on improved deep learning according to the present invention; Figure 2 This is a map showing the coverage area of ​​TDS-1 satellite data. Figure 3 The following are graphs based on ERA5 reanalysis data: (a) is a spatial distribution map of sea ice concentration based on ERA5 reanalysis data, (b) is a spatial distribution map of sea surface temperature based on ERA5 reanalysis data, and (c) is a spatial distribution map of sea surface wind field based on ERA5 reanalysis data. Figure 4 This is a diagram of the convolutional neural network architecture used. Figure 5 A graph showing the change in the loss function during the training process of the Adam optimizer; Figure 6 This is a malformed DDM diagram; Figure 7 The graphs show the changes in screening rate, screening accuracy, and false screening rate under different thresholds. Figure 8 A comparison of sea ice detection accuracy before and after screening for abnormal DDM. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0017] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0018] The following is combined Figures 1-8 The embodiments of the present invention will be described in detail below.

[0019] Example 1 A GNSS-R data processing method based on improved deep learning consists of two stages: a data quality control stage and a model training and optimization stage, such as... Figure 1 As shown, the original time-delay Doppler image (DDM) data is first acquired and preprocessed. Abnormal DDM samples with ADDM values ​​greater than 0.02 are filtered out by calculating ADDM feature values. Then, a stratified random sampling method is used to construct a balanced training set with a seawater to sea ice ratio of 1:1. Next, the high-quality training set is input into the constructed convolutional neural network model for training and optimization. Iterative training is used to ensure the model performance meets requirements, ultimately achieving high-precision monitoring of sea ice.

[0020] Specifically, the following steps are included: S1. Obtain raw DDM data and ERA5 reanalysis data; Level-1B time-delay Doppler image (DDM) data, provided by the TDS-1 satellite, covers areas with latitude greater than 55°, spanning February and November 2018. The data was downloaded from ftp.merrbys.co.uk / Data / L1B / and includes DDM images with a time delay of 128 pixels and a Doppler resolution of 20 pixels. Figure 2 The image shows the coverage area of ​​TDS-1 satellite data.

[0021] Simultaneously, ERA5 reanalysis data (sea ice concentration, sea surface temperature, and sea surface wind field) were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) as a validation benchmark. Figure 3 The data shown is from the ERA5 reanalysis.

[0022] S2. Preprocess and quality control of the raw DDM data obtained in S1; First, Edge values, normalized moment of inertia, and related parameters are extracted from the original DDM data and its metadata file. Time information is then converted into daily second counts and integrated into a NetCDF file. Subsequently, aligned and normalized DDM eigenvalues ​​are calculated, representing the normalized power mean of the DDM within a window of 20 Doppler elements and 40 time-delay elements. The expression is as follows: ; In the formula, i is the Doppler cell index; j is the time delay cell index; This represents the normalized power value of the DDM at the i-th Doppler unit and the j-th time delay unit. Based on this feature, a threshold of ADDM < 0.02 was set to screen out malformed DDMs, and ERA5 reanalysis data was used to complete the labeling of seawater and sea ice categories in the remaining data to achieve data quality control.

[0023] The main reasons for employing the ADDM threshold screening mechanism of this invention are as follows: Figure 7 As shown in the results, extensive experimental verification revealed that as the threshold increases, the screening rate gradually rises and then plateaus; the screening accuracy continues to increase, approaching 100% at a threshold of 0.02; while the false screening rate continuously decreases, approaching 0 at a threshold of 0.02. Considering all three indicators, this invention selects 0.02 as the optimal threshold for screening deformed DDMs. At this threshold, the method maintains high screening accuracy while minimizing the false screening rate, providing a high-quality data foundation for subsequent model training.

[0024] S3. Based on the ERA5 reanalysis data obtained in S1, construct a balanced dataset from the preprocessed DDM data in S2; Based on ERA5 reanalysis data, DDM category labels (seawater / sea ice) were determined. Samples were randomly drawn at a seawater:sea ice ratio of 1:1 to construct training and test sets, as shown in Table 1. Table 1 Dataset Distribution

[0025] Validated using 91,106 independent test samples from November 2018, the method of this invention significantly outperforms the method before screening for malformed DDM in multiple metrics. For example... Figure 8 As shown, the overall accuracy reached a maximum of 96.3%, an improvement of nearly 3.0 percentage points compared to the previous result. Before removing distorted DDMs, the model's average accuracy on each training set was generally low, and it exhibited non-linear fluctuations with the increase or decrease of the sea ice sample ratio in the training set: when the ratio of seawater to sea ice DDMs gradually transitioned from 7:1 to 1:7, the accuracy dropped continuously from 96.3% to 84.5%, with a fluctuation range of 11.8%, indicating that distorted DDMs amplified the bias caused by class imbalance. After removing distorted DDMs, the average accuracy of all training sets increased, and the fluctuation range narrowed significantly (within ±0.8%), with the overall curve becoming flatter, showing that data quality control significantly improved the model's stability and robustness. Among them, the training set G with an exact water-to-ice ratio of 1:1 benefited the most, with its accuracy further increasing from 93.4% to 96.3%, verifying the key role of high-quality, balanced samples in performance.

