A sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints

Abstract

This invention belongs to the field of marine remote sensing information processing technology and discloses a sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints. The invention acquires multi-source heterogeneous GNSS-R data and performs hierarchical grouping and quality control; performs systematic bias and geometric correction on the data to generate a spatiotemporally continuous normalized reflectivity field; constructs a three-layer coherent reflectivity physical model of air-sea ice-seawater; constructs a two-stream physical sensing neural network, embedding the physical model as a differentiable constraint module into the network loss function, and uses the reflectivity field as input to invert sea ice thickness. This invention establishes a refined quality control system for the characteristics of semi-enclosed sea areas, overcoming the limitations of long revisit cycles and low coverage of single satellites. This invention also constructs a high spatiotemporal resolution sea ice monitoring system, solving the problem of instability in inversion of pure data-driven models in thin ice areas, and significantly improving the physical robustness of sea ice thickness inversion.

Description

Technical Field

[0001] This invention belongs to the field of marine remote sensing information processing technology, and relates to a sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints. In particular, it relates to a method for inverting sea ice thickness by using reflection signals from multi-source heterogeneous satellites (FY-3E / F, Tianmu-1, etc.) and combining them with a coherent reflection physical model of the three-layer medium of air-sea ice-seawater. Background Technology

[0002] Sea ice thickness is a key parameter for assessing ice condition levels and ensuring the safety of offshore oil and gas development and shipping. Global Navigation Satellite System Reflectance (GNSS-R) technology utilizes navigation signals (GPS, BDS, etc.) reflected from the sea surface to invert geophysical parameters. With its advantages of low power consumption, all-weather operation, and high spatiotemporal resolution, it has become an emerging method for sea ice monitoring. Currently, GNSS-R-based sea ice thickness inversion techniques are mainly divided into two schools of thought: Technical Route 1: Data-driven inversion based on a single constellation (such as CYGNSS). This method uses CYGNSS data released by NASA to extract observed variables (such as reflectivity and front slope) from the time-delay Doppler image and establishes an inversion model using machine learning algorithms (such as random forest, SVM, or basic CNN). Technical Route 2: Physical inversion based on incoherent scattering models. Following the theory of sea surface wind field inversion, the sea ice surface is treated as a rough surface. The ZV (Zavorotny-Voronovich) model is used to describe the incoherent scattering process of the signal, and the sea ice parameters are fitted using the least squares method.

[0003] Despite the progress made in the above technologies, the following insurmountable drawbacks exist when applied to marine and thin-ice environments: Defect 1: High-latitude heavy ice regions have observation blind spots that cannot be solved by simple multi-satellite stitching. The phenomenon is that the mainstream CYGNSS constellation has an orbital inclination of only 35°, and its specular reflection point can only cover areas up to 38° north and south latitude. The heavy ice region in a certain bay in Liaoning Province is located between 39°N and 41°N, completely outside the observation blind spot. The underlying reason is that current technology is limited by satellite orbital dynamics design, physically unable to cover high latitudes. Although high-inclination satellites have emerged (such as Fengyun-3 E / F and Tianmu-1), the data sources of these heterogeneous satellites vary greatly. Different satellites have different receiver antenna gain patterns, and different navigation systems (such as GPS and BDS) have different transmission powers and signal modulation methods. Current technology lacks a rigorous method for normalizing heterogeneous data. If these multi-source data are directly mixed and input into the model, the systematic biases of different systems will be misinterpreted by the model as changes in sea ice characteristics, resulting in severe patchy artifacts in the inversion results.

[0004] Defect 2: The inversion accuracy for ice thinner than 50cm is extremely low, often resulting in spurious thickness readings. This manifests in the typical thin ice growth period, where the thickness inverted by existing models is often confused with roughness caused by wind and waves. This leads to spurious thickness readings due to wind and waves in ice-free or thin ice areas, or near-zero sensitivity to thickness changes. The underlying cause is that current technology primarily relies on incoherent scattering theory. However, sea ice is mostly one-year-old thin ice with a thickness of 10-50cm, on the same order of magnitude as the 19cm wavelength of L-band signals. At this thickness, when electromagnetic waves penetrate sea ice, the reflected waves from the air-ice interface and the ice-water interface exhibit strong coherent interference. As the thickness changes, the intensity of the reflected signal oscillates periodically, rather than monotonically decaying. Existing models ignore this coherent mechanism and directly apply incoherent models, inevitably leading to failure in thin ice areas.

[0005] Defect 3: Purely data-driven models suffer from poor physical consistency and weak generalization ability. The phenomenon is that existing deep learning inversion methods are mostly end-to-end black-box mappings. In extreme weather or complex sea conditions not covered by training data, the thickness values ​​output by the model often violate thermodynamic laws (e.g., outputting negative or abrupt values). The underlying reason is that neural networks essentially fit data distributions, lacking the constraints of physical laws. During the inversion process, the network is prone to taking shortcuts, overfitting non-core variables such as sea surface roughness and observation geometry, failing to truly learn the intrinsic physical relationship between sea ice dielectric properties and thickness. Once the input data (such as sea conditions) exceeds the training set, the model, lacking physical constraints (such as energy conservation and coherent reflection formulas), will make wild guesses.

[0006] Through in-depth review and analysis of existing technologies both domestically and internationally, the following deep-seated technical shortcomings were found when applying current technologies to sea ice inversion, which are difficult to overcome through conventional methods: (1) Existing incoherent scattering models fail in thin ice scenarios. Current mainstream GNSS-R sea ice inversion technologies (including many physical model-based comparison files) usually use incoherent scattering theory in marine remote sensing (such as the modified ZV model). These models assume that the reflecting surface is rough and the medium thickness is much greater than the signal wavelength, and mainly focus on volume scattering.

[0007] (2) Sea ice is a typical one-year thin ice (usually thin). Its thickness is on the same order of magnitude as the wavelength of GNSS L-band signals (approximately 19 cm). At this scale, electromagnetic waves will exhibit strong coherent interference effects at the air-sea ice and sea ice-seawater interfaces.

[0008] (3) Existing technologies completely ignore this phase interference mechanism. If existing incoherent models are directly applied, the periodic oscillation of reflectivity caused by thickness variation will be misjudged as roughness noise or inversion failure. Existing models cannot solve the thin ice inversion problem by simply adjusting parameters. A completely new coherent reflection physical architecture must be established.

[0009] The challenge of spurious fusion of multi-source heterogeneous data: While there are attempts to monitor sea ice using single high-inclination satellites (such as TechDemoSat-1) or meteorological satellites (such as the FY-3 series), and simple multi-satellite data stitching schemes exist, simple stitching is not true fusion. Due to fundamental differences between FY-3E / F (morning / evening orbit, large satellite platform) and Tianmu-1 (commercial CubeSat, small satellite platform) in receiver antenna gain patterns, thermal noise figures, and the received navigation signal systems (GPS's high power and GLONASS's frequency division multiple access), existing technologies lack a unified physical benchmark calibration method, failing to eliminate this complex inter-platform nonlinear bias. If those skilled in the art simply superimpose multi-source data (conventional techniques), the sea ice thickness field inverted from data from different satellites will exhibit banded artifacts distributed along satellite orbits, disrupting the physical consistency of the data and actually reducing inversion accuracy.

[0010] Purely data-driven models suffer from inherent limitations in physical consistency. Most sea ice retrieval schemes based on deep learning (such as CNNs and random forests) that have emerged in recent years are end-to-end black-box mappings. These methods overly rely on the statistical distribution of the training data without incorporating thermodynamic and electromagnetic laws as constraints. Complex sea conditions (with significant interference from ice-water mixing and sediment turbidity) make it easy for purely data-driven models to fit non-sea ice features (such as wave textures). Existing models are highly prone to producing results that violate physical principles when faced with samples outside the training set (such as extreme cold waves or melting periods) (e.g., outputting a sudden increase in thickness or negative values ​​when temperatures rise). This indicates that simply increasing the number of network layers (a conventional improvement) cannot solve the problem of poor physical robustness; differentiable physical equations must be introduced as constraints for the loss function.

[0011] General interpolation algorithms fail to adapt to complex, semi-enclosed sea areas. Existing data incompleteness often employs inverse distance weighted interpolation or Kriging interpolation, which are based solely on geometric distance. Sea ice distribution in semi-enclosed shallow seas is strongly coupled with seabed topography (water depth) and distance from the shore. Proximity in geometric distance does not equate to similarity in physical environment (e.g., the shallow shoal at the top of a bay in Liaoning Province and the adjacent deep-water channel, despite their proximity, exhibit vastly different ice conditions). Existing interpolation methods forcibly smooth areas with different topographic features, resulting in blurred sea ice edges and loss of texture. The POBI algorithm proposed in this invention, which combines historical trends and topographic similarity, is a targeted solution to this specific geographical challenge, rather than a simple application of general algorithms. Summary of the Invention

