A sea ice type constrained daily sea ice thickness retrieval method and system

Abstract

The application belongs to the technical field of satellite remote sensing and polar sea ice monitoring, and discloses a sea ice type constrained daily sea ice thickness inversion method and system. The method comprises the following steps: obtaining data, quality control and matching a target grid; one-year ice and multi-year ice candidate inversion and grid aggregation; multi-system probability calibration, probability logarithm probability domain fusion and reliability calculation; constructing a boundary prior and calculating an initial fusion weight; spatial regularization of sea ice types and mean-variance propagation; mask, limited interpolation and generation of daily sea ice thickness, proxy uncertainty and quality marks. Through one-year ice and multi-year ice double-candidate thickness inversion and continuous reliability gating, the application avoids directly determining the thickness by hard switching of sea ice types, and reduces the thickness discontinuity near the boundary of sea ice types.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing and polar sea ice monitoring technology, and particularly relates to a method and system for daily sea ice thickness inversion constrained by sea ice type. Background Technology

[0002] Global Navigation Satellite System Reflectance Measurement (GNSS-R) utilizes the echo characteristics formed after L-band signals from navigation satellites are reflected from the sea surface or sea ice surface for remote sensing inversion. This allows for the acquisition of observational information such as delayed Doppler images, specular reflection power, waveform statistics, signal-to-noise ratio, and reflection geometry. This information is related to the dielectric properties of sea ice, surface condition, and propagation conditions, and can provide a data foundation for daily sea ice thickness monitoring.

[0003] The existing GNSS-R sea ice thickness inversion and daily grid product generation processes mainly suffer from the following problems: (1) There are differences between annual ice and multi-year ice in terms of structure, salinity and dielectric propagation characteristics. It is difficult to take into account both types of sea ice by using the same thickness inversion relationship. Making a hard determination of sea ice type first and then switching the thickness model can easily lead to discontinuous results near the sea ice type boundary.

[0004] (2) The probability output, effective sample size and observation quality of different navigation systems are not consistent. Directly averaging the sea ice type probability of multiple systems is difficult to reflect the system reliability, the degree of effective observation support and the discrepancies between systems.

[0005] (3) The results of the inversion along the track have spatial sparsity and missing measurements when forming daily regular grid products; direct smoothing of candidate thickness or interpolation across sea ice type boundaries may destroy the thickness difference on both sides of the boundary; without distinguishing between direct fusion, restricted interpolation and grids that are still missing measurements, it is difficult to trace the source of product processing and its reliability. Summary of the Invention

[0006] To overcome the problems existing in related technologies, this invention provides a method and system for retrieving daily sea ice thickness constrained by sea ice type. Specifically, it relates to a method and system for retrieving daily sea ice thickness in GNSS-R constrained by sea ice type, and also relates to the field of global navigation satellite system reflectance measurement. This invention aims to generate candidate thicknesses for one-year and multi-year ice in the same processing flow, comprehensively utilizing sea ice type probability, boundary priors, observation reliability, and surrogate variance to implement continuous gating, and generating daily sea ice thickness, surrogate uncertainty, and quality-marked grid products. The continuous gating uses a fusion weight of 0 to 1 to continuously adjust the relative contributions of the candidate thicknesses for one-year and multi-year ice, rather than a hard switch between the two.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for inverting daily sea ice thickness constrained by sea ice type, comprising the following steps: S1. Acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, form an effective observation sample set. S2. Input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively to generate one-year ice candidate thickness and multi-year ice candidate thickness. Aggregate them to the target grid using the same observation quality weight and spatial kernel to obtain the grid dual candidate thickness prediction field and the grid effective sample size. S3. Perform system-level calibration on the observation-level multi-year ice probabilities of different navigation systems and aggregate them to the target grid. The system-level calibration refers to correcting the observation-level multi-year ice probabilities of each navigation system separately. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. S4. The uncertainty of sea ice type is obtained from the comprehensive multi-year ice probability, and the surrogate variances of one-year ice and multi-year ice are constructed by combining the basic error of the two thickness networks, the observation reliability, and the candidate thickness difference. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary, where the multi-year ice side is positive, the one-year ice side is negative, and the boundary is 0. The signed boundary distance is determined according to the daily external sea ice type grid. The boundary transition width is determined based on the sea ice type uncertainty, the probability divergence between systems, the sample support degree, and the grid observation reliability, and the multi-year ice boundary prior probability is constructed. The bounded weak ranking prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness is used to obtain the continuous multi-year ice initial fusion weight. The external sea ice type grid is only used as a spatial prior and does not directly determine the final sea ice type or directly select a single candidate thickness. S5. Apply spatial regularization to the initial fusion weights between adjacent grids of the same external sea ice type, and propagate the fusion weights without crossing the boundary between one-year ice and multi-year ice to obtain regularized multi-year ice fusion weights; use the regularized multi-year ice fusion weights to fuse the candidate thicknesses of one-year ice, and at the same time apply a physical correction to the candidate thicknesses of multi-year ice to prevent them from having an inverse difference value that exceeds the allowable range relative to the candidate thicknesses of one-year ice, to obtain the physically corrected candidate thicknesses of multi-year ice, and propagate the surrogate variances of the two inversion networks and the mixed additional variances generated by the difference in candidate thicknesses to obtain the fusion thickness, fusion surrogate variance, and surrogate uncertainty; S6. Apply a mask and non-negative physical constraints to the fused thickness, and perform restricted interpolation using only the directly fused grid that is in the same effective sea ice connectivity region as the missing sea ice grid; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and do not use the interpolated grid as a new interpolation source, generate GNSS-R daily sea ice thickness, surrogate uncertainty and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between the sources of directly fused, restricted interpolation and grids from sea ice regions that are still missing.

[0008] Another object of the present invention is to provide a sea ice type-constrained daily sea ice thickness inversion system, which implements the sea ice type-constrained daily sea ice thickness inversion method, and the system includes: The data acquisition and grid matching module is used to acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, an effective observation sample set is formed. The dual-candidate thickness inversion module is used to input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively, generate one-year ice candidate thickness and multi-year ice candidate thickness, and aggregate them into the target grid using the same observation quality weight and spatial kernel to obtain the grid dual-candidate thickness prediction field and the grid effective sample size; The sea ice type probability and observation reliability calculation module is used to perform system-level calibration of the observation-level multi-year ice probability of different navigation systems and aggregate it to the target grid. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. The reliability-gated fusion module is used to obtain sea ice type uncertainty from the comprehensive multi-year ice probability, and to construct the surrogate variances of one-year ice and multi-year ice by combining the basic errors, observation reliability, and candidate thickness differences of the two thickness networks. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary. The signed boundary distance is determined based on the daily external sea ice type grid, and the boundary transition width is determined based on sea ice type uncertainty, inter-system probability divergence, sample support, and grid observation reliability to construct the multi-year ice boundary prior probability. The bounded weakly ordered prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness yields the continuous multi-year ice initial fusion weights. The weight space regularization and mean-variance propagation module is used to apply space regularization to the initial fusion weights between adjacent grids of the same external sea ice type, propagating the fusion weights without crossing the boundary between one-year ice and multi-year ice, to obtain regularized multi-year ice fusion weights; the one-year ice candidate thickness is fused using the regularized multi-year ice fusion weights, and a physical correction is applied to the multi-year ice candidate thickness to prevent it from having an inverse difference value that exceeds the allowable range relative to the one-year ice candidate thickness, to obtain the physically corrected multi-year ice candidate thickness, and the mixed additional variance generated by the surrogate variance of the two inversion networks and the candidate thickness difference is propagated to obtain the fusion thickness, fusion surrogate variance and surrogate uncertainty; The masking, constrained interpolation, and product generation module is used to apply a mask and non-negative physical constraints to the fused thickness, and to perform constrained interpolation only using directly fused grids that are in the same effective sea ice connectivity region as the missing sea ice grids; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and does not use the already interpolated grids as new interpolation sources, and generates GNSS-R daily sea ice thickness, surrogate uncertainty, and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between directly fused, constrained interpolation, and grid sources that are still missing in the sea ice region.

[0009] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention acquires GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, external sea ice type grids, and mask data for the target sea area; it generates dual candidate thicknesses using a one-year ice thickness inversion network and a multi-year ice thickness inversion network; through multi-navigation system multi-year ice probability calibration, reliability-weighted fusion, and observation reliability calculation, it determines the fusion weights by combining sea ice type boundary priors, surrogate variance, and candidate thickness ranking relationships, and applies sea ice type boundary-preserving spatial regularization to the fusion weights; through candidate thickness fusion, mean-variance propagation, masking, and restricted interpolation within the same sea ice effective connectivity region, it generates daily sea ice thickness, surrogate uncertainty, and quality marker grid products. This invention can avoid thickness discontinuities caused by hard switching of sea ice types, preserve sea ice type boundaries, and improve the reliability and traceability of daily sea ice thickness inversion results.

[0010] Second, by employing dual-candidate thickness inversion for both annual and multi-year ice and continuous reliability gating, the thickness is prevented from being directly determined by hard switching of sea ice types, thus reducing thickness discontinuities near sea ice type boundaries. Through multi-navigation system probabilistic calibration, reliability-weighted fusion, sample support, and system divergence calculation, the fusion weights are made to correspond to observational support and the reliability of different navigation systems. Spatial regularization of the fusion weights is applied only between adjacent grids of the same external sea ice type category, improving spatial consistency while maintaining sea ice type boundaries. Through mean-variance propagation, restricted interpolation within the same effectively connected sea ice region, and quality flag output, the daily sea ice thickness product simultaneously possesses surrogate uncertainty and traceable grid processing status.

[0011] Third, this invention can generate large-scale, daily updated grid products of sea ice thickness, proxy uncertainty, and quality indicators using GNSS-R observation data. By simultaneously generating annual and multi-year candidate ice thicknesses from the same valid observations and fusing them using continuous reliability gating, it can reduce thickness abrupt changes caused by hard switching of sea ice types, reduce outliers in sea ice type boundary areas, and decrease the workload of manual verification. Through multi-navigation system probability calibration, observation reliability calculation, constrained interpolation, and quality indicator output, it can improve the continuity, stability, and source traceability of daily sea ice thickness products. This technical solution can be applied to operational monitoring of sea ice in polar regions, risk assessment of polar shipping and routes, marine environmental forecasting, polar engineering support, climate change research, and sea ice remote sensing data product services. It is expected to shorten the production cycle of sea ice thickness products, reduce the costs of manual screening of multi-source data, model switching, and post-processing, and enhance the application value of GNSS-R satellite observation data.

