A sea ice classification method, a model training method, an electronic device, and a storage medium

Abstract

This disclosure provides a sea ice classification method, relating to the field of artificial intelligence technology, applicable to scenarios such as radar image processing. The specific implementation scheme includes: determining the dual-polarization characteristics of a target sea area; using the dual-polarization characteristics of the target sea area to characterize the multidimensional features of the dual-polarization synthetic aperture radar data of the target sea area; generating simulated full-polarization characteristics of the target sea area using a feature enhancement model based on the dual-polarization characteristics of the target sea area; and generating sea ice classification results for the target sea area using a sea ice classification model based on the simulated full-polarization characteristics of the target sea area; wherein the feature enhancement model and the sea ice classification model are two neural network models obtained through joint training. This disclosure also provides a model training method, an electronic device, and a storage medium.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a sea ice classification method, a model training method, an electronic device, and a storage medium. Background Technology

[0002] As a crucial component of the polar environment, sea ice classification has always been a core area of ​​polar environmental monitoring and scientific research. Detailed classification of sea ice types not only effectively distinguishes its thickness, extent, and concentration but also provides crucial data support for describing the complex dynamic processes of sea ice movement and redistribution. This information has profound significance for polar climate evolution research, polar ecological environment protection, and resource exploration. Furthermore, in practical applications, high-precision sea ice classification results are a prerequisite for ensuring safe polar navigation, supporting polar scientific expeditions, and facilitating logistical support operations at Antarctic research stations.

[0003] Currently, sea ice classification research can be divided into two main paradigms: one is to use the fine-resolution stripmap mode (SM), which has high accuracy but a swath width of only 20-30 kilometers; the other is to use the wide-swath scanning synthetic aperture radar (ScanSAR) or progressive scanning synthetic aperture radar (TOPSAR) mode, which can cover a range of 300 to 500 kilometers, but usually can only provide dual-polarization data with limited polarization information. Summary of the Invention

[0004] In view of this, this disclosure provides a sea ice classification method, a model training method, an electronic device, and a storage medium.

[0005] One aspect of this disclosure provides a sea ice classification method, which includes: determining the dual polarization characteristics of a target sea area; using the dual polarization characteristics of the target sea area to characterize the multidimensional features of dual polarization synthetic aperture radar data of the target sea area; generating simulated full polarization characteristics of the target sea area using a feature enhancement model based on the dual polarization characteristics of the target sea area; and generating sea ice classification results of the target sea area using a sea ice classification model based on the simulated full polarization characteristics of the target sea area; wherein the feature enhancement model and the sea ice classification model are two neural network models obtained through joint training.

[0006] According to embodiments of this disclosure, by utilizing a feature enhancement model to learn the latent mapping between bipolar features and full polarization features, simulated full polarization features of the target sea area are generated based on the bipolar features of the target sea area, and then a sea ice classification model is used to classify the sea ice based on the simulated full polarization features of the target sea area to obtain sea ice classification results. This can improve the accuracy of sea ice classification while maintaining the advantage of large-scale coverage of bipolar data.

[0007] Another aspect of this disclosure provides a model training method, which includes: determining the full polarization features and dual polarization features of a sample sea area; the full polarization features of the sample sea area are used to characterize the multidimensional features of the full polarization synthetic aperture radar data of the sample sea area; the dual polarization features of the sample sea area are used to characterize the multidimensional features of the dual polarization synthetic aperture radar data of the sample sea area; based on the full polarization features of the sample sea area, a teacher feature map and a teacher probability map are generated using a pre-trained teacher network, and the teacher feature map is used as the ground truth of the full polarization features of the sample sea area; the pre-trained teacher network is a neural network model trained using the full polarization features of the sample sea area and the ground truth labels; the teacher feature map is an intermediate feature map extracted by the pre-trained teacher network, the teacher probability map is a probability distribution map of sea ice categories in the sample sea area output by the pre-trained teacher network, and the ground truth labels indicate the sea ice category distribution of the sample sea area; based on the dual polarization features of the sample sea area, The student network's generator generates simulated full polarization features of the sample sea area, serving as the student feature map. Based on the simulated full polarization features and ground truth values ​​of the full polarization features, the student network's discriminator generates corresponding discrimination scores. These discriminant scores include simulated discrimination scores corresponding to the simulated full polarization features of the sample sea area and ground truth discrimination scores corresponding to the ground truth values ​​of the full polarization features. Based on the simulated full polarization features of the sample sea area, the student network's classifier generates a student prediction map. The student prediction map is the probability distribution map of sea ice categories in the sample sea area output by the student network's classifier. Based on the differences between the student prediction map and the teacher probability map, the differences between the student feature map and the teacher feature map, the differences between the student prediction map and the ground truth labels, and the discrimination scores, the student network's generator, discriminator, and classifier are jointly trained. The trained generator and classifier are then used as the feature enhancement model and sea ice classification model, respectively.

[0008] According to embodiments of this disclosure, by constructing a teacher-student distillation architecture, a pre-trained teacher network is used to guide the student network to learn the latent mapping from bipolar features to fully polarized features. Combined with multi-level joint loss function constraints, the generator of the student network can generate simulated fully polarized features based on the input bipolar features. Then, the classifier of the student network is used to achieve high-precision sea ice classification based on the simulated fully polarized features.

[0009] Another aspect of this disclosure provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method described above.

[0010] Another aspect of this disclosure provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the methods described above.

[0011] Another aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0014] Figure 1 This is an exemplary system architecture diagram of a sea ice classification method and apparatus applicable according to an embodiment of the present disclosure;

[0015] Figure 2 This is a flowchart of a sea ice classification method according to an embodiment of the present disclosure;

[0016] Figure 3 This is a schematic diagram of a sea ice classification method according to an embodiment of the present disclosure;

[0017] Figure 4 This is a flowchart of a model training method according to an embodiment of the present disclosure;

[0018] Figure 5 This is a schematic diagram of a model training method according to an embodiment of the present disclosure;

[0019] Figure 6 This is a schematic diagram of a generator according to an embodiment of the present disclosure;

[0020] Figure 7 This is a schematic diagram of a discriminator according to an embodiment of the present disclosure;

[0021] Figure 8 This is an overall block diagram of model training and sea ice classification according to an embodiment of the present disclosure;

[0022] Figure 9This is a block diagram of an electronic device according to an embodiment of the present disclosure, which can apply at least one of a sea ice classification method and a model training method. Detailed Implementation

[0023] The embodiments of this disclosure will now be described with reference to the accompanying drawings. Various details of the embodiments of this disclosure are included to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0024] In the technical solutions disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of data (including but not limited to user personal information) comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and they do not violate public order and good morals.

