A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion

Abstract

The present invention discloses a hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion, including: obtaining CYGNSS L1 GNSS-R data, European Center for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data set, IMERG rainfall product data and WaveWatch III SWH data; preprocessing and quality controlling the acquired data; performing spatiotemporal matching on the data set obtained after quality control, and dividing it into a training set, a test set and a validation set according to a certain ratio; constructing a hybrid deep learning model combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM, and training and validating the model; creating a MultiScaleWaveFormer model deep learning SHAP interpreter, calculating and visualizing the importance of input features; inputting into the trained MultiScaleWaveFormer model to obtain an inversion result; finally comparing the model with the SWH inversion results of the existing model to evaluate the performance of the model. The present invention combines five modules to achieve higher efficiency and functionality, so that the accuracy of spaceborne GNSS-R data inversion of significant wave height is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-research of GNSS reflection measurement and deep learning, and specifically, relates to a hybrid deep learning model construction and output interpretation method for satellite-borne GNSS-R sea surface effective wave height inversion. Background Art

[0002] Significant Wave Height is an important element of ocean dynamic environment monitoring. It is widely used to describe and evaluate the characteristics of ocean waves and is usually used for sea condition monitoring to ensure safe navigation at sea. It is of great significance for ocean shipping and marine resource protection. With the continuous advancement of scientific and technological means, the measurement technology of SWH has also undergone a major transformation from the initial traditional subjective observation to modern high-tech automatic monitoring. This process also reflects that human beings are deepening their exploration of the marine environment and have a deeper understanding of the marine dynamic environment. As for how to more accurately estimate the global ocean significant wave height (SWH), relevant researchers are constantly studying many new methods.

[0003] At present, the feasibility of research on inverting significant wave height based on spaceborne GNSS-R data has been effectively confirmed in terms of empirical models and machine learning models. However, the empirical model inverting significant wave height has the disadvantages of relying on limited input variables, data quality, applicability limitations, ignoring dynamic changes and nonlinear effects, uncertainty of model parameters, computational complexity, and ignoring local features, making it difficult to apply in practice. Machine learning algorithms can easily ignore the key feature information in DDM images in the model input data, which limits the accuracy of significant wave height inversion. Existing deep learning methods also have the disadvantages of ignoring time series characteristics, prone to overfitting, and only being able to make predictions based on current and past information. Although deep learning models can achieve excellent performance in many tasks, such as image classification, natural language processing, and remote sensing data processing, their complex structure also makes these models typical "black box models". This lack of transparency hinders researchers' understanding of the model decision-making process, especially in fields that require high precision and high reliability, such as significant wave height prediction in spaceborne GNSS-R ocean remote sensing. The accuracy of the model is of vital importance, but the interpretability of the model is also not to be ignored. Explanatory deep learning models can not only help us understand the prediction mechanism of the model, but also improve the credibility and controllability of the model. For example, researchers can explain the decision-making process of the model to discover potential patterns in the data or relationships between features, and then optimize the model to improve the output accuracy in practical applications.

[0004] In recent years, with the development of Explanatory Artificial Intelligence (XAI), tools such as SHAP (Shapley Additive Explanations) have begun to be applied to the interpretation of deep learning models. These tools can assign each feature its contribution to the model prediction results, thereby revealing the decision-making mechanism of the deep learning model under different feature inputs. In the inversion of spaceborne GNSS-R significant wave height, deep learning models combined with interpretation methods such as SHAP, InterpretML or Captum can provide a clearer understanding of the key features and influencing factors in significant wave height prediction. This not only helps to optimize the model structure, but also improves the stability and reliability of predictions in practical applications, especially when dealing with extreme ocean environments or rare wave events.

[0005] The convolutional neural network combined with the efficient multi-scale attention module EMA (Efficient Multi-Scale Attention Module) can use the multi-scale parallel sub-network structure to capture the short-range and long-range dependencies in the image, avoid the information loss caused by channel dimensionality reduction, and use the cross-space learning method to capture the pixel-level pairwise relationship, which can improve the aggregation effect of features. Combined with Transformer, its multi-head self-attention mechanism can be used to capture rich features and process the relationship between each position in the input sequence, significantly improving the training efficiency. The bidirectional convolutional long short-term memory network (BiConvLSTM) combined with the two-dimensional convolutional neural network can better understand the input sequence while extracting time-related features and capturing context information; MLP (Multi-Layer Perceptron) is used multiple times for forward propagation and back propagation, and the input data is processed through the weights and activation functions of each layer to finally generate output. It can be seen that the hybrid deep learning method of multi-model fusion can combine the advantages of multiple models to make up for the defects of the single model method during use, thereby improving the accuracy of effective wave height inversion of space-borne GNSS-R data.

[0006] In view of this, the present invention is proposed. Summary of the invention

[0007] In order to overcome the limited applicability of a single empirical model in the process of inversion of sea surface wave height by satellite GNSS-R, the key feature extraction and fusion problems of a single machine learning model, and the limited accuracy improvement caused by this, the present invention provides a hybrid deep learning model construction and output interpretation method for inversion of sea surface effective wave height by satellite GNSS-R. The method combines five modules, namely, a convolutional feature extraction module, an efficient multi-scale attention module EMA (Efficient Multi-Scale Attention Module), a Transformer, a feature relationship reasoning module, and a multi-layer perceptron (MLP). The idea of ​​multi-module fusion aims to effectively integrate multiple functional modules, and the constructed hybrid model is named MultiScaleWaveFormer. The technical solution of the present invention can further improve the ability to achieve high-precision and high-resolution sea surface wave height inversion on a global scale using satellite GNSS-R technology, and promote the development of interpretable hybrid deep learning models in satellite GNSS-R sea surface wave height monitoring.