[0012] Fourth, some existing GNSS-R sea ice thickness inversion schemes focus on specific sea ice types, single thickness inversion models, or simple combinations of multiple inversion results. These schemes struggle to simultaneously address issues such as differing inversion relationships between one-year and multi-year ice, inconsistent probability scales and reliability across multiple navigation systems, discontinuous thickness at sea ice type boundaries, completion of sparse observation areas, and difficulty in tracing product origins within the same daily processing chain. This invention simultaneously generates two candidate thicknesses—one-year and multi-year ice—based on the same valid GNSS-R observations. It calibrates the multi-year ice probabilities for different navigation systems and transforms valid sample support, navigation system support, and inter-system discrepancies into grid observation reliability. Furthermore, it constructs boundary priors and continuous fusion weights that vary with classification uncertainty and observation reliability. Spatial regularization is applied to the fusion weights only, without crossing sea ice type boundaries, and restricted interpolation is performed within the same valid connected region of sea ice after fusion. Therefore, this invention forms a complete technical chain from dual-candidate inversion, multi-system reliability assessment, boundary continuity gating, to uncertainty and quality indicator output, supplementing the technical deficiencies of existing daily GNSS-R sea ice thickness products in terms of boundary continuity, reliability adjustment, and origin tracing.

[0013] Fifth, annual ice and multi-year ice differ in structure, salinity, and dielectric propagation characteristics. Using the same inversion relationship makes it difficult to account for both types of sea ice, while determining the sea ice type first and then hard-switching the thickness model can easily lead to abrupt thickness changes at the type boundary. Meanwhile, GNSS-R observations exhibit spatial sparsity across different dates and regions, and the number of observations, probability output scales, and reliability levels vary across different navigation systems. While existing solutions have made improvements in sea ice type classification, thickness model construction, multi-source data fusion, and spatial interpolation, they still struggle to simultaneously maintain sea ice type boundaries, utilize information from multiple navigation systems, supplement local missing measurements, and distinguish between direct fusion results and interpolation results within the same daily product generation process. This invention addresses these mutually constraining problems collaboratively within a unified technical process through dual-candidate continuous gating, grid reliability based on multi-source evidence, adaptive boundary priors, boundary regularization only for fusion weights, and restricted interpolation and quality flag output.

[0014] Sixth, in traditional processing paths, external sea ice type data is easily used directly as the final classification result for rigidly selecting one-year or multi-year ice thickness models; the outputs of multiple navigation systems are easily averaged; spatial consistency processing also easily smooths the thickness results directly, or allows the already interpolated grid to continue as the interpolation donor for subsequent interpolations. Although the above processing is relatively easy to implement, it may cause thickness jumps near sea ice type boundaries, cross-boundary information propagation, excessive involvement of low-reliability systems, and the gradual spread of interpolation errors. This invention does not directly use external sea ice types as the final model selection result, but only as a spatial boundary prior; it does not perform indiscriminate averaging of results from different navigation systems, but calculates reliability based on system calibration results, sample support, and inter-system divergence; it does not directly smooth the candidate thicknesses of one-year and multi-year ice, but only applies spatial regularization to the fusion weights without crossing type boundaries; at the same time, it stipulates that restricted interpolation only uses the directly fused grid as the donor. Thus, while maintaining the physical differences between sea ice type boundaries and candidate thicknesses, it improves the continuity, reliability, and traceability of daily sea ice thickness products. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the daily sea ice thickness inversion method with sea ice type constraints provided by the present invention; Figure 2 A flowchart illustrating the specific implementation process of this invention, including data preparation, dual-candidate thickness inversion, sea ice type probabilistic fusion, reliability gating, spatial regularization, constrained interpolation, and product generation. Figure 3 This is a schematic diagram of the data flow and functional structure of the GNSS-R daily sea ice thickness inversion system of the present invention; Figure 4 This is a schematic diagram of the scatter plot relationship between the one-year ice thickness inversion model, the multi-year ice thickness inversion model, and the daily fused sea ice thickness results and the corresponding reference data of this invention; (a) is the scatter plot relationship between the predicted thickness and the corresponding CS2 reference thickness on the test subset of the one-year ice thickness inversion network, (b) is the scatter plot relationship between the predicted thickness and the corresponding CS2 reference thickness on the test subset of the multi-year ice thickness inversion network, and (c) is the scatter plot relationship between the final GNSS-R fused sea ice thickness and the AWI CryoSat-2 / SMOS analysis field on the common effective grid; Figure 5The diagram shows the changes in boundary prior, multi-year ice probability, initial fusion weight, regularized fusion weight, candidate thickness of one-year ice, candidate thickness of multi-year ice, hard switching thickness of sea ice type, thickness of the complete method of the present invention, and thickness of the reference field at the same location on the cross section of the boundary between one-year ice and multi-year ice. Among them, (a) represents the boundary prior, the probability of multi-year ice and the fusion weight, and (b) represents the candidate thickness, the hard switching thickness and the thickness of the complete method of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0017] The innovation of this invention lies in addressing the problem of abrupt thickness changes at sea ice type boundaries when using different thickness inversion relationships for annual and multi-year ice. It generates candidate thicknesses for both annual and multi-year ice simultaneously from the same valid GNSS-R observations, directly selecting a single candidate result without hard-determining the sea ice type. It calibrates and fuses the multi-year ice probabilities of four different navigation systems: BeiDou (BDS), GPS, GAL, and GLONASS, and combines this with sample support and inter-system divergence to form grid observation reliability. Furthermore, it uses external sea ice type information as a spatial boundary prior, which, together with the current GNSS-R observation probability and the relative surrogate reliability of the two thickness inversion networks, constitutes a continuous gating weight. Spatial regularization is applied only to the fused weights, without crossing sea ice type boundaries. Finally, after fusing the two candidate thicknesses, the surrogate variance is propagated, and restricted interpolation is applied to the missing grid to generate a daily sea ice thickness grid product with surrogate uncertainty and quality source indicators.

[0018] Example 1 like Figure 1 The daily sea ice thickness inversion method constrained by sea ice type provided in this embodiment of the invention includes the following steps: S1. Acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, form an effective observation sample set. S2. Input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively to generate one-year ice candidate thickness and multi-year ice candidate thickness. Aggregate them to the target grid using the same observation quality weight and spatial kernel to obtain the grid dual candidate thickness prediction field and the grid effective sample size. S3. Perform system-level calibration on the observation-level multi-year ice probabilities of different navigation systems and aggregate them to the target grid. The system-level calibration refers to correcting the observation-level multi-year ice probabilities of each navigation system separately. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. S4. The uncertainty of sea ice type is obtained from the comprehensive multi-year ice probability, and the surrogate variances of one-year ice and multi-year ice are constructed by combining the basic error of the two thickness networks, the observation reliability, and the candidate thickness difference. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary, where the multi-year ice side is positive, the one-year ice side is negative, and the boundary is 0. The signed boundary distance is determined according to the daily external sea ice type grid. The boundary transition width is determined based on the sea ice type uncertainty, the probability divergence between systems, the sample support degree, and the grid observation reliability, and the multi-year ice boundary prior probability is constructed. The bounded weak ranking prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness is used to obtain the continuous multi-year ice initial fusion weight. The external sea ice type grid is only used as a spatial prior and does not directly determine the final sea ice type or directly select a single candidate thickness. S5. Apply spatial regularization to the initial fusion weights between adjacent grids of the same external sea ice type, and propagate the fusion weights without crossing the boundary between one-year ice and multi-year ice to obtain regularized multi-year ice fusion weights; use the regularized multi-year ice fusion weights to fuse the candidate thicknesses of one-year ice, and at the same time apply a physical correction to the candidate thicknesses of multi-year ice to prevent them from having an inverse difference value that exceeds the allowable range relative to the candidate thicknesses of one-year ice, to obtain the physically corrected candidate thicknesses of multi-year ice, and propagate the surrogate variances of the two inversion networks and the mixed additional variances generated by the difference in candidate thicknesses to obtain the fusion thickness, fusion surrogate variance, and surrogate uncertainty; S6. Apply a mask and non-negative physical constraints to the fused thickness, and perform restricted interpolation using only the directly fused grid that is in the same effective sea ice connectivity region as the missing sea ice grid; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and do not use the interpolated grid as a new interpolation source, generate GNSS-R daily sea ice thickness, surrogate uncertainty and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between the sources of directly fused, restricted interpolation and grids from sea ice regions that are still missing.

[0019] In step S1, acquiring GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks for the target date and target sea area includes: Get target date and the set of GNSS-R observation features within the target sea area Sea ice environment and dielectric propagation parameter set and navigation system logo And obtain daily external sea ice type grids. Land shelter and open water body cover The effective sea ice region mask is constructed from the external sea ice type grid and two types of masks. After quality control and target mesh assembly Matching to form a valid observation sample set .

[0020] In step S2, the dual-candidate thickness inversion and mesh aggregation include: combining the effective observation sample set... Each valid observation sample is simultaneously input into a pre-trained one-year ice thickness inversion network. and multi-year ice thickness inversion network The candidate ice thickness along the orbit for one year was obtained. and multi-year ice candidate thickness This forms a dual-candidate thickness prediction set along the track. Based on observation quality weights and space core Aggregate it to The mesh dual-candidate thickness prediction field is obtained. and the effective sample size of the grid ; In step S3, the observation-level multi-year ice probabilities of different navigation systems are calibrated at the system level and aggregated into the target grid, including: Using the GNSS-R sea ice type probabilistic classification model Obtain the target grid for each navigation system. Calibration of multi-year ice probability at the location The comprehensive multi-year ice probability is obtained through probability-log probability domain reliability weighted fusion and grid-level calibration. ;Based on the effective sample size of the grid Obtain sample support A collection of effective navigation systems Obtaining support from navigation systems System Disagreement Penalty Thus, the reliability of grid observations is obtained. ; In step S4, the initial fusion weights of consecutive multi-year ice are obtained, including: according to Determine the signed boundary distance and boundary transition width Constructing multi-year ice boundary prior probabilities Combined with multi-year ice boundary prior probabilities Comprehensive multi-year ice probability Reliability of grid observation Uncertainty about sea ice type Variance of one-year and multi-year ice agents The initial multi-year ice fusion weights are obtained from the candidate thickness ranking relationship. And obtain the initial fusion thickness: ; In the formula, For the target grid The initial fusion thickness is obtained based on the initial multi-year ice fusion weight. For the target grid The annual ice candidate thickness at the location, For the target grid Candidate thicknesses of multi-year ice at the location, all in meters. External sea ice type grid. It only provides spatial priors and is not directly used as the final result for determining sea ice type.