[0012] To address the problems of thin, rapidly changing, and severely interfered sea ice, this invention discloses a multi-level data fusion architecture, particularly a sea ice thickness inversion method based on spatiotemporal fusion of multi-source heterogeneous GNSS-R data and physical mechanism constraints. This method primarily solves the following technical problems: (1) Addressing the inability to unify heterogeneous data: By constructing a cascaded correction system based on an invariant target, the systematic deviation between FY-3E / F and heterogeneous satellites such as Tianmu-1 is eliminated, breaking through the latitudinal limitations of a single constellation and achieving high consistency observation across the entire sea area. (2) Addressing the inapplicability of thin ice inversion mechanisms: Abandoning the general incoherent model, a three-layer medium coherent reflection physical model specifically for thin ice is constructed, accurately describing the interference effects within the thin ice. (3) Addressing the unreliability of black-box models: A physics-data coupling driven architecture is adopted, embedding the physical equations into the network loss function in a differentiable manner, ensuring that the inversion results possess both high accuracy and strict compliance with the physical laws of electromagnetic scattering. The technical solutions are as follows: This invention is implemented as follows: a sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints, comprising the following steps: S1. Acquisition and preprocessing of multi-source heterogeneous data: Acquire spaceborne GNSS-R data covering the sea area, and perform hierarchical grouping and quality control on the data according to the satellite orbit characteristics and the source of reflected signals to construct a heterogeneous observation set; S2. Cascaded correction and spatiotemporal fusion of data: Perform systematic bias correction based on invariant target cross-calibration on heterogeneous observation sets, and geometric correction based on residual lookup table to generate a normalized reflectance field; project the corrected multi-source data onto a regular geographic grid, and use a spatiotemporal interpolation algorithm that integrates historical observation trends and terrain similarity weights to fill in the spatiotemporal gaps and form a spatiotemporally continuous reflectance field. S3. Physical Model Construction: Based on the theory of electromagnetic wave propagation in layered media, a coherent reflection physical model of three-layer media of air-sea ice-seawater suitable for thin ice environment is constructed, and a forward mapping relationship between sea ice thickness and surface reflectivity is established. S4. Physics-Data Coupled Inversion: Construct a dual-stream physical sensing neural network and embed the constructed physical model as a differentiable physical consistency constraint module into the network's loss function; using the obtained spatiotemporal continuous reflectance field as input, the sea ice thickness is inverted by minimizing the difference between the observed reflectance and the theoretical reflectance calculated by the physical model.

[0013] In step S1, the data source for the multi-source heterogeneous data is sampled from 22 satellites of the Fengyun-3E, Fengyun-3F, and Tianmu-1 constellations. Spaceborne GNSS-R L1 level data covering the sea area is acquired, with data sources including 22 satellites from the Fengyun-3E (FY-3E), Fengyun-3F (FY-3F), and Tianmu-1 constellations. The data is then hierarchically grouped according to satellite orbital characteristics (dusk orbit, morning orbit, high-inclination orbit) and reflected signal sources (GPS, BDS, Galileo, GLONASS) to construct a heterogeneous observation set. In step S1, the data is hierarchically grouped based on satellite orbit characteristics and the source of reflected signals; wherein, the satellite orbit characteristics include twilight orbit, morning orbit, and high-inclination orbit, and the source of reflected signals includes GPS, BDS, Galileo, and GLONASS.

[0014] In step S1, quality control includes: (1) Establish a high-precision land-sea mask in the sea area and eliminate observation points within 5km of the preset safe distance threshold from the coastline to eliminate land radio frequency interference; (2) A hierarchical SNR filtering strategy is adopted, and a reference threshold is set for BDS, GPS, Galileo and GLONASS signals respectively; (3) Perform anomaly detection based on trajectory median: The anomaly detection algorithm based on trajectory median is adopted to calculate the median of reflectivity data within a single satellite trajectory, and the median of reflectivity data is calculated by difference with the historical background field reference value at the same location to obtain statistical deviation. If the absolute value of the deviation exceeds the preset statistical threshold, the trajectory is determined to be subject to systematic interference, and a dynamic blacklist is constructed to remove abnormal observations.

[0015] In step S2, data fusion specifically includes: extracting historical observation data of the location of the grid to be completed from the same historical period, calculating the mean, variance, and trend slope of reflectance over time as historical reference features; introducing an environmental similarity factor when calculating the spatial weight of the surrounding effective observation grids, wherein the environmental similarity factor is calculated exponentially based on the differences between the grid to be completed and the surrounding grids in terms of offshore distance and water depth; and weighting and fusing the time reference value based on historical features with the spatial interpolation result based on environmental weights to obtain the final completed value.

[0016] Projection and Gridding: The multi-source heterogeneous scatter data, after cascaded mathematical correction, is projected onto a regular geographic grid of the sea area. Grid Resolution Setting: This embodiment divides the sea area. A high-resolution, regular grid (approximately 5km x 5km). Parameter selection criteria: This resolution setting comprehensively considers the size of the first Fresnel zone of the GNSS-R signal (typically several hundred meters to several kilometers) and the sampling density after multi-satellite networking. If the grid is too dense (e.g....), This can lead to too few effective observation points within a single grid, resulting in significant statistical noise; if the grid is too sparse (e.g., ... If this is done, the texture details of the broken edges of sea ice will be lost. Experiments have verified that... It is the optimal scale for balancing spatial resolution and signal-to-noise ratio inversion. To address the spatiotemporal gaps that still exist after multi-source fusion, a spatial interpolation algorithm based on historical observation behavior is adopted to numerically complete the missing grid using the historical statistical characteristics of the location, thereby generating a spatiotemporally continuous normalized sea surface reflectivity field.

[0017] In step S2, the system bias correction and geometric correction respectively include: Systematic bias correction is based on the invariant target cross-calibration theory. Using low-tilt constellation data as a benchmark, the reflectance statistical characteristics of heterogeneous observation sets in the invariant target region are calculated as a global benchmark value. The median reflectance of each observation group in the region is calculated, and the difference between the median and the global benchmark value is calculated as a systematic bias constant, which is then linearly subtracted. Geometric correction involves dividing the data, after systematic bias correction, into grid bins based on the antenna line-of-sight and azimuth angles. The mean residual value within each bin is calculated to construct a two-dimensional residual lookup table. Based on the geometric angles of the observation points, the nonlinear bias introduced by the antenna gain model error is subtracted using a lookup table method, generating a normallyized reflectivity field with unified physical meaning. .

[0018] Selecting calm, offshore waters far from the observation area as the invariant target region, the statistical analysis was conducted on the [number]th [item]. Median reflectance of each observation group in the invariant target region Calculate the median reflectance. Compared with global baseline value deviation and the original observation data Perform linear correction, the expression is: ; In the formula, This is the first-corrected reflectance after systematic bias correction. The original satellite observation reflectance, For the first The systematic bias constant of each observation group; The reflectivity residuals are binned and statistically analyzed according to the antenna line-of-sight angle and azimuth angle, and a two-dimensional correction lookup table is constructed to eliminate the observation errors introduced by satellite attitude and antenna gain.

[0019] The reflectivity residuals are binned and statistically analyzed according to the antenna line-of-sight angle and azimuth angle, and a two-dimensional correction lookup table is constructed to eliminate the observation errors introduced by satellite attitude and antenna gain. The geometric correction specifically includes: adjusting the corrected data... According to the antenna line of sight and azimuth Perform gridded binning, and construct a two-dimensional lookup table by calculating the mean residual value within each angle bin. Furthermore, the nonlinear deviation introduced by the satellite antenna gain model error is eliminated by a lookup table method, the expression of which is: ; In the formula, For the final generated normalized reflectance, and These are the antenna line-of-sight angle and azimuth angle corresponding to the observation point, respectively. This is a function for a two-dimensional residual lookup table.

[0020] The innovation lies in using the calm sea surface in the open ocean as an invariant value in nature to construct the LUT in reverse.

[0021] In step S2, the spatiotemporal fusion and gap filling of multi-source data includes: The multi-source heterogeneous scatter data, after cascaded correction, is projected onto a regular geographic grid of the sea area. For the spatiotemporal gaps still existing after projection fusion, an improved spatiotemporal interpolation algorithm based on historical observation behavior is used to numerically complete the missing grid using the historical statistical characteristics of that location. The specific steps are as follows: (a) High-dimensional extraction of historical statistical features: for any grid point to be completed Backtrack the valid observation set for the same location over the past 5 years (same season, 7-day windows before and after); calculate the historical mean of this set. As a climatological baseline, historical variance is calculated. To assess uncertainty, the least squares method was used to extract the slope of the reflectance trend as a function of years. Construct historical reference values ​​that include long-term trend corrections. The expression is: ; In the formula, This represents the time offset of the current moment relative to the center of the historical window; the introduction of the trend slope effectively solves the lag error caused by simply using the historical mean. (b) Spatial weight construction based on terrain similarity (spatial dimension improvement): utilizing surrounding effective observation grids Treating the completion points When performing spatial interpolation, calculate the spatial correlation weights. Traditional inverse distance weighted (IDW) only considers geometric distance, while this invention introduces a terrain similarity factor. Considering that sea ice distribution is significantly influenced by water depth and distance from shore (terrain features), the weighting formula is defined as follows: ; A terrain similarity factor is introduced, which is calculated by exponential decay based on the differences in distance from the shore and water depth between the grid to be completed and the surrounding grids. The weight calculation formula is defined as follows: ; In the formula, For grid With grid The geometric Euclidean distance between them For grid Relative to grid Terrain similarity weights, and They are grids and grid The distance from the shore, and They are grids and grid The water depth, and These are the characteristic attenuation scales for distance from the shore and water depth, respectively; This formula shows that even if two points are geometrically close, their weights will be exponentially suppressed if their terrain environments (such as one being deep and the other shallow) are vastly different, thus avoiding incorrect interpolation.

[0022] (c) Spatiotemporal joint dynamic completion: The calculated historical reference values ​​are used to complete the calculation. Spatial interpolation results based on terrain weights Weighted fusion is performed to obtain the final complete value; by introducing historical trends and topographic constraints, a normalized reflectivity field with continuous spatiotemporal distribution and conforming to the physical distribution law of sea ice is generated.