[0008] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0009] A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion includes the following steps:

[0010] Step S1, obtaining CYGNSS L1-level spaceborne GNSS-R data, European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis dataset, IMERG rainfall product data, and WaveWatch III (WW3) SWH data;

[0011] Step S2, preprocessing and quality control of the acquired data;

[0012] Step S3, further performing spatiotemporal matching on the data set obtained after quality control, and then dividing it into a training set, a test set, and a validation set according to a certain ratio;

[0013] Step S4, construct a hybrid deep learning model for spaceborne GNSS-R sea surface significant wave height inversion combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM, and train and verify the MultiScaleWaveFormer model;

[0014] Step S5, create a MultiScaleWaveFormer model deep learning SHAP interpreter to calculate and visualize the importance of input features;

[0015] Step S6, input the test set data into the trained MultiScaleWaveFormer model to obtain the inversion result;

[0016] Step S7, compare the SWH inversion results of this model with those of the existing model to evaluate the performance of the new model.

[0017] Furthermore, the CYGNSS L1-level satellite-borne GNSS-R observation data in step S1 includes L1b power DDM (power_analog), bistatic radar cross section (BRCS), effective scattering area (eff_scatter), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), DDM acquisition time (ddm_timestamp_utc), longitude and latitude of the specular reflection point (sp_lon, sp_lat) and incident angle (sp_inc_angle);

[0018] The European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data set includes wind speed, wind direction, water depth and SWH data: the ERA5 wind speed is the composite wind speed of the U component and the V component 10 m above the sea surface; the wind direction data is the wind direction data (10 m wind direction) provided in the ERA5 reanalysis data; the water depth data is the water depth (Model bathymetry) provided in the ERA5 product; the significant wave height data is the significant wave height of combined windwaves and swell data (Significant the height of combined windwaves and swell) provided in the ERA5 reanalysis;

[0019] The IMERG rainfall data is the IMERG final operation (IMERG-F) rainfall intensity.

[0020] Furthermore, step S2 includes the following sub-steps:

[0021] Step S2.1, extract observation variable parameters from the onboard GNSS-R data and calculate characteristic observation values, including NBRCS, LES, receiver antenna gain, reflection point longitude, reflection point latitude and incident angle characteristic parameters;

[0022] Step S2.2 uses some quality control flags to remove data of lower quality. These flags include: if the absolute value of the spacecraft roll is between 1 and 30 degrees, the pitch is between 1 and 10 degrees, or the yaw is between 1 and 5 degrees, then set (Bit 03: small_sc_attitude_err); if the specular reflection point is not over land, then set (Bit 11: sp_over_land); if the specular reflection point is within 25 kilometers of land, then set (Bit 12: sp_very_near_land); if the calculated specular reflection point delay row based on zero is less than 6 or greater than 10, then set (Bit 19: brcs_ddm_sp_bin_delay_error); if the calculated specular reflection point Doppler column based on zero is less than 4 or greater than 6, then set (Bit 20: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 30: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 31: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 32: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 33: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 34: brcs_ddm_sp_bin_dopp_error); if the 3x used to calculate ddm_nbrcs is greater than 4, then set (Bit 35: brc 5brcs area has a negative value, it is set (Bit 21: neg_brcs_value_used_for_nbrcs); mirror point does not exist (Bit23: sp_non_existent_error); insufficient blackbody data for calibration (Bit 26: bb_framing_error); insufficient data to properly calibrate the NBRCS for the GPS satellites for this sample (Bit 28: low_quality_gps_ant_knowledge); spacecraft altitude is outside the nominal altitude range. The nominal altitude is defined as 490 km to 550 km (Bit 29: sc_altitude_out_of_nominal_range); set when an anomaly is observed on a CYGNSS or GPS satellite when this sample is collected. The data produced by these samples is inconsistent with typical CYGNSS results (Bit 30: anomalous_sampling_period).

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] In step S3.1, the CYGNSS GNSS-R L1 data are spatially and temporally matched with the ECMWF ERA5 wind speed data, ERA5 wind direction data, ERA5 water depth data, ERA5 SWH data, WaveWatch III significant wave height data, and IMERG rainfall data to obtain the matched data set.

[0025] Step S3.2, randomly divide the matched data set into three parts: training set, validation set and test set according to the ratio of 60%, 15% and 25%. The training set is used to train the network model, the validation set is used to supervise the model training, and the test set is used to avoid overfitting.

[0026] Furthermore, the hybrid deep learning model for inversion of space-borne GNSS-R sea surface significant wave height combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM described in step S4 mainly includes five modules: the first network module is a convolutional feature extraction module, which takes power_analog, BRCS and effective scattering area (eff_scatter) images as input, and uses a two-dimensional convolutional neural network to automatically and effectively extract spatial feature information related to the sea surface significant wave height from the power_analog, BRCS and effective scattering area (eff_scatter) images.

[0027] The second network module is the Efficient Multi-Scale Attention Module (EMA), which uses a multi-scale parallel sub-network structure to capture short-range and long-range dependencies in the image, reorganizes some channels into batch dimensions to avoid information loss caused by channel dimensionality reduction, and uses cross-space learning methods to fuse the outputs of parallel sub-networks to capture pixel-level pairwise relationships and improve feature aggregation. The efficient attention mechanism of the EMA module can significantly improve the performance of image classification and target detection tasks while using fewer parameters and reducing computational overhead. The module output of EMA can be expressed as:

[0028]

[0029] Here we assume that the input feature map is X∈R C×H×W , where C is the number of channels, H is the height, and W is the width. The multi-scale extracted features are recorded as X s , s is the number of scales, GAP represents the global average pooling process, which is used to obtain the global context information of each scale feature, A s =σ(W a ·GAP(X s )+b a ) represents the attention weight, W a and b a is a learnable parameter and σ represents the activation function.

[0030] The third module introduces the Transformer deep learning architecture, uses the multi-head attention mechanism, calculates attention in parallel in multiple different representation subspaces, captures rich features, uses the self-attention mechanism to process the relationship between each position in the input sequence, and allows parallel processing of all positions in the sequence, significantly improving training efficiency. Multi-head self-attention can be expressed as:

[0031] MultiHead(Q,K,V)=Concat(head1,head2,...,head n )W O (2)

[0032] The output for each header is:

[0033]

[0034] In formulas (2) and (3), Q represents the query vector (the input feature that currently needs to be paid attention to); K represents the key vector (each input feature in the dataset); and V represents the value vector (the information corresponding to each key). represents the output weight matrix, The learned weight matrix, d k is the dimension of each head.