[0021] This formula is used to perform continuous weighted fusion of candidate thicknesses for one-year ice and candidate thicknesses for multi-year ice. The coefficients are not fixed empirical coefficients, nor are they directly equal to the probability of a single classification. Instead, they are jointly determined by information such as the uncertainty of sea ice type mentioned above. Using the above formula, the traditional hard switching between one-year ice (FYI) and multi-year ice (MYI) is replaced with continuous fusion; sudden jumps in thickness results are avoided when the classification probability is close to 0.5; the dominance of the current classification probability on the fusion result is reduced when the reliability of GNSS-R observations is low; and the contribution ratio is adaptively adjusted when there are differences in the reliability of the two networks. This ensures a continuous transition between candidate thicknesses of one-year ice and multi-year ice in the sea ice type boundary region, reducing abrupt thickness changes caused by model switching. Since FYI and MYI use different thickness inversion relationships or inversion models, the two types of inversion results may produce significant thickness differences in the sea ice type boundary region. This method reduces abrupt thickness changes in the boundary region and improves the spatial continuity of the inversion results by continuously fusing the two types of candidate thicknesses.

[0022] In step S5, the mixed additional variance generated by propagating the surrogate variance of the two inversion networks and the candidate thickness difference includes: the annual ice candidate thickness. Multi-year ice candidate thickness Comprehensive multi-year ice probability and boundary prior probabilities All effective fused mesh sets Above, to By implementing spatial regularization to preserve sea ice type boundaries, regularized multi-year ice fusion weights are obtained. Physical corrections were made to the candidate thickness of multi-year ice. ,use Fusion and The fusion thickness is obtained by propagating the surrogate variance of the two networks and the mixed additional variance generated by the candidate thickness difference. Fusion Agent Variance and agent uncertainty .

[0023] In step S6, generating the GNSS-R daily sea ice thickness, surrogate uncertainty, and quality indicator grid products for the target date includes: For fusion thickness To implement nonnegative physical constraints, first apply... Then, directly fused grid sets within the same effectively connected sea ice region are utilized. Constrained interpolation was performed on the missing sea ice grid to obtain a set of successfully interpolated grids. Even after interpolation, the measured mesh set is still missing. Generate target date GNSS-R daily sea ice thickness output field Agent uncertainty output field and quality marks Grid products; Both the one-year ice thickness inversion network and the multi-year ice thickness inversion network include a delayed Doppler graph coding branch and a multi-source auxiliary parameter coding branch. The inputs for the one-year ice thickness inversion include specular reflection parameters, waveform statistical parameters, sea ice environment parameters, and dielectric propagation parameters. The inputs for the multi-year ice thickness inversion include reflection geometry and position parameters, sea ice environment parameters, and GNSS-R signal statistical parameters. The features in the inputs include at least the observed specular power reflectivity, Fresnel reflection features, simplified attenuation surrogate features, distance from shore, and navigation signal wavelength. The one-year ice thickness inversion network is trained using the physical consistency loss constraint between theoretical reflectivity and observed specular power reflectivity. The multi-year ice thickness inversion network uses a conditional domain discriminator with navigation system identifiers as domain labels to reduce the feature distribution differences between different navigation systems.

[0024] For example, the target mesh The effective set of along-track observations is denoted as Using observation quality weights and space core The candidate thicknesses along the track for one-year ice and multi-year ice were weighted and aggregated separately: ; In the formula, For the first Candidate thicknesses of annual or multi-year ice observed along orbit. For the corresponding grid candidate thickness; Index for candidate thickness categories of sea ice It means a year's worth of ice. This represents multi-year ice, and the effective sample size of the grid is [missing information]. ; using the pre-determined low quantile values ​​from the training data and high quantile ,and The sample support is determined as follows: ; In the formula, For the target grid The sample support, with a value range of . , To be Cut off to closed interval The function, i.e. ; The observation quality weights are determined by the original quality flags, signal-to-noise ratio, and observation geometry; when If the value is empty or the weighted denominator does not meet the preset positive value condition, the corresponding candidate mesh thickness will be marked as missing.

[0025] Weighted averaging and interval normalization are classic mathematical processing methods. However, the improvement of this invention lies in the fact that the same valid GNSS-R observations simultaneously generate two candidate thicknesses, FYI and MYI; and the two candidate thicknesses use the same observation set. Observation quality weight and space core Grid aggregation is performed; observation quality weights consider original quality indicators, signal-to-noise ratio, and observation geometry simultaneously; the effective sample size of the grid is not only used for statistics but is transformed into sample support and continues to participate in the calculation of grid observation reliability and gating weights. Through the above formulas, consistent spatial observation support is ensured for the candidate thicknesses of FYI and MYI; inconsistent effective regions are avoided due to separate screening or interpolation of the two candidate grids; the impact of distant, low signal-to-noise ratio, or poor observation geometry samples is reduced; and sparse observation grids are treated conservatively in subsequent fusion, rather than being treated the same as high-density grids.

[0026] For the target grid A set of navigation systems with effective probabilities First, system-level calibration and grid aggregation are performed on the observation-level multi-year ice probability of each navigation system to obtain the first... The probability of multi-year ice formation in the grid of a navigation system To ensure that the subsequent probability-log probability transformation is defined, we first perform probability truncation, expressed as: ; In the formula, Let be the probability cutoff constant, which satisfies A dimensionless small positive number is used to avoid the truncation probability taking 0 or 1, which would cause the subsequent probability logarithmic probability transformation to diverge. Define probability logarithmic chance transformation The initial fusion probability of multiple systems is: ; Probability truncation, log-probability transformation, and sigmoid functions are traditional mathematical tools. However, this invention improves upon these by forming the following processing chain: subsystem observation probability, subsystem calibration, intra-system grid aggregation, reliability-weighted log-probability domain fusion, and grid-level recalibration. It goes beyond simply arithmetically averaging the probabilities of BDS, GPS, GAL, and GLO. Different navigation systems differ in signal characteristics, frequencies, observation geometry, and model calibration effectiveness. Therefore, this invention does not perform a simple arithmetic average of the probabilities of each navigation system; instead, it performs a weighted fusion in the log-probability domain according to predetermined system reliability weights.

[0027] Through the above formula, this invention eliminates the problem of inconsistent probability scales and confidence levels among different navigation systems; enables navigation systems with higher reliability to make a greater contribution; probability truncation prevents 0 or 1 from causing undefined logarithmic probabilities or numerical divergence; fusion in the logarithmic probability domain can preserve the evidence direction and strength of each system's probability relative to 0.5; grid-level calibration further corrects the probability bias caused by spatial aggregation after observation-level calibration. Located in the open interval General probability variables within, For the target grid The preliminary multi-year ice probability, obtained through multi-navigation system reliability weighting and before grid-level calibration, is... It is an S-shaped function. This is the probability preprocessing transformation function. The non-negative navigation system reliability weights are determined based on a pre-validated dataset, and these weights satisfy: ; Calculation is performed only when the sum of the reliability weights of the navigation systems participating in the fusion of the current grid is greater than 0. ;right Grid-level calibration is performed to obtain the comprehensive multi-year ice probability. .make The total number of navigation systems participating in the processing is given by the navigation system support level: ; In the formula, The total number of navigation systems that can participate in sea ice probability processing, as pre-defined; When at least two navigation systems are effective, first calculate the reliability-weighted average of the truncation probabilities of each system. : ; Recalculate the system divergence penalty. : ; The larger the system divergence penalty, the lower the consistency of the multi-year ice probability among the navigation systems; the grid observation reliability is: ; In the formula, For the observation reliability bias term, , and It is a non-negative adjustment parameter. The normalized system divergence penalty quantity satisfies ;when At that time, the system's divergences are unpredictable, making... and ;when When the value is empty, the combined multi-year ice probability and observation reliability of the corresponding grid are marked as missing.

[0028] The improvements in quantity normalization and weighted deviation calculation in this invention are: simultaneous use of effective sample support; the number of effective navigation systems; multi-system probability divergence; and the establishment of specific conservative rules for single navigation systems and no navigation systems; thereby forming observation reliability specifically for multi-system GNSS-R grid probabilistic fusion. .

[0029] Through the above formula, this invention distinguishes between multiple navigation systems supporting each other and only one system providing results; when the number of effective navigation systems decreases or the observation quality deteriorates, the reliability of grid observation decreases accordingly; at the same time, it avoids misjudging a single-system grid as highly consistent due to the inability to calculate differences between systems; the stronger the probability conflict among multiple systems, the lower the reliability; the more sufficient the effective samples and effective systems, the higher the reliability. Continue to adjust the boundary width, GNSS-R probabilistic evidence strength, and proxy variance.

[0030] Gridded external sea ice type products in the target grid The annual / multi-year ice category is denoted as When two types of grids exist simultaneously in the target sea area, the sea ice type boundary is determined based on the changes in the four neighboring categories, and the signed boundary distance is calculated. This increases the uncertainty of sea ice type. The boundary transition width is determined based on the uncertainty of sea ice type, system divergence penalty, sample support, and grid observation reliability. ; In the formula, The reference boundary transition width, in units of same, , , and The parameter is a non-negative dimensionless adjustment parameter; the prior probability of the multi-year ice boundary is: ; The basic signed distance and sigmoid methods are soft boundary construction methods. The improvement of this invention lies in the boundary width. It is not a fixed value, but rather determined by the uncertainty of sea ice type. Systemic divergence Sample support Reliability of grid observation To be determined jointly.

[0031] The boundary transition width is determined adaptively, dynamically adjusted based on the uncertainty of sea ice type, system divergence, sample support, and grid observation reliability in different regions. Therefore, different regions can correspond to different boundary transition widths, such as 75 km or 25 km. Using the above formula, when classification uncertainty or system divergence is large, the boundary transition zone is expanded to avoid premature hard decisions; when samples are sufficient and observations are reliable, the transition zone is narrowed to maintain a clear boundary. This allows the role of external boundary priors to dynamically change with current observation conditions, rather than using a fixed buffer distance.

[0032] When the target sea area has only one valid sea ice type and no boundary is formed, let the signed boundary distance of each valid grid be... This sets the boundary distance difference factor in the subsequent spatial adjacency weights to 1; the multi-year ice category grid is set to... One year ice category grid take ,in, A grid that is valid for annual ice candidate thickness, multi-year ice candidate thickness, combined multi-year ice probability, and boundary prior probability is defined as follows: ; To differentiate the subsystem calibration probability The cutoff values ​​for the combined multi-year ice probability and the boundary prior probability are defined as follows: ; Apply two thickness inversion networks to the target mesh. The candidate thickness difference at the location is denoted as: ; for Define the proxy error amplification factor: ; The variance of agents for one-year-old ice and multi-year-old ice is determined by the following formula: ; As can be seen from the above formula, the surrogate variances for one-year ice and multi-year ice are not posterior variances in a strictly statistical sense. Instead, they are relative quality indices constructed based on the fundamental errors of the FYI and MYI thickness inversion networks on the test dataset, combined with the reliability of current grid observations, the uncertainty of sea ice type, and the difference between the two candidate thicknesses. These indices are used to characterize the relative reliability of the two candidate thicknesses.

[0033] This invention extends global verification error to the current grid; increases surrogate variance when observations are unreliable or the two networks are severely divergent; and allows two thickness inversion networks to have different sensitivities to the same quality factors.

