Sea ice concentration retrieval method and system based on physical information guidance, and medium

Abstract

The application discloses a sea ice density inversion method and system based on physical information guidance and a medium. The method comprises the following steps: obtaining a corresponding spatial sea ice density map by using a pre-trained sea ice density inversion model; performing feature extraction of microwave scattering maps, microwave radiation maps and meteorological element maps in multiple scales; for each scale of microwave scattering features, microwave radiation features and meteorological element features, guiding microwave radiation feature correction by using meteorological element features and microwave scattering features to realize physical information guided cross-modal feature fusion; extracting spatial information maps and decoding the fused latent feature maps step by step to obtain corresponding spatial sea ice density maps. The application aims to fully model the physical dependence relationship among radiation, scattering and environmental driving factors, realize the spatio-temporal collaboration among features, improve the ability of learning the continuity and structural properties of sea ice distribution and the physical consistency and spatial generalization ability of sea ice density inversion results.

Description

Technical Field

[0001] This invention relates to the field of sea ice concentration detection technology, specifically to a sea ice concentration inversion method, system, and medium based on physical information guidance. Background Technology

[0002] Sea ice concentration is a crucial parameter describing the state of sea ice in a given body of water, defined as the proportion of sea ice area within a specific region. It is a primary parameter in passive sea ice remote sensing and a vital input and boundary condition for various numerical weather prediction models. Therefore, accurately obtaining sea ice concentration is essential for monitoring Arctic sea ice conditions and ensuring safe navigation in the Arctic region. Currently, sea ice concentration is primarily obtained through several methods: on-site observations from fixed coastal stations, drifting ice stations, and ships; satellite-based remote sensing; and ice condition maps created by ice condition analysts according to professional technical standards. Among these, satellite remote sensing is the most widely used method due to its extensive spatial coverage and consistent temporal monitoring capabilities. Satellite remote sensing technology can be broadly categorized into optical remote sensing and microwave remote sensing, the latter further divided into active and passive microwave systems. Optical remote sensing inverts sea ice concentration by utilizing the difference in spectral reflectance between sea ice and open water, with ice exhibiting higher albedo. This difference allows for the use of thresholding and lattice-based algorithms for sea ice concentration inversion. However, its strong dependence on weather conditions limits its reliable use under cloud cover and polar night conditions. Passive microwave remote sensing has become the primary technology for monitoring Arctic sea ice concentration due to its all-weather capability. The most widely used algorithms for retrieving sea ice concentration using microwave brightness temperature (SIC) data include the NASATeam (NT) algorithm, the Bootstrap algorithm, the NT2 algorithm, and the ASI algorithm. The NT and Bootstrap algorithms are primarily based on low-frequency channels (e.g., 19 GHz and 37 GHz), while the NT2 and ASI algorithms utilize higher-frequency channels (e.g., 89 GHz). Each type of algorithm exhibits different performance characteristics and limitations. For example, while low-frequency channels are less sensitive to atmospheric interference, they offer limited spatial resolution (typically 25 km), which restricts the ability of related algorithms to resolve small-scale sea ice features. Conversely, high-frequency channels offer improved spatial resolution (approximately 12.5 km), but are more susceptible to atmospheric influences such as water vapor, cloud liquid water, and surface winds, thus introducing greater uncertainty to SIC retrieval under complex meteorological conditions. As an active microwave remote sensing instrument, synthetic aperture radar (SAR) can provide high-resolution sea ice images and reveal detailed structural information, even during complex ice conditions and melting periods. Due to their different scattering characteristics, water and ice typically produce distinct backscattered signals, enabling SAR to distinguish them. Therefore, existing techniques have incorporated convolutional neural networks (CNNs) into sea ice concentration inversion tasks, directly retrieving sea ice concentration from dual-polarization SAR images using an end-to-end learning approach, and employing manually interpreted ice maps as labels, laying the methodological foundation for future research. Existing techniques also disclose the use of sea ice concentration products retrieved from passive microwave radiometers as training labels to overcome the scarcity of manually labeled data, significantly enhancing the scalability and automation potential of the method. However, most existing inversion methods are based on the traditional U-Net architecture for building their network structures.These methods simply stitch together features from different sensors, failing to adequately model the physical dependencies between radiation, scattering, and environmental driving factors. They also ignore the spatiotemporal co-structure between features, limiting the physical consistency and spatial generalization ability of the inversion results. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method, system and medium for sea ice concentration inversion based on physical information, addressing the above-mentioned problems in the prior art. This invention aims to fully model the physical dependencies between radiation, scattering and environmental driving factors, achieve spatiotemporal synergy between features, improve the ability to learn the continuity and structural properties of sea ice distribution, as well as the physical consistency and spatial generalization ability of sea ice concentration inversion results, and generate more detailed and physically more realistic outputs that reflect the continuous transition between open water, thin ice and solidified ice.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A physically-guided sea ice concentration inversion method includes using a pre-trained sea ice concentration inversion model (PIMS-Net) to obtain the corresponding spatial sea ice concentration map from the microwave scattering map, microwave radiation map, and meteorological element map of the area to be detected.

[0006] S101, features are extracted from microwave scattering map, microwave radiation map and meteorological element map respectively to obtain microwave scattering features, microwave radiation features and meteorological element features at multiple scales;

[0007] S102, for each scale of microwave scattering characteristics, microwave radiation characteristics and meteorological element characteristics, the microwave radiation characteristics are guided by the meteorological element characteristics and microwave scattering characteristics to achieve cross-modal feature fusion guided by physical information.

[0008] S103, extract spatial information map from the fusion latent feature map obtained by cross-modal feature fusion at the last scale;

[0009] S104. The extracted spatial information map is sequentially connected with the fused potential feature map obtained by fusing cross-modal features of other scales and then decoded to obtain the final spatial sea ice concentration map.

[0010] Optionally, in step S101, feature extraction of the microwave scattering map, microwave radiation map, and meteorological element map to obtain microwave scattering features, microwave radiation features, and meteorological element features at multiple scales is achieved through three encoders, including: encoders It is used to extract microwave scattering features at multiple scales from microwave scattering maps; encoder It is used to extract microwave radiation features at multiple scales from microwave radiation maps; encoder It is used to extract meteorological feature characteristics at multiple scales from meteorological feature maps; the three encoders each include multiple coding layers connected in sequence, and a downsampling module is provided between adjacent coding layers.

