Sea ice thickness inversion method and system based on GNSS-R observation data

Abstract

This application discloses a method and system for sea ice thickness inversion based on GNSS-R observation data, relating to the field of satellite remote sensing technology. The method includes: acquiring GNSS-R data and reference labels from at least two satellites; constructing an input feature set containing physical features of delayed Doppler images, geographical location, time, and satellite identifiers; constructing a multi-task neural network including a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; performing end-to-end joint training using an uncertainty-weighted loss function to achieve collaborative optimization of thickness and ice type; inputting the preprocessed data to be inverted into the trained network; simultaneously outputting sea ice thickness estimates and ice type classification results; and generating a gridded product. This invention significantly improves inversion accuracy and autonomy, and enhances the model's generalization ability and spatiotemporal coverage through physically guided multi-task learning and multi-satellite data fusion.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing technology, and in particular to a method and system for inverting sea ice thickness based on GNSS-R observation data. Background Technology

[0002] Arctic sea ice is a crucial indicator and regulator of the global climate system, making high-precision, high-spatiotemporal coverage monitoring of its thickness essential. Currently, the mainstream technologies for monitoring Arctic sea ice thickness and their limitations are as follows:

[0003] 1. Active microwave remote sensing (e.g., CryoSat-2 satellite): This satellite carries a synthetic aperture interferometric radar altimeter, which measures the freeboard height of sea ice and calculates the sea ice thickness based on the principle of hydrostatic equilibrium. The advantage of this method is its strong ability to detect multi-year ice (MYI, i.e., thick ice with a rough surface). However, its main drawbacks are: long revisit periods leading to insufficient spatiotemporal coverage, and the high dependence on climatological data for snow depth correction, introducing significant uncertainties.

[0004] 2. Passive microwave remote sensing (such as the SMOS satellite): This satellite utilizes the thermal radiation sensitivity of the sea ice-seawater interface by an L-band radiometer, making it particularly suitable for the inversion of thin ice (<0.5 m) and capable of daily coverage. However, its main drawbacks are that L-band signals are easily saturated with thick ice, making it unable to effectively detect multi-year ice, and data quality degrades during polar nights.

[0005] 3. Fusion products (such as CS2SMOS): This method integrates CryoSat-2 and SMOS data through optimal interpolation algorithms and is currently a widely used reference dataset. However, its fundamental drawback is that it is still limited by the spatiotemporal coverage defects of the source data, and there are significant interpolation uncertainties in the Arctic central region and the fusion boundary. Essentially, it is a post-processed fusion product rather than a direct observation inversion result.

[0006] 4. Single-task machine learning methods based on GNSS-R: These methods utilize the delayed Doppler image (DDM) features of GNSS-R (Global Navigation Satellite System Reflectance Measurement) technology to retrieve sea ice thickness using machine learning models (such as neural networks). For example, some methods use the physical characteristics of the DDM (such as peak signal-to-noise ratio, NBRCS, etc.) as input to directly train a regression model to output the thickness value. However, existing technologies have the following key drawbacks:

[0007] (1) Insufficient utilization of ice type information and error propagation: The DDM scattering characteristics of first-year ice (FYI) and multi-year ice (MYI) are completely different. FYI is mainly coherent scattering, and the DDM shows a high peak value and narrow broadening; MYI is mainly incoherent scattering, and the DDM shows a low peak value and broadening waveform. Traditional methods either completely ignore ice type information, resulting in the model being unable to distinguish the scattering modes of different ice types, thus limiting the inversion accuracy; or adopt a "pipeline" approach, that is, first use an independent classifier or rely on external ice type products to identify the ice type, and then input the identification result as a feature into the thickness inversion model. This approach has the problems of one-way dependence and error accumulation. If the ice type classification is wrong in the first step, it will directly lead to a systematic deviation in the thickness inversion in the second step, and it cannot be corrected subsequently.