[0025] Synthetic Aperture Radar (SAR), especially Polarized Synthetic Aperture Radar (PolSAR), has become the most mainstream and effective tool for monitoring polar sea ice. Compared with traditional optical remote sensing, SAR has all-weather, all-time imaging capabilities, and can penetrate the frequent clouds and fog and long polar nights in the polar regions. PolSAR goes a step further, by transmitting and receiving electromagnetic waves with a combination of horizontal (H) and vertical (V) polarization, it can capture richer backscattering characteristics of ground features.

[0026] With the maturation of polarization decomposition technology, researchers can decompose polarization information into scattering components corresponding to different physical mechanisms (such as surface scattering, volume scattering, and secondary reflection). These physical mechanisms are highly correlated with different types of sea ice, significantly enhancing the robustness of sea ice segmentation and classification algorithms. Currently, polar research institutions use PolSAR technology annually to acquire a large number of polar sea ice images to support scientific research and operational needs.

[0027] In recent years, the deep integration of deep learning (DL) with PolSAR data has become a mainstream trend for improving classification performance. Convolutional Neural Networks (CNNs) automate feature extraction through multi-layer nonlinear transformations, overcoming the limitations of manually designing features. Researchers initially built simple CNN models for datasets such as Sentinel-1. Subsequently, to improve classification accuracy, deeper structures such as ResNets were introduced and combined with Long Short-Term Memory Networks (LSTMs) to learn the spatiotemporal evolution characteristics of sea ice, significantly improving classification accuracy. For fully polarimetric datasets such as GF-3, deep learning models can more effectively capture complex polarimetric features.

[0028] Despite significant advancements in classification techniques, serious challenges remain in practical applications. For example, the lack of polarization information, particularly in dual-polarization modes (such as HH / HV), results in incomplete scattering descriptors (e.g., Pauli decomposition features), making it difficult for models to accurately distinguish between ice types with similar backscattering intensities (e.g., young ice YI and one-year-old ice FYI). Previous methods often treated full-polarization and dual-polarization classification tasks separately. While some studies attempted to recover missing parameters through polarization reconstruction, these methods often introduced substantial estimation errors. High-precision classification relies on full-polarization SM data that cannot cover large areas, while the coverage advantage of dual-polarization modes comes at the cost of classification accuracy. Due to the physical limitations of radar payload hardware, this trade-off between accuracy and coverage cannot currently be fundamentally resolved through purely hardware-based methods.

[0029] In view of this, this disclosure provides a sea ice classification method, a model training method, an electronic device, and a storage medium that can compensate for missing polarization information and improve the accuracy of sea ice classification while maintaining the advantages of wide-swath observation.

[0030] Figure 1 This is a schematic diagram of an exemplary system architecture for applying a sea ice classification method and apparatus according to an embodiment of this disclosure. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0031] like Figure 1As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0032] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0033] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0034] It should be noted that the sea ice classification method provided in this embodiment can generally be executed by server 105. Correspondingly, the sea ice classification device provided in this embodiment can generally be located in server 105. The sea ice classification method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the sea ice classification device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105.

[0035] Figure 2 This is a flowchart of a sea ice classification method according to an embodiment of the present disclosure.

[0036] like Figure 2 As shown, the method 200 may include operations S210 to S230.

[0037] In operation S210, the dual polarization characteristics of the target sea area are determined.

[0038] In this embodiment of the disclosure, the dual-polarization characteristics of the target sea area are used to characterize the multidimensional features of the dual-polarization synthetic aperture radar data of the target sea area. The target sea area refers to the polar ocean region currently being observed.

[0039] During the operation of S220, based on the dual polarization characteristics of the target sea area, a feature enhancement model is used to generate simulated full polarization characteristics of the target sea area.

[0040] In this embodiment of the disclosure, the feature enhancement model is a trained neural network model. For example, a generative adversarial network (GAN) is trained, and the trained generator is used as the feature enhancement model. This GAN can be a progressive generative adversarial network (ProGAN) or other GANs.

[0041] In operation S230, based on the simulated full polarization characteristics of the target sea area, the sea ice classification results for the target sea area are generated using a sea ice classification model.

[0042] In this embodiment of the disclosure, the sea ice classification model is a neural network model obtained by joint training with a feature enhancement model. For example, during the training of ProGAN, a classifier is jointly trained, and this classifier is used to classify the generated simulated full polarization features. The trained classifier is then used as the sea ice classification model.

[0043] Through the embodiments of this disclosure, a feature enhancement model is used to learn the latent mapping between bipolar features and full polarization features. Based on the bipolar features of the target sea area, simulated full polarization features of the target sea area are generated. Then, a sea ice classification model is used to classify the sea ice based on the simulated full polarization features of the target sea area to obtain the sea ice classification result. This can improve the accuracy of sea ice classification while maintaining the advantage of large-scale coverage of bipolar data.

[0044] In this embodiment of the disclosure, the dual-polarization characteristics of the target sea area can be obtained by multi-source data fusion and feature extraction of dual-polarization synthetic aperture radar data of the target sea area.

[0045] In some embodiments of operation S210 described above, determining the dual-polarization characteristics of the target sea area includes: determining dual-polarization synthetic aperture radar (SAR) data of the target sea area; determining the polarization entropy component using eigenvalue decomposition based on the dual-polarization SAR data of the target sea area to obtain the target polarization entropy characteristics; extracting the dual-polarization intensity component from the dual-polarization SAR data of the target sea area to obtain the target intensity characteristics; and performing feature fusion on the target polarization entropy characteristics and the target intensity characteristics to obtain the dual-polarization characteristics of the target sea area.

[0046] Among them, the target polarization entropy feature and the target intensity feature can be fused by channel concat, and the resulting dual polarization feature of the target sea area is a multi-dimensional feature vector with a size of 3×3.