[0034] in, The base error of the thickness inversion network on the test dataset is expressed in meters. These are the non-negative dimensionless coefficients calibrated for annual ice and multi-year ice networks, respectively. The candidate thickness difference scale is in meters; It is a dimensionless nonnegative quantity, from which we obtain The variance is non-negative, expressed in square meters. The evidence items related to sea ice type and the relative reliability of the inversion network, as well as the evidence items related to candidate thickness ranking, are as follows: ; ; ; The initial fusion thickness is obtained from the initial multi-year ice fusion weights: ; In the formula, For the target grid Initial fusion thickness at the location; This is the initial weight bias term. , , and Non-negative evidence weighting The variance stability constant is expressed in square meters; the above parameters, boundary transition width parameter, and surrogate variance coefficient are all determined by the test dataset or cross-validation. The variance ratio term represents the relative surrogate reliability of the two thickness inversion networks, and the candidate thickness difference term represents the bounded weak ranking prior that the candidate thickness of multi-year ice is not less than the candidate thickness of one-year ice; the weak ranking prior does not independently determine the fusion weights.

[0035] The initial multi-year ice fusion weights of this invention comprehensively consider external space priors, current GNSS-R probabilities and observation reliability, relative surrogate variances of the two thickness inversion networks, and bounded candidate thickness ranking priors.

[0036] The fused mesh set that effectively combines the candidate thickness of one-year ice, the candidate thickness of multi-year ice, the combined probability of multi-year ice, and the boundary prior probability is denoted as […]. Its index set is denoted as For four neighboring target meshes and donor grid Non-zero spatial adjacency weights are only set when both belong to the same external sea ice type category; For example, the set of indices for effectively fused meshes is denoted as For adjacent target grids and donor grid Calculate spatial adjacency weights: ; In the formula, For indicator functions, The scale is the boundary distance difference and the unit is the same as the scale. same, The initial fusion thickness difference scale is in meters; the fusion weights between adjacent grids only participate in spatial regularization when they belong to the same external sea ice type category. Grids with zero boundary distance still participate in the regularization on the corresponding side according to their external sea ice type category. The set of undirected edges formed by spatially adjacent grids in the effectively fused grid is denoted as […]. ,make For the target grid The initial weights are preserved. Given the spatial regularization strength, the feasible region for fusion weights is: ; The spatial regularization objective function and its solution are: ; ; In the formula, ,and To avoid an inverse difference exceeding the allowable range between the multi-year ice candidate thickness and the one-year ice candidate thickness, a physical correction is implemented for the multi-year ice candidate thickness: ; In the formula, The allowable thickness difference, in meters, is determined through a test dataset or cross-validation; when At that time, a physical correction quality flag for multi-year ice candidate thickness was set. The fusion thickness and fusion surrogate variance were respectively: ; ; The surrogate uncertainty of directly fused meshes is: ; The spatial regularization only constrains the fusion weights and does not directly smooth the two candidate thicknesses; the surrogate uncertainty is used for product quality characterization and is not interpreted as a strict statistical confidence interval.

[0037] The technical improvement of this invention lies in its target and adjacency relationship construction: only the FYI / MYI fusion weights are regularized, not the sea ice thickness; only adjacent grids with the same external sea ice type are established with regularized connections; the adjacency strength is controlled by both the signed boundary distance difference and the initial fusion thickness difference; the adjacency weights across one-year / multi-year ice boundaries are directly set to 0; the fusion weights are limited to... The initial weight preservation term prevents spatial regularization from completely covering the current observational evidence. Through the above formula, isolated weight spots caused by observational noise within similar sea ice regions are suppressed; FYI and MYI candidate thicknesses are not directly smoothed to avoid weakening the true thickness gradient; weight propagation across sea ice type boundaries is prevented; and spatial consistency and the original gating evidence of the current grid are balanced.

[0038] In step S6, after masking, the target mesh is divided into a land mesh set, an open water mesh set, a sea ice effective mesh set, and other invalid mesh sets outside the sea ice effective area; the sea ice effective mesh is further divided into a directly fused mesh set. Successful constrained interpolation grid set Even after interpolation, the measured mesh set is still missing. The sets do not overlap with each other and together constitute a complete partition of the target grid.

[0039] For sea ice missing measurement target grid its neighborhood Only include directly merged grids within a limited search radius that are located in the same sea ice effective connectivity region as the missing grid; only when Perform restricted interpolation if the expression is non-empty and the following condition holds true; otherwise, assign the grid to the appropriate grid. : ; In the formula, For the neighborhood Summation index of directly fused internal mesh, For missing sea ice grid The restricted interpolation neighborhood includes only those within a finite search radius and those... Directly merged grids located within the same effectively connected sea ice region For the target grid With donor grid Grid distance between It is a non-negative kernel that monotonically decreases with distance; The distance decay interpolation weights and interpolation mean are as follows: ; In the formula, For the constrained interpolation neighborhood Inner Direct Blending Mesh For the target grid Normalized interpolation weights, For the target grid The constrained interpolation thickness at the location is in meters. To directly merge meshes The fusion thickness at the point is expressed in meters. The distance between grid cells; The kernel is a non-negative kernel that monotonically decreases with distance; the interpolation proxy variance is: ; in, An additional surrogate variance is added to the interpolation, which increases with increasing interpolation distance and decreasing number of effective neighboring grids; the surrogate uncertainty for successfully constrained interpolation grids is: ; The sea ice thickness and proxy uncertainty fields of the final product are uniformly denoted as: and For directly fused meshes and successfully constrained interpolation meshes, the following are respectively: ; ; For grids with missing sea ice measurements, land grids, and other invalid grids, the output fields for sea ice thickness and surrogate uncertainty should be filled with preset missing values. For open water grids, the sea ice thickness output field should be 0, and the surrogate uncertainty output field should be filled with preset missing values. Only the original values ​​of the effective sample size, comprehensive multi-year ice probability, grid observation reliability, and navigation system combined identifier should be output for directly fused grids; for other grids, the corresponding fields should be filled with preset missing values. The quality indicators should at least distinguish between direct fusion, restricted interpolation, sea ice areas with missing measurements, land, open water, outside the effective sea ice area, non-negative truncation, and multi-year ice candidate thickness physical correction states. The surrogate uncertainty is used for product quality characterization and is not interpreted as a strict statistical confidence interval.

[0040] As can be seen, the main improvements of this invention lie in the interpolation object, interpolation source, and constraints. First, FYI / MYI candidate thickness fusion is completed, and then the final fused thickness is interpolated; FYI and MYI candidate grids are not interpolated separately before fusion; only the directly fused grid is used as the interpolation donor; the interpolation donor must be located in the same sea ice effective connectivity region as the grid to be interpolated; grids that do not meet the neighborhood and weight conditions continue to be missing measurements; the interpolation variance simultaneously propagates the neighborhood surrogate variance, the neighborhood thickness dispersion, and the distance / neighbor number penalty; the interpolation grid is identified separately by a quality indicator, and no fake GNSS-R probability and sample support information are output. Through the above formulas, local holes formed by along-track observations are filled; thickness propagation is avoided across land, open water, or unconnected sea ice regions; the interpolation results are prevented from being misinterpreted as direct GNSS-R inversion results; higher surrogate uncertainty is given for grids with large interpolation distances or sparse neighborhoods; and the source of the final product remains traceable.

[0041] Example 2 This invention provides a daily sea ice thickness inversion system constrained by sea ice type, including a data acquisition and grid matching module, a dual-candidate thickness inversion module, a sea ice type probability and observation reliability calculation module, a reliability gating fusion module, a weight space regularization and mean-variance propagation module, and a mask, constrained interpolation and product generation module. The six modules communicate with each other and coordinately execute the processing steps S1 to S6 to generate daily sea ice thickness, surrogate uncertainty and quality marker grid products.

[0042] The data acquisition and grid matching module is used to acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, an effective observation sample set is formed. The dual-candidate thickness inversion module is used to input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively, generate one-year ice candidate thickness and multi-year ice candidate thickness, and aggregate them into the target grid using the same observation quality weight and spatial kernel to obtain the grid dual-candidate thickness prediction field and the grid effective sample size; The sea ice type probability and observation reliability calculation module is used to perform system-level calibration of the observation-level multi-year ice probability of different navigation systems and aggregate it to the target grid. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. The reliability-gated fusion module is used to obtain sea ice type uncertainty from the comprehensive multi-year ice probability, and to construct the surrogate variances of one-year ice and multi-year ice by combining the basic errors, observation reliability, and candidate thickness differences of the two thickness networks. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary. The signed boundary distance is determined based on the daily external sea ice type grid, and the boundary transition width is determined based on sea ice type uncertainty, inter-system probability divergence, sample support, and grid observation reliability to construct the multi-year ice boundary prior probability. The bounded weakly ordered prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness yields the continuous multi-year ice initial fusion weights. The weight space regularization and mean-variance propagation module is used to apply space regularization to the initial fusion weights between adjacent grids of the same external sea ice type, propagating the fusion weights without crossing the boundary between one-year ice and multi-year ice, to obtain regularized multi-year ice fusion weights; the one-year ice candidate thickness is fused using the regularized multi-year ice fusion weights, and a physical correction is applied to the multi-year ice candidate thickness to prevent it from having an inverse difference value that exceeds the allowable range relative to the one-year ice candidate thickness, to obtain the physically corrected multi-year ice candidate thickness, and the mixed additional variance generated by the surrogate variance of the two inversion networks and the candidate thickness difference is propagated to obtain the fusion thickness, fusion surrogate variance and surrogate uncertainty; The masking, constrained interpolation, and product generation module is used to apply a mask and non-negative physical constraints to the fused thickness, and to perform constrained interpolation only using directly fused grids that are in the same effective sea ice connectivity region as the missing sea ice grids; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and does not use the already interpolated grids as new interpolation sources, and generates GNSS-R daily sea ice thickness, surrogate uncertainty, and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between directly fused, constrained interpolation, and grid sources that are still missing in the sea ice region.

[0043] like Figure 2This invention provides a method for inverting daily sea ice thickness constrained by sea ice type, comprising: This implementation method completes data acquisition and grid matching, dual-candidate thickness inversion, calculation of sea ice type probability and observation reliability for multiple navigation systems, boundary priors and initial gating, spatial regularization and mean-variance propagation for the same sea ice type, and masking, constrained interpolation, and daily product generation in sequence according to S1 to S6. The target date is March 31, 2024, and the target grid is a 0.25° latitude and longitude grid, with a latitude range of 50°N to 90°N and a longitude range of [missing information]. 180° to 179.75°, with a grid shape of 161×1440.