[0008] (2) Lack of physical process modeling: Traditional methods usually treat sea ice thickness as a static, global mapping problem. However, sea ice thickness has significant spatial heterogeneity. Different sea areas have different thermodynamic and dynamic processes, resulting in large differences in thickness, and the thickness increases over time. Existing technologies fail to explicitly incorporate these geophysical priors into the model, such as geographical location and temporal information, leading to insufficient generalization ability of the model, especially when facing unseen regions or seasons, such as from winter to spring, where the inversion accuracy will drop significantly.

[0009] (3) Limited single-satellite observation coverage: Most existing GNSS-R sea ice inversions are based on single satellite data. Due to satellite orbit limitations, there are inherent observation gaps, making it impossible to achieve high-frequency, seamless coverage of the Arctic region.

[0010] Therefore, in view of the above-mentioned problems in the existing technology, how to solve the problems of one-way dependence on ice type information and error propagation, achieve synergistic optimization of ice type classification and thickness inversion, effectively embed physical prior knowledge such as the spatial heterogeneity of sea ice thickness and seasonal evolution law into the inversion model, improve the generalization ability and physical consistency of the model, break through the coverage bottleneck of single-satellite observation, achieve effective fusion of data from multiple satellites and system bias correction, and improve the spatiotemporal coverage capability of the Arctic region have become urgent problems to be solved. Summary of the Invention

[0011] In view of this, to address the shortcomings of existing Arctic sea ice thickness monitoring technologies, particularly the deficiencies of GNSS-R-based machine learning methods in ice type-thickness inversion coordination, physical prior fusion, and observation coverage, this paper proposes a joint inversion method and system for spaceborne GNSS-R sea ice parameters based on multi-task learning and physical prior coding. This high-precision, highly autonomous, and high-coverage sea ice thickness inversion method achieves coordinated optimization of ice type classification and thickness inversion, avoiding error propagation; it embeds physical prior knowledge such as spatial heterogeneity and seasonal evolution patterns into the model, improving generalization ability and physical consistency; and it overcomes the bottleneck of single-satellite observation coverage, enabling multi-satellite data fusion and system bias correction.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] In a first aspect, the present invention provides a method for inverting sea ice thickness based on GNSS-R observation data, comprising the following steps:

[0014] S1. Data Acquisition and Feature Construction: Acquire GNSS-R observation data from at least two satellites and corresponding reference data on sea ice thickness and ice type, and perform quality control and feature construction to build a model input feature set that includes physical features of delayed Doppler images, geographical location features, temporal features, and satellite identification features;

[0015] S2. Model Construction and Joint Training: Construct a multi-task neural network that includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; Using the model input feature set and the reference data, perform end-to-end joint training on the multi-task neural network through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks;

[0016] S3. Thickness Inversion: The GNSS-R observation data to be inverted is preprocessed and input into the trained multi-task neural network. The thickness regression task head outputs the sea ice thickness at each observation point and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

[0017] As a further aspect of the present invention, the physical characteristics of the delayed Doppler image include peak signal-to-noise ratio, normalized signal-to-noise ratio, normalized bistatic radar cross section, reflectivity, hysteresis spread signal, differential hysteresis spread signal, antenna gain, and incident angle.

[0018] As a further aspect of the present invention, the geographical location feature is the latitude and longitude of the observation point; the time feature is the year-day cycle; and the satellite identification feature is a satellite identifier used to distinguish data sources from different satellites.

[0019] As a further aspect of the present invention, the reference data for sea ice thickness and ice type corresponding to the GNSS-R observation data includes:

[0020] Thickness label: The thickness label is obtained by matching the mirror points of the GNSS-R observation data with the external sea ice thickness grid product in space and time, and the thickness value within the grid is used as the supervision label for the regression task;

[0021] Ice type label: obtained by matching the mirror points of the GNSS-R observation data with the external sea ice type grid product in space and time. The ice type category within the grid is used as the supervision label for the classification task, and samples with uncertain labels are removed; wherein, the ice type category includes annual ice and multi-year ice.