[0047] Currently, wide-swath sea ice monitoring often employs dual-polarization models (such as HH / HV), lacking a fully polarized descriptor (such as polarization entropy) capable of characterizing the randomness of physical scattering. This results in severely limited classification accuracy when distinguishing between young ice (YI) and one-year-old ice (FYI) with extremely similar backscattering intensities, due to a lack of discriminative criteria. The embodiments disclosed in this paper explicitly introduce a polarization entropy component as a physical guide, quantifying the randomness of the scattering process. This deeply integrates prior knowledge of the physical scattering mechanism with dual-polarization intensity data, thereby enhancing the model's ability to identify ice species with similar physical properties from a multi-dimensional feature space.

[0048] In some embodiments of operation S210 described above, determining the dual-polarization synthetic aperture radar (DSAR) data of the target sea area includes: acquiring the original DSAR image of the target sea area; performing radiometric calibration on the original DSAR image of the target sea area to obtain a calibrated image of the target sea area; performing filtering processing on the calibrated image of the target sea area to obtain a filtered image of the target sea area; and performing multi-temporal feature alignment on the filtered image of the target sea area to register images at different times to the same spatial coordinate system, thereby obtaining the DSAR data of the target sea area.

[0049] The filtering process can be speckle filtering or other filtering methods.

[0050] The embodiments disclosed herein preprocess the original dual-polarized images of the target sea area by performing radiometric calibration, filtering, and multi-temporal feature alignment, which can obtain high-quality standardized data and improve the accuracy of subsequent sea ice classification.

[0051] Figure 3 This is a schematic diagram of a sea ice classification method according to an embodiment of the present disclosure.

[0052] like Figure 3As shown, for the preprocessed dual-polarization synthetic aperture radar data of the target sea area (referred to as dual-polarization data in the figure), the dual-polarization features of the target sea area are obtained by extracting the polarization entropy component and concatenating it with the original intensity component. These dual-polarization features are then input into a feature enhancement model trained on a ProGAN-based generator for feature enhancement, resulting in simulated full-polarization features of the target sea area. Following the above operations, the process may further include: generating multi-scale enhanced features of the target sea area using a sea ice classification model based on the simulated full-polarization features; performing superpixel segmentation of the target sea area in the spatial domain based on the multi-scale enhanced features to obtain a superpixel mask; and constructing a superpixel conditional random field and performing energy minimization inference based on the sea ice classification results, multi-scale enhanced features, and the superpixel mask to obtain the post-processed classification results. Furthermore, the accuracy of the post-processed classification results can be evaluated.

[0053] In this embodiment of the disclosure, in order to address the problems of severe speckle noise interference and the lack of spatial continuity and edge fineness in the classification results under the wide-swath monitoring mode of 300 km to 500 km, superpixel segmentation combined with Conditional Random Field (CRF) is used for spatial semantic refinement and edge optimization. This not only effectively suppresses common noise in wide-swath images, but also ensures that while maintaining the advantages of large-scale coverage, a high-quality sea ice distribution map with classification accuracy comparable to the fully polarimetric strip mode is output.

[0054] In some specific implementations, the sea ice classification model may include a multi-scale feature extraction layer, a feature cascade layer, a spatiotemporal attention mechanism layer, a classification probability generation layer, and a multi-scale enhanced feature generation layer. The operation described above, "generating multi-scale enhanced features of the target sea area using the sea ice classification model based on the simulated full polarization features of the target sea area," includes: extracting local features and contextual features at each level using the multi-scale feature extraction layer of the sea ice classification model based on the simulated full polarization features of the target sea area, thus obtaining multi-scale features of the target sea area; performing feature cascade using the feature cascade layer of the sea ice classification model based on the multi-scale features of the target sea area, thus obtaining multi-scale cascaded features of the target sea area; extracting features using the spatiotemporal attention mechanism layer of the sea ice classification model based on the multi-scale cascaded features of the target sea area, thus obtaining spatiotemporal attention features of the target sea area; and generating multi-scale enhanced features of the target sea area using the multi-scale enhanced feature generation layer of the sea ice classification model based on the spatiotemporal attention features of the target sea area. Furthermore, based on the spatiotemporal attention characteristics of the target sea area, the probability distribution map of sea ice categories in the target sea area can be generated using the classification probability generation layer of the sea ice classification model, which is the sea ice classification result of the target sea area.

[0055] In this embodiment, a multi-scale spatiotemporal attention prediction mechanism is introduced to address the challenges of extremely wide image coverage (300-500 km), complex ice texture, and susceptibility to speckle noise. This mechanism combines contextual information within a sliding window for multi-scale fusion prediction. By adaptively allocating spatial and channel weights to the feature map, it effectively suppresses wide-area background noise while simultaneously capturing local sea ice details and broad physical contextual dependencies. This network inference based on global topological relationships overcomes the limitations of traditional pixel-level decision-making, significantly improving the spatiotemporal consistency and overall noise resistance of the classification results.

[0056] In some specific implementations, algorithms such as Simple Linear Iterative Clustering (SLIC) are used for superpixel segmentation. The aforementioned operation, "based on the multi-scale enhancement features of the target sea area, performs superpixel segmentation of the target sea area in the spatial domain to obtain a superpixel mask of the target sea area," includes: transforming the multi-scale enhancement features of the target sea area to a uniform color space to obtain a five-dimensional feature vector of the target sea area; the five-dimensional feature vector includes a brightness component, a red-green color component, a yellow-blue color component, a pixel horizontal coordinate, and a pixel vertical coordinate; initializing several cluster centers based on the five-dimensional feature vector of the target sea area and adjusting each cluster center to the position of minimum gradient; assigning cluster labels to each pixel in the neighborhood of each cluster center based on the weighted distance between each pixel and the five-dimensional feature vector of the cluster center; updating the cluster centers and repeating the iteration until convergence to obtain the superpixel mask of the target sea area.