[0011] Optionally, in step S102, the microwave radiation feature correction is guided by meteorological element features and microwave scattering features to achieve physical information-guided cross-modal feature fusion through a cross-modal convolutional fusion block. Implement the cross-modal convolutional fusion block Processing of microwave scattering characteristics, microwave radiation characteristics, and meteorological element characteristics includes:

[0012] S201 combines microwave scattering features, microwave radiation features, and meteorological element features of the same scale along the channel dimension to form a composite feature map;

[0013] S202 passes the aggregated feature map through a 3×3 convolutional layer, followed by batch normalization and activation function activation to facilitate nonlinear transformation and structural recombination to obtain a recombined feature map;

[0014] S203 fuses the reconstructed feature map with the original input cross-modal convolutional block via residual connections. The microwave radiation features are stitched together along the channel dimension to obtain a fused potential feature map.

[0015] Optionally, extracting the spatial information map from the fused latent feature map obtained by fusing cross-modal features at the last scale in step S103 refers to passing the fused latent feature map obtained by fusing cross-modal features at the last scale through a bidirectional convolutional long short-term memory network module. Capture spatial information maps of sea ice concentration fields.

[0016] Optionally, the bidirectional convolutional long short-term memory network module The spatial information map captured in the sea ice concentration field includes: S301, which is the fused latent feature map at the last scale obtained by cross-modal feature fusion. Apply a sliding window operation to divide it into a series of overlapping tiles in the horizontal and vertical directions, using a fixed tile size. To generate two ordered, containing From the sequence of tiles, we obtain the horizontal tile sequence. and vertical tile sequence ;

[0017] S302, Transform the horizontal sequence and vertical sequence Treating it as a pseudo-time series, we used a bidirectional convolutional long short-term memory network to capture the spatial continuity in the sea ice concentration field to obtain an enhanced horizontal patch sequence. and vertical tile sequence ;

[0018] S303, the enhanced horizontal tile sequence and vertical tile sequence After being spatially reassembled through a tiling operation, the data is input into the boundary fusion layer. The boundary fusion layer first performs feature reorganization and boundary information aggregation in the local neighborhood through a 3×3 convolution that keeps the channel dimension unchanged. Then, it uses two-dimensional batch normalization to stabilize the feature distribution. Finally, it uses ReLU nonlinear activation to enhance the representation capability and introduces sparsity to improve the discriminability of boundary details and the fusion effect.

[0019] S304. The fused feature map obtained after boundary fusion layer processing is restored to the original spatial layout using 1×1 convolution, thereby obtaining the spatial information map in the sea ice concentration field. Spatial information map of the sea ice concentration field and fusion of latent feature maps Both have the same spatial resolution.

[0020] Optionally, in step S104, when concatenating and decoding the extracted spatial information map with the fused latent feature map obtained by fusing the other scales of cross-modal features to obtain the final spatial sea ice concentration map, this includes passing the fused latent feature map obtained by fusing the extracted spatial information map with the other scales of cross-modal features through a decoder network. The decoder network performs hierarchical fusion, connection, and decoding to obtain the corresponding spatial sea ice concentration map. It includes multiple decoding layers. Each decoding layer has a connection module and an upsampling module at its input. The connection module is used to concatenate the fused latent feature map or spatial information map with the feature map output by the upsampling module in the channel dimension and use it as the input of the decoding layer. The input feature of the upsampling module is the output feature of the previous decoder or the spatial information map, and the last decoder layer outputs a spatial sea ice concentration map.

[0021] Optionally, the microwave scattering map is a C-band synthetic aperture radar image, which includes the backscattering coefficients of two channels: horizontal transmission and horizontal reception polarization (HH) and horizontal transmission and vertical reception polarization (HV). The microwave radiation map is a multi-band microwave radiometer data map, and each band includes two channels: horizontal polarization mode H and vertical polarization mode V. The meteorological element map is a numerical weather prediction data distribution image from the reanalysis data of the weather forecast center, which includes the total cloud liquid water content (tclw), surface radiation temperature (skt), and the rotated 10-meter height northerly wind component (v10m_rotated). Before step S101, the microwave scattering map, microwave radiation map, and meteorological element map are resized to achieve the same resolution, and then the microwave scattering map, microwave radiation map, and meteorological element map with the same resolution are spatially aligned.

[0022] Optionally, the loss function used during training of the sea ice concentration inversion model is:

[0023] ;

[0024] in, For loss function, For the sample size, and These are the height and width of the image, respectively. For the first In the nth sample Weight of each pixel and The first In the nth sample The predicted pixel value and label value of the nth pixel; where the nth pixel is the predicted pixel value and label value. In the nth sample The calculation of the weight of each pixel includes:

[0025] The first In the nth sample Label value of each pixel Equal width compartments:

[0026] ;

[0027] in, For the first A set of boxes; K is the number of boxes;

[0028] Calculate the training set at the th Pixel frequency of each sub-box :

[0029] ;

[0030] in, This is an indicator function; the value is 1 if the condition within the parentheses is met, and 0 otherwise.

[0031] Calculate the first In the nth sample Weights of each pixel:

[0032] ;

[0033] in, For binning indexing functions, A constant used to prevent division by zero in the denominator. It is an adjustable exponent, and has the following properties:

[0034] ;

[0035] in, Indicates to Round down.

[0036] The present invention also provides a physics-guided sea ice concentration inversion system, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the physics-guided sea ice concentration inversion method.

[0037] The present invention also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the physical information-guided sea ice concentration inversion method by a processor.

[0038] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The sea ice concentration inversion method based on physical information of the present invention includes using a pre-trained sea ice concentration inversion model to obtain the corresponding spatial sea ice concentration map; extracting features at multiple scales from microwave scattering maps, microwave radiation maps, and meteorological element maps; for each scale of microwave scattering features, microwave radiation features, and meteorological element features, using meteorological element features and microwave scattering features to guide microwave radiation feature correction to achieve physical information-guided cross-modal feature fusion; decoding the fused potential feature map obtained from cross-modal feature fusion to obtain the corresponding spatial sea ice concentration map. The present invention can fully model the physical dependencies between radiation, scattering, and environmental driving factors, achieve spatiotemporal synergy between features, improve the ability to learn the continuity and structural properties of sea ice distribution, and enhance the physical consistency and spatial generalization ability of sea ice concentration inversion results. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0040] Figure 2This is a schematic diagram of the network structure of the sea ice concentration inversion model (PIMS-Net) in an embodiment of the present invention.

[0041] Figure 3 This is a schematic diagram of the network structure of the cross-modal convolutional fusion block in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram of the network structure of the bidirectional convolutional long short-term memory network module in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram showing the ranking of the relative contributions of various features to the inversion performance in an embodiment of the present invention.