[0022] As a further aspect of the present invention, quality control and feature construction are performed on the acquired GNSS-R observation data and corresponding reference data, including:

[0023] Abnormal observation samples with a peak signal-to-noise ratio below 2 dB and an incident angle greater than 65° in the GNSS-R observation data and reference data were removed;

[0024] Z-score standardization is applied to the continuous numerical features in the input feature set of the model.

[0025] The satellite identification features are encoded using one-hot encoding.

[0026] The preprocessed dataset is divided into training, validation, and test sets in chronological order.

[0027] As a further aspect of the present invention, in step S1, the model input feature set further includes at least one of the following environmental covariates: cumulative freezing degree days, snow depth data, and sea ice drift speed.

[0028] As a further aspect of the present invention, in step S1, the satellites include at least two of the following: Fengyun-3E, Fengyun-3F, Fengyun-3G, or satellites carrying GNSS-R payload data. Multi-satellite data fusion is achieved by expanding the dimensions of the satellite identification features.

[0029] As a further embodiment of the present invention, the shared feature extraction layer consists of three fully connected layers, each followed by a batch normalization layer, an activation function, and a dropout layer, and is responsible for extracting a general scattering feature representation that is related to both ice type and thickness from the input features of the model input feature set.

[0030] As a further embodiment of the present invention, the thickness regression task head and the ice type classification task head are connected to the shared feature extraction layer. The thickness regression task head consists of a fully connected layer with 32 neurons and an output layer with 1 neuron, used to map the shared features to continuous sea ice thickness values. The ice type classification task head consists of a fully connected layer with 32 neurons and an output layer with 2 neurons, used to map the shared features to class probabilities of FYI / MYI.

[0031] As a further embodiment of the present invention, the multi-task neural network further includes at least one snow depth regression task head, which is connected to the shared feature extraction layer for performing snow depth regression tasks.

[0032] As a further aspect of the present invention, in step S2, the loss function is an uncertainty-weighted loss function that automatically balances the weights of the tasks, and its expression is:

[0033] ;

[0034] in, This represents the total loss during model training. The loss for the thickness regression task is specifically the mean squared error between the estimated and actual sea ice thickness. The loss for the ice type classification task is specifically the cross-entropy loss between the ice type category probability distribution predicted by the model and the true label. and These are the uncertainty parameters that the model can learn during training, representing the inherent uncertainties of the thickness regression task and the ice type classification task, respectively. The items are used as weights to automatically balance the contribution of different task losses to the total loss. The term serves as a regularization term to prevent the uncertainty parameter from becoming too large.

[0035] As a further aspect of the present invention, in step S2, the multi-task neural network further includes a system bias correction branch, which is used to input the satellite identification features after one-hot encoding into a 16-neuron fully connected layer, and concatenate the output with the output of the shared feature extraction layer, and then send it into the thickness regression task head and the ice type classification task head to correct the system bias between different satellite data.

[0036] As a further embodiment of the present invention, the shared feature extraction layer adopts a convolutional neural network architecture or a Transformer architecture. The convolutional neural network architecture is used to process the model input feature set containing two-dimensional delayed Doppler image data; the Transformer architecture is used to capture long-distance dependencies between features using a self-attention mechanism.

[0037] Secondly, the present invention also provides a sea ice thickness inversion system based on GNSS-R observation data, comprising:

[0038] The data preprocessing module is used to acquire GNSS-R observation data from at least two satellites and corresponding sea ice thickness and ice type reference data, and to perform quality control and feature construction; it constructs a model input feature set that includes physical features of delayed Doppler images, geographic location features, temporal features, and satellite identification features;

[0039] The model training module is used to construct a multi-task neural network that includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; using the model input feature set and the reference data, the multi-task neural network is jointly trained end-to-end through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks;