[0057] The core purpose of superpixel segmentation is to eliminate the isolated, "salt and pepper" noise often generated by pixel-level classification, while preserving the true physical boundaries of sea ice to the greatest extent possible. Its principle is to break away from the traditional independent grid pixel inference model, introducing algorithms such as SLIC and combining them with multi-scale enhanced features to pre-aggregate spatially adjacent pixels with highly similar physical backscattering characteristics into a single unit. Real-world polar sea ice exists as "ice floes" or "icebergs," and this oversegmentation mechanism forces the model to use "superpixel blocks" that conform to objective physical morphology as the smallest inference unit. This not only significantly reduces the computational complexity of subsequent image retouching from the bottom up but also naturally anchors the geometric contours of sea ice, effectively preventing blurring and fragmentation of classification edges at the boundaries between different ice types or ice-water mixtures.

[0058] In some specific implementations, edge optimization of sea ice classification results is performed by constructing a superpixel-level conditional random field (CRF) model. The operation described above, "constructing a superpixel CRF and performing energy minimization inference based on the sea ice classification results of the target sea area, multi-scale enhancement features, and a superpixel mask, to obtain the post-processed classification result," includes: converting the pixel-level sea ice classification result into a preliminary superpixel-level classification probability based on the superpixel mask, as a univariate potential term; determining the feature similarity between adjacent superpixels based on the superpixel mask and multi-scale enhancement features, as a binary potential term; constructing the energy function of the CRF based on the univariate and binary potential terms; performing energy minimization inference on the energy function to obtain the category label of each superpixel; and mapping the category label of each superpixel back to the pixel level to obtain the post-processed classification result.

[0059] Edge optimization using superpixel conditional random fields (CRFs) can compensate for the inherent limitations of feedforward neural networks, such as insufficient sharpness at heterogeneous sea ice interfaces and subtle classification discrepancies within homogeneous sea ice regions. The principle is to construct a superpixel-level CRF model and utilize an "energy penalty mechanism" for global topological inference. Specifically, the preliminary sea ice classification results output by the sea ice classification model are used as the base confidence level, while spatial proximity and polarization feature differences (such as polarization parameter gradients) between superpixels are extracted as mutual constraints. During iterative inference, if spatially adjacent superpixels with highly similar radar features are assigned different sea ice categories by the model, the system applies a significant energy penalty, thus forcibly smoothing out anomalous misclassifications within homogeneous sea ice. Conversely, at the boundaries of truly different sea ice species, due to the strong radar feature step phenomenon, the penalty mechanism automatically fails, stopping the smoothing and achieving precise "sharpening" of the classification boundary. This refined correction based on contextual joint constraints ensures that the sea ice classification results across a wide swath possess extremely high spatial coherence and edge fidelity.

[0060] As can be understood, the sea ice classification method of this disclosure has been explained above. The model training method of this disclosure will be explained below.

[0061] Figure 4 This is a flowchart of a model training method according to an embodiment of the present disclosure.

[0062] like Figure 4 As shown, the method 400 may include operations S410 to S460.

[0063] In operation S410, the full polarization and dual polarization characteristics of the sample sea area are determined.

[0064] In this embodiment of the disclosure, the full polarization characteristics of the sample sea area are used to characterize the multidimensional characteristics of the full polarization synthetic aperture radar (SAP) data of the sample sea area; the dual polarization characteristics of the sample sea area are used to characterize the multidimensional characteristics of the dual polarization SAP data of the sample sea area. Specifically, the full polarization SAP data is full polarization strip mode data, and the dual polarization SAP data is dual polarization wide-swath mode data.

[0065] In the S420 operation, based on the full polarization characteristics of the sample sea area, a pre-trained teacher network is used to generate a teacher feature map and a teacher probability map, and the teacher feature map is used as the true value of the full polarization characteristics of the sample sea area.

[0066] In this embodiment of the disclosure, the pre-trained teacher network is a neural network model trained using the full polarization features of the sample sea area and the real labels; the teacher feature map is the intermediate feature map extracted by the pre-trained teacher network, the teacher probability map is the probability distribution map of sea ice categories in the sample sea area output by the pre-trained teacher network, and the real labels indicate the distribution of sea ice categories in the sample sea area.

[0067] When operating S430, based on the bipolarity characteristics of the sample sea area, the generator of the student network is used to generate simulated full polarization characteristics of the sample sea area as the student feature map.

[0068] In this embodiment of the disclosure, the generator of the student network may include an encoder and a decoder constructed using a progressive growth strategy, wherein the downsampling module of the encoder and the upsampling module of the decoder are structurally symmetrical, and skip connections are used between the same resolution layers of the encoder and the decoder.

[0069] In some embodiments of operation S430 described above, generating simulated full-polarization features of the sample sea area using the generator of the student network based on the dual-polarization features of the sample sea area includes: generating each level of encoded feature maps of the sample sea area layer by layer using each downsampling module of the encoder based on the dual-polarization features of the sample sea area; generating each level of decoded feature maps of the sample sea area layer by layer using each upsampling module of the decoder based on the each level of encoded feature maps of the sample sea area; and determining the decoded feature map output by the last upsampling module of the decoder as the simulated full-polarization features of the sample sea area.

[0070] In operating S440, based on the simulated full polarization characteristics and the true values ​​of full polarization characteristics of the sample sea area, the discriminator of the student network generates the corresponding discrimination scores.

[0071] In this embodiment of the disclosure, the discrimination score includes the simulated discrimination score corresponding to the simulated full polarization features of the sample sea area and the true discrimination score corresponding to the ground truth value of the full polarization features of the sample sea area. The discriminator of the student network may include a multi-scale downsampling layer, a mini-batch standard deviation layer, and a global average pooling layer, wherein the multi-scale downsampling layer includes multiple downsampling residual blocks that grow synchronously with the generator of the student network.

[0072] In some embodiments, in some implementations of the above operation S440, generating a corresponding discrimination score using the discriminator of the student network based on the simulated full polarization features and the true values ​​of the full polarization features of the sample sea area includes: extracting multi-scale discrimination features using a multi-scale downsampling layer based on the feature map to be discriminated to obtain a first feature map; the feature map to be discriminated is either the simulated full polarization features or the true values ​​of the full polarization features of the sample sea area; determining second-order statistics using a mini-batch standard deviation layer based on the feature map to be discriminated in the current batch and concatenating them to the first feature map to obtain a second feature map; and determining the confidence score of the feature map to be discriminated using a global average pooling layer based on the second feature map to obtain a discrimination score.