[0023] In step S3, the physical model construction specifically includes: Based on the theory of electromagnetic wave propagation in layered media, a coherent reflection physical model suitable for thin ice environments is constructed. The sea ice scenario is abstracted into a three-layer structure of air (medium 1), sea ice (medium 2), and seawater (medium 3). GNSS signals undergo reflection and transmission at each interface, and their equivalent reflection coefficients are... With sea ice thickness The relationship is expressed as follows: ; In the formula, For sea ice thickness, $ represents the Fresnel reflection coefficient of electromagnetic waves at the air-sea ice interface and the sea ice-seawater interface, respectively. The vertical wavenumber is the signal propagation frequency within the sea ice medium. This represents the roughness attenuation factor. Embedding electromagnetic wave propagation theory as a constraint term into the neural network loss function is another innovative aspect of this invention.

[0024] In step S4, the physical-data coupling inversion specifically includes: A coupled architecture consisting of an inversion master network and a physical mapping module is constructed. The inversion master network uses normalized reflectance maps to predict sea ice thickness. The physical mapping module, based on a physical model, maps the predicted thickness values ​​back to theoretical reflectance. By minimizing the difference between observed and theoretical reflectance, a physical consistency loss function is constructed to force the network to converge in the solution space under the constraints of physical laws.

[0025] The physical consistency loss function is defined as: ; ; In the formula, This is the total loss function; This is the mean square error loss between the inverted thickness and the measured true value; These are the weighting coefficients for the physical constraint terms; Loss due to physical constraints; This represents the number of samples in the training batch. Normalized observed reflectance; The sea ice thickness predicted by the network; To predict the thickness of the network The theoretical reflectivity is calculated by substituting it into the physical model.

[0026] This parameter is not a fixed empirical value. In this embodiment, a strategy of initialization based on orders of magnitude and dynamic annealing adjustment is used to determine it. Specifically, it includes two stages: determination of the initial value (based on orders of magnitude balancing): due to data loss (usually) (order of magnitude) and physical loss (Depending on the reflectance value, it may be) The magnitudes of the two loss terms may differ numerically. To avoid one loss term dominating the gradient direction, an initial ratio is calculated through pre-training experiments. Initialize to: This ensures that both contribute to the gradient at the beginning of training. Dynamic adjustment strategy (cross-validation and annealing): To strengthen the corrective effect of physical constraints in the later stages of training, this embodiment introduces a dynamic weight mechanism. During training, the weights gradually increase as the epoch increases. The value. Simultaneously, 5-Fold Cross Validation is used to... A grid search is performed to find the final upper limit value. Experiments show that when When the values ​​are taken in the interval [0.1, 2.0], the model has the lowest RMSE on the validation set and does not exhibit overfitting. Therefore, the final value selected is... As the optimal trade-off point, it is minimized using the backpropagation algorithm. This ensures that the inversion results conform to the coherent reflection mechanism of electromagnetic waves.

[0027] Another objective of this invention is to provide a sea ice thickness inversion system based on multi-source heterogeneous fusion and physical constraints for implementing the aforementioned sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints. The system comprises: The multi-source heterogeneous data hierarchical acquisition and grouping module is used to acquire L1 level data from FY-3E, FY-3F and the Tianmu-1 constellation (22 satellites), and group them according to orbital characteristics (dawn / dusk / high inclination) and signal system (GPS / BDS / Galileo / GLONASS) to establish a full-coverage observation set; A statistically based cascaded mathematical correction module is used to eliminate uninterpretable black-box processing and establish a cross-calibration system based on trajectory statistics for quality control and on an invariant target (calm sea surface in the open ocean). Through explicit mathematical calculations, it eliminates inter-system biases and antenna nonlinear errors, and generates a normalized reflectivity field with unified physical meaning. The three-layer medium physical modeling and constraint module is used to derive the coherent reflection physical formula applicable to thin ice and establish the nonlinear mapping relationship between sea ice thickness and reflectivity. The physics-data coupling-driven inversion module is used to construct a two-stream deep neural network and embeds the physical model as a differentiable constraint module into the loss function. By minimizing the difference between observed reflectance and theoretical reflectance, the network is forced to output sea ice thickness that conforms to physical laws.

[0028] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention proposes a method for inverting sea ice thickness. Firstly, it acquires multi-system (GPS / BDS / Galileo / GLONASS) reflection signals captured by multiple receiving platforms (including FY-3F, FY-3E, and Tianmu-1), constructing a unified quality control and cross-calibration system to eliminate observational biases between different receiver platforms and navigation signal systems. Secondly, it reconstructs a high spatiotemporal resolution normalized reflectivity field using machine learning fusion technology. Finally, it designs a dual-stream deep neural network model embedded with thermodynamic and dynamic constraints. This invention fully utilizes the compatibility of new-generation satellites such as Tianmu-1 with multiple navigation systems, significantly improving the spatiotemporal coverage and inversion accuracy of sea ice monitoring.

[0029] Secondly, this invention overcomes the limitations of long revisit cycles and low coverage of single satellites, constructing a high spatiotemporal resolution sea ice monitoring system. Existing technologies typically rely on a single GNSS-R satellite for inversion, resulting in insufficient spatiotemporal coverage in small areas. This invention proposes a multi-source heterogeneous data spatiotemporal fusion technology. Addressing the signal system differences among different navigation satellite systems (GPS / BDS / Galileo / GLONASS), a normalization processing flow based on invariant target cross-calibration is designed, and a machine learning fusion algorithm is introduced to achieve unified processing of multi-constellation reflection signals. In particular, it utilizes the coverage advantage of the FY-3E / F twilight orbit over high-latitude heavy ice areas and the high revisit rate advantage of the Tianmu-1 constellation, forming a complementary observation system with wide spatial coverage and high temporal density. In contrast, the existing CYGNSS constellation, limited by a 35° orbital inclination, has an effective observation point density close to zero at the top of a certain bay in Liaoning (north of 40°N), forming a permanent blind zone. By introducing FY-3E / F (morning / evening orbit) and Tianmu-1 (high inclination), this invention increases the average daily number of observation points in this blind zone from 0 to over 1500, completely eliminating high-latitude monitoring blind spots. Regarding revisit cycles: existing single satellites (such as FY-3E) typically have revisit cycles of over 12 hours for the same location, failing to capture the rapid drift of sea ice under tidal influences. This invention, through multi-satellite network fusion, shortens the average revisit cycle of the sea area to less than 4 hours, enabling the capture of the dynamic changes in the sea ice edge line with tidal fluctuations within a single day.

[0030] Third, this invention solves the problem of instability in inversion of pure data-driven models in thin ice regions, significantly improving the physical robustness of sea ice thickness inversion. Addressing the characteristics of thin sea ice and its susceptibility to the effects of sea surface roughness aliasing, this invention abandons the traditional black-box model and innovatively constructs an inversion architecture driven by both physical mechanisms and deep learning. By introducing physical constraint mechanisms of thermodynamic growth (cumulative freezing degree days FDD) and kinetic drift, this invention adds a physical consistency regularization term to the loss function, forcing the model output to conform to the natural physical laws of sea ice growth. This effectively suppresses spurious thicknesses and outliers caused by environmental noise, greatly improving the model's generalization ability under complex sea conditions. Quantitative proof materials (based on ablation experiments): To demonstrate the effectiveness of the physical constraints, we conducted A / B tests with and without physical constraints (Group A: pure data-driven DNN; Group B: the physically coupled network of this invention): Performance in thin ice zones less than 15cm: In nascent thin ice zones, the Group A model, lacking guidance from physical mechanisms, easily misclassifies roughness caused by sea surface waves as sea ice, outputting a large number of false thickness values ​​or even negative values ​​(non-physical phenomena). In contrast, Group B (this invention) introduces a coherent reflection physical model as a loss function constraint, reducing the average absolute error in thin ice zones from 5.34cm to 2.15cm under the same sea conditions, without exhibiting any negative value outputs that violate physical principles.

[0031] Extreme sample generalization: During the extreme thawing period when training data was scarce, the inversion accuracy of Group A model dropped by 60%, indicating failure; while Group B, thanks to the universality of the physical equations, only dropped by 15% in inversion accuracy, maintaining high robustness, proving that the invention is not just fitting data, but truly learning physical laws.

[0032] Fourth, this invention establishes a refined quality control system for the characteristics of semi-enclosed sea areas. Addressing the problem of severe land-based signal contamination in these areas, this invention establishes a stringent multi-level quality control process. By combining high-precision land-sea masking (removing data within 5km of the shore), dynamic signal-to-noise ratio threshold screening, and a blacklist mechanism based on trajectory anomaly detection, it effectively eliminates inferior data affected by land-based radio frequency interference and mixed pixels, ensuring high reliability of sea ice thickness inversion from the data source.

[0033] Fifth, this invention can ensure the safety of energy extraction and reduce operation and maintenance costs: Traditional ice condition monitoring relies on manual inspections or expensive SAR satellite imagery, with long revisit cycles. This invention utilizes low-cost GNSS-R signals to achieve high-frequency monitoring in less than 4 hours, providing real-time sea ice thickness warnings and drift trajectory predictions for offshore drilling platforms, guiding them to take timely icebreaking measures. It is estimated that this can reduce annual losses from work stoppages and facility maintenance costs for oil and gas companies due to sea ice disasters. Optimizing winter shipping routes and improving logistics efficiency: The high-precision sea ice thickness distribution map generated by this invention can directly serve maritime departments and shipping companies for icebreaker scheduling and dynamic route planning for merchant ships, avoiding areas with thick ice. During periods of high sea ice activity, this can significantly reduce ship fuel consumption and grounding risks, and improve port throughput efficiency.