[0040] The inversion product generation module is used to input the preprocessed GNSS-R observation data to be inverted into the trained multi-task neural network, and the thickness regression task head outputs the sea ice thickness of each observation point, and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

[0041] Compared with existing technologies, the present invention provides a method and system for joint inversion of spaceborne GNSS-R sea ice parameters based on multi-task learning and physical prior coding, which has the following advantages:

[0042] 1. This invention uses a multi-task learning framework to place thickness regression and ice type classification tasks in a shared feature extraction layer for end-to-end joint training. This allows the two tasks to promote each other and optimize synergistically during the optimization process. Thickness information helps to distinguish ice types with ambiguous scattering characteristics, while accurate ice type context information can significantly improve the thickness estimation of the corresponding category. This mechanism avoids error accumulation, thus achieving higher accuracy than traditional methods overall, especially in the inversion of multi-year ice.

[0043] 2. The multi-task network constructed in this invention can output high-precision sea ice thickness while simultaneously and autonomously outputting high-confidence sea ice type classification results. By inputting satellite identifiers as a key feature into the model and designing a systematic bias correction branch, the model can internally distinguish and correct systematic biases from different satellite payloads, achieving effective fusion and joint inversion of multi-source GNSS-R data. This weaves the observation data of multiple satellites into a denser observation network, thereby greatly filling the gaps in single-satellite observations and improving the spatial continuity and temporal sampling frequency of the inversion products.

[0044] In summary, this invention is applicable to thickness inversion, especially significantly improving the accuracy of multi-year ice inversion, greatly reducing dependence on external ice type products, achieving autonomous inversion, effectively improving spatiotemporal coverage through binary star fusion, and better adapting to changes in different regions and seasons through physical prior coding. The model output conforms to the basic laws of sea ice thermodynamics and dynamics, and has the characteristics of high accuracy, strong autonomy, wide coverage, strong generalization ability, and good physical consistency.

[0045] These or other aspects of the invention will become more apparent from the following description of embodiments. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. In the drawings:

[0047] Figure 1 This is a flowchart of a sea ice thickness inversion method based on GNSS-R observation data according to the present invention.

[0048] Figure 2 This is an end-to-end flowchart of the sea ice thickness inversion method based on GNSS-R observation data according to the present invention, from data input to product generation.

[0049] Figure 3 This is a sub-flowchart of physical prior coding and feature input in a sea ice thickness inversion method based on GNSS-R observation data according to the present invention.

[0050] Figure 4 This is a sub-flowchart of system bias correction and multi-satellite fusion in a sea ice thickness inversion method based on GNSS-R observation data according to the present invention. Detailed Implementation

[0051] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0052] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0053] See Figures 1 to 4 As shown, one embodiment of the present invention provides a sea ice thickness inversion method based on GNSS-R observation data, comprising the following steps:

[0054] Step S10, Data Acquisition and Feature Construction: Acquire GNSS-R observation data from at least two satellites and corresponding reference data on sea ice thickness and ice type, and perform quality control and feature construction to construct a model input feature set that includes physical features of delayed Doppler images, geographical location features, temporal features, and satellite identification features.

[0055] In this embodiment, the satellites include at least two of the following: Fengyun-3E, Fengyun-3F, Fengyun-3G, or satellites carrying GNSS-R payload data. Multi-satellite data fusion is achieved by expanding the dimensions of the satellite identification features. For example, in this embodiment, GNSS-R data from the GNOS payloads of FY-3E and FY-3F satellites, along with corresponding reference data, are acquired, and the data undergoes quality control and feature construction. Specifically, the satellite data can be obtained from the GNOS payload data of FY-3E and FY-3F satellites between October 1, 2024, and March 31, 2025, from which a model input feature set containing delayed Doppler image physical features, geographic location features, temporal features, and satellite identification features is extracted.