[0044] Figure 6 This is a box plot of the mean square error (MSE) for seven sea ice density ranges in an embodiment of the present invention.

[0045] Figure 7 This shows the distribution of sea ice concentration samples in each dataset in this embodiment of the invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0047] like Figure 1 As shown, the physical information-guided sea ice concentration inversion method in this embodiment includes using a pre-trained sea ice concentration inversion model (PIMS-Net) to obtain the corresponding spatial sea ice concentration map based on the microwave scattering map, microwave radiation map, and meteorological element map of the area to be detected.

[0048] S101, features are extracted from microwave scattering map, microwave radiation map and meteorological element map respectively to obtain microwave scattering features, microwave radiation features and meteorological element features at multiple scales;

[0049] S102, for each scale of microwave scattering characteristics, microwave radiation characteristics and meteorological element characteristics, the microwave radiation characteristics are guided by the meteorological element characteristics and microwave scattering characteristics to achieve cross-modal feature fusion guided by physical information.

[0050] S103, extract spatial information map from the fusion latent feature map obtained by cross-modal feature fusion at the last scale;

[0051] S104. The extracted spatial information map is sequentially connected with the fused potential feature map obtained by fusing cross-modal features of other scales and then decoded to obtain the final spatial sea ice concentration map.

[0052] To address memory limitations and facilitate efficient model optimization, the microwave scattering map, microwave radiation map, and meteorological element map of the detected area were further divided into 512×512 pixel tiles.

[0053] In this embodiment, the microwave scattering pattern can be defined as: The microwave scattering map is a C-band synthetic aperture radar (SAR) image, which includes the backscattering coefficients of two channels: horizontally transmitted horizontally received polarimetric (HH) and horizontally transmitted vertically received polarimetric (HV). Specifically, the Sentinel-1 ultrawide swath mode medium-resolution Level-1 ground range detection (GRDM) product is used, with each image covering an area of ​​400×400 square kilometers and an initial pixel spacing of 40×40 meters (downsampled to 80 meters in the pre-trained version). The backscattering values ​​have been calibrated from [-30, +10] and converted to a normalized linear scale of [-1, 1]. Since the ultrawide swath mode SAR image is generated by combining five subbands in the azimuth direction, slight radiometric inconsistencies may exist between adjacent subbands, especially in the HV channel. Therefore, the Sentinel-1 SAR noise correction algorithm can be used for SAR noise correction.

[0054] In this embodiment, the microwave radiation pattern can be defined as: ,in Indicates spatial dimension, The channel number is indicated. The microwave radiometer map is a multi-band microwave radiometer data map, and each band includes two channels: horizontal polarization mode H and vertical polarization mode V. In this embodiment, the AMSR2 microwave radiometer dataset from Sentinel-1 is used, which contains microwave radiometer data that has been spatially aligned to the geometry of Sentinel-1. This microwave radiometer data covers seven different frequency bands, each including two channels: horizontal polarization mode H and vertical polarization mode V; therefore, the number of channels is 14. To ensure temporal and spatial correspondence with a given SAR image, the strip selection window for the AMSR2 microwave radiometer data is ±3.5 hours before and after the strip acquisition time of the Sentinel-1 SAR image to ensure geographic overlap. These selected AMSR2 microwave radiometer data are then resampled using Gaussian weighted interpolation onto a reduced-resolution Sentinel-1 pixel grid to obtain the AMSR2 microwave radiometer dataset with a spatial resolution of 2 kilometers.

[0055] In this embodiment, the meteorological element map can be defined as: ,Include Several environmental parameters. The meteorological element map is a distribution image of numerical weather forecast data from the reanalysis data of the weather forecast center, including total cloud liquid water content (tclw), surface radiation temperature (skt), and the rotated 10-meter height northerly wind component (v10m_rotated), i.e.: The meteorological element map uses numerical weather prediction data provided by the Weather Forecasting Center. It is spatially resampled using Gaussian weighted interpolation to match the geometry of Sentinel-1, using the same method as the AMSR2 microwave radiometer dataset, and the pixel spacing is the same as the AMSR2 microwave radiometer dataset.

[0056] Before step S101, the microwave scattering map, microwave radiation map, and meteorological element map are resized to achieve the same resolution, and then spatially aligned. When using the pre-trained sea ice concentration inversion model (PIMS-Net) to obtain the corresponding spatial sea ice concentration map of the detected area, all data from the microwave scattering map, microwave radiation map, and meteorological element map are spatially aligned and stitched along the channel dimension to form the final input tensor. ,in This represents the total number of samples, with each sample containing three data sources.

[0057] In this embodiment, the sea ice concentration inversion model (PIMS-Net) aims to learn from multiple input sources. The mapping is used to estimate the optimal sea ice concentration prediction. Each of them This represents a spatial sea ice concentration map, approximating the corresponding ground truth label. During the training phase, the corresponding sea ice concentration label is represented as... This dataset is derived from ice condition maps. Ice condition maps are created by sea ice analysts using SAR images supplemented by other multi-source satellite data. The SAR images are segmented into a series of visually similar sub-regions, and each sub-region is labeled with a corresponding single code, providing high-resolution sea ice information. This dataset includes ice condition maps produced by Greenland and the Canadian Ice Service. The main information contained in the ice condition maps includes sea ice concentration, ice floe size, and sea ice development status. This study uses its sea ice concentration data as the ground truth label for the model. In the ice condition maps, sea ice concentration is divided into 11 categories (0-10), where 0 and 10 represent pure water and pure ice, respectively, 1 represents a sea ice concentration of 10%, and so on. The resolution of the ice condition maps is the same as that of the SAR images.

[0058] To accurately retrieve high-resolution sea ice concentration from multi-source observations, this embodiment proposes a deep learning framework based on physical information guidance and spatial continuity perception for the sea ice concentration retrieval model (PIMS-Net). In step S101, feature extraction is performed on the microwave scattering map, microwave radiation map, and meteorological element map to obtain microwave scattering features, microwave radiation features, and meteorological element features at multiple scales. This is achieved through three encoders, including: encoders... It is used to extract microwave scattering features at multiple scales from microwave scattering maps; encoder It is used to extract microwave radiation features at multiple scales from microwave radiation maps; encoder This is used to extract meteorological feature characteristics at multiple scales from meteorological feature maps; each of the three encoders includes multiple sequentially connected coding layers, with downsampling modules between adjacent coding layers. The coding process of each modality's data source across multiple coding layers can be represented as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] in, Indicates the corresponding data source in the th Layer fusion latent feature map, This represents the deepest fusion potential feature map. The points in the subscripts take values ​​of a, s, or m, representing three modes: microwave scattering features, microwave radiation features, and meteorological element features, respectively. The subscript i is the sample index. For example... Figure 2 As shown, in this embodiment, each of the three encoders includes five sequentially connected coding layers. The five coding layers are denoted as: Encoder #1 Layer 1 to Encoder #1 Layer 5. The five coding layers are denoted as: Encoder #2 Layer 1 to Encoder #2 Layer 5. The five coding layers are denoted as: Encoder #3 Layer 1 to Encoder #3 Layer 5.