[0034] Sixth, this invention enables sea ice monitoring using a domestically produced multi-source heterogeneous GNSS-R satellite network: Existing international sea ice inversion methods are primarily based on NASA's single CYGNSS constellation. This invention pioneers a heterogeneous data fusion system adapted to my country's FY-3E / F (morning / evening orbit) and Tianmu-1 (a commercial high-inclination constellation), utilizing domestic multi-satellite resources for high-latitude, high-spatial-resolution sea ice monitoring. Based on the coherent scattering mechanism, this invention employs a thin ice inversion model, specifically addressing the long-standing lack of a dedicated inversion model for the first-year thin ice (<50cm) characteristic of a certain sea area. The coherent reflection physical model constructed in this invention, involving a three-layer medium of air, sea ice, and seawater, solves the theoretical gaps in existing incoherent models regarding the description of thin ice interference effects, providing a new theoretical paradigm for global mid-to-low latitude seasonal sea ice monitoring.

[0035] Seventh, this invention solves the problem of the incompatibility between high-latitude observation blind spots and high-frequency monitoring in semi-enclosed sea areas: For a long time, sea ice monitoring has been limited by satellite orbit design, resulting in either small coverage areas (low-inclination satellites cannot cover a certain bay in Liaoning) or long revisit cycles (polar-orbiting satellites only revisit 1-2 times per day). This invention, through the spatiotemporal complementary fusion of heterogeneous constellations, successfully achieves seamless coverage and sub-sun level continuous monitoring of the entire sea area of ​​a certain sea.

[0036] Eighth, existing technologies generally assume that sea ice surfaces are rough and should be inverted using incoherent theories such as the ZV model. This invention, through experiments and mechanistic analysis, proves that in thin ice scenarios, coherent interference effects are the dominant factor, thus breaking the conventional thinking of blindly applying incoherent models and establishing a new understanding that thin ice requires coherent models. Traditional views hold that deep learning lacks interpretability and is unsuitable for rigorous physical parameter inversion. This invention, by embedding differentiable physical equations into the neural network loss function, achieves explicit constraints of physical laws on data-driven models, proving that deep learning can be deeply coupled with physical mechanisms and overcoming the prejudice that purely data-driven models are unreliable and unphysical. Attached Figure Description

[0037] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints provided in this embodiment of the invention; Figure 2 This is a diagram of the overall technical architecture for sea ice thickness inversion based on multi-source heterogeneous data fusion and physical mechanism constraints provided in this embodiment of the invention. Figure 3 This is a flowchart illustrating the specific implementation of multi-source heterogeneous data cascade mathematical correction and spatiotemporal fusion provided in this embodiment of the invention; Figure 4 This is a block diagram of the physically constrained dual-stream network structure provided in an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] The innovation of this invention lies mainly in its systematic breakthrough in addressing two major pain points of existing technologies: numerous blind spots in data coverage and the inability to interpret the physical models. Innovation Point 1: A spatiotemporal fusion system for multi-source heterogeneous data based on invariant target cascade correction was constructed. This addresses the shortcomings of existing technologies: traditional single data sources (such as CYGNSS) are limited by orbits, resulting in permanent observation blind spots and long revisit cycles in high-latitude regions of certain seas; while simple multi-satellite stitching schemes suffer from severe inter-system biases due to significant differences in signal systems and antenna characteristics between different satellites (such as FY-3E / F and Tianmu-1). This invention innovatively introduces a cascaded mathematical correction strategy, establishing a benchmark transfer system based on calm sea surfaces in the open ocean (invariant target). By eliminating nonlinear biases between heterogeneous data and combining it with a POBI interpolation algorithm based on terrain similarity, the problems of incomplete coverage and inconsistent accuracy of single data sources are solved. This allows the system to integrate the advantages of twilight orbits, morning orbits, and high-inclination commercial satellites, achieving, for the first time, sub-solar level high-consistency full-coverage monitoring in the sea area.

[0040] Innovation Point Two: This invention proposes a physics-data coupled, differentiable inversion architecture, addressing the shortcomings of existing technologies. Existing purely data-driven models (such as conventional deep learning) are essentially black-box mappings, lacking physical mechanism constraints. In thin ice scenarios, such models easily confuse sea surface roughness with sea ice thickness signals, leading to false thickness or negative values ​​that violate physical principles, and exhibiting extremely poor generalization ability in extreme sea conditions not covered by training data. This invention abandons the general incoherent scattering assumption and constructs a coherent reflection model of the air-sea ice-seawater three-layer medium specifically for thin ice, embedding it as a constraint term in the deep learning loss function in the form of a differentiable operator. By embedding a differentiable physical model as a constraint in the deep learning loss function, a deep integration of data-driven and physical mechanisms is achieved. This architecture utilizes the gradient flow generated by physical equations to guide network parameter updates, forcing the model to learn the coherent interference laws of electromagnetic waves, thereby breaking through the generalization bottleneck of purely empirical models in thin ice inversion and ensuring that the inversion results possess high physical robustness and interpretability.

[0041] Example 1: A sea ice thickness inversion method based on spatiotemporal fusion of multi-source heterogeneous data and physical-data dual-driven approach like Figure 1 As shown, this embodiment provides a complete sea ice thickness inversion technology solution. This solution combines the observation advantages of domestic Fengyun meteorological satellites and commercial micro-nano satellite constellations, solves the heterogeneous data compatibility problem through cascaded mathematical correction, and utilizes deep neural networks embedded with physical mechanisms to solve the problem of inaccurate thin ice inversion.

[0042] S1. Acquisition and preprocessing of multi-source heterogeneous data: Acquire spaceborne GNSS-R data covering the sea area, and perform hierarchical grouping and quality control on the data according to the satellite orbit characteristics and the source of reflected signals to construct a heterogeneous observation set; S2. Cascaded correction and spatiotemporal fusion of data: Perform systematic bias correction based on invariant target cross-calibration on heterogeneous observation sets, and geometric correction based on residual lookup table to generate a normalized reflectance field; project the corrected multi-source data onto a regular geographic grid, and use a spatiotemporal interpolation algorithm that integrates historical observation trends and terrain similarity weights to fill in the spatiotemporal gaps and form a spatiotemporally continuous reflectance field. S3. Physical Model Construction: Based on the theory of electromagnetic wave propagation in layered media, a coherent reflection physical model of three-layer media of air-sea ice-seawater suitable for thin ice environment is constructed, and a forward mapping relationship between sea ice thickness and surface reflectivity is established. S4. Physics-Data Coupled Inversion: Construct a dual-stream physical sensing neural network and embed the constructed physical model as a differentiable physical consistency constraint module into the network's loss function; using the obtained spatiotemporal continuous reflectance field as input, the sea ice thickness is inverted by minimizing the difference between the observed reflectance and the theoretical reflectance calculated by the physical model.

[0043] The specific process of this embodiment is as follows: Figure 2 As shown, it mainly includes the following four core steps: Step 1: Acquisition, stratification, and refined cleaning of multi-source heterogeneous data; 1. Data Source Acquisition: To achieve high spatiotemporal coverage of the sea area, this embodiment constructs a multidimensional dataset containing satellite observation data and auxiliary environmental data. This invention specifically selects the following complementary satellite constellation combinations to overcome the limitations of single data sources in sea area observations: (1) Spaceborne GNSS-R L1 level observation data: Dawn / Sunset Orbit Group (Complementary Space Coverage): Fengyun-3E (FY-3E) and Fengyun-3F (FY-3F) satellites were selected. Traditional low-inclination GNSS-R satellites (such as CYGNSS, with an inclination of approximately 35°) struggle to effectively cover the heavy ice region in the northern part of a certain sea (north of 38°N), such as the top of a bay in Liaoning. FY-3E, as the world's first civilian tawn / sunset orbit meteorological satellite, working in conjunction with the morning orbit FY-3F, can stably cover high-latitude regions.

[0044] High-inclination orbit group (complementary temporal resolution): Selected satellites 01-22 of the Tianmu-1 constellation. Sea ice exhibits rapid formation, dissipation, and drift, requiring monitoring methods with extremely high temporal resolution. As a constellation composed of multiple satellites, Tianmu-1 possesses the advantage of a high revisit rate.

[0045] Analysis of the combined advantages (physical enhancement at the data source level): This heterogeneous combination of dawn / dusk orbits and high-inclination constellations constitutes a spatiotemporally complementary observation system: FY-3E / F provides spatial fallback, ensuring no high-latitude marginal sea areas such as the Liaodong Bay are missed; while Tianmu-1 provides temporal encryption, ensuring the ability to capture the rapid evolution of sea ice within a short period. The combination of these two technologies solves the problems of incomplete and inaccurate data in sea ice inversion from the data source itself.

[0046] (2) Auxiliary environment data (corresponding) Figure 2 ERA5 Input): ERA5 reanalysis data is obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF), specifically including sea surface air temperature at 2 meters, sea surface salinity, and sea surface wind speed data, which are used for the calculation of dielectric constants and the extraction of thermodynamic parameters in subsequent physical models.

[0047] 2. Construction of a hierarchical grouping and indexing system for heterogeneous data; To address the management chaos and matching challenges of multi-source heterogeneous data (involving more than 20 satellites and 4 signal systems), this embodiment designs a dual management architecture of hierarchical grouping and coding indexing, providing an efficient data scheduling foundation for subsequent cascaded correction.

[0048] (1) Construct a three-dimensional feature index coding system: Assign a unique structured identifier code to each observation data packet. The code contains feature information in three dimensions, and the format is defined as: [platform type]-[orbit feature]-[signal system].

[0049] Platform type segment: Identifies the receiver carrier. For example, "TM" represents the Tianmu-1 constellation, and "FY" represents the Fengyun-3 series.