[0056] Among them, the satellites are not limited to FY-3E and FY-3F, but can be seamlessly extended to other domestically produced satellites carrying GNSS-R payloads, such as FY-3G for precipitation measurement, or the dedicated GNSS-R constellation planned in the future. By increasing the dimension of satellite ID features, larger-scale constellation data fusion can be achieved to build a polar monitoring system with high spatiotemporal resolution.

[0057] In this embodiment, the physical characteristics of the delayed Doppler image include peak signal-to-noise ratio, normalized signal-to-noise ratio, normalized bistatic radar cross section, reflectivity, hysteresis spread signal, differential hysteresis spread signal, antenna gain, and incident angle. The geographical location characteristics are the latitude and longitude of the observation point; the time characteristic is the day-in-year; and the satellite identification characteristic is a satellite identifier used to distinguish data sources from different satellites. Specifically, the satellite identifier (Satellite_ID) is used as a feature input to the model, encoded through branching or directly as a feature. This allows the model to automatically learn and correct system biases between different satellite payloads during training, achieving effective data fusion. The protection scope can cover methods for joint training and bias correction of multi-satellite GNSS-R data using the aforementioned identification features.

[0058] The aforementioned physical features of delayed Doppler images, geographic location features, temporal features, and satellite identifier features constitute the model input feature set, and each observation sample can extract the above 13-dimensional features. Using geographic location (latitude, longitude) as a proxy variable for spatial heterogeneity and the Day of Year as a proxy variable for seasonal evolution, along with other DDM physical features, as input features of the model, endows the model with the ability to perceive physical processes. This can cover methods that use the aforementioned physical priors (such as cumulative freezing days) as input features for sea ice thickness inversion.

[0059] In some embodiments, the model input feature set further includes at least one of the following environmental covariates: cumulative freezing days, snow depth data, and sea ice drift velocity. The cumulative freezing days are calculated based on 2-meter air temperature from reanalysis data (such as ERA5), serving as a more accurate thermodynamic proxy variable than annual freezing days. The snow depth data are introduced OSISAF snow depth products or CryoSat-2 snow depth climatological data, assisting the model in understanding the attenuation and scattering effects of snow cover on GNSS-R signals. Sea ice drift velocity serves as a supplementary proxy variable for the dynamic history.

[0060] In this embodiment, the reference data for sea ice thickness and ice type corresponding to the GNSS-R observation data includes:

[0061] Thickness label: The thickness label is obtained by matching the mirror points of the GNSS-R observation data with external sea ice thickness grid products (such as 25km grid) in space and time, and the thickness value within the grid is used as the supervision label for the regression task.

[0062] Ice type label: The ice type label is obtained by matching the mirror points of the GNSS-R observation data with external sea ice type grid products (such as 10km grid) in space and time. The ice type category within the grid is used as the supervision label for the classification task, and samples with uncertain labels are removed. The ice type category includes annual ice and multi-year ice.

[0063] In this embodiment, quality control and feature construction are performed on the acquired GNSS-R observation data and corresponding reference data, including:

[0064] Abnormal observation samples with a peak signal-to-noise ratio below 2 dB and an incident angle greater than 65° in the GNSS-R observation data and reference data were removed;

[0065] Z-score standardization is applied to the continuous numerical features in the input feature set of the model.

[0066] The satellite identification features are encoded using one-hot encoding.

[0067] The preprocessed dataset is divided into training, validation, and test sets in chronological order.

[0068] For example, the dataset is divided as follows: the training set is 60% from October 2024 to January 2025; the validation set is 20% from February 2025; and the test set is 20% from March 2025.

[0069] Step S20, Model Construction and Joint Training: Construct a multi-task neural network that includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; using the model input feature set and the reference data, perform end-to-end joint training on the multi-task neural network through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks.

[0070] In this step, the shared feature extraction layer consists of three fully connected layers with 256, 128, and 64 neurons respectively. Each layer is followed by a batch normalization layer, an activation function (ReLU), and a dropout layer (p=0.2), which is responsible for extracting a general scattering feature representation that is related to both ice type and thickness from the input features of the model input feature set.