[0063] In this embodiment, step S102 utilizes meteorological element features and microwave scattering features to guide microwave radiation feature correction, thereby achieving physical information-guided cross-modal feature fusion through a cross-modal convolutional fusion block. accomplish, Figure 2 The five coding layers corresponding to the three encoders are configured with five cross-modal convolutional fusion blocks. Cross-modal convolutional fusion block The ability to integrate information from all data sources can be represented functionally as follows:

[0064] ;

[0065] in, For the first Cross-modal convolutional fusion blocks of layers The resulting fusion latent feature map.

[0066] like Figure 3 As shown, in this embodiment, the cross-modal convolutional fusion block Processing of microwave scattering characteristics, microwave radiation characteristics, and meteorological element characteristics includes:

[0067] S201 combines microwave scattering features, microwave radiation features, and meteorological element features of the same scale along the channel dimension to form a composite feature map;

[0068] S202 passes the aggregated feature map through a 3×3 convolutional layer, followed by batch normalization and nonlinear activation to facilitate nonlinear transformation and structural recombination to obtain a recombined feature map;

[0069] S203 fuses the reconstructed feature map with the original input cross-modal convolutional block via residual connections. The microwave radiation features are stitched together along the channel dimension to obtain a fused potential feature map.

[0070] In this embodiment, the cross-modal convolutional fusion block When processing microwave scattering features, microwave radiation features, and meteorological element features, a physics-guided feature correction mechanism is introduced. This mechanism uses meteorological element features and microwave scattering features to guide the correction of microwave radiation features, thereby achieving representation learning that is more in line with physical principles. Specifically, at different feature scales, the features of the three types of data are first concatenated along the channel dimension to form an aggregated feature map. This composite feature map is then passed through a 3×3 convolutional layer, followed by batch normalization and activation functions to promote nonlinear transformation and structural reorganization. The resulting output is fused with the original microwave radiation features through residual connections to produce enhanced fused features. This architecture integrates cross-modal physical constraints while preserving the integrity of the initial information, thereby improving the robustness and interpretability of the learned feature representation.

[0071] In this embodiment, step S103, extracting the spatial information map from the fused latent feature map obtained by fusing cross-modal features at the last scale, refers to passing the fused latent feature map obtained by fusing cross-modal features at the last scale through a bidirectional convolutional long short-term memory network module. A spatial information map capturing the sea ice concentration field can be represented as:

[0072] ;

[0073] in, This is a spatial information map of the sea ice concentration field. It is a bidirectional convolutional long short-term memory network module. For the first The fused latent feature map output by the deepest layer is the fused latent feature map obtained by fusing cross-modal features at the last scale. Long Short-Term Memory (LSTM) networks are widely used to model sequence dependencies in time-series data. Convolutional Long Short-Term Memory (ConvLSTM) networks extend the standard LSTM by replacing the fully connected operations in the standard LSTM with convolutional operations, enabling it to effectively capture the spatiotemporal dynamics of structured data (such as video or weather sequences). Inspired by this capability, this embodiment employs a bidirectional convolutional long short-term memory network module that includes a bidirectional convolutional long short-term memory network (Bi-ConvLSTM). To enhance spatial continuity modeling in sea ice concentration fields, a bidirectional convolutional long short-term memory (Bi-ConvLSTM) network is employed. While ConvLSTM was originally designed for time series, this embodiment repurposes it for modeling spatial dependencies by treating latent feature maps as ordered spatial sequences. Specifically, this embodiment crops a large-scale remote sensing image into a series of sub-images and inputs them into two ConvLSTM modules, one horizontally and one vertically, to model contextual dependencies. This allows the network to aggregate spatial context from different regions, thereby enhancing the smoothness and coherence of the final sea ice concentration inversion. Figure 4 As shown, this embodiment utilizes a bidirectional convolutional long short-term memory network module. The spatial information map capturing the sea ice concentration field includes:

[0074] S301, for cross-modal feature fusion, obtain the fused latent feature map at the last scale. Apply a sliding window operation to divide it into a series of overlapping tiles in the horizontal and vertical directions, using a fixed tile size. To generate two ordered, containing From the sequence of tiles, we obtain the horizontal tile sequence. and vertical tile sequence ;

[0075] S302, Transform the horizontal sequence and vertical sequence Treating it as a pseudo-time series, we used a bidirectional convolutional long short-term memory network to capture the spatial continuity in the sea ice concentration field to obtain an enhanced horizontal patch sequence. and vertical tile sequence ;

[0076] S303, the enhanced horizontal tile sequence and vertical tile sequence After being spatially reassembled through a tiling operation, the data is input into the boundary fusion layer. The boundary fusion layer first performs feature reorganization and boundary information aggregation in the local neighborhood through a 3×3 convolution that keeps the channel dimension unchanged. Then, it uses two-dimensional batch normalization to stabilize the feature distribution. Finally, it uses ReLU nonlinear activation to enhance the representation capability and introduces sparsity to improve the discriminability of boundary details and the fusion effect.

[0077] S304: After processing the boundary fusion layer, the fused feature maps are stitched together and then a 1×1 convolution is used to restore the original spatial layout, thus obtaining the spatial information map in the sea ice concentration field. Spatial information map of the sea ice concentration field and fusion of latent feature maps Both have the same spatial resolution.