[0050] Orbital characteristic segment: Identifies the satellite's orbital geometry. For example, "INC45" represents a 45-degree high-inclination orbit (Tianmu), and "SSO" represents a dawn / sunset sun-synchronous orbit (FY-3E / F).

[0051] Signal type segment: Identifies the source of reflected signals. For example, "G" represents GPS L1, "C" represents BDS B1I / B1C, "E" represents Galileo E1, and "R" represents GLONASS G1.

[0052] For example, the code TM-INC53-C clearly indicates "the BeiDou reflected signal received by the Tianmu-1 satellite in a 53-degree inclination orbit".

[0053] (2) Establish a multidimensional association mapping table: Based on the above unique encoding, construct a dynamic association database of heterogeneous data and metadata. This mapping table establishes an index link between each group of data and the following key physical parameters: Orbit parameter mapping: Associates the satellite ephemeris, orbital altitude, and revisit period information corresponding to this group; Instrument parameter mapping: Associated with the corresponding receiver antenna gain pattern ID and thermal noise figure; Geometric parameter mapping: associating the effective incident angle range of this group with the distribution characteristics of specular reflection points in a certain sea area.

[0054] By establishing this indexing system, the system can directly call up similar data in batches through index key values ​​when performing subsequent cross-calibration, without having to traverse the entire dataset, which significantly improves the retrieval efficiency and processing speed of massive heterogeneous data.

[0055] Furthermore, this refined hierarchical grouping strategy is the physical prerequisite for the high-precision execution of system bias correction in the subsequent step S2. Its technical principle lies in: (1) Homogenization of error characteristics: Different satellite orbits (such as the dawn-dusk orbit of FY-3E and the high-inclination orbit of Tianmu-1) correspond to completely different receiver antenna gain patterns and thermal noise levels; while different signal sources (such as GPS L1 and BDS B1I) correspond to different transmit powers and signal bandwidths. If not grouped, the error distribution of the mixed data will exhibit complex multi-peak characteristics, making it impossible to extract a uniform deviation value.

[0056] (2) Targeted calculation: By grouping, this invention ensures that the systematic bias of the observation data is approximately a stable constant within each specific subset (e.g., the Tianmu-BDS group). This makes the invariant target cross-calibration theory in step S2 valid, thus enabling the calculation and subtraction of the unique bias fingerprint for each group, achieving accurate alignment of heterogeneous data.

[0057] 3. Refined quality control: A three-stage filtration strategy is implemented to address the characteristics of a semi-enclosed sea area. (1) First-level spatial filtering (land-sea mask): A high-precision land-sea mask is established to eliminate observation points within a preset safe distance threshold from the coastline (set to 5km in this embodiment). The physical basis for setting this distance threshold is that the signal received by the spaceborne GNSS-R does not come from a geometric point on the sea surface, but from a blaze area with a certain area. For the Fengyun-3E satellite with an orbital altitude of about 800km, although the diameter of the first Fresnel zone is only a few hundred meters, the reflectivity of land (especially cities or wetlands) to L-band signals is much higher than that of the sea surface (strong reflection source). When the specular reflection point is too close to the coastline, the strong land signal will enter the receiving window through the sidelobe of the antenna pattern or as clutter, producing a ringing effect and seriously polluting the weak reflection signal of sea ice. According to the measured statistical analysis, within 5km of the coastline, there is a significant non-physical jump in the correlation power of the reflected waveform. Therefore, setting 5km as the safe threshold can completely eliminate the signal pollution caused by land radio frequency interference and multipath effect, and ensure the purity of the inversion reference.

[0058] (2) Set different signal-to-noise ratio filtering thresholds for different signal systems (GPS L1, BDS B1I / B1C, Galileo E1, GLONASS G1).

[0059] The physical and statistical basis for setting the differentiation threshold is as follows: Physical differences: Different navigation systems exhibit significant variations in satellite transmission power, signal modulation methods (such as BPSK and BOC modulation), and spreading gain. For example, BDS IGSO / GEO satellites typically have higher grounding power, while GPS satellite signals are relatively weaker. Using a uniform SNR threshold would lead to excessive rejection of valid data from weak signal systems or an inability to effectively intercept noisy data from strong signal systems.

[0060] Statistical Determination Method: The threshold is not a fixed empirical value, but is determined through preliminary experiments and statistics. This embodiment selects three consecutive months of historical observation data from the sea area to construct histograms of the signal-to-noise ratio (SNR) distribution for each signal system under both pure noise background (such as land-blocked areas) and effective signal background. Based on the statistical principle that the false alarm rate is less than 5%, the upper limit boundary of the noise distribution is extracted as the benchmark threshold for each system (e.g., based on statistical calculations, the threshold for BDS signal is set to -4.3 dB, and for GPS signal to -2.7 dB).

[0061] (3) Three-level statistical filtering (trajectory blacklist): An anomaly detection algorithm based on trajectory median is adopted, which specifically includes two core steps: Construction of Dynamic Historical Background Field Reference Values: To eliminate misjudgments introduced by seasonal sea ice changes, this embodiment abandons static mean and adopts a time-sliding window and spatial gridding strategy to construct a dynamic climatological background field. Specifically, historical data from the past 5 years (e.g., 2020-2024) are selected. It should be noted that the 5-year time window length is an adjustable parameter (usually 3-10 years), and its specific value is not fixed. This embodiment selects 5 years as the preferred value based on a balance between sample statistical significance and climatological non-stationarity: if the window is too short (e.g., <3 years), insufficient sample size leads to high statistical noise; if the window is too long (e.g., >10 years), it cannot accurately capture the recent trend of sea ice decline caused by global warming. In practical applications, the variance of the background field under different window lengths can be calculated, and the minimum number of years with stable variance can be selected as the optimal time window. The sea area is divided into... The geographic grid is used to calculate the average reflectance within each grid, generating a climatological background field that dynamically evolves over time.

[0062] Adaptive determination of statistical threshold: Calculating the median reflectance of all observation points within a single satellite trajectory. And extract the reference value of the corresponding position of the trajectory in the background field. Calculate the deviation between the two. To determine the judgment threshold, this embodiment adopts a method based on the normal distribution assumption. Rule: Statistically analyze all trajectory deviations across the entire sea area. Calculate the probability distribution and its standard deviation. Set anomaly detection threshold. (Corresponding to approximately a 99.7% confidence interval). Judgment logic: If the deviation of a certain trajectory... If the trajectory is found to be affected by systematic interference from non-sea ice factors (such as satellite attitude maneuvers, solar flare pollution, or receiver thermal noise abrupt changes), it is considered an outlier and the entire trajectory is added to the dynamic blacklist and removed from the subsequent inversion process.

[0063] Step 2: Cascaded correction and spatiotemporal fusion based on mathematical physics methods; like Figure 3 As shown, in order to unify data from different satellites and antenna gains to the same physical reference, this step abandons black-box processing and adopts a two-step cascaded mathematical correction method: Level 1: Inter-system Deviation Linear Correction To unify data from different satellites and antenna gains to a common physical reference, this step abandons black-box processing and employs a two-step cascaded mathematical correction method. First, rigorous reference selection and deviation calculation are performed: (1) Quantitative selection criteria for the invariant target area: An open sea area in the central Pacific Ocean far from land (shore distance > 200 km) was selected as the calibration field. Environmental screening was performed using ERA5 reanalysis data, retaining only quasi-mirror observation points with sea surface wind speeds less than 3 m / s (preferably less than 2 m / s) and significant wave heights less than 0.5 m. Under these conditions, the influence of sea surface roughness on the signal is minimized, and its radar cross section tends to a physically stable value, which can be regarded as a natural calibration source.

[0064] (2) Global baseline value The method for determining the reference standard is as follows: An adaptive in-system calibration strategy is adopted. Since the data sources used in this invention (FY-3E / F and Tianmu-1) all have relative biases and there is no external absolute calibration source, a virtual physical reference is constructed: Effective observation data from all observation groups (covering all orbits and signal systems) within the invariant target area (outer sea, low wind speed) are aggregated into a complete sample pool. The median reflectance of this complete sample pool is calculated and defined as the global reference value. This method forces the distribution centers of all heterogeneous observation data to align with the center of the entire set, thereby eliminating relative systematic biases between different satellite platforms (such as FY-3E and Tianmu-1) and ensuring data consistency during subsequent fusion processing.

[0065] Level Two: Antenna and Geometric Nonlinearity Correction. To eliminate nonlinear errors introduced by antenna gain model errors and satellite attitude changes, a two-dimensional residual lookup table is constructed. Considering the significant differences in observation data density between different satellites at different angles (e.g., dense data in the center region of the antenna line of sight, and sparse data in the edge region), this invention abandons the traditional fixed-step binning and adopts an adaptive grid binning strategy, ensuring correction accuracy while reducing the storage overhead and lookup time of the lookup table. (1) Adaptive binning construction strategy: Set a statistical threshold for the distribution density of observed data. The data after the first stage of correction Mapped to antenna line-of-sight angle (scope ) and azimuth (scope On a two-dimensional plane.

[0066] Fine-grained binning: in densely distributed central regions (e.g.) ), using high-resolution meshes (such as This allows for precise capture of minute distortion details in antenna gain; Coarse-grained binning: In sparsely distributed edge regions, adjacent grids are automatically merged, using a low-resolution grid (e.g., ...). ), until the number of valid samples within that grid meets the statistical significance requirement (sample number). This variable granularity strategy effectively avoids overfitting noise caused by insufficient samples in sparse regions, while significantly compressing the size of the lookup table.