[0071] The thickness regression task head and the ice type classification task head are connected to the shared feature extraction layer. The thickness regression task head consists of a 32-neuron fully connected layer (ReLU activation) and a 1-neuron output layer (linear activation), used to map the shared features to continuous sea ice thickness values. The ice type classification task head consists of a 32-neuron fully connected layer (ReLU activation) and a 2-neuron output layer (Softmax activation), used to map the shared features to FYI / MYI class probabilities.

[0072] This embodiment constructs a shared feature extraction layer and connects the thickness regression task head and the ice type classification task head to achieve end-to-end joint training and collaborative optimization of the two tasks. This is a fundamental innovation that distinguishes it from traditional "pipeline" or "single-task" methods. The scope of protection of the sea ice thickness inversion method in this embodiment can cover any method that applies this framework to GNSS-R sea ice parameter inversion, as well as the neural network structure that implements this framework.

[0073] In some embodiments, the multi-task neural network further includes at least one snow depth regression task head, which is connected to the shared feature extraction layer to perform the snow depth regression task. The physical coupling of ice type, thickness, and snow depth is stronger, and joint optimization is expected to bring greater gains.

[0074] In this embodiment, the loss function is an uncertainty-weighted loss function that automatically balances the weights of the tasks, and its expression is:

[0075] ;

[0076] in, This represents the total loss during model training. The loss for the thickness regression task is specifically the mean squared error between the estimated and actual sea ice thickness. The loss for the ice type classification task is specifically the cross-entropy loss between the ice type category probability distribution predicted by the model and the true label. and These are the uncertainty parameters that the model can learn during training, representing the inherent uncertainties of the thickness regression task and the ice type classification task, respectively. The items are used as weights to automatically balance the contribution of different task losses to the total loss. The term serves as a regularization term to prevent the uncertainty parameter from becoming too large.

[0077] In this embodiment, the optimizer is Adam with an initial learning rate of 0.001 in the training strategy; the early stopping strategy uses the monitoring and validation set loss with a patience of 15 epochs; the batch size is 256; and the number of samples in FYI and MYI is balanced by weighted sampling.

[0078] In some embodiments, the multi-task neural network further includes a system bias correction branch, which inputs the satellite identification features after one-hot encoding into a 16-neuron fully connected layer, concatenates the output with the output of the shared feature extraction layer, and then sends it into the thickness regression task head and the ice type classification task head to correct the system bias between different satellite data.

[0079] In some embodiments, the shared feature extraction layer employs a convolutional neural network (CNN) architecture or a Transformer architecture. The CNN architecture is used to process the model input feature set containing two-dimensional delayed Doppler image data. If the input includes not only scalar features but also two-dimensional image data of the DDM, a CNN can be used as a shared feature extractor to automatically learn the spatial spectral features of the DDM image. The Transformer architecture is used to capture long-distance dependencies between features using a self-attention mechanism.

[0080] Step S30, Thickness Inversion: After preprocessing the GNSS-R observation data to be inverted, input it into the trained multi-task neural network. The thickness regression task head outputs the sea ice thickness at each observation point and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

[0081] In this step, the trained model is applied to new observational data. For all eligible DDM observations for a given day, the model is fed forward and propagated, using only the output of the thickness regression head to obtain the sea ice thickness at each observation point. These observations are then averaged over a 25km polar azimuth 3D grid to generate a grid product of Arctic sea ice thickness for that day. Simultaneously, the probability results from the ice type classification head can be output to generate the corresponding ice type distribution map.

[0082] This invention significantly improves the accuracy of ice thickness inversion, especially multi-year ice thickness inversion, achieves end-to-end ice type classification, reduces dependence on external products, enhances the physical consistency and generalization ability of the model, and improves spatial coverage in the Arctic region. Through an innovative physics-guided multi-task learning framework and a dual-satellite fusion strategy, this invention systematically addresses several key shortcomings of existing GNSS-R sea ice thickness inversion technologies, achieving a technical solution with higher accuracy, greater autonomy, and wider coverage.