[0078] set up This indicates the number of layers extracted from the last layer of the encoder. The fused latent feature map of each sample, i.e., the fused latent feature map obtained by fusing cross-modal features at the last scale. To achieve spatial continuity perception of sea ice concentration, this embodiment... Apply a sliding window operation to divide it into a series of overlapping tiles in the horizontal and vertical directions, using a fixed tile size. To generate two ordered, containing A sequence of tiles:

[0079] ;

[0080] ;

[0081] ;

[0082] Each tile The local spatial structure and characteristic channels are preserved. To model long-range dependencies within the spatial field, this embodiment will... and Treat it as a pseudo-time series and input it into Bi-ConvLSTM:

[0083] ;

[0084] ;

[0085] Bi-ConvLSTM processes sequences in both the forward and reverse directions to capture contextual dependencies between adjacent tiles. Let This indicates that after incorporating both forward and reverse space dependencies, the first... The updated feature representation of each patch. The internal computation of a unidirectional ConvLSTM unit follows a standard formula:

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] in, For input gate output, Output for the forget gate. In cellular state, For output gate output, In hidden state, For the first The updated feature representation of each patch, Input weights into the input gate. Input weights for the forget gate. Input weights for candidate cell states. Input weights to the output gate, The hidden state weights of the input gate. Hidden state weights for the forget gate The hidden state weights are used to represent the candidate cell state. The hidden state weights of the output gate. For input gate bias, To offset the forget gate, Bias for candidate cell states. For output gate bias, This represents the cell state at the previous moment. This is the hidden state from the previous moment. It is the sigmoid activation function. For convolution operations, For element-wise multiplication, It is the hyperbolic tangent activation function.

[0092] In Bi-ConvLSTM, the forward and reverse sequences are processed separately, and then aggregated by concatenation:

[0093] ;

[0094] here, and Representing time steps The hidden state passed by ConvLSTM in both forward and reverse directions.

[0095] Once all tiles in the sequence have been processed, this embodiment yields two enhanced tile sequences. and Finally, the two enhanced sequences are spatially reassembled through a tiling operation and then processed by a boundary fusion layer. A 1×1 convolution is then used to restore the feature map to its original spatial layout. This yields a refined representation: a spatial information map of the sea ice concentration field. It maintains the same spatial resolution as when inputting the Bi-ConvLSTM, and is then input to the decoder for high-resolution sea ice concentration reconstruction.

[0096] In step S104 of this embodiment, when the extracted spatial information map is sequentially connected with the fused latent feature map obtained by fusing the cross-modal features of other scales and then decoded to obtain the final spatial sea ice concentration map, this includes passing the fused latent feature map obtained by fusing the extracted spatial information map with the cross-modal features of other scales through a decoder network. The decoder network performs hierarchical fusion, connection, and decoding to obtain the corresponding spatial sea ice concentration map. Includes multiple decoding layers, for example Figure 2 There are five decoding layers, denoted as Decoder Layer 1 to Decoder Layer 5. Each decoding layer has a connection module and an upsampling module at its input. The connection module is used to concatenate the fused latent feature map or spatial information map with the feature map output from the upsampling module along the channel dimension, and use the result as the input to the decoding layer. The input features of the upsampling module are the output features or spatial information map of the previous decoder, and the last decoder layer outputs a spatial sea ice concentration map, which can be represented as:

[0097] ;

[0098] in, This is the spatial sea ice concentration map for sample i. Therefore, the overall forward propagation process of the sea ice concentration inversion model (PIMS-Net) in this embodiment can be formalized as follows:

[0099] ;

[0100] here, , and There are three encoders respectively. , and These are microwave scattering map, microwave radiation map, and meteorological element map, respectively. , and These are the fused latent feature maps obtained by fusing cross-modal features at the last scale from the three encoders. For cross-modal convolutional fusion blocks, , and The three encoders are respectively The fused latent feature map obtained by fusing cross-modal features at various scales. It is a bidirectional convolutional long short-term memory network module. For decoder networks, These are the training parameters for the three encoders. These represent cross-modal convolutional fusion modules. Bidirectional convolutional long short-term memory network module and decoder network The training parameters.

[0101] Considering the imbalance of sea ice concentration samples and the complexity of sea ice edge changes, the sea ice concentration inversion model (PIMS-Net) in this embodiment uses weighted mean square error (MSE) as the loss function:

[0102] ;

[0103] in, For loss function, For the sample size, and These are the height and width of the image, respectively. For the first In the nth sample Weight of each pixel and The first In the nth sample The predicted pixel value and label value (representing the predicted and actual values ​​of sea ice density) of each pixel, and have As an optional implementation, in this embodiment, the first... In the nth sample The weights of each pixel are defined based on the statistical distribution of different densities in the training set, with greater weights assigned to edge regions so that the model pays more attention to these areas. The calculation includes: s

[0104] (1) The first In the nth sample Label value of each pixel Equal width compartments:

[0105] ;

[0106] in, For the first The set of sub-boxes; K is the number of sub-boxes, which is 10 in this embodiment.

[0107] (2) Calculate the training set at the th Pixel frequency of each sub-box :

[0108] ;

[0109] in, This is an indicator function; the value is 1 if the condition within the parentheses is met, and 0 otherwise.

[0110] (3) Calculate the first In the nth sample Weights of each pixel:

[0111] ;

[0112] in, For binning indexing functions, A constant used to prevent division by zero in the denominator. It is an adjustable exponent, and has the following properties:

[0113] ;

[0114] in, Indicates to Round down.

[0115] To verify the performance of the physics-guided sea ice concentration inversion method (Sea Ice Concentration Inversion Model (PIMS-Net)) in this embodiment, the experiments of the Sea Ice Concentration Inversion Model (PIMS-Net) in this embodiment were carried out using PyTorch. The model optimizer was Adam, the initial learning rate was set to 0.0001, and a cosine annealing learning rate adjustment strategy was adopted. The batch size was set to 16, and the training period was set to 50.

[0116] This embodiment selects the U-Net model as the baseline model to evaluate the performance difference in sea ice concentration inversion between the results obtained from the U-Net model and the sea ice concentration inversion model (PIMS-Net) of this embodiment. Mean squared error (MSE) and coefficient of determination (COP) are selected. As an evaluation indicator:

[0117] ;

[0118] ;

[0119] in, Indicates the first In the first scenario The predicted sea ice concentration estimate per pixel, This represents the actual label value of the same pixel. This represents the average value of the true labels for all pixels. The final results are shown in Table 1.

[0120] Table 1: Training Model and Performance Summary

[0121]

[0122] Table 1 lists the highest performance achieved by the U-Net model and the sea ice concentration inversion model (PIMS-Net) of this embodiment in the test scenario for predicting sea ice concentration. Table 1 shows the highest performance achieved by the two architectures in the test scenario. According to the objective evaluation metrics in Table 1, the sea ice concentration inversion model (PIMS-Net) of this embodiment outperforms the U-Net model in the sea ice concentration inversion task, with a 2.2% improvement in the coefficient of determination and a 33.7% reduction in the mean squared error.