[0044] In step S1, the target date is obtained. and the set of GNSS-R observation features within the target sea area Sea ice environment and dielectric propagation parameter set Navigation system identifier The daily external sea ice type grid provided by the Ocean and Sea Ice Satellite Application Facility (OSI SAF) Land shelter and open water body cover The sea ice effective area mask is constructed from the external sea ice type grid and two types of masks. GNSS-R observation characteristics include delayed Doppler plots, signal-to-noise ratio (SNR), kurtosis, Fresnel reflection characteristics, distance from shore, reflection geometry parameters, and navigation signal wavelength; sea ice environment and dielectric propagation parameters include salinity, temperature, snow depth, and wind speed. Finite value checks are performed on the input data to remove samples with unsatisfactory SNR, incident angles greater than 40°, or missing key variables. The data undergoes quality control and target grid assembly. Matching to form a valid observation sample set The observation quality weights are then formed based on the original quality indicators, signal-to-noise ratio, and observation geometry. .

[0045] External sea ice type, land, and open water data are first converted to unit spherical 3D coordinates, and then mapped to the target grid through nearest neighbor search. Non-overlapping spatial categories are formed by land grids, open water grids, annual ice grids, multi-year ice grids, and other invalid grids outside the sea ice effective zone. External sea ice type is only used to determine the sea ice effective zone, sea ice type boundaries, and spatial priors; it does not directly replace the GNSS-R sea ice type probabilistic classification results.

[0046] This implementation uses the AWI CryoSat-2 / SMOS weekly analysis product covering the period from March 25, 2024 to March 31, 2024. The analyzed field variables... This data is used directly for comparison with the GNSS-R results from March 31, 2024, at the same location reference field, without applying additional time interpolation to the analysis field for that week. This analysis field is a continuous reference field formed by analyzing CryoSat-2 and SMOS observations; it is not considered a strict true value and is not used for sea ice type gating, spatial regularization, or restricted interpolation. Uncertainty variables... The quality information was only read as a reference product and was not used in the common grid screening, error calculation or weighted evaluation of this embodiment.

[0047] In step S2, each valid GNSS-R observation sample is simultaneously input into a pre-trained one-year ice thickness inversion network and a multi-year ice thickness inversion network to obtain candidate one-year ice thickness and candidate multi-year ice thickness along the orbit, respectively. Both networks include delayed Doppler. Figure 2 The network employs a dimensional encoding branch and an auxiliary parameter encoding branch. The features from these two branches are concatenated and input into the thickness regression output layer. During the training phase, the multi-year ice network incorporates a gradient inversion layer and a conditional domain discriminator labeled with BDS, GPS, GAL, and GLO navigation systems.

[0048] This implementation uses the parameters of the trained one-year ice thickness inversion network and multi-year ice thickness inversion network for inference, without retraining the model on the target date. Thickness labels are sourced from the CryoSat-2 product.

[0049] When performing target mesh aggregation on candidate thicknesses along the track, the target mesh is... The effective set of along-track observations is denoted as , No. The quality weight of each observation is denoted as The space core is denoted as ; The distance between the observation location and the grid center decreases monotonically, with a search radius of 25 km. For , No. The candidate thickness of annual or multi-year ice observed along the orbit is denoted as The candidate mesh thickness is calculated using the following formula: ; In the formula, For the first Candidate thicknesses of annual or multi-year ice observed along orbit. For the corresponding grid candidate thickness; Index for candidate thickness categories of sea ice It means a year's worth of ice. This represents multi-year ice, and the effective sample size of the grid is [missing information]. ; and Representing grids respectively One-year ice candidate thickness and multi-year ice candidate thickness. When If the value is empty or the weighted denominator does not meet the positive value condition, the corresponding candidate thickness is marked as missing; the effective sample size of the grid is denoted as... The candidate thickness field is obtained directly from the aggregation of inference results along the track; this implementation does not read the FYI or MYI kriging interpolation grid.

[0050] The training process of the annual ice thickness inversion network can also incorporate a three-layer medium physical consistency loss. This loss calculates the theoretical reflectivity based on sea ice thickness, snow depth, incident angle, salinity, temperature, and dielectric parameters. and the observed mirror power reflectivity Forming a robust error term: ; In the formula, This represents the thickness regression loss. Represents the robust error function. The physical consistency loss coefficient is represented. In one embodiment of the aforementioned one-year ice thickness inversion network, the training of the one-year ice thickness inversion network may incorporate the aforementioned physical consistency loss; however, the one-year ice thickness inversion model used in this application experiment employs mean squared error regression loss and does not enable this additional physical consistency loss. Therefore, the results in Tables 1 to 3 are only used to illustrate the dual-candidate continuous gating, fusion weight space regularization, and restricted interpolation, and are not used to prove the independent technical effect of the aforementioned physical consistency loss.

[0051] In step S3, the observation-level multi-year ice probabilities of the four navigation systems—BDS, GPS, GAL, and GLO—are read respectively. The probabilities of each system are first calibrated using system-level logistic regression and grid aggregation, and then fused in the logarithmic probability domain according to pre-calibrated non-negative system reliability weights. For the grid... A set of navigation systems with effective probabilities , No. The multi-year ice probability of a navigation system's grid is denoted as . Its cutoff value is denoted as The initial fusion probability of multiple systems is: ; In the formula, For probability logarithmic odds transformation, As a reliability weight for the navigation system, It is an sigmoid function. For The comprehensive multi-year ice probability is obtained after grid-level calibration. The effective sample size of the grid is Utilizing pre-determined low quantile values ​​from the training process. and high quantile Determine sample support; this implementation method pre-selects the sample size distribution of the calibration grid as of March 30, 2024. , : ; make The total number of navigation systems that can participate in the processing; navigation system support. Reliability-weighted average of the truncation probabilities of each system System Disagreement Penalty They are respectively: ; ; System divergence penalty Located within the range of 0 to 1, grid observation reliability Determine using the following formula: ; In the formula, For the observation reliability bias term, , and This is a non-negative adjustment parameter. When season and And continue calculating ;when When empty, the probability of ice over many years will be considered. and grid observation reliability Marked as missing test.

[0052] In step S4, the composition of sea ice types in the target sea area is determined. When the target sea area contains both annual ice type grids and multi-year ice type grids, the sea ice type boundaries are determined based on the four-neighbor class changes of the OSI SAF annual ice and multi-year ice grids, and the grid is calculated. Signed boundary distance to the boundary The multi-year ice side is taken as positive, the one-year ice side as negative, and the boundary is taken as 0. Let the sea ice type uncertainty be... The boundary transition width is determined based on sea ice type uncertainty, system divergence penalty, sample support, and grid observation reliability. And calculate the prior probability of multi-year ice boundary. : ; In the formula, the transition width of the reference boundary is... Take 50km, , , and This is a non-negative adjustment parameter. When the target sea area contains only one effective sea ice type and no category boundary is formed, no boundary distance transformation is performed, and the distance of each effective grid is adjusted accordingly. Only includes seasonal ice categories. It only includes seasonal ice categories for one year. ,in The cutoff values ​​for the combined multi-year ice probability and the boundary prior probability are denoted as follows: and The candidate thickness difference is denoted as .for Agent error amplification factor and agent variance Determine using the following formula: ; ; In the formula, To correspond to the basic error on the thickness inversion network test set, , and It is a non-negative dimensionless coefficient. Take 0.50m. This gives the annual ice agent variance. and the variance of ice agents over the years Sea ice type evidence item Relative reliability of inversion network and evidence items for candidate thickness ranking They are respectively: ; ; Initial multi-year ice fusion weights and initial fusion thickness They are respectively: ; ; In the formula, This is the initial weight bias term. , , and Non-negative evidence weighting is the variance stability constant. The candidate thickness ranking evidence term adopts a bounded hyperbolic tangent form and is only used as a weak ranking prior, not to independently determine the final fusion weight.

[0053] In step S5, the fused mesh set where the candidate thicknesses of one-year ice, multi-year ice, multi-year ice probability, and boundary prior probability are all valid is denoted as . Its index set is denoted as For four neighboring grids and Non-zero spatial adjacency weights are only set when both belong to the same external sea ice type category: ; In the formula, For indicator functions, Take 75km, Take 0.35m; Spatial adjacency weight across annual / multi-year ice boundaries Strictly set to 0. When the target sea area contains only one type of sea ice, each effective grid... Even with the boundary distance difference attenuation term set to 1, spatial regularization still applies. The set of undirected edges formed by spatially adjacent grids is denoted as... ,make To preserve the coefficients for the initial weights, Taking 0.65, the feasible region for fusion weights is denoted as... Then the spatial regularization objective function and its solution are: ; ; In the formula, ,and This implementation uses projection iteration to solve the box-constrained optimization problem.

[0054] To limit the inverse difference between multi-year ice candidate thickness and one-year ice candidate thickness that exceeds the allowable range, the allowable thickness difference is first used. After correcting for the multi-year ice candidate thickness, the fusion thickness and fusion surrogate variance are calculated: ; ; ; In the formula, To directly fuse thickness, To integrate the surrogate variance, the surrogate uncertainty of the directly integrated grid is denoted as... Its dimension is m. The surrogate uncertainty is output as a product quality field and is not interpreted as a statistical confidence interval.

[0055] In step S6, a land mask, an open water mask, and a sea ice effective zone mask are applied first, and a non-negative constraint is applied to the direct fusion thickness. For missing meshes within the sea ice effective zone, donors are selected only from the directly fused meshes that are in the same four-neighbor sea ice effective connectivity region. The sea ice effective connectivity region is denoted as... The interpolation search radius is denoted as Missing grid donor neighborhood Defined as: ; In this embodiment A range of 75km is used, and the number of directly merged neighbors is at least 3. Existing interpolated grids are not used as subsequent interpolation sources; if the number of neighbors is insufficient, the interpolation kernel weights do not meet the positive value condition, or the interpolation result is not a finite value, the missing-measured grid remains missing. This implementation sets the distance attenuation kernel... The distance decay interpolation weights and interpolation mean are as follows: ; In the formula, Represents the spherical distance between grid cells. Interpolation proxy variance. Determine using the following formula: ; In the formula, this embodiment takes Its dimension is m²; the surrogate uncertainty of the successfully constrained interpolation grid is denoted as . The dimension is m. Sea ice grids that do not meet the constrained interpolation conditions remain incomplete.

[0056] The directly merged mesh set, the successfully constrained interpolation mesh set, and the mesh set still missing measurements after interpolation are respectively denoted as... , and Final sea ice thickness output field and proxy uncertainty output field Write it as follows: ; ; The final product outputs include sea ice thickness, proxy uncertainty, effective grid sample size, integrated multi-year ice probability, grid observation reliability, navigation system combination identifier, and quality mark. Bits 0 to 7 of the quality flag represent, respectively, direct fusion, restricted interpolation, missing measurements in sea ice areas, land, open water, outside the effective sea ice area, non-negative truncation, and physical correction for candidate thickness of multi-year ice. The sample size, combined multi-year ice probability, grid observation reliability, and navigation system combination flag retain their original values ​​only in directly fused grids; for restricted interpolation grids and other grids, the corresponding fields are filled with missing measurements.