[0083] The sea ice thickness inversion method based on GNSS-R observation data of this invention uses a multi-task learning framework to place thickness regression and ice type classification tasks in a shared feature extraction layer for end-to-end joint training. This allows the two tasks to promote each other and optimize synergistically during the optimization process. Thickness information helps to distinguish ice types with ambiguous scattering characteristics, and accurate ice type context information can significantly improve the thickness estimation of the corresponding category. This mechanism avoids error accumulation, thus achieving higher accuracy than traditional methods overall, especially in the inversion of multi-year ice.

[0084] The multi-task network constructed in this invention can output high-precision sea ice thickness while simultaneously and autonomously outputting high-confidence sea ice type classification results. By inputting satellite identifiers as a key feature into the model and designing a systematic bias correction branch, the model can internally distinguish and correct systematic biases from different satellite payloads, achieving effective fusion and joint inversion of multi-source GNSS-R data. This weaves the observation data of multiple satellites into a denser observation network, thereby greatly filling the gaps in single-satellite observations and improving the spatial continuity and temporal sampling frequency of the inversion products.

[0085] It should be understood that although the above description follows a certain order, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, some steps in this embodiment may include multiple steps or multiple stages, which are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of steps or stages in other steps.

[0086] In one embodiment of the present invention, a sea ice thickness inversion system based on GNSS-R observation data is also provided for performing the steps of the above method. The sea ice thickness inversion system includes:

[0087] The data preprocessing module is used to acquire GNSS-R observation data from at least two satellites and corresponding sea ice thickness and ice type reference data, and to perform quality control and feature construction; it constructs a model input feature set that includes physical features of delayed Doppler images, geographic location features, temporal features, and satellite identification features;

[0088] The model training module is used to construct a multi-task neural network that includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; using the model input feature set and the reference data, the multi-task neural network is jointly trained end-to-end through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks;

[0089] The inversion product generation module is used to input the preprocessed GNSS-R observation data to be inverted into the trained multi-task neural network, and the thickness regression task head outputs the sea ice thickness of each observation point, and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

[0090] Based on the above features, the sea ice thickness inversion system based on GNSS-R observation data of the present invention is used to execute the steps of the sea ice thickness inversion method based on GNSS-R observation data in the above embodiments, which will not be repeated here.

[0091] In summary, this invention is applicable to thickness inversion, especially significantly improving the accuracy of multi-year ice inversion, greatly reducing dependence on external ice type products, achieving autonomous inversion, effectively improving spatiotemporal coverage through binary star fusion, and better adapting to changes in different regions and seasons through physical prior coding. The model output conforms to the basic laws of sea ice thermodynamics and dynamics, and has the characteristics of high accuracy, strong autonomy, wide coverage, strong generalization ability, and good physical consistency.

[0092] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for inverting sea ice thickness based on GNSS-R observation data, characterized in that, Includes the following steps: Acquire GNSS-R observation data from at least two satellites and corresponding reference data on sea ice thickness and ice type, and perform quality control and feature construction to build a model input feature set that includes physical features of delayed Doppler images, geographical location features, temporal features and satellite identification features; A multi-task neural network is constructed, which includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head. Using the model input feature set and the reference data, the multi-task neural network is jointly trained end-to-end through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks. The GNSS-R observation data to be inverted is preprocessed and then input into the trained multi-task neural network. The thickness regression task head outputs the sea ice thickness at each observation point and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

2. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 1, characterized in that, The physical characteristics of the delayed Doppler image include peak signal-to-noise ratio, normalized signal-to-noise ratio, normalized bistatic radar cross section, reflectivity, hysteresis spread signal, differential hysteresis spread signal, antenna gain, and incident angle.

3. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 2, characterized in that, The geographical location feature is the latitude and longitude of the observation point; the time feature is the year-day; and the satellite identification feature is a satellite identifier used to distinguish data sources from different satellites.

4. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 3, characterized in that, The reference data for sea ice thickness and ice type corresponding to the GNSS-R observation data include: Thickness label: The thickness label is obtained by matching the mirror points of the GNSS-R observation data with the external sea ice thickness grid product in space and time, and the thickness value within the grid is used as the supervision label for the regression task; Ice type label: obtained by matching the mirror points of the GNSS-R observation data with the external sea ice type grid product in space and time. The ice type category within the grid is used as the supervision label for the classification task, and samples with uncertain labels are removed; wherein, the ice type category includes annual ice and multi-year ice.

5. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 1, characterized in that, Quality control and feature construction are performed on the acquired GNSS-R observation data and corresponding reference data, including: Abnormal observation samples with a peak signal-to-noise ratio below 2 dB and an incident angle greater than 65° in the GNSS-R observation data and reference data were removed; Z-score standardization is applied to the continuous numerical features in the input feature set of the model. The satellite identification features are encoded using one-hot encoding. The preprocessed dataset is divided into training, validation, and test sets in chronological order.

6. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 1, characterized in that, The shared feature extraction layer consists of three fully connected layers, each followed by a batch normalization layer, an activation function, and a dropout layer. It is responsible for extracting a general scattering feature representation that is related to both ice type and thickness from the input features of the model input feature set.

7. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 6, characterized in that, The thickness regression task head and the ice type classification task head are connected to the shared feature extraction layer. The thickness regression task head consists of a fully connected layer with 32 neurons and an output layer with 1 neuron, used to map the shared features to continuous sea ice thickness values. The ice type classification task head consists of a fully connected layer with 32 neurons and an output layer with 2 neurons, used to map the shared features to class probabilities of FYI / MYI.

8. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 1, characterized in that, The loss function is an uncertainty-weighted loss function that automatically balances the weights of the tasks. Its expression is: ； in, This represents the total loss during model training. The loss for the thickness regression task is specifically the mean squared error between the estimated and actual sea ice thickness. The loss for the ice type classification task is specifically the cross-entropy loss between the ice type category probability distribution predicted by the model and the true label. and These are the uncertainty parameters that the model can learn during training, representing the inherent uncertainties of the thickness regression task and the ice type classification task, respectively. The items are used as weights to automatically balance the contribution of different task losses to the total loss. The term serves as a regularization term to prevent the uncertainty parameter from becoming too large.

9. The sea ice thickness inversion method based on GNSS-R observation data as described in claim 1, characterized in that, The multi-task neural network also includes a system bias correction branch, which inputs the satellite identification features after one-hot encoding into a 16-neuron fully connected layer, concatenates the output with the output of the shared feature extraction layer, and then sends it into the thickness regression task head and the ice type classification task head to correct the system bias between different satellite data.

10. A sea ice thickness inversion system based on GNSS-R observation data, characterized in that, For performing the sea ice thickness inversion method based on GNSS-R observation data as described in any one of claims 1-9, the sea ice thickness inversion system comprises: The data preprocessing module is used to acquire GNSS-R observation data from at least two satellites and corresponding sea ice thickness and ice type reference data, and to perform quality control and feature construction, constructing a model input feature set that includes delayed Doppler image physical features, geographic location features, time features and satellite identification features; The model training module is used to construct a multi-task neural network that includes a shared feature extraction layer, a thickness regression task head, and an ice type classification task head; using the model input feature set and the reference data, the multi-task neural network is jointly trained end-to-end through a loss function to simultaneously optimize the sea ice thickness regression and sea ice ice type classification tasks; The inversion product generation module is used to input the preprocessed GNSS-R observation data to be inverted into the trained multi-task neural network, and the thickness regression task head outputs the sea ice thickness of each observation point, and simultaneously outputs the ice type classification results to generate the corresponding ice type distribution map.

Images