[0026] S4. Design a convolutional neural network model and train it; This invention designs a lightweight convolutional neural network model suitable for DDM image processing, specifically including the following steps: S401. Design the network structure of the convolutional neural network model: The three-layer convolutional neural network structure adopted is as follows: Figure 4 As shown.

[0027] Figure 4 The number of layers in the convolutional neural network structure is described in Table 2.

[0028] Table 2 Convolutional Neural Network Structure Parameters

[0029] S402. Initializing the parameters of the convolutional neural network model: This invention uses the Kaiming initialization method. By setting the initial values ​​of the weights, the network can maintain stable gradient propagation during training, thereby avoiding the problems of gradient vanishing or exploding. For the ReLU activation function, the standard deviation of the weights W is initialized as follows: ; In the formula, The number of input neurons in the current layer.

[0030] S403. Configuring the optimizer for the convolutional neural network model: The Adam optimizer is used, combining the advantages of the momentum and RMSProp optimizers. It dynamically adjusts the learning rate by calculating the exponential moving average of the first moment (mean) and second moment (uncentered variance) of the gradient, as shown in the following expression: The first-order moment estimate is expressed as follows: ; In the formula, Let the first-order momentum be at the current moment; The first-order momentum at the previous moment; The gradient at the current time step; The decay coefficient of first-order momentum; The second-order moment estimate is expressed as follows: ; In the formula, Let be the second momentum at the current moment; This is the second momentum of the previous moment; The element-wise square of the current gradient; The decay coefficient of the second momentum; The first-order moment correction is expressed as follows: ; In the formula, Let the first-order momentum be at the current moment; The first-order momentum corrected at the current moment; The decay coefficient of the first-order momentum at the current moment; The second-order moment correction is expressed as follows: ; In the formula, Let be the second momentum at the current moment; The corrected second-order momentum at the current moment; The decay coefficient of the second momentum at the current moment; The parameter is updated using the following expression: ; In the formula, These are the model parameters at the current moment; These are the updated model parameters; The learning rate; It is a small constant that prevents division by zero.

[0031] The Adam optimizer combines momentum mechanisms and adaptive learning rate adjustment strategies, significantly improving the stability and convergence speed of model training. Specifically, as follows... Figure 5 As shown.

[0032] S404. Constructing the Loss Function: This invention uses the Cross-Entropy Loss function as the optimization objective, which can effectively measure the difference between the model's predicted probability distribution and the true label distribution. The expression is as follows: ; In the formula, Number of categories; This is a real label; To predict the probability of a convolutional neural network model belonging to the i-th class, This represents the loss function.

[0033] By minimizing the cross-entropy loss, the model can gradually improve classification accuracy. During training, minimizing the cross-entropy loss is equivalent to maximizing the log-likelihood of the correct class, thereby driving the model parameters to be updated in the direction of maximizing classification accuracy.

[0034] S405. Suppressing Overfitting in Convolutional Neural Network Models Using Regularization Strategies: This invention employs a combination of L2 regularization and Dropout. By randomly dropping some neurons with a certain probability during training, it effectively suppresses overfitting of the model to the training data. This invention introduces a Dropout layer into the model, setting a dropout probability after the first two fully connected layers. =0.2, to significantly weaken the co-adaptation relationship between neurons. A lower dropout probability is set after the last fully connected layer near the output. =0.1, while maintaining a certain regularization strength, retains more effective information for the final classification of seawater and sea ice.

[0035] S5. Verify and deploy the designed convolutional neural network model; To comprehensively evaluate model performance, this invention uses four metrics for verification: accuracy, precision, recall, and F1 score. Accuracy represents the proportion of correctly classified samples in all test samples and is a direct indicator of the model's overall classification performance. Precision measures the accuracy of the model when predicting a positive class (sea ice), i.e., the proportion of samples predicted as sea ice that are actually sea ice. Recall measures the proportion of samples that are actually positive (sea ice) that are correctly predicted, i.e., the proportion of real sea ice samples that are correctly identified. The F1 score, as a harmonic mean that comprehensively considers precision and recall, is suitable for imbalanced datasets, can balance the influence of these two metrics, and provides a more comprehensive evaluation of model performance.

[0036] This invention has broad applicability and promotional value. It adopts a modular design concept, decoupling the core algorithm from the hardware platform. It can be applied to various GNSS-R tasks such as TDS-1, CYGNSS, and FY-3E. It can be quickly migrated and deployed without modifying the core algorithm, which greatly reduces the adaptation cost and usage threshold between different platforms.