[0123] To quantitatively analyze the relative contributions of each of the three types of features—microwave scattering characteristics, microwave radiation characteristics, and meteorological element characteristics—to the inversion performance used in this embodiment, an importance ranking analysis was conducted, and the results are as follows: Figure 5 As shown. Figure 5 In the diagram, "AMSR2 brightness temperature" represents microwave radiation characteristics, "SAR backscattering" represents microwave scattering characteristics, and "ERA5 meteorological elements" represents meteorological element characteristics; SAR_HH and SAR_HV represent the two channels, horizontal transmission and horizontal reception polarization HH and horizontal transmission and vertical reception polarization HV, respectively; BT_10.7V represents the vertical polarization mode V of the 10.7GHz band of the microwave radiation map, BT_6.9H represents the horizontal polarization mode H of the 6.9GHz band of the microwave radiation map, and so on for other bands; tclw is the total cloud liquid water content, skt is the surface radiation temperature, and v10m_rotated is the northerly wind component at a height of 10 meters after rotation. Figure 5 The results show that, among the microwave radiation features, the vertical polarization brightness temperatures from the microwave radiometer at 10.7 GHz and 18.7 GHz have the greatest impact on the model's predictions. This is followed by the backscattering coefficients from the two SAR polarization modes, which also contribute significantly. In contrast, meteorological features have the lowest importance, indicating that their impact on the sea ice concentration inversion model (PIMS-Net) output is more limited.

[0124] To quantitatively analyze the accuracy of the sea ice concentration inversion model (PIMS-Net) in predicting the seasonal inversion of sea ice concentration in this embodiment, the monthly performance of the sea ice concentration inversion model (PIMS-Net) and the U-Net model was tested in this embodiment, and the results are shown in Table 2.

[0125] Table 2: Monthly performance comparison of the sea ice concentration inversion model (PIMS-Net) and the U-Net model

[0126]

[0127] Table 2 shows that the sea ice concentration inversion model (PIMS-Net model) consistently outperformed the U-Net model in all months, demonstrating its improved accuracy in sea ice concentration inversion. Both models exhibited similar seasonal patterns, with optimal performance in winter and declining performance during the summer melting season. The PIMS-Net model peaked its performance in January (0.973), February (0.961), and March (0.953), indicating robust performance under extensive ice cover. A secondary peak was observed in November (0.955), corresponding to the early freezing period. The performance decline observed in December (significant in both the PIMS-Net and U-Net models) likely marks a period of increased inversion difficulty during the early to mid-freezing phase. This period involves complex processes influenced by wind and ocean currents, such as ice ridges and interglacial channel formation, leading to high spatial heterogeneity in ice thickness and concentration. The recovery in January suggests that once a stable, solidified winter ice sheet is established, sea ice conditions become easier for the model to predict. Notably, the performance advantage of the PIMS-Net model is most pronounced in the mid-to-late melting season. Although the accuracy of both models declines from June to September, the PIMS-Net model maintains a significantly higher coefficient of determination value—particularly in August (0.827 vs 0.566) and September (0.791 vs 0.577)—indicating its enhanced ability to capture the complex dynamics of active melting and minimum ice periods. In July, the PIMS-Net model also achieved a much higher R² (0.827) than the U-Net model (0.742). These results demonstrate that the PIMS-Net model not only maintains strong performance under stable freezing conditions but also provides significant improvements in simulating seasonal transitions and melting-related processes, while the traditional U-Net model exhibits clear limitations in these areas. To intuitively evaluate the performance of the proposed method in sea ice concentration estimation across different seasons, a challenging scenario was specifically selected from the test dataset for comparison. The results clearly show that the model performs better in winter when sea ice conditions are stable. Conversely, during the melting season, the model performs poorly, with turbulent currents mixing ice and water to create swirling textures. Notably, when the scene includes large areas of low sea ice concentration, the PIMS-Net model produces relatively better results in September than in August. This indicates that the PIMS-Net model demonstrates high accuracy and reliability in open water estimation. Analysis of the feature importance heatmaps for different periods reveals significant seasonal variations in the main input channels. From November to May of the following year, microwave radiometer brightness temperature data is the main contributor; conversely, from June to October, SAR backscattering signals become the most influential feature. This shift is likely due to the more dynamic and morphologically complex sea ice conditions in summer, resulting in richer texture details.Therefore, high-resolution SAR imagery can effectively capture small-scale structures, thereby improving the model's ability to reconstruct detailed sea ice patterns. The overall importance of meteorological variables in the meteorological element map is relatively low, but their contribution increases significantly in June and July, indicating their benefit to improving the accuracy of sea ice concentration inversion during these months. Notably, among the multiple frequencies of the microwave radiometer, only the 10.7 GHz and 18.7 GHz vertical polarization channels show significant importance. When arranged individually, the influence of other frequency bands is limited, suggesting substantial information redundancy among microwave radiometer data at different frequencies.

[0128] Furthermore, statistical analysis was performed on the inversion accuracy of the sea ice concentration inversion model (PIMS-Net) in this embodiment for different sea ice concentration ranges, and the following results were generated: Figure 6 The box plot of mean squared error (MSE) is shown. The MSE distribution for different sea ice concentration ranges indicates that the inversion error is significantly higher for low to medium concentrations than for pure water and pure ice. This suggests that the proposed method performs better in learning open water and pure ice samples, but struggles more with low to medium concentration samples. This is because the dataset exhibits significant sample imbalance, with pure water and pure ice samples dominating, while medium concentration samples constitute only a small portion. The distribution of sea ice concentration samples in each dataset is shown below. Figure 7 As shown, a significant sample imbalance is evident. Therefore, the sea ice concentration retrieval model (PIMS-Net) performs poorly when learning these intermediate concentration samples, leading to the error trend observed in the figure. The importance of different features in sea ice concentration classification varies significantly across different concentration ranges. Some brightness temperature features (such as BT_10.7V, BT_18.7V, and BT_23.8V) show consistently high importance in the high concentration range, indicating that these frequency bands are particularly effective in distinguishing high-concentration sea ice conditions. In contrast, some low-frequency brightness temperature channels (e.g., BT_6.9H, BT_7.3H) show stronger correlation in the low-concentration range, but their importance decreases significantly or approaches zero under high-concentration conditions, indicating limited sensitivity in the high-concentration range. ERA5 meteorological variables show high importance in the intermediate-to-high concentration range, meaning that atmospheric conditions provide valuable contextual information, enhancing the retrieval accuracy of sea ice in the intermediate-to-high concentration range. SAR signals contribute significantly in the intermediate-concentration range, but their influence weakens in extreme cases such as extremely low and extremely high concentrations. This model indicates that microwave scattering maps are most informative in distinguishing transitional or heterogeneous ice conditions, where surface texture and spatial variability are more pronounced. Overall, features from different sensors and data sources exhibit complementary behavior across different density ranges. Microwave radiographs dominate in identifying high-density sea ice, meteorological variables enhance performance in the medium-to-high density range, and microwave scattering maps provide crucial input in the medium density range, collectively supporting a more robust and adaptive inversion model.