[0035] The fourth module is the feature relationship reasoning module, which combines the bidirectional convolutional long short-term memory network (BiConvLSTM) with the two-dimensional convolutional neural network to extract time-related features and capture context information to better understand the input sequence. Suppose the input sequence is X = (x1, x2, ..., x T ), where x t is the input at time t. In BiConvLSTM, we use two LSTM layers, forward and reverse, to perform convolution on the input. Let the convolution operation be C, which can be expressed as:

[0036]

[0037] in, Represents the feature map processed by the convolution layer, W c is the convolution kernel weight, b c For bias.

[0038] Use the forward ConvLSTM to process the input sequence from t = 1 to t = T, and the reverse ConvLSTM to process the input sequence from t = T to t = 1. The final output is obtained by concat the outputs of these two directions:

[0039]

[0040] The final output of BiConvLSTM can be expressed as:

[0041] y t =W y ·h t +b y (6)

[0042] Among them, W y is the final output weight matrix, b y is the output bias.

[0043] The fifth network module uses a multi-layer perceptron (MLP) to process the input data through the weights and activation functions of each layer and finally generate output. For neurons from the input layer to the lth layer, the input x and weight W (l) and bias b (l) The calculation formula can be expressed as:

[0044] z (l) =W (l) ·x+b (l) (7)

[0045] Here, W (l) represents the weight matrix of the lth layer, b (l) represents the bias vector of the lth layer. Finally, the activation function f is applied to get the output of this layer:

[0046] a (l) =f(z (l) ) (8)

[0047] The training of the MultiScaleWaveFormer hybrid deep learning model described in step S4 selects the improved S-Huber loss function of the present invention. The loss function introduces a scaling parameter h on the basis of the Huber loss and combines the weighted sum of the mean square error (MSE). The formula is:

[0048] S-Huber(y true ,y pred )=h×Huber(y true ,y pred )+(1-h)×MSE(y true ,y pred ) (9)

[0049] Among them, Huber, as a loss function for dealing with outliers, is defined as:

[0050]

[0051] When |y true -y pred When |≤δ, the error is small and the loss is mainly in square form (similar to mean square error MSE). true -y pred When |>δ, the error is large and the loss adopts a linear form to reduce the impact of outliers on the overall model.

[0052] The mean square error (MSE) is defined as:

[0053]

[0054] In formulas (9), (10), and (11), h is a scaling factor between 0 and 1, which is used to distribute weights between Huber loss and MSE. true ,y pred ) is the standard Huber loss, MSE(y true ,y pred ) is the mean square error loss, y true is the true value, y pred is the predicted value, δ is the threshold for adjusting abnormal sensitivity, and N is the number of samples.

[0055] Furthermore, step S5 of creating a MultiScaleWaveFormer model deep learning SHAP interpreter and calculating and visualizing the importance of input features specifically includes the following sub-steps:

[0056] Step S5.1, install the SHAP (SHapley Additive exPlanations) library in the deep learning environment tensorflow framework;

[0057] Step S5.2, select 1000 sample input data as the benchmark data set (background data) for creating the MultiScaleWaveFormer model SHAP deep learning interpreter;

[0058] Step S5.3, use the interpreter to calculate the Shapley value for the model's input data. The Shapley value represents the contribution of each input feature to the prediction result; the Shapley value The calculation formula can be expressed as:

[0059] Among them, n(i) represents the number of times sample i is selected, represents the predicted mean of all samples, f(x ij ) is the jth feature x of the predicted sample i ijThe prediction contribution to sample i, K is the number of features of sample i.

[0060] Step S5.4, visualize the SHAP value. Plot the SHAP value of a single sample to explain how the model makes a specific prediction; visualize the global importance of features to see the importance of the overall features; visualize feature dependencies to see how a single feature affects the prediction and interacts with other features through dependency graphs. The SHAP value of each sample can be expressed as:

[0061] Shap value=explainer(X) (13)

[0062] Among them, explainer(X) is the explainer created based on the MultiScaleWaveFormer model, and X represents the sample.

[0063] Furthermore, step S6 is specifically implemented as follows: a test set containing 25% of the data is used in the verified model to obtain an output result.

[0064] Furthermore, the performance evaluation indicators selected in step S7 are: root mean square error (RMSE), bias (Bias), mean absolute percentage error (MAPE) and Pearson correlation coefficient (CC), and the relevant calculation formulas are as follows:

[0065]

[0066] Bias=E[F(x)]-A(x) (15)

[0067]

[0068]

[0069] In the above formulas (14), (15), (16) and (17), m is the number of data samples, y is i,M is the significant wave height estimated by the model, y i,T is the significant wave height value from the reference dataset, y i,M and i,T The average value of E[F(x)] is the expected value of the model for a specific input x, A(x) is the true value, and A t is the true value at time t, F t is the predicted value at time t, and |·| represents an absolute value operation.

[0070] After adopting the above technical scheme, the present invention has the following beneficial effects compared with the prior art.

[0071] The method of the present invention combines five modules, namely, a convolutional feature extraction module, an efficient multi-scale attention module EMA (Efficient Multi-Scale Attention Module), a Transformer, a feature relationship reasoning module and a multi-layer perceptron (MLP). The idea of ​​multi-module fusion is to effectively integrate multiple functional modules and further enhance the flexibility of the method, thereby achieving higher efficiency and functionality, and further improving the accuracy of effective wave height inversion from space-borne GNSS-R data.

[0072] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are part of this application and are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute an improper limitation of the present invention. Obviously, the drawings described below are only some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0074] Figure 1 is a flow chart of the method of the present invention;

[0075] Figure 2 A structural diagram of the MultiScaleWaveFormer hybrid deep learning model for spaceborne GNSS-R effective wave height inversion provided in an embodiment of the present invention;

[0076] Figure 3 A module structure diagram of efficient multi-scale attention provided in an embodiment of the present invention;

[0077] Figure 4 Result graph of SHAP values ​​of power DDM (power_analog), bistatic radar cross section (BRCS), and effective scattering area (eff_scatter) calculated by the deep learning SHAP interpreter of the MultiScaleWaveFormer model provided in an embodiment of the present invention;

[0078] Figure 5 A SHAP feature importance histogram, a bee swarm diagram, and a SHAP waterfall diagram calculated by the MultiScaleWaveFormer model deep learning SHAP interpreter provided in an embodiment of the present invention;

[0079] Figure 6 A scatter density diagram of significant wave heights and ERA5 significant wave heights inverted by different machine learning and deep learning models provided in an embodiment of the present invention;

[0080] Figure 7 A scatter density diagram of significant wave heights and WW3 significant wave heights inverted by different machine learning and deep learning models provided in an embodiment of the present invention;

[0081] Figure 8 It is a statistical diagram of the inversion accuracy of the ERA5 data and different models in different SWH intervals provided in the embodiment of the present invention;

[0082] Fig. 9 It is a statistical diagram of the inversion accuracy of the WW3 data and different models in different SWH intervals provided in the embodiment of the present invention.