[0037] This invention provides an innovative approach to solving the problem of sample imbalance in the field of remote sensing. The 1:1 balanced training strategy is not only applicable to sea ice monitoring, but can also be extended to multiple remote sensing application fields such as soil moisture inversion and vegetation monitoring, which has important methodological significance and promotional value.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A GNSS-R data processing method based on improved deep learning, characterized in that, The method comprises the following steps: S1, obtaining original DDM data and ERA5 reanalysis data, wherein the ERA5 reanalysis data comprises sea ice density, sea surface temperature and sea surface wind field; S2, preprocessing and quality control of the original DDM data obtained in S1; S3, constructing a balanced data set based on the DDM data preprocessed in S2 and the ERA5 reanalysis data obtained in S1; S4, designing a convolutional neural network model and training the same; S5, verifying and deploying the designed convolutional neural network model.

2. The GNSS-R data processing method based on improved deep learning according to claim 1, characterized in that, S2 specifically comprises: The Edgea value, normalized rotational inertia and related parameters are extracted from the original DDM data and its metadata file, the time information is converted, then integrated and stored in a NetCDF file, and finally the aligned normalized DDM characteristic value ADDM is calculated, and the expression is as follows: ； where i is the Doppler bin index; j is the delay bin index; is the normalized power value of DDM at the i-th Doppler bin and the j-th delay bin. Based on the characteristic value, the threshold of ADDM<0.02 is set to exclude abnormal DDM, and the sea water and sea ice category labeling of the remaining data is completed using the ERA5 reanalysis data, and the data quality control is completed.

3. The GNSS-R data processing method based on improved deep learning according to claim 1, characterized in that: The balanced data set construction in S3 is a DDM category label determined based on the ERA5 reanalysis data, and the training set and test set are constructed by randomly sampling samples in the proportion of sea water: sea ice = 1:

1.

4. The GNSS-R data processing method based on improved deep learning according to claim 1, characterized in that, S4 specifically comprises: S401, designing the network structure of the convolutional neural network model; S402, initializing the parameters of the convolutional neural network model; S403, configuring the optimizer of the convolutional neural network model; S404, constructing a loss function; S405, using a regularization strategy to suppress overfitting of the convolutional neural network model.

5. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that: The network structure of the convolutional neural network model in S401 adopts a three-layer convolutional neural network structure, including an input layer, a convolutional layer 1, a pooling layer 1, a convolutional layer 2, a pooling layer 2, a convolutional layer 3, a fully connected layer 1, a Dropout layer and an output layer.

6. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that, The specific content of initializing the parameters of the convolutional neural network model in S402 is as follows: The Kaiming initialization method is used to set the initial value of the weight, so that the network maintains stable gradient propagation during training, avoiding the problem of gradient vanishing or explosion; for the ReLU activation function, the standard deviation of the weight W is initialized as follows: ； In the formula, The number of current layer input neurons.

7. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that, The specific content of configuring the optimizer of the convolutional neural network model in S403 is as follows: The Adam optimizer is used to dynamically adjust the learning rate by calculating the exponential moving average of the first and second moments of the gradient, and the expression is as follows: First moment estimation, expression as follows: ； wherein is the first order momentum at the current time instant; is the first order momentum at the previous time instant; is the gradient at the current time instant; is the decay coefficient of the first order momentum; Second moment estimation, expression as follows: ； wherein is the second order momentum at the current time instant; is the second order momentum at the previous time instant; is the element-wise square of the current gradient; is the decay coefficient for the second order momentum; First moment correction, expression as follows: ； In the formula, is the first order momentum at the current time instant; is the first order momentum after correction at the current time instant; is the decay coefficient of the first order momentum at the current time instant; Second moment correction, expression as follows: ； In the formula, is the second order momentum at the current time instant; is the second order momentum at the current time instant after correction; is the decay coefficient of the second order momentum at the current time instant; Parameter update, expression as follows: ； In the formula, is the convolutional neural network model parameter at the current time; is the updated convolutional neural network model parameter; is the learning rate; is a small constant to prevent division by zero.

8. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that, The specific content of constructing a loss function in S404 is as follows: The cross-entropy loss function is used as the optimization objective to measure the difference between the predicted probability distribution of the convolutional neural network model and the true label distribution, and the expression is as follows: ； wherein is the number of classes; is the true label; is the probability predicted by the convolutional neural network model that the i-th class belongs to, denotes the loss function.

9. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that, The specific content of using a regularization strategy to suppress overfitting of the convolutional neural network model in S405 is as follows: The L2 regularization and the Dropout are combined, neurons are randomly discarded with a set probability in the training process, and the overfitting phenomenon of the convolutional neural network model to the training data is inhibited; a Dropout layer is introduced into the convolutional neural network model, and the dropout probability is set after the first two full connection layers , the coadaptation relationship between neurons is weakened; the dropout probability is set after the full connection layer of the last layer close to the output layer , the regularization strength is maintained, and effective information is retained for the final classification of seawater and sea ice.

10. The GNSS-R data processing method based on improved deep learning according to claim 4, characterized in that: The accuracy, precision, recall and F1 score are used to verify the convolutional neural network model designed in S4.

Images