[0057] Figure 3 The sequential relationship between multi-source input, dual-candidate thickness and reliability gating, fusion propagation, constrained interpolation, and product output is shown in the form of system data flow.

[0058] like Figure 4 Figures (a) to (c) show the test clusters of the one-year ice thickness inversion network and the multi-year ice thickness inversion network, as well as the grid points at the same location of the final fused product and the AWI analysis field. Figure 4 Figures (a) to (c) are used to illustrate the numerical correspondence between the thickness fields and are not intended as a basis for comparing the performance of other methods. All the above fields together constitute the GNSS-R daily sea ice thickness grid product for the target date. Figure 4 Figures (a) and (b) are used to illustrate the regression relationship between the two candidate thickness networks. Figure 4 Figure (c) illustrates the numerical correspondence between the final fused product and the reference field at the same location. The dashed line in the figure represents the 1:1 reference line, and the color represents the sample or grid point density.

[0059] The corresponding GNSS-R daily sea ice thickness inversion system includes a data acquisition and grid matching module, a dual-candidate thickness inversion module, a sea ice type probability and observation reliability calculation module, a reliability gating fusion module, a weight space regularization and mean-variance propagation module, as well as a mask, restricted interpolation and product generation module; each module transmits data in sequence and executes the above S1 to S6.

[0060] The parameters described in this specification are intended to illustrate a specific executable setup. While keeping the input, processing relationships, and output fields of each step unchanged, the search radius, spatial kernel scale, regularization coefficients, and iteration termination conditions can be preset based on the target date, observation density, and grid resolution.

[0061] Application Experiment: Data Isolation and Parameter Freezing Explanation. This application experiment is used to illustrate the execution process and results of the described method on representative dates and is not intended as a rigorous independent test. The existing FYI thickness network uses a 70% / 15% / 15% random partitioning of samples at the sample level according to navigation system hierarchy. On March 31, 2024, there were 46,608 samples, with 32,804, 6,963, and 6,841 samples falling into the training, validation, and test subsets, respectively. The MYI thickness network training also uses a 70% / 15% / 15% random partitioning of samples within each navigation system and a fixed random seed of 42 for verification. On March 31, 2024, there were 6,019 samples, with 4,320, 843, and 856 samples falling into the training, validation, and test subsets, respectively. No rule was set to completely exclude samples based on date. The sea ice type classification model was trained using data from March 12 to 29, 2024, validated and calibrated using data from March 30, 2024, and March 31, 2024 was used as the classification test date; the system-level probability calibration, system reliability weights and grid-level probability calibration were all determined using the validation data from March 30, 2024.

[0062] The current application experiments do not retrain the two thickness networks or sea ice type classification models. Gating, surrogate variance, and spatial regularization parameters are fixed by existing parameter sensitivity analysis or spatial regularization scans. The low and high quantile values ​​of sample support are pre-set to 1 and 4 based on the sample size distribution of the calibration grid on March 30, 2024. The system divergence is directly limited to the range of 0 to 1 according to the aforementioned definition, without further target date quantile normalization. Therefore, subsequent application experiment results are presented as comparisons with reference fields at the same location and are not considered independent validations.

[0063] This application experiment directly reuses the fusion intermediate values ​​archived on March 31, 2024, without retraining the one-year or multi-year ice thickness inversion network. The main comparison only compares the hard switching of sea ice types with the complete method of this invention; the soft fusion without spatial regularization is only used as an auxiliary ablation for spatial regularization. All methods share the same one-year ice candidate thickness, multi-year ice candidate thickness, combined multi-year ice probability, physical correction rules, target grid, spatial mask, reference data, and common evaluation grid.

[0064] It can be seen that the technical improvements, experimental data, and the technical effects supported by this invention include: Technical improvement point 1: Dual candidate continuous reliability gating replaces 0.5 hard switching. Application experiment 1, the main data are MAE, RMSE, Bias and R of the global region, boundary region and classification uncertainty region. The positive effect is to reduce the sudden change in candidate thickness source caused by hard switching, and to improve the boundary and classification uncertainty regions. Technical improvement point 2: Spatial regularization is applied only to the fusion weights. Application experiment 2: The main data include boundary error and spatial dispersion of fusion weights (TVa). The positive effect is to reduce isolated weight fluctuations in similar sea ice regions and not directly smooth candidate thicknesses. Technical improvement point 3: first fusion and then constrained interpolation. Application experiment 3. The main data are: number of direct grids and interpolation grids, error and surrogate uncertainty. Positive effects include: filling the local gaps after direct inversion and distinguishing the product source and credibility. Technical improvement point 4: surrogate uncertainty and quality indicators; application experiment 3, the main data include average surrogate uncertainty and quality main category; positive effects include: making direct fusion, constrained interpolation and missing grid traceable, and avoiding the interpretation of interpolation as direct fusion.

[0065] Application Experiment 1: Comparison between hard switching of sea ice types and the complete method of this invention.

[0066] To verify the effectiveness of this invention's use of continuous reliability gating to replace hard switching of sea ice types, a sea ice type hard switching method based on a 0.5 probability threshold was set up as a comparison, while maintaining the same candidate thicknesses for one-year ice, multi-year ice, combined multi-year ice probability, physical correction rules, target mesh, spatial mask, and reference field. At that time, the hard switching method selects the physically corrected candidate thickness of multi-year ice. ;when Selecting candidate ice thicknesses for one year .

[0067] ; The complete method of this invention uses regularized weight fusion. Perform sequential gated fusion on the two candidate thicknesses: ; The main comparison is statistically analyzed only on the common grid where both methods and the AWI CryoSat-2 / SMOS joint analysis fields are effective in the directly fused grid. Unrestricted interpolation results are not used. Evaluation metrics include MAE, RMSE, Bias, and Pearson correlation coefficient. and effective common grid number After verifying the data source and model training process, the AWI weekly analysis field was not used as the input for gating, spatial regularization, or restricted interpolation in this round; however, since the target date sample participated in the sample-level random partitioning of the two thickness networks, the following results are presented as a comparison with the reference field at the same location, and are not referred to as independent verification.

[0068] The evaluation region includes all common, effective, directly fused grids, the boundary region within 50 km of the external one-year glacier-multi-year glacier boundary, and the classification uncertainty region with a comprehensive multi-year glacier probability between 0.4 and 0.6. The boundary region and the classification uncertainty region are defined as follows: ; .

[0069] Table 1 shows the comparison results of hard switching of sea ice types and the complete method of this invention in different evaluation areas. Table 1. Comparison of sea ice type hard switching and the complete method of this invention in different evaluation areas.

[0070] On 28,374 commonly valid direct fusion grids as of March 31, 2024, the MAE and RMSE of the complete method of this invention were 0.1493m and 0.2049m, respectively, lower than the 0.1648m and 0.2380m of the sea ice type hard switching method. On 2,392 sea ice type boundary grids, the RMSE of the complete method of this invention was 0.2735m, while that of the hard switching method was 0.3590m; on 1,246 grids with uncertain classification, the RMSE of the complete method of this invention was 0.2521m, while that of the hard switching method was 0.3830m. These results indicate that within this embodiment and the commonly evaluated grid range, the continuous reliability gating of this invention can reduce the in-situ reference field error compared to the hard switching method based on a 0.5 threshold. The improvement is mainly concentrated on sea ice type boundaries and grids with uncertain classification results; this conclusion cannot be extrapolated to other dates, sea areas, or conditions.

[0071] Application Experiment 2: Auxiliary Ablation Analysis with Fusion Weight Space Regularization.

[0072] To examine the effect of applying spatial regularization only to the fusion weights, unregistered soft fusion is retained as an auxiliary ablation. Unregistered soft fusion uses the initial fusion weights. : ; Construct a set of four-neighbor undirected edges with the same external sea ice type on all valid direct fusion meshes. And calculate the spatial dispersion of the fusion weights: ; The auxiliary ablation results for the fusion weight space regularization are shown in Table 2. Table 2. Auxiliary ablation results using weight space regularization.

[0073] The per-grid error changes little in the boundary region before and after spatial regularization, but the spatial dispersion of the fused weights decreases from 0.042621 to 0.021540; both states are statistically analyzed on 43290 adjacent edges of the same sea ice type. This result indicates that spatial regularization in this embodiment is mainly used to reduce isolated weight fluctuations within regions of the same external sea ice type and improve weight spatial consistency, rather than directly smoothing candidate thicknesses for one-year and multi-year ice, nor is it solely aimed at significantly reducing per-grid error across the entire region.

[0074] Application Experiment 3: Source Analysis of Direct Fusion and Constrained Interpolation Products Constrained interpolation is used to supplement locally missing grids after direct fusion and does not replace GNSS-R direct fusion evidence. As required by the method in step S6, constrained interpolation only uses the directly fused grids within the same sea ice effective connectivity area as the donor, and the interpolated grids are not used as new donors to continue spreading outward; the interpolated grids do not output effective sample size, integrated multi-year ice probability, grid observation reliability, and navigation system combination identifier, and are distinguished from the directly fused grids by quality indicators.

[0075] Table 3. Reference field consistency results of grids from different product sources

[0076] Within the same reference field range in this embodiment, the MAE of the directly fused grid is 0.1493m, and the MAE of the constrained interpolation grid is 0.1933m; their average surrogate uncertainties are 0.2487m and 0.3585m, respectively. The error and average surrogate uncertainty of the constrained interpolation grid are both higher than those of the directly fused grid, which is consistent with its product completion source and lower confidence level; the surrogate uncertainty is only used as a quality field and is not interpreted as a strict statistical confidence interval.

[0077] Boundary Profile Description: A latitudinal profile traversing the outer annual glaciation-multi-glaciation boundary was selected from the results as of March 31, 2024. This profile contains 54 continuous, effective, directly fused grids, with the composite multi-glaciation probability ranging from 0.1385 to 0.8850. The median difference in candidate thicknesses is 0.1821 m. Selection criteria prioritized traversing the outer annual glaciation-multi-glaciation boundary, preferentially passing through areas with a composite multi-glaciation probability close to or exceeding 0.5, exhibiting visible differences in candidate thicknesses, and having a large number of continuous effective grids; the method did not follow the minimum error approach. Figure 5 As shown in Figures (a) and (b), Figure (a) plots the boundary prior, the combined multi-year ice probability, the initial fusion weight, and the regularized fusion weight, while Figure (b) plots the candidate thickness of one-year ice, the candidate thickness of multi-year ice after physical correction, the thickness of sea ice type hard switching, the thickness of the complete method of this invention, and the reference field at the same location.