[0083] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0085] Example 1

[0086] In order to verify the effectiveness of the method proposed in this paper, CYGNSS L1-level spaceborne GNSS-R observation data covering all days from 2019 to 2020, European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data, IMERG rainfall data, and WaveWatch III significant wave height data were collected for experiments. At the same time, the experimental results of this invention were compared and verified with machine learning methods (bagged trees and artificial neural networks) and deep learning methods (BiLSTM, Transformer, and DCNN).

[0087] The basic configuration of the experimental platform for building the MultiScaleWaveFormer model is as follows:

[0088] Table: Configuration of the experimental platform

[0089]

[0090] A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion includes the following steps:

[0091] Step S1, obtaining CYGNSS L1 GNSS-R data, European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis dataset, IMERG rainfall product data and WaveWatch III (WW3) SWH data;

[0092] Step S2, preprocessing and quality control of the acquired data;

[0093] Step S3, further performing spatiotemporal matching on the data set obtained after quality control, and then dividing it into a training set, a test set, and a validation set according to a certain ratio;

[0094] Step S4, construct a hybrid deep learning model for spaceborne GNSS-R sea surface significant wave height inversion combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM, and train and verify the MultiScaleWaveFormer model;

[0095] Step S5, create a MultiScaleWaveFormer model deep learning SHAP interpreter to calculate and visualize the importance of input features;

[0096] Step S6, input the test set data into the trained MultiScaleWaveFormer model to obtain the inversion result;

[0097] Step S7, compare the SWH inversion results of this model with those of the existing model to evaluate the performance of the new model.

[0098] As an implementation method of this embodiment, Figure 1 The steps of constructing a hybrid deep learning model and interpreting the output of the method for inversion of sea surface significant wave height using spaceborne GNSS-R are intuitively demonstrated.

[0099] As an implementation manner of this embodiment, in step S1, the CYGNSS L1-level satellite-borne GNSS-R observation data includes L1b power DDM (power_analog), bistatic radar cross section (BRCS), effective scattering area (eff_scatter), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), DDM acquisition time (ddm_timestamp_utc), longitude and latitude (sp_lon, sp_lat) and incident angle (sp_inc_angle) of the mirror reflection point; the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data set includes wind speed, wind direction, water depth and SWH data, the ERA5 wind speed is the composite wind speed of the U component and the V component 10m above the sea surface, the wind direction data is the wind direction data (10m wind direction) provided in the ERA5 reanalysis data, and the water depth data is the water depth (Model 10) provided in the ERA5 product. The significant wave height data are the significant wave height of combined wind waves and swell data provided in the ERA5 reanalysis; and the IMERG rainfall data are the rainfall intensity of the IMERG final operation (IMERG-F).

[0100] As an implementation method of this embodiment, in step S2, firstly, observation variable parameters are extracted from the onboard GNSS-R data, and characteristic observation values ​​are calculated, including NBRCS, LES, receiver antenna gain, reflection point longitude, reflection point latitude and incident angle characteristic parameters, etc., and then some quality control flags are used to delete low-quality data. These flags include: if the absolute value of the spacecraft roll is between 1 and 30 degrees, the pitch is between 1 and 10 degrees, or the yaw is between 1 and 5 degrees, then set (Bit 03: small_sc_attitude_err); if the mirror reflection point is not on land, then set (Bit 11: sp_over_land); if the mirror reflection point is within 25 kilometers of the land, then set (Bit 12: sp_very_near_land); if the calculated delay line of the mirror reflection point from zero is less than 6 or greater than 10, then set (Bit 19: brcs_ddm_sp_bin_delay_error); if the calculated specular reflection point has a zero-based Doppler column less than 4 or greater than 6, it is set (Bit 20: brcs_ddm_sp_bin_dopp_error); if any of the numerical values ​​in the 3x 5brcs region used to calculate ddm_nbrcs has a negative value, it is set (Bit 21: neg_brcs_value_used_for_nbrcs); the specular point does not exist (Bit23: sp_non_existent_error); insufficient blackbody data for calibration (Bit 26: bb_framing_error); not enough data to properly calibrate the NBRCS for this sample of GPS satellites (Bit 28: low_quality_gps_ant_knowledge); the spacecraft altitude is outside the rated altitude range. The nominal altitude is defined as 490 km to 550 km (Bit 29: sc_altitude_out_of_nominal_range); set when an anomaly is observed on a CYGNSS or GPS satellite when this sample is collected. The data produced by these samples is inconsistent with typical CYGNSS results (Bit 30: anomalous_sampling_period).

[0101] As an implementation method of this embodiment, in step S3, the CYGNSS L1-level satellite-borne GNSS-R data is spatiotemporally matched with the ECMWF ERA5 wind speed data, ERA5 wind direction data, ERA5 water depth data, ERA5 SWH data, WaveWatch III significant wave height data, and IMERG rainfall data to obtain a matched data set. The matched data set is randomly divided into three parts: a training set, a validation set, and a test set according to the ratio of 60%, 15%, and 25%. The training set is used to train the network model, the validation set is used to supervise the model training, and the test set is used to avoid overfitting.