[0067] (2) Lookup table correction execution: Calculate the mean of the residual data in each adaptive grid box and construct a non-uniform two-dimensional lookup table. During the calibration phase, the geometric angles of the current observation point are used as a basis. The corresponding deviation value is obtained between adaptive grids using a bilinear interpolation algorithm, and nonlinear subtraction is then performed. ; This step generates a normalized reflectivity field with unified physical meaning and free from instrument effects. .

[0068] Step 3: Construct a coherent reflection physical model of the three-layer medium of air, sea ice, and seawater; To provide physical constraints for the neural network, this invention constructs a coherent reflection physical model suitable for thin ice environments. The choice of a coherent model over a traditional incoherent model is based on the unique physical properties of sea ice: unlike the multi-year thick ice in polar seas (typically >1.5m thick, mainly exhibiting incoherent volume scattering), sea ice is typically seasonal first-year ice, with a relatively thin thickness (<0.5m). Since the wavelength of the GNSS L-band signal (approximately 19cm) is on the same order of magnitude as the sea ice thickness, significant phase interference occurs between the reflected waves from the air-sea ice interface and the sea ice-seawater interface when the signal penetrates the sea ice. In this case, if a traditional incoherent model (i.e., simply adding the reflected energy of each layer) is used, the periodic oscillation characteristics caused by thickness variations will be ignored, leading to extremely large inversion errors. Therefore, this embodiment must abstract the detection scenario into a three-layer structure of air (medium 1), sea ice (medium 2), and seawater (medium 3). Using the ERA5 air temperature and salinity data obtained in step one, the complex dielectric constant of the sea ice is calculated using a hybrid dielectric model. Derivation of sea ice thickness With surface reflectivity Forward mapping function: ; The physical meaning of each parameter and the specific methods for obtaining them are as follows: (Quantity to be inverted): Sea ice thickness. During the training phase, this value is derived from the predicted output of the inversion network. Provided; during the verification phase, the corresponding measured true value. and (Fresnel reflection coefficients) represent the Fresnel reflection coefficients of electromagnetic waves at the air-sea ice interface and the sea ice-seawater interface, respectively. Obtained by calculation based on Fresnel's equations. Specific calculation depends on the incident angle. and the complex permittivity of each medium ( ). (Dielectric constant of sea ice) and The (seawater dielectric constant) is calculated using a hybrid dielectric model, with sea surface temperature and salinity provided by ERA5 reanalysis data as input parameters. This embodiment employs the Vant hybrid dielectric model to dynamically calculate the complex dielectric constant based on the brine volume content within the sea ice.

[0069] (Vertical wavenumber): The vertical wavenumber of the signal propagating within the sea ice medium, characterizing the phase delay. Acquisition method: Calculation formula is... . The wavelength of the GNSS signal (approximately 19cm for GPS L1 band) is used. The angle of incidence of the signal during satellite observation (which can be calculated from ephemeris geometry).

[0070] (Roughness Attenuation Factor): Characterizes the effect of sea ice surface micro-roughness on the scattering attenuation of coherent reflected energy. It is estimated using the Beckmann-Spizzichino model. . This represents the root-mean-square height of the sea ice surface. In this embodiment, Set as an empirical constant (e.g., take...) (corresponding to the typical roughness of smooth new ice), or it can be used as a learnable parameter of the network to be automatically optimized during the inversion process.

[0071] This formula will be embedded into the neural network as a non-trainable physical layer, as described above in the physical model formula. In this embodiment, it is not merely a theoretical derivation, but rather implemented in code using a differentiable operator layer within an automatic differentiation framework (such as PyTorch or TensorFlow). Its technical advantage lies in maintaining mathematical connectivity between this physical model layer and its preceding deep neural network. During the backpropagation phase, the physical consistency loss function... gradient with respect to theoretical reflectivity This can be further propagated as a gradient of sea ice thickness using the chain rule. This data is then fed back into the weight parameters of the neural network. It is this differentiable implementation that bridges the gap between physical equations and the data-driven network, allowing physical laws to directly guide the updating of neural network parameters and achieving true physics-data coupling.

[0072] Step 4: Two-stream physical sensing neural network driven by physical-data coupling; This embodiment establishes the following: Figure 4 The dual-stream physical sensing neural network shown contains two feature extraction branches and a physical constraint module.

[0073] 1. Network Input Layer; Branch 1 Input (Image Stream): The input is a cropped time-delayed Doppler image or a spatiotemporally fused reflectance map slice, containing sea ice texture and scattering features. Branch 2 Input (Vector Stream): The input is a normalized one-dimensional feature vector, including auxiliary physical parameters (ERA5 air temperature, cumulative freezing degree day FDD, ocean current velocity) and observation geometric parameters (satellite incident angle, azimuth angle).

[0074] 2. Two-stream feature extraction backbone; the inversion main network aims to extract the implicit sea ice thickness features from the input normalized reflectance field. Although this embodiment uses a two-stream CNN-MLP architecture as an example, in practical applications, this main network can adopt a variety of deep learning architectures capable of extracting high-dimensional spatiotemporal features: Scalability of architecture selection: In addition to the basic convolutional neural network, the U-Net architecture (which uses skip connections to preserve high-frequency texture details and is suitable for pixel-by-pixel thickness regression), Vision Transformer (which uses a self-attention mechanism to capture long-distance spatial dependencies), or ConvLSTM (a convolutional long short-term memory network used to extract evolution patterns in time series) can also be used. Principle of spatiotemporal feature extraction: Regardless of the specific architecture used, the core logic is: 1. Spatial texture extraction: using convolutional kernels or attention modules to capture the normalized reflectance field. The system employs several techniques: 1) identifying local texture differences (e.g., the broken texture of the ice-water mixing zone versus the smooth texture of continuous sea ice); 2) physical environment fusion: deeply fusing the extracted visual features with the physical environment vectors (temperature, salinity, and current velocity) input from the auxiliary branches. This allows the network to not only see the image features of the sea surface but also understand the thermodynamic background of sea ice growth, thus enabling it to distinguish sea ice with similar reflectivity but at different growth stages during inversion.

[0075] (1) First branch (convolutional encoder): The image input is processed by multi-layer convolutional neural network and pooling layer to extract high-dimensional spatial texture features of sea ice surface. Then, the feature transfer is enhanced by dense connection blocks and finally flattened globally.

[0076] (2) Second branch (physical feature encoder): a multi-layer fully connected network is used to process the environment and geometric vectors to extract the implicit features of thermodynamics and dynamics.

[0077] (3) Feature fusion layer: The output feature vectors of the two branches are concatenated to form a comprehensive feature vector that integrates visual texture and physical environment information.

[0078] 3. Inversion regression and physical constraints; (1) Regression prediction: The fused features are passed through a fully connected regression layer to directly output the predicted sea ice thickness value. .

[0079] (2) Physical consistency loss function: In order to solve the problem of poor generalization of black-box models, the following loss function is designed to guide training: ; In the formula, The data-driven term calculates the mean square error between the predicted thickness and the measured true value (such as SMOS or OIB data). These are physical constraints that affect the network's predictions. Substitute into the physical model described in step three Calculate the theoretical reflectivity and compare it with the actual satellite-observed reflectivity. Differences: ; By minimizing the loss function through backpropagation, the network is forced to converge in the solution space constrained by physical laws, thereby eliminating spurious thickness values ​​that violate the coherent reflection mechanism of electromagnetic waves.

[0080] To clearly illustrate the implementation process of this technical solution, it is necessary to clearly distinguish between the two stages of offline training and online inference: Offline Training Phase: Data Preparation: Construct a paired dataset containing historical observations and ground truth values. The input is the historical normalized reflectance field processed in steps S1-S2. Environmental parameters; tagged with high-precision sea ice thickness products from the same period (such as SMOS or OIB aerial survey data). Closed-loop optimization: The input data is fed into the inversion master network to obtain the predicted thickness. On the one hand, calculation Difference from truth labels (data loss) On the other hand, The data is then passed to the subsequent physical mapping module to calculate the theoretical reflectivity. and with the input end Calculate physical differences (physical losses) Gradient update: based on total loss The backpropagation algorithm is used to simultaneously update the weight parameters in the inversion main network. During this stage, the physical model guides the network's learning, forcing it to avoid shortcuts and noise, and to learn only features that conform to physical laws.

[0081] Online Inference Phase: Model Deployment: After training, save the weight parameters of the inversion main network. Note: At this point, the physical mapping module has completed its constraint function and is removed during the inference phase, no longer participating in the calculation. Fast Inversion: For newly acquired real-time satellite observation data, only preprocessing steps S1-S2 are needed to generate the model. By directly inputting the trained inversion main network, the sea ice thickness inversion results can be output in milliseconds after a single forward propagation. Technical effect: This strategy of coupling physics during training and purely data-driven during inference ensures that the model has high robustness to physical mechanisms while retaining the extremely high processing efficiency of deep learning models in business applications.

[0082] Adaptive handling for typical complex operating conditions: To enable those skilled in the art to better understand how this invention addresses the complex marine environment in actual operation, the following detailed descriptions are provided for several typical operating conditions: Scenario 1: Monitoring of the broken sea ice zone at the top of Liaomou Bay (addressing spatial heterogeneity); In early winter or during the ice-melting period, due to the effects of tides and waves, the sea ice at the top of Liaomou Bay is often broken, resulting in severe ice-water mixing. At this time, a single pixel may contain both sea ice and seawater (mixed pixels), and the spatial distribution is extremely uneven. The POBI interpolation algorithm in step S2 of this invention plays a crucial role. The system automatically identifies the offshore distance and water depth characteristics of the area, utilizing terrain similarity weights (… The algorithm automatically suppresses interpolation weights from deep water areas (which are typically ice-free or have little ice), using only surrounding observation points belonging to the same shallow water topography for completion. This avoids erroneously smoothing the low reflectivity of deep water areas to shallow water areas, thus accurately preserving the edge texture of broken ice floes and preventing the averaging and disappearance of small ice pieces.