[0078] The profile results show that hard switching of sea ice type directly changes the source of candidate thickness when the overall multi-year ice probability crosses the 0.5 threshold; initial soft fusion weakens this switching through continuous weights, and spatial regularization only acts on the fusion weights without directly smoothing the two candidate thicknesses, thereby reducing isolated weight fluctuations while maintaining the physical differences of candidate thicknesses.

[0079] Figure 5 This is a schematic diagram showing the changes in boundary prior, comprehensive multi-year ice probability, initial fusion weight, regularized fusion weight, candidate thickness, hard switching thickness, and thickness of the complete method of this invention on the cross-section of the boundary between one-year ice and multi-year ice in an embodiment of the present invention.

[0080] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for inverting daily sea ice thickness constrained by sea ice type, characterized in that, The method includes the following steps: S1. Acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, form an effective observation sample set. S2. Input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively to generate one-year ice candidate thickness and multi-year ice candidate thickness. Aggregate them to the target grid using the same observation quality weight and spatial kernel to obtain the grid dual candidate thickness prediction field and the grid effective sample size. S3. Perform system-level calibration on the observation-level multi-year ice probabilities of different navigation systems and aggregate them to the target grid. The system-level calibration refers to correcting the observation-level multi-year ice probabilities of each navigation system separately. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. S4. The uncertainty of sea ice type is obtained from the comprehensive multi-year ice probability, and the surrogate variances of one-year ice and multi-year ice are constructed by combining the basic error of the two thickness networks, the observation reliability, and the candidate thickness difference. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary, where the multi-year ice side is positive, the one-year ice side is negative, and the boundary is 0. The signed boundary distance is determined according to the daily external sea ice type grid. The boundary transition width is determined based on the sea ice type uncertainty, the probability divergence between systems, the sample support degree, and the grid observation reliability, and the multi-year ice boundary prior probability is constructed. The bounded weak ranking prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness is used to obtain the continuous multi-year ice initial fusion weight. The external sea ice type grid is only used as a spatial prior and does not directly determine the final sea ice type or directly select a single candidate thickness. S5. Apply spatial regularization to the initial fusion weights between adjacent grids of the same external sea ice type, and propagate the fusion weights without crossing the boundary between one-year ice and multi-year ice to obtain regularized multi-year ice fusion weights; use the regularized multi-year ice fusion weights to fuse the candidate thicknesses of one-year ice, and at the same time apply a physical correction to the candidate thicknesses of multi-year ice to prevent them from having an inverse difference value that exceeds the allowable range relative to the candidate thicknesses of one-year ice, to obtain the physically corrected candidate thicknesses of multi-year ice, and propagate the surrogate variances of the two inversion networks and the mixed additional variances generated by the difference in candidate thicknesses to obtain the fusion thickness, fusion surrogate variance, and surrogate uncertainty; S6. Apply a mask and non-negative physical constraints to the fused thickness, and perform restricted interpolation using only the directly fused grid that is in the same effective sea ice connectivity region as the missing sea ice grid; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and do not use the interpolated grid as a new interpolation source, generate GNSS-R daily sea ice thickness, surrogate uncertainty and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between the sources of directly fused, restricted interpolation and grids from sea ice regions that are still missing.

2. The daily sea ice thickness inversion method constrained by sea ice type according to claim 1, characterized in that, In step S1, acquiring GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grid, and land and open water masks for the target date and target sea area includes: acquiring the target date and the set of GNSS-R observation features within the target sea area Sea ice environment and dielectric propagation parameter set and navigation system logo And obtain daily external sea ice type grids. Land shelter and open water body cover The effective sea ice region mask is constructed from the external sea ice type grid and two types of masks. After quality control and target mesh assembly Matching to form a valid observation sample set .

3. The daily sea ice thickness inversion method constrained by sea ice type according to claim 2, characterized in that, Effective observation sample set Each valid observation sample is input into a pre-trained one-year ice thickness inversion network. and multi-year ice thickness inversion network The candidate ice thickness along the orbit for one year was obtained. and multi-year ice candidate thickness This forms a dual-candidate thickness prediction set along the track. ; Based on observation quality weights and space core Aggregate to The mesh dual-candidate thickness prediction field is obtained. and the effective sample size of the grid ; Both the one-year ice thickness inversion network and the multi-year ice thickness inversion network include a delayed Doppler graph coding branch and a multi-source auxiliary parameter coding branch. The one-year ice thickness inversion input includes specular reflection parameters, waveform statistical parameters, sea ice environment parameters, and dielectric propagation parameters. The multi-year ice thickness inversion input includes reflection geometry and position parameters, sea ice environment parameters, and GNSS-R signal statistical parameters. The features in the input include at least the observed specular power reflectivity, Fresnel reflection features, simplified attenuation surrogate features, distance from shore, and navigation signal wavelength. The one-year ice thickness inversion network is trained using the physical consistency loss constraint between theoretical reflectivity and observed specular power reflectivity. The multi-year ice thickness inversion network uses a conditional domain discriminator with navigation system identifiers as domain labels to reduce the feature distribution differences between different navigation systems. target grid The effective set of along-track observations is denoted as Using observation quality weights and space core Weighted aggregation of candidate thicknesses along the rail for one-year ice and multi-year ice respectively: ； In the formula, For the first Candidate thicknesses of annual or multi-year ice observed along orbit. For the corresponding grid candidate thickness; Index for candidate thickness categories of sea ice It means a year's worth of ice. This represents multi-year ice, and the effective sample size of the grid is [missing information]. ; using the pre-determined low quantile values ​​from the training data and high quantile ,and The sample support is determined as follows: ； In the formula, For the target grid The sample support, with a value range of . , To be Cut off to closed interval The function, i.e. ; The observation quality weight Determined by the original quality specifications, signal-to-noise ratio, and observation geometry; when If the value is empty or the weighted denominator does not meet the preset positive value condition, the corresponding candidate mesh thickness will be marked as missing.

4. The daily sea ice thickness inversion method constrained by sea ice type according to claim 3, characterized in that, In step S3, the observation-level multi-year ice probabilities of different navigation systems are calibrated at the system level and aggregated into the target grid, including: using the GNSS-R sea ice type probability classification model. Obtain the target grid for each navigation system. Calibration of multi-year ice probability at the location The comprehensive multi-year ice probability is obtained through probability-log probability domain reliability weighted fusion and grid-level calibration. ;Based on the effective sample size of the grid Obtain sample support A collection of effective navigation systems Obtaining support from navigation systems System Disagreement Penalty The reliability of grid observations is obtained. ; For the target grid A set of navigation systems with effective probabilities System-level calibration and grid aggregation were performed on the observation-level multi-year ice probability of each navigation system to obtain the first... The probability of multi-year ice formation in the grid of a navigation system Probability truncation is performed, and the expression is: ； In the formula, This is the probability cutoff constant; Define probability logarithmic chance transformation The initial fusion probability of multiple systems is: ； In the formula, Located in the open interval General probability variables within, For the target grid The preliminary multi-year ice probability, obtained through multi-navigation system reliability weighting and before grid-level calibration, is... It is an S-shaped function. This is the probability preprocessing transformation function. The non-negative navigation system reliability weights are determined based on a pre-validated dataset, and these weights satisfy: ； Calculation is performed only when the sum of the reliability weights of the navigation systems participating in the fusion of the current grid is greater than 0. ;right Grid-level calibration is performed to obtain the comprehensive multi-year ice probability. ;make The total number of navigation systems participating in the processing is given by the navigation system support level: ； In the formula, The total number of navigation systems participating in sea ice probability processing, as pre-defined; When at least two navigation systems are effective, calculate the reliability-weighted average of the truncation probabilities of each system. : ； Recalculate the system divergence penalty. : ； The larger the system divergence penalty, the lower the consistency of the multi-year icing probability among the navigation systems; the reliability of grid observations is: ； In the formula, For the observation reliability bias term, All are non-negative adjustment parameters. The normalized system divergence penalty quantity satisfies ;when At that time, the system's divergences are unpredictable, making... and ;when When the value is empty, the combined multi-year ice probability and observation reliability of the corresponding grid are marked as missing.

5. The daily sea ice thickness inversion method constrained by sea ice type according to claim 4, characterized in that, In step S4, the initial fusion weights of consecutive multi-year ice are obtained, including: according to Determine the signed boundary distance and boundary transition width Constructing multi-year ice boundary prior probabilities Combined with multi-year ice boundary prior probabilities Comprehensive multi-year ice probability Reliability of grid observation Uncertainty about sea ice type One-year ice agent variance and the variance of ice agents over the years The initial multi-year ice fusion weights are obtained from the candidate thickness ranking relationship. And obtain the initial fusion thickness: ； In the formula, For the target grid The initial fusion thickness is obtained based on the initial multi-year ice fusion weight. For the target grid The annual ice candidate thickness at the location, For the target grid Candidate thickness of multi-year ice at the location; External sea ice type grid It only provides spatial priors and is not directly used as the final result for determining sea ice type.

6. The daily sea ice thickness inversion method constrained by sea ice type according to claim 5, characterized in that, The external sea ice type grid For gridded external sea ice type products in the target grid For the annual / multi-year ice category, when two types of grids exist simultaneously in the target sea area, the sea ice type boundary is determined based on the category changes of the four neighboring regions, and the signed boundary distance is calculated. ; The uncertainty of sea ice type is The boundary transition width is determined based on the uncertainty of sea ice type, system divergence penalty, sample support, and grid observation reliability. ； In the formula, All are non-negative dimensionless adjustment parameters. The reference boundary transition width; The prior probability of the multi-year ice boundary is: ； When the target sea area has only one valid sea ice type and no boundary is formed, let the signed boundary distance of each valid grid be... This sets the boundary distance difference factor in the subsequent spatial adjacency weights to 1; the multi-year ice category grid is set to... One year ice category grid take ,in, The grid that is valid for annual ice candidate thickness, multi-year ice candidate thickness, combined multi-year ice probability, and boundary prior probability is defined as follows: ； In the formula, To effectively integrate mesh sets, For target mesh set The index is The target grid, This refers to the set of all target grids within the target sea area. and Target mesh One-year ice candidate thickness and multi-year ice candidate thickness. For the target grid The overall multi-year ice probability at the location, For the target grid Prior probability of multi-year ice boundary at the location; To differentiate the subsystem calibration probability The cutoff values ​​for the combined multi-year ice probability and the boundary prior probability are defined as follows: ； In the formula, For the target grid Comprehensive multi-year ice probability Cut off to The value after that, For the target grid Prior probability of multi-year ice boundary Cut off to The value after; Apply two thickness inversion networks to the target mesh. The candidate thickness difference at the location is denoted as: ； In the formula, For the target grid The absolute difference between the candidate thickness of multi-year ice and the candidate thickness of one-year ice, in meters; for Define the proxy error amplification factor: ； The variance of agents for one-year-old ice and multi-year-old ice is determined by the following formula: ； In the formula, The base error of the thickness inversion network on the test dataset is expressed in meters. These are the non-negative dimensionless coefficients calibrated for annual ice and multi-year ice networks, respectively. The candidate thickness difference scale is in meters; It is a dimensionless nonnegative quantity, from which we obtain The variance is non-negative, in square meters; the evidence items for sea ice type and the relative reliability of the inversion network and the candidate thickness ranking are as follows: ； ； ； In the formula, For the target grid The sea ice type evidence term is formed by the boundary prior probability and the combined multi-year ice probability. For the target grid The relative reliability evidence term is formed by the relative relationship between the variances of the two candidate thicknesses and the ranking relationship of the candidate thicknesses. For probability logarithmic chance transformation, it is defined as , To comprehensively consider the probability of ice over many years The cutoff value, This is the initial weight bias term. Non-negative evidence weighting The variance stability constant is expressed in square meters. The initial fusion thickness is obtained from the initial multi-year ice fusion weights: ； In the formula, For the target grid The initial fusion thickness at that location.