[0073] When operating the S450, a student prediction map is generated using a classifier from a student network based on the simulated full polarization characteristics of the sample sea area.

[0074] In this embodiment of the disclosure, the student prediction graph is the probability distribution map of sea ice categories in the sample sea area output by the classifier of the student network, i.e., the sea ice classification result of the sample sea area. The classifier of the student network may include a multi-scale feature extraction layer, a feature cascade layer, a spatiotemporal attention mechanism layer, a classification probability generation layer, and a multi-scale enhanced feature generation layer.

[0075] In some embodiments of operation S450 described above, generating a student prediction map using a classifier of a student network based on the simulated full polarization characteristics of the sample sea area includes: extracting local features and contextual features at each level using a multi-scale feature extraction layer of the classifier to obtain multi-scale features of the sample sea area based on the simulated full polarization characteristics of the sample sea area; performing feature concatenation using a feature concatenation layer of the classifier to obtain multi-scale concatenated features of the sample sea area based on the multi-scale concatenated features of the sample sea area; extracting features using a spatiotemporal attention mechanism layer of the classifier to obtain spatiotemporal attention features of the sample sea area based on the spatiotemporal attention features of the sample sea area; and generating a probability distribution map of sea ice categories of the sample sea area using a classification probability generation layer of the classifier, based on the spatiotemporal attention features of the sample sea area, as the student prediction map. Furthermore, multi-scale enhanced features of the sample sea area can also be generated using a multi-scale enhanced feature generation layer of the classifier based on the spatiotemporal attention features of the sample sea area.

[0076] In operating S460, based on the differences between the student's predicted map and the teacher's probability map, the differences between the student's feature map and the teacher's feature map, the differences between the student's predicted map and the true label, and the discrimination score, the generator, discriminator, and classifier of the student network are jointly trained, and the trained generator and classifier are used as the feature enhancement model and the sea ice classification model, respectively.

[0077] In this embodiment, loss information can be determined using various loss functions based on the differences between the student's predicted map and the teacher's probability map, the differences between the student's feature map and the teacher's feature map, the differences between the student's predicted map and the true label, and the discrimination score. The parameters of the student network can be adjusted based on this loss information to train the student network.

[0078] Through the embodiments of this disclosure, a heterogeneous teacher-student distillation architecture is constructed. By utilizing a pre-trained teacher network, the student network is guided to learn the latent mapping from limited observations to the fully polarimetric scattering mechanism. Combined with multi-level joint loss function constraints, the model can implicitly learn complex fully polarimetric semantic logic, effectively avoiding error propagation caused by explicit reconstruction, and achieving higher fidelity cross-modal knowledge transfer.

[0079] In some embodiments, the operation of pre-training the teacher network may be included before the above-described operation S420. For example: the full polarization features of the sample sea area are input into the teacher network to obtain a teacher prediction map; the teacher prediction map is a probability distribution map of sea ice categories in the sample sea area output by the teacher network; based on the difference between the teacher prediction map and the true label, the teacher network is trained to obtain a pre-trained teacher network. Further, the teacher network may be a ResNet-101 network.

[0080] As can be understood, the model training method of this disclosure has been described above, and the feature extraction method of this disclosure will be described below.

[0081] In some embodiments of operation S410 described above, determining the full polarization characteristics and dual polarization characteristics of the sample sea area includes: determining full polarization synthetic aperture radar (SAP) data and dual polarization SAP data of the sample sea area; determining the Pauli decomposition components based on the full polarization SAP data of the sample sea area to obtain the full polarization characteristics of the sample sea area; determining the polarization entropy components based on the dual polarization SAP data of the sample sea area to obtain the sample polarization entropy characteristics; extracting the dual polarization intensity components from the dual polarization SAP data of the sample sea area to obtain the sample intensity characteristics; and performing feature fusion on the sample polarization entropy characteristics and the sample intensity characteristics to obtain the dual polarization characteristics of the sample sea area.

[0082] In some specific implementations, the full polarimetric synthetic aperture radar (SAP) data and dual polarimetric synthetic aperture radar (DPS) data of the sample sea area are obtained by preprocessing the acquired raw polarimetric images of the sample sea area. The aforementioned operation "determining the full polarimetric synthetic aperture radar (SAP) data and dual polarimetric synthetic aperture radar (DPS) data of the sample sea area" includes: acquiring the raw polarimetric images of the sample sea area; the raw polarimetric images of the sample sea area are either raw full polarimetric synthetic aperture radar (SAP) images or raw dual polarimetric synthetic aperture radar (DPS) images of the sample sea area; performing radiometric calibration on the raw polarimetric images of the sample sea area to obtain calibrated images of the sample sea area; performing filtering on the calibrated images of the sample sea area to obtain filtered images of the sample sea area; performing multi-temporal feature alignment on the filtered images of the sample sea area, registering images from different times to the same spatial coordinate system to obtain the polarimetric data of the sample sea area; the polarimetric data of the sample sea area are either full polarimetric synthetic aperture radar (SAP) data or dual polarimetric synthetic aperture radar (DPS) data of the sample sea area.

[0083] In this embodiment, after preprocessing the original fully polarized stripe mode data and dual-polarized wide-swath mode data such as radiometric calibration, speckle filtering, and feature alignment, the polarization entropy component is explicitly calculated using eigenvalue decomposition or an approximate estimation algorithm for wide-swath dual-polarized data. Polarization entropy, as a core physical indicator for quantifying the randomness of electromagnetic wave scattering, can effectively reflect the surface roughness and internal structural complexity of sea ice at the mechanistic level. Through this rapid extraction mechanism, the system successfully transforms the complex physical scattering mechanism into physical guidance features that can be directly processed by deep learning networks, providing a core discrimination basis for subsequently resolving the classification ambiguity problem of backscattering intensity between similar ice species such as young ice (YI) and one-year ice (FYI).