[0083] Operating Scenario 2: Monitoring the rapid freezing process during cold wave outbreaks (addressing high temporal dynamics); Sea ice is frequently affected by Siberian cold waves, and can undergo dramatic growth (thickness increase of over 10 cm) within 24 hours. Traditional polar-orbiting satellites only visit 1-2 times per day, easily missing this rapid growth process. This invention triggers a multi-source collaborative mode. Utilizing the high-frequency revisit advantage of the Tianmu-1 constellation (22 satellites), combined with the time-layered indexing in step S1, the system can synthesize multiple sea ice thickness snapshots (e.g., one every 4 hours) within a single day. This invention can completely capture the rapid evolution trajectory of sea ice from its initial formation to its full thaw, providing sub-day-level thickness growth early warning for offshore operating platforms.

[0084] Scenario 3: Interference from wet snow / surface melting during the spring ice melt season (addressing changes in dielectric properties); As spring temperatures rise, sea ice melts, forming a water film or covering it with wet snow. The presence of liquid water drastically alters the dielectric constant, leading to an abnormally high reflectivity, which ordinary models easily misinterpret as thick ice. The physical model in step S3 of this invention automatically intervenes. The system dynamically adjusts the complex dielectric constant of sea ice at the physical layer based on input auxiliary meteorological parameters (temperature > 0°C). The calculation model (introducing a high water content model) uses a physical constraint layer to forcibly correct the inversion results, distinguishing between the differences in physical signals indicating thickening and wetting, ensuring that the thickness inverted during the ice melt period shows a correct thinning trend, rather than an erroneous thickening trend.

[0085] Through the above steps, this embodiment realizes the end-to-end inversion from multi-source heterogeneous raw data to high-precision, physically interpretable sea ice thickness products.

[0086] Example 2: The sea ice thickness inversion system based on multi-source heterogeneous fusion and physical constraints provided in this embodiment of the invention includes: The multi-source heterogeneous data hierarchical acquisition and grouping module is used to acquire L1 level data from FY-3E, FY-3F and the Tianmu-1 constellation (22 satellites), and group them according to orbital characteristics (dawn / dusk / high inclination) and signal system (GPS / BDS / Galileo / GLONASS) to establish a full-coverage observation set; A statistically based cascaded mathematical correction module is used to eliminate uninterpretable black-box processing and establish a cross-calibration system based on trajectory statistics for quality control and on an invariant target (calm sea surface in the open ocean). Through explicit mathematical calculations, it eliminates inter-system biases and antenna nonlinear errors, and generates a normalized reflectivity field with unified physical meaning. The three-layer medium physical modeling and constraint module is used to derive the coherent reflection physical formula applicable to thin ice and establish the nonlinear mapping relationship between sea ice thickness and reflectivity. The physics-data coupling-driven inversion module is used to construct a two-stream deep neural network and embeds the physical model as a differentiable constraint module into the loss function. By minimizing the difference between observed reflectance and theoretical reflectance, the network is forced to output sea ice thickness that conforms to physical laws.

[0087] Example 3: Refined Early Warning Application System for Sea Ice Disasters in Oilfields To demonstrate the inventiveness and practical application value of the technical solution of this invention, this embodiment shows the specific application process of the method in the operation support system of a certain offshore oil and gas field (such as PL19-3 oil field).

[0088] An oil field located in the southern part of a sea faces the risk of sea ice impact during winter. The work platform needs to decide whether to stop drilling operations or evacuate personnel based on the thickness of the sea ice. Traditional methods relying on manual visual inspection or helicopter patrols are limited by weather and nighttime conditions, making 24 / 7 support impossible. This invention, as the core algorithm engine, is deployed on the server of the oil field's land-based control center, executing the following fully automated process: Step A: The multi-satellite data real-time access system receives FY-3E / F and Tianmu-1 L1 level data streams forwarded from ground stations in real time.

[0089] Step B: Physical-data coupled solution. Using the dual-stream physical sensing neural network trained by this invention, high-resolution (5km grid) thickness inversion is performed on the sea area within a radius of 50km centered on the oilfield platform.

[0090] Step C: The work window judgment and early warning system sets a safety threshold (e.g., a yellow alarm is triggered if the thickness is continuously >20cm).

[0091] Step D: The command issuance system automatically generates a "Sea Ice Thickness Distribution Heat Map" and a "12-Hour Ice Condition Trend Report," which are then pushed to the platform dispatcher's terminal. Based on this, the dispatcher issues a 4-hour advance notice order to suspend hoisting operations. Compared to leasing icebreakers for 24 / 7 escort, this system, through accurate thickness inversion, helps oilfields deploy icebreaking forces only when truly needed, saving operation and maintenance costs and effectively avoiding accidents such as damage to jacket structures or breakage of submarine cables caused by sudden ice floes.

[0092] To verify the effectiveness of the sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints proposed in this invention, this experiment selected the Liaomou Bay area in January 2024 (the sea ice peak period) as the experimental scenario and carried out a comparative experiment with existing technologies.

[0093] 1. Experimental verification and effect comparison analysis; Experimental data source: The sea ice thickness inversion method of this invention was used to integrate the reflection signals of a total of 24 satellites, including FY-3E / F (morning / evening orbit) and Tianmu-1 (high inclination orbit).

[0094] True benchmark: High-resolution sea ice thickness products from Sentinel-1 SAR (Synthetic Aperture Radar) passing through the region during the same period were selected as the true benchmark. SAR data has extremely high spatial resolution and is widely recognized as the standard for sea ice remote sensing.

[0095] Evaluation metrics: Root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficient (R), and spatiotemporal coverage were selected as quantitative evaluation metrics.

[0096] 2. Comparison of scheme designs; To highlight the technical advantages of this invention, the following three sets of comparative experiments were conducted: Option 1, Existing Technology / Benchmark Group: Using only single constellation data, employing traditional inverse distance weighted interpolation, and using the standard random forest algorithm for inversion (without physical constraints).

[0097] Option 2, Ablation Experimental Group: Using the multi-source heterogeneous fusion data of the present invention, but without introducing the physical mechanism constraints in steps S3 and S4, only using a pure data-driven deep neural network for inversion.

[0098] Option 3, the method of this invention: fully adopts the multi-source heterogeneous fusion, improved POBI interpolation, and physical-data coupling driven dual-stream inversion network proposed in this invention.

[0099] 3. Experimental Results and Data Analysis; (1) Comparison of spatiotemporal coverage: The average daily effective observation grid coverage of the three schemes in a certain bay area (39°N-41°N) was statistically analyzed: Option 1 (existing technology): Due to the limitations of CYGNSS orbital inclination and the single signal, the daily coverage is only 32.4%, and there are large observation blind spots in the high latitude region north of 40°N.

[0100] Option 3 (the present invention): Thanks to the multi-source heterogeneous fusion technology of FY-3E / F and Tianmu-1 introduced in steps S1 and S2, the daily coverage rate is significantly improved to 91.6%, achieving seamless coverage of the entire sea area of ​​Liaomou Bay.

[0101] Effect analysis: This demonstrates that the heterogeneous data fusion and cascade correction technology of the present invention can effectively solve the problem of incomplete sea coverage in existing GNSS-R technology.

[0102] (2) Comparison of inversion accuracy data: The inversion results of the three schemes were compared with the Sentinel-1 true values ​​point by point, and the statistical results are shown in Table 1.

[0103] Table 1. Comparison of accuracy data for different inversion methods

[0104] Compared to Solution 1 (existing technology): the RMSE of the method of the present invention (Solution 3) is reduced from 9.82cm to 2.87cm, and the accuracy is improved by about 70%. This is mainly due to the introduction of multi-source data, which increases the sample size, and the improved POBI algorithm, which effectively fills in the spatiotemporal gaps by utilizing historical trends and terrain similarity, avoiding the smoothing effect error caused by traditional IDW interpolation.

[0105] Compared to Scheme 2 (purely data-driven): In thin ice areas (thickness <15cm), Scheme 2 is prone to non-physical phenomena (overfitting noise) such as inverted values ​​being too high or negative. This invention, however, introduces a physical consistency loss function in step S4 (…). By forcing the network output to conform to the thermodynamic and electromagnetic scattering mechanisms, the RMSE was further reduced by 2.47 cm.

[0106] Terrain adaptability demonstration: In the shallow waters of a certain bay in Liaoning (water depth <10m), the error of Scheme 1 is as high as 12cm or more; while the present invention uses POBI interpolation based on terrain similarity weight in step S2, and the error in this area is controlled within 3.5cm, which proves that the present invention has significant adaptability advantages for complex terrain sea areas.

[0107] 4. Specific comparative analysis of data fusion strategies: In order to further verify the effectiveness of the spatiotemporal interpolation algorithm based on historical observation trends and terrain similarity weights in step S2, this experiment compared the impact of different interpolation strategies on the reconstruction quality of the normalized reflectivity field while keeping the inversion network model unchanged.

[0108] (1) Comparison settings: Control group A (spatial interpolation only): The traditional inverse distance weighted algorithm is used, which only uses the spatial neighborhood information of the observation data of the day to complete the data, without considering historical trends, and the spatial weights depend only on geometric distance.

[0109] Control group B (ordinary spatiotemporal interpolation): Historical means are introduced for filling, but terrain similarity constraints are not considered during spatial interpolation (i.e., ).

[0110] This invention group (POBI algorithm): fully adopts the fusion of historical trend items and terrain similarity weights ( The interpolation strategy.