[0129] To intuitively evaluate and compare the performance of the PIMS-Net model and the traditional U-Net model in sea ice concentration inversion, several challenging scenarios were selected from the test dataset for analysis. The results show that the U-Net model has lower accuracy, producing incorrect predictions where continuous ice is misclassified as open water. In contrast, the PIMS-Net model achieves superior inversion accuracy, thanks to its effective integration of the spatial continuity structure of sea ice. Although a slight overestimation of sea ice concentration was observed in the third scenario, the PIMS-Net model effectively avoids the critical error of classifying sea ice as open water, demonstrating the effectiveness of the cross-modal convolutional fusion block and the bidirectional convolutional long short-term memory network module. Improved effectiveness.

[0130] Fjords are typically a particularly challenging case for sea ice concentration inversion, primarily due to topographic shading and mixed pixel effects along narrow and geometrically complex channels. In two test cases of ice-free fjord scenarios, PIMS-Net accurately identified the fjord as open water, while the conventional U-Net introduced incorrect sea ice.

[0131] Furthermore, a test was conducted on a scene in Scorsese-Bissau, eastern Greenland, acquired on February 20, 2021. Due to surface winds, the roughness of the open water surface increased significantly, resulting in blurred SAR textures between the open water and sea ice. Nevertheless, the PIMS-Net model demonstrated strong robustness in this challenging scene. The PIMS-Net model predictions showed a strong correlation with the reference ice map, identifying high sea ice concentrations. A clear boundary was also present between the sea ice edge and the open water. Scorsese-Bissau is known for its intershore fixed ice, which is present in this scene. Sea ice in the narrow fjords was detected and matched relatively well with the concentration in the reference ice map. Additionally, a test was conducted on a scene in the Denmark Strait in southeastern Greenland, acquired on October 7, 2021. This scene includes sea ice near the meandering coastline, numerous islands, and varying ice concentrations. The predictions were very similar to the reference map. The location of the sea ice was accurate, but the quality of the concentration predictions varied. Some drastic changes in sea ice are lost in homogeneous polygons, which may be because the PIMS-Net model overemphasizes spatial continuity, thus suppressing its ability to capture local drastic changes.

[0132] To better verify the contributions of this embodiment, an ablation study was conducted on Arctic sea ice data. This embodiment mainly investigates the impact of the proposed bidirectional convolutional long short-term memory (BSL-S) network module and decoder network on the final results. This embodiment removes the BSL-S module from the proposed method and replaces the decoder network with a classification module as the baseline, achieving a determination coefficient of 90.8%. Then, this embodiment adds the BSL-S module to the baseline. The (Bi-ConvLSTM) and decoder networks were compared. Finally, the scores were updated by combining the bidirectional convolutional long short-term memory network module and the decoder network module. The ablation experiment results are shown in Table 3.

[0133] Table 3: Ablation Experiment Results

[0134]

[0135] Ablation studies can determine which components are more important for improving sea ice mapping performance and which have the least impact. See Table 3 for examples; the ablation study shows that removing the bidirectional convolutional long short-term memory network module... Multiple encoders can significantly impact a model's scores on the coefficient of determination and mean squared error (MSE). For example, removing the bidirectional convolutional long short-term memory network module... This would decrease the coefficient of determination by 1.2% and increase the mean squared error by 0.0026. Similarly, removing the multiple encoders significantly reduced model accuracy, decreasing the coefficient of determination by 2.0% and increasing the mean squared error by 0.0045.

[0136] In summary, to address the limitations of the traditional U-Net framework in multi-source data fusion and capturing the spatial continuity of sea ice, this embodiment investigates two issues: multi-source data fusion methods and spatial continuity capture. To address these issues, this embodiment implements a spatial continuity-aware PIMS-Net model for sea ice estimation using multi-source data fusion. Results show that the multi-encoder structure and bidirectional convolutional long short-term memory (LSTM) network module enable the PIMS-Net model to make more reliable predictions in low sea ice concentration regions. Furthermore, the PIMS-Net model in this embodiment also implements a spatial continuity-aware deep learning approach for sea ice estimation using multi-source data fusion. Notably, this embodiment is the first to apply a bidirectional convolutional LTM network module to sea ice concentration inversion, effectively integrating multi-source satellite data. Comprehensive experiments on existing datasets demonstrate the effectiveness of the PIMS-Net model in this embodiment. The PIMS-Net model in this embodiment achieves a mean squared error of 0.0138 and a determination coefficient of 93.5%, significantly outperforming traditional U-Net-based models. From a seasonal analysis perspective, the model performs best in winter. As temperatures rise in spring, sea ice melting increases the difficulty of model predictions. The model performs worst in summer. This seasonal variation reflects the significant impact of sea ice dynamics on model accuracy. In summer, the increase in melt pools complicates the backscattered signal from the sea ice surface, making inversion more difficult and reducing model accuracy. Analysis of model accuracy across different sea ice concentration ranges shows that the method in this embodiment primarily improves inversion performance in low sea ice concentration areas, especially in open water. Ablation studies indicate that the multi-encoder structure and bidirectional convolutional long short-term memory network module of the PIMS-Net model in this embodiment are crucial for improving sea ice mapping performance. In summary, this demonstrates the significant potential of using the PIMS-Net model in this embodiment to fuse multi-source satellite data to improve sea ice concentration inversion accuracy. The PIMS-Net model in this embodiment effectively addresses the limitation of insufficient spatial correlation extraction in traditional semantic segmentation models, thereby improving the accuracy of sea ice concentration inversion.