[0084] After acquiring the basic physical parameters, a multi-dimensional feature recombination technique is used to deeply couple the extracted polarization entropy physical priors with the original dual polarization intensity components (such as HH and HV polarization channels). This process constructs high-dimensional feature vectors, organically combining quantitative indicators reflecting the intensity of ground object echoes with prior laws characterizing physical scattering mechanisms, forming a complete sea ice feature description system. This multi-source feature fusion strategy not only enriches the model's input information from a multi-dimensional spatial perspective but also explicitly introduces polarization physical mechanisms into feature expression, enabling the input data to more accurately depict the subtle physical differences between different ice species. This operation lays a solid feature foundation for subsequent real-time cross-modal feature enhancement using the pre-defined ProGAN generator and for achieving refined sea ice classification with high spatial consistency.

[0085] In some specific implementations, the above operation "based on the fully polarimetric synthetic aperture radar data of the sample sea area, using Pauli decomposition to determine the Pauli decomposition components, and obtaining the fully polarimetric characteristics of the sample sea area" includes: performing Pauli decomposition on the polarization scattering matrix in the fully polarimetric synthetic aperture radar data of the sample sea area to obtain odd-order scattering components, even-order scattering components, and volume scattering components; determining the power of the odd-order scattering components as the sample surface scattering characteristics; determining the power of the even-order scattering components as the sample dihedral scattering characteristics; determining the power of the volume scattering components as the sample volume scattering characteristics; and performing feature fusion on the sample surface scattering characteristics, sample dihedral scattering characteristics, and sample volume scattering characteristics to obtain the fully polarimetric characteristics of the sample sea area.

[0086] In this embodiment of the disclosure, by performing Pauli decomposition on the fully polarimetric synthetic aperture radar data of the sample sea area, the rich scattering mechanisms in the fully polarimetric synthetic aperture radar data are systematically encoded into computer-processable feature representations, providing high-quality input for teachers to learn the physical discrimination rules of sea ice online.

[0087] As can be understood, the feature extraction method of this disclosure has been explained above, and the following will combine it with... Figure 5 The multi-level loss function disclosed herein is explained.

[0088] Figure 5 This is a schematic diagram of a model training method according to an embodiment of the present disclosure.

[0089] like Figure 5 As shown: In the teacher branch, for the fully polarimetric synthetic aperture radar data of the sample sea area (referred to as fully polarimetric data in the figure), Pauli decomposition is used to determine the fully polarimetric features, and a ResNet-101 network is used for feature extraction and classification to obtain the teacher feature map and the teacher probability map. The teacher feature map serves as the ground truth of the fully polarimetric features in subsequent operations, and the teacher probability map indicates the probability distribution of sea ice categories, serving as a soft label for subsequent operations. In the student branch, for the dual polarimetric synthetic aperture radar data of the sample sea area (referred to as dual polarimetric data in the figure), the dual polarimetric features are determined by extracting polarization entropy and performing channel stitching. A generator is used to reconstruct and enhance the dual polarimetric features to obtain simulated fully polarimetric features. On the one hand, a classifier is used to classify the simulated fully polarimetric features to obtain the student prediction map. On the other hand, a discriminator is used to discriminate between the simulated fully polarimetric features and the ground truth of the fully polarimetric features to obtain the corresponding discrimination scores. The simulated fully polarimetric features also serve as the student feature map in subsequent operations, and the student prediction map indicates the probability distribution of sea ice categories.

[0090] In some embodiments of operation S460 described above, jointly training the generator, discriminator, and classifier of the student network based on the differences between the student's predicted map and the teacher's probability map, the differences between the student's feature map and the teacher's feature map, the differences between the student's predicted map and the real label, and the discrimination score, to obtain the trained generator and classifier as the feature enhancement model and the sea ice classification model respectively, includes: determining the response layer distillation loss using Kolb-Leibler (KL) divergence based on the differences between the student's predicted map and the teacher's probability map; determining the feature layer consistency loss using mean squared error constraints based on the differences between the student's feature map and the teacher's feature map; determining the classification task supervision loss using cross-entropy based on the differences between the student's predicted map and the real label; determining the adversarial generation loss based on the differences between the real discrimination score and the simulated discrimination score and the gradient penalty term; calculating the total loss based on the response layer distillation loss, the feature layer consistency loss, the classification task supervision loss, and the adversarial generation loss; and jointly training the generator, discriminator, and classifier of the student network based on the total loss, to obtain the trained generator and classifier as the feature enhancement model and the sea ice classification model respectively.

[0091] In some specific implementations, the Wasserstein GAN with gradient penalty (WGAN-GP) loss can be used to determine the adversarial generation loss. The operation described above, "determining the adversarial generation loss based on the difference between the real and simulated discriminant scores and the gradient penalty term," includes: linearly interpolating the ground truth and simulated full polarization features of the sample sea area to obtain interpolated samples; inputting the interpolated samples into the discriminator of the student network to obtain the discriminant scores of the interpolated samples; determining the gradient norm of the discriminant scores of the interpolated samples relative to the interpolated samples, and constructing a gradient penalty term based on the difference between the gradient norm and one; and determining the adversarial generation loss based on the difference between the real and simulated discriminant scores and the gradient penalty term.

[0092] To ensure the fidelity of cross-modal knowledge transfer and the accuracy of classification results, this disclosure constructs a multi-level joint loss function to perform end-to-end optimization of the model. Overall Loss Function Loss due to response layer distillation Feature layer consistency loss Counter-generation loss and the loss of supervision of classification tasks It is composed of weighted combinations, and its overall mathematical expression is shown in equation (1).

[0093] Equation (1)

[0094] in, , , , These are the hyperparameter weights for each loss term, used to balance the training gradients for cross-modal mapping and sea ice classification tasks.

[0095] The specific definitions and physical constraints of each sub-loss are explained below.

[0096] (1) Distillation loss of response layer Calculating student prediction graphs using KL divergence Probability graph of teachers The differences between them. This study guides the student network to exhibit a classification tendency consistent with that of the full-polarization expert when processing bipolar data.

[0097] Equation (2)

[0098] Where i is the pixel index in the feature map, P T,i Let P be the value of the i-th pixel in the teacher probability graph. S,i Predict the value of the i-th pixel in the image for the student.

[0099] (2) Feature layer consistency loss Introducing a feature similarity metric to constrain student feature map F S Teacher Feature Map F T To maintain spatial and semantic consistency and ensure that cross-modal mappings are not distorted, this disclosure employs mean square error (MSE) as a constraint.