7. The daily sea ice thickness inversion method constrained by sea ice type according to claim 1, characterized in that, In step S5, the mixed additional variance generated by propagating the surrogate variance of the two inversion networks and the candidate thickness difference includes: the annual ice candidate thickness. Multi-year ice candidate thickness Comprehensive multi-year ice probability and boundary prior probabilities All effective fused mesh sets Above, to By implementing spatial regularization to preserve sea ice type boundaries, regularized multi-year ice fusion weights are obtained. Physical corrections were made to the candidate thickness of multi-year ice. ,use Fusion and The fusion thickness is obtained by propagating the surrogate variance of the two networks and the mixed additional variance generated by the candidate thickness difference. Fusion Agent Variance and agent uncertainty .

8. The daily sea ice thickness inversion method constrained by sea ice type according to claim 7, characterized in that, Let the set of valid merged mesh indices be denoted as For adjacent target grids and donor grid Calculate spatial adjacency weights: ； In the formula, For adjacent target grids With donor grid The non-negative, dimensionless adjacency weights used to fuse weights that are regularized in the weight space. and Target mesh and donor grid External sea ice type category, and Target mesh and donor grid Signed boundary distance at the location, and Target mesh and donor grid Initial fusion thickness at the location, For indicator functions, The scale is the boundary distance difference and the unit is the same as the scale. same, Let be the initial fusion thickness difference scale in meters; fusion weights between adjacent grids only participate in spatial regularization when they belong to the same external sea ice type category; grids with zero boundary distance still participate in regularization on the corresponding side according to their external sea ice type category; the set of undirected edges formed by spatially adjacent grids in the effectively fused grid is denoted as . ,make For the target grid The initial weights are preserved. Given the spatial regularization strength, the feasible region for fusion weights is: ； In the formula, The feasible region for the multi-year ice fusion weight vector. The multi-year ice fusion weight vector to be optimized. For the target grid The weighted component of multi-year ice melting at the location, To effectively integrate mesh sets The corresponding set of grid indices; The spatial regularization objective function and its solution are: ； ； In the formula, Let the space regularization objective function be... For the target grid Maintain initial fusion weights The positive coefficient, Target mesh and donor grid The weighting of multi-year ice fusion needs to be optimized. To effectively merge the set of undirected edges formed by spatially adjacent grids in the grid, For the edge The corresponding spatial adjacency weights, In the feasible region Internal envoy The regularized multi-year ice fusion weight vector that achieves the minimum value; in, ,and To avoid an inverse difference exceeding the allowable range between the multi-year ice candidate thickness and the one-year ice candidate thickness, a physical correction is implemented for the multi-year ice candidate thickness: ； In the formula, For the target grid The physical correction for candidate thicknesses of multi-year ice, in meters. The allowable thickness difference, in meters, is determined through a test dataset or cross-validation; when At that time, a physical correction quality flag for multi-year ice candidate thickness was set; the fusion thickness and fusion surrogate variance were respectively: ； ； In the formula, For the target grid The fusion thickness at the point is expressed in meters. For the target grid The variance of the fusion agent at the location is expressed in square meters. For the target grid The weight of multi-year ice fusion after regularization The physical correction for candidate thicknesses of multi-year ice is given, in meters. and The proxy variances for one-year ice candidate thickness and multi-year ice candidate thickness are respectively, both in square meters; The surrogate uncertainty of directly fused meshes is: ； In the formula, To directly fuse the target mesh The non-negative agency uncertainty at the point is determined by the fused agency variance. The result is truncated and expressed in meters. It is used only to characterize the relative quality level of results from different grids and is not interpreted as a strict statistical standard uncertainty or confidence interval.

9. The daily sea ice thickness inversion method constrained by sea ice type according to claim 1, characterized in that, In step S6, generating the GNSS-R daily sea ice thickness, surrogate uncertainty, and quality indicator grid products for the target date includes: For fusion thickness To implement nonnegative physical constraints, first apply... Then, directly fused grid sets within the same effectively connected sea ice region are utilized. Constrained interpolation was performed on the missing sea ice grid to obtain a set of successfully interpolated grids. Even after interpolation, the measured mesh set is still missing. Generate target date GNSS-R daily sea ice thickness output field Agent uncertainty output field and quality marks Grid products; After masking, the target mesh is divided into a land mesh set, an open water mesh set, a sea ice effective mesh set, and other invalid mesh sets outside the sea ice effective area; the sea ice effective mesh is further divided into a directly merged mesh set. Successful constrained interpolation grid set Even after interpolation, the measured mesh set is still missing. The sets do not overlap with each other and together constitute a complete partition of the target grid; For sea ice missing measurement target grid its neighborhood Only include directly merged grids within a limited search radius that are located in the same sea ice effective connectivity region as the missing grid; only when Perform restricted interpolation if the expression is non-empty and the following condition holds true; otherwise, assign the grid to the appropriate grid. : ； In the formula, For the neighborhood Summation index of directly fused internal mesh, For the missing sea ice target grid The restricted interpolation neighborhood includes only those within a finite search radius and those... Directly merged grids located within the same effectively connected sea ice region For the target grid With donor grid Grid distance between It is a non-negative kernel that monotonically decreases with distance; The distance decay interpolation weights and interpolation mean are as follows: ； In the formula, For the constrained interpolation neighborhood Inner Direct Blending Mesh For the target grid Normalized interpolation weights, For the target grid The constrained interpolation thickness at the location is in meters. To directly merge meshes The fusion thickness at the point is expressed in meters. The distance between grid cells; The interpolation proxy variance is: ； In the formula, To successfully interpolate the target mesh Interpolation proxy variance at the location, To directly merge meshes The variance of the fusion agent at the location, An additional surrogate variance is added to the interpolation, which increases with increasing interpolation distance and decreasing number of effective neighboring grids; the surrogate uncertainty for successfully constrained interpolation grids is: ； In the formula, To successfully interpolate the target mesh The non-negative surrogate uncertainty at the point is determined by the interpolation surrogate variance. The result is truncated and expressed in meters, used to characterize the relative quality level of the interpolation result; The sea ice thickness and proxy uncertainty fields of the final product are uniformly denoted as: and For directly fused meshes and successfully constrained interpolation meshes, the following are respectively: ； ； In the formula, For the target grid The final sea ice thickness output field at the location, For the target grid The final proxy uncertainty output field at the location, To directly integrate the surrogate uncertainty of the grid, To successfully surrogate the uncertainty of the constrained interpolation grid, To directly merge mesh sets, For successful constrained interpolation mesh set.

10. A daily sea ice thickness inversion system constrained by sea ice type, characterized in that, The system implements the daily sea ice thickness inversion method constrained by any one of claims 1 to 9, and the system comprises: The data acquisition and grid matching module is used to acquire GNSS-R observation characteristics, sea ice environment and dielectric propagation parameters, navigation system identifiers, daily external sea ice type grids, and land and open water masks within the target date and target sea area. After quality control and target grid matching, an effective observation sample set is formed. The dual-candidate thickness inversion module is used to input the same valid observation sample into the pre-trained one-year ice thickness inversion network and the multi-year ice thickness inversion network respectively, generate one-year ice candidate thickness and multi-year ice candidate thickness, and aggregate them into the target grid using the same observation quality weight and spatial kernel to obtain the grid dual-candidate thickness prediction field and the grid effective sample size; The sea ice type probability and observation reliability calculation module is used to perform system-level calibration of the observation-level multi-year ice probability of different navigation systems and aggregate it to the target grid. Then, after probability log probability domain reliability weighted fusion and grid-level calibration, the comprehensive multi-year ice probability is obtained. The grid observation reliability is calculated based on the effective sample size of the grid, the number of effective navigation systems, and the probability divergence between systems. The reliability-gated fusion module is used to obtain sea ice type uncertainty from the comprehensive multi-year ice probability, and to construct the surrogate variances of one-year ice and multi-year ice by combining the basic errors, observation reliability, and candidate thickness differences of the two thickness networks. The signed boundary distance is the distance from the target grid to the one-year ice-multi-year ice boundary. The signed boundary distance is determined based on the daily external sea ice type grid, and the boundary transition width is determined based on sea ice type uncertainty, inter-system probability divergence, sample support, and grid observation reliability to construct the multi-year ice boundary prior probability. The bounded weakly ordered prior constructed by combining the boundary prior probability, the comprehensive multi-year ice probability, the grid observation reliability, and the relative size of the multi-year ice candidate thickness and the one-year ice candidate thickness yields the continuous multi-year ice initial fusion weights. The weight space regularization and mean-variance propagation module is used to apply space regularization to the initial fusion weights between adjacent grids of the same external sea ice type, propagating the fusion weights without crossing the boundary between one-year ice and multi-year ice, to obtain regularized multi-year ice fusion weights; the one-year ice candidate thickness is fused using the regularized multi-year ice fusion weights, and a physical correction is applied to the multi-year ice candidate thickness to prevent it from having an inverse difference value that exceeds the allowable range relative to the one-year ice candidate thickness, to obtain the physically corrected multi-year ice candidate thickness, and the mixed additional variance generated by the surrogate variance of the two inversion networks and the candidate thickness difference is propagated to obtain the fusion thickness, fusion surrogate variance and surrogate uncertainty; The masking, constrained interpolation, and product generation module is used to apply a mask and non-negative physical constraints to the fused thickness, and to perform constrained interpolation only using directly fused grids that are in the same effective sea ice connectivity region as the missing sea ice grids; wherein, the effective sea ice connectivity region is a continuous grid region formed within the sea ice effective region mask according to the four-neighbor connectivity relationship; and does not use the already interpolated grids as new interpolation sources, and generates GNSS-R daily sea ice thickness, surrogate uncertainty, and quality indicator grid products for the target date, wherein the quality indicator at least distinguishes between directly fused, constrained interpolation, and grid sources that are still missing in the sea ice region.

Images