[0102] As an implementation method of this embodiment, the hybrid deep learning model for inversion of spaceborne GNSS-R sea surface significant wave height combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM described in step S4 mainly includes five modules: Figure 2 The structure diagram of the MultiScaleWaveFormer hybrid deep learning model for spaceborne GNSS-R effective wave height inversion is specifically shown, from which the distribution of each module and the detailed data processing flow can be seen. The first network module is the convolution feature extraction module, which takes power_analog, BRCS and effective scattering area (eff_scatter) images as input, and uses a two-dimensional convolutional neural network to automatically and effectively extract spatial feature information related to sea surface effective wave height from power_analog, BRCS and effective scattering area (eff_scatter) images.

[0103] The second network module is the Efficient Multi-Scale Attention Module (EMA), Figure 3 The specific structure of this module is shown, showing a multi-scale parallel sub-network structure. This network can capture short-range and long-range dependencies in the image, reorganize some channels into batch dimensions, and avoid information loss caused by channel dimensionality reduction. The figure points out the output process of the cross-space learning method in the fusion parallel sub-network. This method can capture pixel-level pairwise relationships and improve the aggregation effect of features. In short, the efficient attention mechanism of the EMA module can significantly improve the performance of image classification and target detection tasks while using fewer parameters and reducing computational overhead. The module output of EMA can be expressed as:

[0104]

[0105] Here we assume that the input feature map is X∈R C×H×W , where C is the number of channels, H is the height, and W is the width. The multi-scale extracted features are denoted by X s, s is the number of scales, GAP represents the global average pooling process, which is used to obtain the global context information of each scale feature, A s =σ(W a ·GAP(X s )+b a ) represents the attention weight, W a and b a is a learnable parameter and σ represents the activation function.

[0106] The third module introduces the Transformer deep learning architecture. Figure 2 It shows a more detailed structural composition, including the use of a multi-head attention mechanism to calculate attention in parallel in multiple different representation subspaces to capture rich features, the use of a random deactivation function Dropout after the attention output to prevent overfitting, the use of Add for addition when performing residual connections on the input to help the gradient be effectively transferred, and the layer normalization of the results after the Add operation to maintain stability. In the Transformer deep learning architecture, the Dense fully connected layer is used for feature transformation, and the flattening operation Flatten is used to flatten high-dimensional features into one dimension to facilitate subsequent Dense fully connected layer processing. The self-attention mechanism (Self-Attention) of the Transformer deep learning architecture processes the relationship between each position in the input sequence, allowing all positions in the sequence to be processed in parallel, significantly improving training efficiency. The multi-head self-attention in the structure can be expressed by the formula:

[0107] MultiHead(Q,K,V)=Concat(head1,head2,...,head n )W O (2)

[0108] The output for each header is:

[0109]

[0110] In formulas (2) and (3), Q represents the query vector (the input feature that currently needs to be paid attention to); K represents the key vector (each input feature in the dataset); and V represents the value vector (the information corresponding to each key). represents the output weight matrix, The learned weight matrix, d k is the dimension of each head.

[0111] Attached Figure 2The feature relationship reasoning module presented combines the Bidirectional Convolutional Long Short-Term Memory (BiLSTM) with a two-dimensional convolutional neural network. First, the Reshape layer reshapes the shape of the input data to adapt it to the processing of the subsequent network layer. Subsequently, the model uses a two-layer bidirectional recurrent neural network (Bidirectional) to simultaneously capture the forward and backward information of the sequence data to enhance the understanding of the context. Next, the multi-dimensional features are flattened into one dimension through the Flatten layer to facilitate the subsequent Dense layer to process. Multi-layer Dense is used to gradually extract and combine features to improve the expressive power of the model. In order to prevent overfitting, the model uses a Dropout layer after some Dense layers to enhance the generalization performance of the model. Finally, the results are output through the output layer. The feature relationship reasoning module can extract time-related features and capture contextual information to better understand the input sequence. Let the input sequence be X = (x1, x2, ..., x T ), where x t is the input at time t. In the bidirectional LSTM, we use two LSTM layers, forward and reverse, to perform convolution on the input. Let the convolution operation be C, which can be expressed as:

[0112]

[0113] in, Represents the feature map processed by the convolution layer, W c is the convolution kernel weight, b c For bias.

[0114] Use the forward LSTM to process the input sequence from t = 1 to t = T, and the reverse LSTM to process the input sequence from t = T to t = 1. The final output is obtained by concat the outputs of these two directions:

[0115]

[0116] The final output of BiConvLSTM can be expressed as:

[0117] y t =W y ·h t +b y (6)

[0118] Among them, W y is the final output weight matrix, b y is the output bias.

[0119] Attached Figure 2The Multi-Layer Perceptron (MLP) module shown combines the features automatically extracted by the network and the features extracted manually. It performs multiple nonlinear transformations on the input through the Dense layer and combines the features from different sources using the Multiply operation. This module processes the input data through the weights and activation functions of each layer and finally generates output. For neurons from the input layer to the lth layer, the input x and weight W (l) and bias b (l) The calculation formula can be expressed as:

[0120] z (l) =W (l) ·x+b (l) (7)

[0121] Here, W (l) represents the weight matrix of the lth layer, b (l) represents the bias vector of the lth layer. Finally, the activation function f is applied to get the output of this layer:

[0122] a (l) =f(z (l) ) (8)

[0123] The training of the MultiScaleWaveFormer hybrid deep learning model described in step S4 selects the improved S-Huber loss function of the present invention. The loss function introduces a scaling parameter h on the basis of the Huber loss and combines the weighted sum of the mean square error (MSE). The formula is:

[0124] S-Huber(y true ,y pred )=h×Huber(y true ,y pred )+(1-h)×MSE(y true ,y pred ) (9)

[0125] Among them, Huber, as a loss function for dealing with outliers, is defined as:

[0126]

[0127] When |y true -y pred When |≤δ, the error is small and the loss is mainly in square form (similar to mean square error MSE). true -y pred When |>δ, the error is large and the loss adopts a linear form to reduce the impact of outliers on the overall model.

[0128] The mean square error (MSE) is defined as:

[0129]

[0130] In formulas (9), (10), and (11), h is a scaling factor between 0 and 1, which is used to distribute weights between Huber loss and MSE. true ,y pred ) is the standard Huber loss, MSE(y true ,y pred ) is the mean square error loss, y true is the true value, y pred is the predicted value, δ is the threshold for adjusting abnormal sensitivity, and N is the number of samples.