[0100] Equation (3)

[0101] (3) Adversarial generation loss To improve the detail and texture of the features generated by ProGAN, WGAN-GP loss is used, thereby enhancing the training stability of the model and preventing mode collapse.

[0102] Equation (4)

[0103] Where, λ gp The gradient penalty coefficient is... It exhibits dual-polarization characteristics. For the true value of the fully polarized characteristic, For generator, For discriminator, These are interpolation samples.

[0104] (4) Loss of monitoring classified tasks Based on the real label y, the standard cross-entropy loss is calculated to provide the network with a definite direction for task optimization, ensuring that the model has a solid basic classification ability.

[0105] Equation (5)

[0106] in, This is the value of the i-th pixel in the actual label.

[0107] The embodiments disclosed herein optimize the weights of the aforementioned multiple loss functions, enabling the model to maintain the monitoring advantage of a 300-500 km swath while achieving high-fidelity feature extrapolation and high-precision sea ice classification.

[0108] As can be understood, the multi-level loss function of this disclosure has been explained above. The following will combine... Figure 5 This document provides an explanation of the teacher and student networks disclosed herein.

[0109] like Figure 5 As shown, the teacher network aims to construct an expert model with high-dimensional polarization semantic extraction capabilities, serving as the "performance ceiling" for the student network's learning. This disclosure employs a deep residual network, ResNet-101, as the basic backbone of the teacher network, and performs offline pre-training using fully polarized (strip mode) data. This expert network, by processing high-resolution images containing complete Pauli decomposition features, can capture subtle physical scattering differences in sea ice under different polarization bases and encode them into high-order semantic feature maps. During distillation, the teacher network keeps its parameters frozen, providing expert-level guidance information in the fully polarized modality to the student network through the class probability distribution (teacher probability map) output by its prediction layer and the feature maps extracted by the intermediate convolutional layers (teacher feature maps).

[0110] like Figure 5 As shown, the student network aims to recover missing fully polarized semantic information through cross-modal latent mapping under limited dual-polarized input conditions. This disclosure introduces a Progressive Generative Adversarial Network (ProGAN) as the basic framework of the student network, achieving high-fidelity polarization feature enhancement through a coarse-to-fine feature evolution process.

[0111] During the data input phase, the student network does not rely solely on the original images. Instead, it uses a multidimensional feature reconstruction module to concatenate the dual polarization intensities (HH, HV) with the pre-computed polarization entropy components, forming a multidimensional input tensor of H×W×3. This design deeply couples the prior physical randomness with the echo intensity features at the front end, providing explicit physical mechanism constraints for the generator during feature mapping.

[0112] The generator employs a symmetrical encoder-decoder topology and introduces a progressive resolution growth strategy. The encoder extracts multi-scale spatial features through layer-by-layer downsampling, while the decoder restores spatial dimensions through upsampling and convolution operations. Simultaneously, skip connections are introduced between layers of the same resolution to fuse feature maps of the same dimension, ensuring that the generated fully polarimetric enhanced feature map retains the sharpness of sea ice edges. In progressive training, the generator starts with extremely low-resolution convolutional blocks, and as the number of training epochs increases, the system smoothly "grows" higher-resolution convolutional modules. Linear interpolation is used for transitions between new and old layers, effectively avoiding oscillations in the early stages of deep network training and ensuring the smoothness of the semantic mapping from bipolar to fully polarimetric. The generator's final output is a simulated fully polarimetric semantic feature vector, which, in the feature space, represents the fully polarimetric feature map output by the teacher network.

[0113] The discriminator, through an adversarial game mechanism, forces the enhanced features output by the generator to approximate the distribution of real fully polarized data. The discriminator employs an inverse structure that grows synchronously with the generator, downsampling the input at multiple spatial scales to simultaneously capture both macroscopic ice distribution and microscopic textures in wide-area imagery. Furthermore, a mini-batch standard deviation layer is introduced at the end of the discriminator to enhance the model's perception of sea ice category diversity by calculating the statistical properties of the feature maps, effectively preventing the mode collapse problem common in adversarial training.

[0114] Figure 6 This is a schematic diagram of a generator according to an embodiment of the present disclosure.

[0115] like Figure 6 As shown, the generator employs a symmetrical encoder-decoder topology, aiming to recover cross-modal fully polarimetric semantic information under limited dual-polarization input conditions. In the encoding stage, a series of stride convolutional layers are used to perform multi-scale dimensionality reduction on the initial H×W×3 feature tensor, which is a fusion of dual-polarization intensity and polarization entropy, to extract deep scattering features of sea ice under different receptive fields. In the decoding stage, symmetrical transposed convolutional layers are used to progressively restore the spatial resolution of the feature maps. Skip connections between layers of the same resolution are used to fuse feature maps of the same resolution, ensuring that the generated fully polarimetric enhanced feature maps accurately map the complex fully polarimetric physical mechanisms while preserving the sharpness of the sea ice's geometric contours.

[0116] Figure 7 This is a schematic diagram of a discriminator according to an embodiment of the present disclosure.

[0117] like Figure 7As shown, the discriminator employs a fully convolutional integrator classifier structure. Through multiple downsampled residual blocks that grow synchronously with the generator, it densely evaluates the distribution of feature maps across multiple spatial scales to determine whether they are ground truth fully polarimetric features from the teacher network or simulated fully polarimetric features from the generator. To enhance the stability of the training process and suppress mode collapse, the discriminator integrates a mini-batch standard deviation layer. This layer enhances the model's ability to perceive the diversity of different sea ice categories (such as young ice and one-year-old ice) by calculating the second-order statistics of the feature map sequence. Finally, a global average pooling layer outputs a confidence score. An adversarial game mechanism forces the generator's output to closely approximate the true fully polarimetric observation data in terms of probability distribution, thus ensuring high fidelity of feature enhancement under wide swath monitoring.