[0137] Those skilled in the art will understand that the technical solutions provided by this invention can take the form of methods, systems, or computer program products. For example, this invention can provide a physically-guided sea ice concentration inversion system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the physically-guided sea ice concentration inversion method. This invention can provide a computer-readable storage medium storing a computer program or instructions programmed or configured to execute the physically-guided sea ice concentration inversion method via a processor. This invention can provide a computer program product including a computer program or instructions programmed or configured to execute the physically-guided sea ice concentration inversion method via a processor. Furthermore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this invention can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for retrieving sea ice concentration based on physical information, characterized in that, This includes using a pre-trained sea ice concentration retrieval model to obtain the corresponding spatial sea ice concentration map from the microwave scattering map, microwave radiation map, and meteorological element map of the area to be detected. S101, features are extracted from microwave scattering map, microwave radiation map and meteorological element map respectively to obtain microwave scattering features, microwave radiation features and meteorological element features at multiple scales; S102, for each scale of microwave scattering characteristics, microwave radiation characteristics and meteorological element characteristics, the microwave radiation characteristics are guided by the meteorological element characteristics and microwave scattering characteristics to achieve cross-modal feature fusion guided by physical information. S103, extract spatial information map from the fusion latent feature map obtained by cross-modal feature fusion at the last scale; S104, the extracted spatial information map is sequentially connected with the fused potential feature map obtained by fusing cross-modal features of other scales and decoded to obtain the final spatial sea ice concentration map; Step S103, extracting the spatial information map from the fused latent feature map obtained by fusing cross-modal features at the last scale, refers to passing the fused latent feature map obtained by fusing cross-modal features at the last scale through a bidirectional convolutional long short-term memory network module. Capturing spatial information maps of sea ice concentration fields, including: S301, for cross-modal feature fusion, obtain the fused latent feature map at the last scale. Apply a sliding window operation to divide it into a series of overlapping tiles in the horizontal and vertical directions, using a fixed tile size. To generate two ordered, containing From the sequence of tiles, we obtain the horizontal tile sequence. and vertical tile sequence ; S302, Transform the horizontal sequence and vertical sequence Treating it as a pseudo-time series, we used a bidirectional convolutional long short-term memory network to capture the spatial continuity in the sea ice concentration field to obtain an enhanced horizontal patch sequence. and vertical tile sequence ; S303, the enhanced horizontal tile sequence and vertical tile sequence After being spatially reassembled through a tiling operation, the data is input into the boundary fusion layer. The boundary fusion layer first performs feature reorganization and boundary information aggregation in the local neighborhood through a 3×3 convolution that keeps the channel dimension unchanged. Then, it uses two-dimensional batch normalization to stabilize the feature distribution. Finally, it uses ReLU nonlinear activation to enhance the representation capability and introduces sparsity to improve the discriminability of boundary details and the fusion effect. S304. The fused feature map obtained after boundary fusion layer processing is restored to the original spatial layout using 1×1 convolution, thereby obtaining the spatial information map in the sea ice concentration field. Spatial information map of the sea ice concentration field and fusion of latent feature maps Both have the same spatial resolution.

2. The sea ice concentration inversion method based on physical information as described in claim 1, characterized in that, In step S101, feature extraction is performed on the microwave scattering map, microwave radiation map, and meteorological element map to obtain microwave scattering features, microwave radiation features, and meteorological element features at multiple scales. This is achieved through three encoders, including: encoders It is used to extract microwave scattering features at multiple scales from microwave scattering maps; encoder It is used to extract microwave radiation features at multiple scales from microwave radiation maps; encoder It is used to extract meteorological feature characteristics at multiple scales from meteorological feature maps; the three encoders each include multiple coding layers connected in sequence, and a downsampling module is provided between adjacent coding layers.

3. The sea ice concentration inversion method based on physical information as described in claim 1, characterized in that, In step S102, meteorological element features and microwave scattering features are used to guide microwave radiation feature correction to achieve physical information-guided cross-modal feature fusion through a cross-modal convolutional fusion block. Implement the cross-modal convolutional fusion block Processing of microwave scattering characteristics, microwave radiation characteristics, and meteorological element characteristics includes: S201 combines microwave scattering features, microwave radiation features, and meteorological element features of the same scale along the channel dimension to form a composite feature map; S202 passes the aggregated feature map through a 3×3 convolutional layer, followed by batch normalization and activation function activation to facilitate nonlinear transformation and structural recombination to obtain a recombined feature map; S203 fuses the reconstructed feature map with the original input cross-modal convolutional block via residual connections. The microwave radiation features are stitched together along the channel dimension to obtain a fused potential feature map.

4. The sea ice concentration inversion method based on physical information as described in claim 1, characterized in that, In step S104, when the extracted spatial information map is sequentially connected with the fused latent feature map obtained by fusing the cross-modal features of other scales and then decoded to obtain the final spatial sea ice concentration map, this includes passing the fused latent feature map obtained by fusing the extracted spatial information map with the cross-modal features of other scales through a decoder network. The decoder network performs hierarchical fusion, connection, and decoding to obtain the corresponding spatial sea ice concentration map. It includes multiple decoding layers. Each decoding layer has a connection module and an upsampling module at its input. The connection module is used to concatenate the fused latent feature map or spatial information map with the feature map output by the upsampling module in the channel dimension and use it as the input of the decoding layer. The input feature of the upsampling module is the output feature of the previous decoder or the spatial information map, and the last decoder layer outputs a spatial sea ice concentration map.

5. The sea ice concentration inversion method based on physical information as described in claim 1, characterized in that, The microwave scattering map is a C-band synthetic aperture radar image, which includes the backscattering coefficients of two channels: horizontal transmission and horizontal reception polarization (HH) and horizontal transmission and vertical reception polarization (HV). The microwave radiation map is a multi-band microwave radiometer data map, and each band includes two channels: horizontal polarization mode H and vertical polarization mode V. The meteorological element map is a numerical weather prediction data distribution image from the reanalysis data of the weather forecasting center, which includes the total cloud liquid water content (tclw), surface radiation temperature (skt), and the rotated 10-meter height northerly wind component (v10m_rotated). Before step S101, the microwave scattering map, microwave radiation map, and meteorological element map are resized to achieve the same resolution, and then the microwave scattering map, microwave radiation map, and meteorological element map with the same resolution are spatially aligned.

6. The sea ice concentration inversion method based on physical information as described in claim 1, characterized in that, The loss function used during training of the sea ice concentration inversion model is: ； in, For loss function, For the sample size, and These are the height and width of the image, respectively. For the first In the nth sample Weight of each pixel and The first In the nth sample The predicted pixel value and label value of the nth pixel; where the nth pixel is the predicted pixel value and label value. In the nth sample The calculation of the weight of each pixel includes: The first In the nth sample Label value of each pixel Equal width compartments: ； in, For the first A set of boxes; K is the number of boxes; Calculate the training set at the th Pixel frequency of each sub-box : ； in, This is an indicator function; the value is 1 if the condition within the parentheses is met, and 0 otherwise. Calculate the first In the nth sample Weights of each pixel: ； in, For binning indexing functions, A constant used to prevent division by zero in the denominator. It is an adjustable exponent, and has the following properties: ； in, Indicates to Round down.

7. A sea ice concentration inversion system guided by physical information, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the physical information-guided sea ice concentration inversion method according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the physical information-guided sea ice concentration inversion method according to any one of claims 1 to 6.

Images