[0131] As an implementation of this embodiment, the creation of the MultiScaleWaveFormer model deep learning SHAP interpreter and the calculation and visualization of the importance of input features described in step S5 specifically include the following sub-steps:

[0132] Step S5.1, install the SHAP (SHapley Additive exPlanations) library in the deep learning environment tensorflow framework;

[0133] Step S5.2, select 1000 sample input data as the benchmark data set (background data) for creating the MultiScaleWaveFormer model SHAP deep learning interpreter;

[0134] Step S5.3, use the interpreter to calculate the Shapley value for the model's input data. The Shapley value represents the contribution of each input feature to the prediction result; the Shapley value The calculation formula can be expressed as:

[0135]

[0136] Among them, n(i) represents the number of times sample i is selected, represents the predicted mean of all samples, f(x ij ) is the jth feature x of the predicted sample i ij The prediction contribution to sample i, K is the number of features of sample i.

[0137] Step S5.4, visualize the SHAP value. Plot the SHAP value of a single sample to explain how the model makes a specific prediction; visualize the global importance of features to see the importance of the overall features; visualize feature dependencies to see how a single feature affects the prediction and interacts with other features through dependency graphs. The SHAP value of each sample can be expressed as:

[0138] Shap value=explainer(X) (13)

[0139] Among them, explainer(X) is the explainer created based on the MultiScaleWaveFormer model, and X represents the sample.

[0140] Attached Figure 4 The SHAP value results of power DDM (power_analog), bistatic radar cross section (BRCS), and effective scattering area (eff_scatter) calculated by the deep learning SHAP interpreter of the MultiScaleWaveFormer model are shown. Figure 5 In the figure, from left to right are a bar chart showing the mean of the absolute SHAP values, a beeswax chart showing the SHAP values, and a waterfall chart showing the SHAP values ​​for a single prediction. The bar chart illustrates the features that the model considers to be the most important for predicting the significant wave height, and quantifies the importance of each feature. It can be seen from the figure that the influence of the features shown on the left on the model's prediction of the significant wave height decreases from top to bottom. Among them, the reflection point location, especially the latitude (sp_lat), has the greatest influence on the model's prediction of the significant wave height, and the rainfall (Rainrate) has the least influence on the model's prediction of the significant wave height. The beeswax chart provides an intuitive way to see how the model prediction is produced by the combined effect of various features, and which features have the most influence on the model prediction, and sort them according to their average influence on the model prediction (that is, the average of the absolute values ​​of the SHAP values). The most important features appear at the top of the chart, and the least important features appear at the bottom. Therefore, it is not difficult to see that sp_lat is the most important feature affecting the prediction of the significant wave height, while Rainrate has a smaller influence on the prediction of the significant wave height. The waterfall chart of SHAP value is used to display the SHAP value of a single prediction. Each row shows whether each feature contributes positively (red) or negatively (blue) to the model prediction. The topmost mark f(x) = 2.062 in the figure is the model's prediction value for the significant wave height. The lower right corner E[f(X)] = 0.597 represents the model's baseline value, which is the average prediction value (global average) of the model for the significant wave height when no feature influence is considered. For example, it can be seen from the figure that WS (wind speed) contributes the most to the model prediction, with a SHAP value of -0.13, indicating that it reduces the prediction result by 0.13 and makes a negative contribution to the model prediction. sp_lon has a greater positive impact on the model prediction, with a SHAP value of 0.07, indicating that it improves the prediction result by 0.07. In addition, other features also contribute positively or negatively to the prediction results to varying degrees.

[0141] As an implementation of this embodiment, step S6 is specifically implemented as follows: a test set containing 25% of the data is used in a verified model to obtain an output result.

[0142] As an implementation of this embodiment, the performance evaluation indicators selected in step S7 are: root mean square error (RMSE), bias (Bias), mean absolute percentage error (MAPE) and Pearson correlation coefficient (CC), and the relevant calculation formulas are as follows:

[0143]

[0144] Bias=E[F(x)]-A(x) (15)

[0145]

[0146]

[0147] In the above formulas (14), (15), (16), and (17), m is the number of data samples, y is i,M is the significant wave height estimated by the model, y i,T is the significant wave height value from the reference dataset, y i,M and i,T The average value of E[F(x)] is the expected value of the model for a specific input x, A(x) is the true value, and A t is the true value at time t, F t is the predicted value at time t, and |·| represents an absolute value operation.

[0148] The four indicators of all models compared with ERA5 data and WW3 data are calculated respectively, and the accuracy statistics are shown in Tables 1 and 2.

[0149] Table 1 Comparison of the accuracy of significant wave height inverted by different models with ERA5 data

[0150]

[0151] Table 2 Comparison of the accuracy of significant wave height inverted by different models with WW3 data

[0152]

[0153]

[0154] It can be seen from Table 1 and Table 2 that when the proposed model MultiScaleWaveFormer is compared with the ERA5 data, compared with the five models of bagged tree, ANN, DCNN, BiLSTM and Transformer, the RMSE is reduced by 11.75%, 19.51%, 11.48%, 21.87% and 13.82%, respectively, the MAPE is reduced by 2.6%, 24.41%, 3.52%, 17.72% and 12.31%, respectively, and the correlation coefficient is improved by 2.2%, 4. ..49%, 2.2%, 4.49% and 2.2%; when compared with WW3 data, compared with the five models of bagged trees, ANN, DCNN, BiLSTM and Transformer, RMSE decreased by 7.59%, 13.92%, 8.06%, 13.92% and 7.59%, MAPE decreased by 3.18%, 19.03%, 3.7%, 12.21% and 9.28%, and the correlation coefficient increased by 2.25%, 4.6%, 2.25%, 4.6% and 2.25%, respectively. Figure 6 , Attachment Figure 7 The scatter density plots of the effective wave heights inverted by different machine learning and deep learning models and the effective wave heights of ERA5 and WW3 respectively, and the above conclusions can also be drawn from the figures.

[0155] Attached Figure 8 , Attachment Fig. 9 These are the inversion accuracy statistics of ERA5 data, WW3 data and different models in different SWH intervals. It can be seen from the figure that in the effective wave height range of 1-7m, the MultiScaleWaveFormer model has obvious advantages over other models in terms of RMSE, Bias, correlation coefficient and MAPE.