[0111] (2) Experimental results and analysis of physical phenomena (see Table 2 below): In order to quantify the accuracy of the completion, we adopted the artificial hole-cutting experiment: randomly covered 10% of the known observation points as the verification set, and compared the difference between the reconstructed value and the real observation value.

[0112] Table 2 Comparison of Reconstruction Accuracy and Physical Properties of Different Interpolation Strategies

[0113] Results Analysis: Gap Filling Capability: Control group A was limited by the coverage range of the satellite trajectory on that day, and could not effectively fill in observation blind spots with a diameter exceeding 50km, resulting in a fragmented reflectivity field. In contrast, this invention utilizes historical observation trends as prior knowledge to successfully achieve seamless coverage of the entire sea area.

[0114] Reduced edge error (core advantage): In the shallow waters at the top of a bay in Liaoning Province, water depth varies drastically. Control groups A and B, relying solely on geometric distance, incorrectly averaged the values ​​of the deep water (low reflectivity) and the shallow water (high reflectivity) areas (smoothing effect), resulting in higher RMSE in the edge areas. The terrain similarity weight introduced in this invention acts as a physical barrier, ensuring that interpolation calculations are performed only within homogeneous areas with similar terrain features, thus perfectly preserving the physical gradient and texture details of sea ice distribution.

[0115] In summary, the sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints proposed in this invention represents a significant improvement over existing technologies in addressing the two key technical challenges of difficult heterogeneous fusion of marine data and low accuracy in thin ice inversion. Experimental data demonstrate that this invention not only substantially improves the spatiotemporal coverage of monitoring but also ensures the physical reliability and high accuracy of the inversion results.

[0116] 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 sea ice thickness based on multi-source heterogeneous fusion and physical constraints, characterized in that, The method includes the following steps: S1. Acquisition and preprocessing of multi-source heterogeneous data: Acquire spaceborne GNSS-R data covering the sea area, and perform hierarchical grouping and quality control on the data according to the satellite orbit characteristics and the source of reflected signals to construct a heterogeneous observation set; S2. Cascaded correction and spatiotemporal fusion of data: Perform systematic bias correction based on invariant target cross-calibration on heterogeneous observation sets, and geometric correction based on residual lookup table to generate a normalized reflectance field; project the corrected multi-source data onto a regular geographic grid, and use a spatiotemporal interpolation algorithm that integrates historical observation trends and terrain similarity weights to fill in the spatiotemporal gaps and form a spatiotemporally continuous reflectance field. S3. Physical Model Construction: Based on the theory of electromagnetic wave propagation in layered media, a coherent reflection physical model of three-layer media of air-sea ice-seawater suitable for thin ice environment is constructed, and a forward mapping relationship between sea ice thickness and surface reflectivity is established. S4. Physics-Data Coupled Inversion: Construct a dual-stream physical sensing neural network and embed the constructed physical model as a differentiable physical consistency constraint module into the network's loss function; using the obtained spatiotemporal continuous reflectance field as input, the sea ice thickness is inverted by minimizing the difference between the observed reflectance and the theoretical reflectance calculated by the physical model.

2. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S1, the data source for the multi-source heterogeneous data is the data collected from 22 satellites of the Fengyun-3E, Fengyun-3F and Tianmu-1 constellations.

3. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S1, the data is hierarchically grouped based on satellite orbit characteristics and the source of reflected signals; the satellite orbit characteristics include twilight orbit, morning orbit, and high-inclination orbit, and the source of reflected signals includes GPS, BDS, Galileo, and GLONASS signals.

4. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S1, quality control includes: Establish a high-precision land-sea mask in the sea area to eliminate observation points within a preset safe distance threshold from the coastline in order to eliminate land radio frequency interference; A hierarchical SNR filtering strategy is adopted, and benchmark thresholds are set for BDS, GPS, Galileo, and GLONASS signals respectively; Perform anomaly detection based on trajectory median: The anomaly detection algorithm based on trajectory median is used to calculate the median of reflectivity data within a single satellite trajectory, and the median of reflectivity data is calculated by difference with the historical background field reference value at the same location to obtain the statistical deviation. If the absolute value of the deviation exceeds the preset statistical threshold, the trajectory is determined to be subject to systematic interference, and a dynamic blacklist is constructed to remove abnormal observations.

5. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S2, the system bias correction and geometric correction respectively include: Systematic bias correction is based on the invariant target cross-calibration theory. It calculates the reflectance statistical characteristics of heterogeneous observation sets in the invariant target region as a global benchmark value, calculates the median reflectance of each observation group in the region, calculates the difference between the median and the global benchmark value as a systematic bias constant, and performs linear subtraction. Geometric correction involves dividing the data, after systematic bias correction, into grid bins based on the antenna line-of-sight and azimuth angles. The mean residual value within each bin is calculated to construct a two-dimensional residual lookup table. Based on the geometric angles of the observation points, the nonlinear bias introduced by the antenna gain model error is subtracted using the lookup table method, generating a normallyized reflectivity field with unified physical meaning. .

6. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 5, characterized in that, System bias correction specifically includes: Selecting calm, offshore waters far from the observation area as the invariant target region, the statistical analysis was conducted on the [number]th [item]. Median reflectance of each observation group in the invariant target region Calculate the median reflectance. Compared with global baseline value deviation and the original observation data Perform linear correction, the expression is: ； In the formula, This is the first-corrected reflectance after systematic bias correction. For the original satellite observation reflectivity, For the first The systematic bias constant of each observation group; The reflectivity residuals are binned and statistically analyzed according to the antenna line-of-sight angle and azimuth angle, and a two-dimensional correction lookup table is constructed to eliminate the observation errors introduced by satellite attitude and antenna gain. The geometric correction specifically includes: adjusting the corrected data... According to the antenna line of sight and azimuth Perform gridded binning, and construct a two-dimensional lookup table by calculating the mean residual value within each angle bin. Furthermore, the nonlinear deviation introduced by the satellite antenna gain model error is eliminated by a lookup table method, the expression of which is: ； In the formula, For the final generated normalized reflectance, and These are the antenna line-of-sight angle and azimuth angle corresponding to the observation point, respectively. This is a function for a two-dimensional residual lookup table.

7. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S2, the spatiotemporal fusion and gap filling of multi-source data includes: The multi-source heterogeneous scatter data, after cascaded correction, is projected onto a regular geographic grid of the sea area. For the spatiotemporal gaps still existing after projection fusion, an improved spatiotemporal interpolation algorithm based on historical observation behavior is used to numerically complete the missing grid using the historical statistical characteristics of that location. The specific steps are as follows: (a) High-dimensional extraction of historical statistical features: for any grid point to be completed Backtrack the valid observation set for the same historical period at this location; calculate the historical mean of this set. As a climatological baseline, historical variance is calculated. To assess uncertainty, the slope of the reflectance trend as a function of year is extracted using the least squares method. Construct historical reference values ​​that include long-term trend corrections. The expression is: ； In the formula, This represents the time offset of the current moment relative to the center of the historical window. (b) Spatial weight construction based on terrain similarity: utilizing surrounding effective observation grids Treating the completion points When performing spatial interpolation, calculate the spatial correlation weights. The expression is: ； In the formula, For grid With grid The geometric Euclidean distance between them For grid Relative to grid The terrain similarity weight; A terrain similarity factor is introduced, which is calculated exponentially based on the differences between the grid to be completed and the surrounding grids in terms of distance from the shore and water depth. The weight calculation formula is defined as follows: ； In the formula, and They are grids and grid The distance from the shore, and They are grids and grid The water depth, and These are the characteristic attenuation scales for distance from the shore and water depth, respectively; (c) Spatiotemporal joint dynamic completion: The calculated historical reference values ​​are used to complete the calculation. Spatial interpolation results based on terrain weights Weighted fusion is performed to obtain the final complete value; by introducing historical trends and topographic constraints, a normalized reflectivity field with continuous spatiotemporal distribution and conforming to the physical distribution law of sea ice is generated.

8. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S3, the physical model construction specifically includes: Based on the theory of electromagnetic wave propagation in layered media, a coherent reflection physical model suitable for thin ice environments is constructed; the detection scene is abstracted into a layered structure of air, sea ice and seawater, and a nonlinear forward mapping function between sea ice thickness and surface reflectivity is established as a physical mechanism constraint equation. Equivalent reflection coefficient With sea ice thickness The relation is: ； In the formula, For sea ice thickness, These are the Fresnel reflection coefficients of electromagnetic waves at the air-sea ice interface and the sea ice-seawater interface, respectively. The vertical wavenumber is the signal propagation frequency within the sea ice medium. This is the roughness attenuation factor.

9. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 1, characterized in that, In step S4, the physical-data coupling inversion specifically includes: A coupled architecture consisting of an inversion master network and a physical mapping module is constructed. The inversion master network uses normalized reflectance maps to predict sea ice thickness. The physical mapping module, based on a physical model, maps the predicted thickness values ​​back to theoretical reflectance. By minimizing the difference between observed and theoretical reflectance, a physical consistency loss function is constructed to force the network to converge in the solution space under the constraints of physical laws.

10. The sea ice thickness inversion method based on multi-source heterogeneous fusion and physical constraints according to claim 9, characterized in that, The physical consistency loss function is defined as: ； ； In the formula, For the total loss function, To account for the mean square error loss between the inverted thickness and the measured true value, These are the weighting coefficients for the physical constraint terms. Loss due to physical constraints, This represents the number of samples in the training batch. Normalized observed reflectance, The sea ice thickness predicted by the network; To predict the thickness of the network The theoretical reflectivity is calculated by substituting it into the physical model.

Images