[0156] In summary, the implementation scheme of the present invention not only uses an efficient multi-scale attention module to perform information fusion at different feature scales to reduce computational complexity, but also combines the CNN-Transformer-BiConvLSTM hybrid deep learning model with the SHAP interpreter for feature importance evaluation, enhancing the interpretability of the model, and the combination of multiple models significantly improves the inversion performance of the model. The advantage of this method is that it overcomes the problem of limited accuracy improvement of a single empirical model and a single machine learning model in the process of inversion of sea surface wave height from satellite-borne GNSS-R, effectively integrates multiple functional modules, and further enhances the flexibility of the method, thereby achieving higher efficiency and functionality.

[0157] Among the many space-borne GNSS-R sea surface effective wave height inversion methods, this deep learning method that combines multiple models innovatively performs fine-grained analysis of features, intuitively displays the role of different input features in the model, utilizes the advantages of different models, and improves the performance of the overall model. From this point of view, the development of space-borne GNSS-R sea surface effective wave height inversion methods will develop in a more efficient, flexible and interpretable direction with the development of multi-model combinations.

[0158] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not used to limit the present invention. Any technician familiar with this patent can make some changes or modify the technical contents suggested above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the solution of the present invention.

Claims

1. A hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion, characterized in that: The following steps are involved: Step S1, obtaining CYGNSS L1-level spaceborne GNSS-R observation data, European Centre for Medium-Range Weather Forecasts ERA5 reanalysis dataset, IMERG rainfall product data and WaveWatch III SWH data; Step S2, preprocessing and quality control of the acquired data; Step S3, further performing spatiotemporal matching on the data set obtained after quality control, and then dividing it into a training set, a test set, and a validation set according to a certain ratio; Step S4, construct a hybrid deep learning model for spaceborne GNSS-R sea surface significant wave height inversion combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM, and train and verify the MultiScaleWaveFormer model; Among them, the hybrid deep learning model for spaceborne GNSS-R sea surface significant wave height inversion combining efficient multi-scale attention and CNN-Transformer-BiConvLSTM is constructed, which includes five modules: the first module is the convolution feature extraction module, the second module is the efficient multi-scale attention module, the third module is the Transformer deep learning architecture module, the fourth module is the feature relationship reasoning module, and the fifth module is the multi-layer perceptron; Step S5, creating a MultiScaleWaveFormer model deep learning SHAP interpreter to calculate and visualize the importance of input features; Step S6, input the test set data into the trained MultiScaleWaveFormer model to obtain the inversion result; Step S7, compare the SWH inversion results of this model with those of the existing model to evaluate the performance of the new model.

2. According to claim 1, a hybrid deep learning model construction and output interpretation method for space-borne GNSS-R sea surface significant wave height inversion is characterized in that: The CYGNSS L1-level satellite-borne GNSS-R observation data includes L1b power DDM, bistatic radar cross section BRCS, effective scattering area, normalized bistatic radar cross section NBRCS, leading edge slope LES, signal-to-noise ratio SNR, receiver antenna gain, DDM acquisition time, longitude and latitude of the specular reflection point, and incident angle; The European Centre for Medium-Range Weather Forecasts ERA5 reanalysis data set includes wind speed, wind direction, water depth and significant wave height SWH data; wherein the wind speed data is the composite wind speed of the U component and the V component 10 m above the sea surface, the wind direction data is the wind direction data provided in the ERA5 reanalysis data; the water depth data is the water depth provided in the ERA5 product, and the significant wave height SWH data is the significant wave height data of the combined wind wave and swell provided in the ERA5 reanalysis; The IMERG rainfall product data is the final operating rainfall intensity of IMERG.

3. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 1 is characterized in that: Step S2 includes the following sub-steps: Step S2.1, calculating characteristic observation values ​​from the onboard GNSS-R data, including calculating the average peak power DDMA from the power_analog image, and calculating the trailing edge slope of the normalized integrated delay waveform, the sum of the leading edge waveform values ​​NIDW_LEWS, and the sum of the trailing edge waveform values ​​NIDW_TEWS from the BRCS image; Step S2.2, use the quality control flag variables in the CYGNSS L1-level spaceborne GNSS-R observation data to delete low-quality data. These flags include: if the absolute value of the spacecraft roll is between 1 and 30 degrees, the pitch is between 1 and 10 degrees, or the yaw is between 1 and 5 degrees, set Bit 03: small_sc_attitude_err; if the mirror reflection point is not over land, set Bit 11: sp_over_land; if the mirror reflection point is within 25 kilometers of the land, set Bit 12: sp_very_near_land; if the calculated mirror reflection point delay row from zero is less than 6 or greater than 10, set Bit 19: brcs_ddm_sp_bin_delay_error; if the calculated mirror reflection point Doppler column based on zero is less than 4 or greater than 6, set Bit 20: brcs_ddm_sp_bin_dopp_error; set if any of the values ​​in the 3x5brcs region used to calculate ddm_nbrcs has a negative value Bit 21: neg_brcs_value_used_for_nbrcs; the mirror point does not exist Bit 23: sp_non_existent_error; insufficient blackbody data for calibration Bit 26: bb_framing_error; there is not enough data to properly calibrate the NBRCS for the GPS satellites for this sample Bit 28: low_quality_gps_ant_knowledge; the spacecraft altitude is outside the nominal altitude range, which is defined as 490 km to 550 km Bit29: sc_altitude_out_of_nominal_range; set when anomalies are observed on CYGNSS or GPS satellites when this sample is collected, and the data produced by these samples is inconsistent with typical CYGNSS results Bit 30: anomalous_sampling_period.

4. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 2 is characterized in that: Step S3 includes the following sub-steps: Step S3.1, perform spatiotemporal matching of the CYGNSS L1-level spaceborne GNSS-R observation data with the European Centre for Medium-Range Weather Forecasts ERA5 wind speed data, ERA5 wind direction data, ERA5 water depth data, ERA5 SWH data, WaveWatch III SWH data, and IMERG rainfall data to obtain a matched data set; In step S3.2, the matched data set is randomly divided into three parts: training set, validation set and test set according to the ratio of 60%, 15% and 25%. The training set is used to train the network model, the validation set is used to supervise the model training, and the test set is used to avoid overfitting.

5. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 1 is characterized in that: In step S4: The first network module is a convolution feature extraction module, which takes power_analog, BRCS and effective scattering area images as input. It uses a two-dimensional convolutional neural network to automatically and effectively extract spatial feature information related to the effective wave height of the sea surface from power_analog, BRCS and effective scattering area images. The second network module is the efficient multi-scale attention module EMA, which uses a multi-scale parallel sub-network structure to capture short-range and long-range dependencies in the image, reorganizes some channels into batch dimensions to avoid information loss caused by channel dimensionality reduction, and uses cross-space learning methods to fuse the outputs of parallel sub-networks to capture pixel-level pairwise relationships and improve feature aggregation. The efficient attention mechanism of the EMA module can significantly improve the performance of image classification and target detection tasks while using fewer parameters and reduce computational overhead; the module output of EMA can be expressed as: Here we assume that the input feature map is X∈R C×H×W , where C is the number of channels, H is the height, and W is the width; X s is the multi-scale feature extraction, s is the number of scales, GAP represents the global average pooling process, which is used to obtain the global context information of each scale feature, A s =σ(W a ·GAP(X s )+b a ) represents the attention weight, W a and b a is a learnable parameter, σ represents the activation function; The third module introduces the Transformer deep learning architecture, uses the multi-head attention mechanism to calculate attention in parallel in multiple different representation subspaces, captures rich features, uses the self-attention mechanism to process the relationship between each position in the input sequence, and allows all positions in the sequence to be processed in parallel, significantly improving training efficiency; multi-head self-attention can be expressed by the formula: MultiHead(Q,K,V)=Concat(head1,head2,...,head n )W O (2) The output for each header is: In formulas (2) and (3), Q represents the query vector; K represents the key vector; V represents the value vector. represents the output weight matrix, The learned weight matrix, d k is the dimension of each head; The fourth module is the feature relationship reasoning module, which combines the bidirectional convolutional long short-term memory network with the two-dimensional convolutional neural network to extract time-related features and capture context information to better understand the input sequence; let the input sequence be X = (x1, x2, ..., x T ), where x t is the input at time t. In BiConvLSTM, we use two ConvLSTM layers, forward and reverse, to perform convolution on the input. Let the convolution operation be C, which can be expressed as: in, Represents the feature map processed by the convolution layer, W c is the convolution kernel weight, b c is bias; Use the forward ConvLSTM to process the input sequence from t=1 to t=T, and the reverse ConvLSTM to process the input sequence from t=T to t=1: The final output of BiConvLSTM can be expressed as: y t =W y ·h t +b y (6) Among them, W y is the final output weight matrix, b y is the output bias; The fifth network module uses a multi-layer perceptron to process the input data through the weights and activation functions of each layer and finally generate output. For neurons from the input layer to the lth layer, the input x and weight W (l) and bias b (l) The calculation formula can be expressed as: z (l) =W (l) ·x+b (l) (7) Here, W (l) represents the weight matrix of the lth layer, b (l) represents the bias vector of the lth layer. Finally, the activation function f is applied to get the output of this layer: a (l) =f(z (l) ) (8) The training of the MultiScaleWaveFormer hybrid deep learning model described in step S4 selects the improved S-Huber loss function of the present invention. The loss function introduces a scaling parameter h on the basis of the Huber loss and combines the weighted sum of the mean square error MSE. The formula is: S-Huber(and true ,and pred )=h×Huber(y true ,and pred )+(1-h)×MSE(y true ,and pred ) (9) Among them, Huber, as a loss function for dealing with outliers, is defined as: When |y true -y pred When |≤δ, the error is small and the loss is mainly in square form. true -y pred When |>δ, the error is large and the loss adopts a linear form to reduce the impact of outliers on the overall model; The mean square error MSE is defined as: In formulas (9), (10), and (11): h is a scaling factor between 0 and 1, which is used to distribute weight between Huber loss and MSE. Huber(y true ,y pred ) is the standard Huber loss, MSE(y true ,y pred ) is the mean square error loss, y true is the true value, y pred is the predicted value, δ is the threshold for adjusting abnormal sensitivity, and N is the number of samples.

6. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 1 is characterized in that: Step S5 includes the following sub-steps: Step S5.1, install the SHAP library in the deep learning environment tensorflow framework; Step S5.2, select 1000 sample input data as the benchmark dataset for creating the MultiScaleWaveFormer model SHAP deep learning interpreter; Step S5.3, using the interpreter to calculate the Shapley value for the model's input data, where the Shapley value represents the contribution of each input feature to the prediction result; Shapley value The calculation formula can be expressed as: Among them, n(i) represents the number of times sample i is selected, represents the predicted mean of all samples, f(x ij ) is the jth feature x of the predicted sample i ij The prediction contribution to sample i, K is the number of features of sample i; Step S5.4, visualize the SHAP value, plot the SHAP value of a single sample to explain how the model makes a specific prediction; visualize the global importance of features to see the importance of the overall features; visualize feature dependencies, see how a single feature affects the prediction through the dependency graph and interacts with other features; the SHAP value of each sample can be expressed as: Shap value=explainer(X) (13) Among them, explainer(X) is the explainer created based on the MultiScaleWaveFormer model, and X represents the sample.

7. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 1 is characterized in that: Step S6 is specifically implemented as follows: a test set containing 25% of the data is used in the verified model to obtain an output result.

8. The hybrid deep learning model construction and output interpretation method for spaceborne GNSS-R sea surface significant wave height inversion according to claim 1 is characterized in that: The performance evaluation indicators selected in step S7 are: root mean square error RMSE, bias, mean absolute percentage error MAPE and Pearson correlation coefficient. The relevant calculation formulas are as follows: Bias=E[F(x)]-A(x) (15) Where m is the number of data samples, y i,M is the significant wave height estimated by the model, y i,T is the significant wave height value from the reference dataset, y i,M and i,T The average value of E[F(x)] is the expected value of the model for a specific input x, A(x) is the true value, and A t is the true value at time t, F t is the predicted value at time t, and |·| represents an absolute value operation.

Images