[0118] In summary, addressing the performance bottlenecks of limited swath width in fully polarized strip patterns and insufficient classification accuracy in dual-polarized wide-swath patterns during polar sea ice monitoring, this embodiment introduces polarization entropy components as physical guidance to explicitly quantify the randomness of the scattering process. This effectively compensates for the lack of discriminative basis when using dual-polarized data to identify similar ice species such as young ice and one-year-old ice. During offline training, this embodiment constructs a heterogeneous teacher-student network architecture. Utilizing fully polarized expert knowledge, through KL divergence response layer distillation and feature layer consistency constraints, it guides the student network to learn the latent mapping from limited observations to the fully polarized scattering mechanism. This avoids explicit polarization reconstruction errors while achieving cross-modal knowledge transfer. Combining Progressive Generative Adversarial Networks (ProGANs) and spatiotemporal attention mechanisms, the model achieves high-fidelity feature enhancement and ensures the continuity of classification results across spatiotemporal scales. During online inference, spatial semantic refinement is performed using superpixel-level conditional random fields (CRF) to ensure that the system outputs high-precision sea ice classification results with confidence evaluation while maintaining the advantage of a swath width of 300 to 500 kilometers.

[0119] Figure 8 This is an overall block diagram of model training and sea ice classification according to an embodiment of this disclosure. The following is in conjunction with... Figure 8 The training phase and inference phase of the embodiments of this disclosure will be described separately.

[0120] like Figure 8 As shown, the training phase mainly includes the following operations:

[0121] 1. Synchronous preprocessing of multi-source SAR data: Radiometric calibration, filtering and multi-temporal feature alignment are performed on the full polarization synthetic aperture radar images and dual polarization synthetic aperture radar images of the sample sea area to provide a high-quality collaborative training set for subsequent cross-modal knowledge transfer.

[0122] 2. Polarization entropy physical prior feature extraction: For dual-polarization data, the polarization entropy features are explicitly calculated and fused with the dual-polarization intensity data. The randomness of the physical scattering mechanism is used to enhance the model's ability to distinguish between young ice and one-year-old ice.

[0123] 3. Construction of Heterogeneous Teacher-Student Distillation Architecture: Construct a heterogeneous model with a fully polarized network as the teacher network and a bipolarized network as the student network. Through a knowledge transfer mechanism, the student network implicitly learns the complex fully polarized scattering logic.

[0124] 4. Cross-modal feature latent mapping training: The Progressive Generative Adversarial Network (ProGAN) is used to learn the implicit mapping from bipolar observations to the fully polarized semantic space, thereby achieving feature enhancement while avoiding physical parameter reconstruction errors.

[0125] 5. Multi-level loss function joint optimization: By constraining the classification logic distribution through KL divergence and combining feature layer consistency loss with truth supervision, end-to-end model fine-tuning is completed for wide-span monitoring tasks.

[0126] like Figure 8 As shown, the reasoning phase mainly includes the following operations:

[0127] 1. Loading wide-swath TOPSAR image data: Input dual-polarization TOPSAR image data streams covering 300 km to 500 km as dual-polarization synthetic aperture radar images of the target sea area, and complete preprocessing to ensure monitoring effectiveness covers a large active sea area.

[0128] 2. Polarization-guided feature recombination: Simultaneously calculate the polarization entropy components of the current observation area and construct a multi-dimensional input vector to provide necessary physical scattering prior constraints for inference.

[0129] 3. Real-time generation of enhanced classification feature maps: The pre-built ProGAN generator is called to perform cross-modal enhancement processing on the input and output high-order enhanced feature maps with full polarization semantic information in real time as simulated full polarization features.

[0130] 4. Multi-scale spatiotemporal attention fusion prediction: The fixed classifier is invoked and combined with the context information in the sliding window to perform multi-scale fusion prediction, so as to obtain preliminary sea ice classification results. This effectively suppresses the noise commonly found in wide-swath monitoring and improves the spatiotemporal consistency of classification results.

[0131] 5. Superpixel CRF Refinement and Output: Superpixel segmentation combined with conditional random fields is used to optimize the edges and suppress speckle in the preliminary sea ice classification results, and finally outputs high-precision post-processed classification results and performs accuracy evaluation.

[0132] This application addresses the limitations of existing sea ice classification technologies and offers significant advantages in resolving practical business bottlenecks: By introducing polarization entropy physical priors, it completes the dimension representing the randomness of scattering, significantly improving the classification accuracy of young ice versus one-year-old ice without increasing hardware load; Utilizing the progressive growth strategy of ProGAN ensures the stability of the feature mapping process, and by implicitly learning the full polarization semantic logic, it effectively avoids the diffusion of reconstruction errors, making the feature representation of wide-swath data closer to the level of real full polarization experts; Addressing the problems of severe noise and fragmented edges in wide-swath images, it utilizes superpixel CRF to achieve global consistency optimization, breaking the limitations of pixel-level decision-making and ensuring that the classification results have extremely high physical morphological integrity and edge fidelity, making the output quality of the wide-swath monitoring mode comparable to the high-precision narrow-swath strip mode.

[0133] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0134] Figure 9 A block diagram of an electronic device that can be used to implement the methods of embodiments of the present disclosure is illustrated schematically. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0135] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0136] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0137] The computing unit 901 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as at least one of the sea ice classification method and the model training method. For example, in some embodiments, at least one of the sea ice classification method and the model training method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by computing unit 901, it can perform one or more steps of at least one of the sea ice classification method and model training method described above. Alternatively, in other embodiments, computing unit 901 can be configured to perform at least one of the sea ice classification method and model training method by any other suitable means (e.g., by means of firmware).

[0138] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0139] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0140] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM) or flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0141] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) monitor or a liquid crystal display (LCD)); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0142] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0143] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0144] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0145] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method of sea ice classification, characterized by, include: Determine the dual-polarization characteristics of the target sea area; The dual-polarization characteristics of the target sea area are used to characterize the multidimensional features of the dual-polarization synthetic aperture radar data of the target sea area. Based on the dual polarization characteristics of the target sea area, a simulated full polarization characteristic of the target sea area is generated using a feature enhancement model; Based on the simulated full polarization characteristics of the target sea area, the sea ice classification results of the target sea area are generated using a sea ice classification model. The feature enhancement model and the sea ice classification model are two neural network models obtained through joint training.

2. The sea ice classification method of claim 1, wherein, Determine the dual polarization characteristics of the target sea area, including: Determine the target sea area using dual-polarization synthetic aperture radar data; Based on the dual-polarization synthetic aperture radar data of the target sea area, the polarization entropy components are determined by eigenvalue decomposition to obtain the target polarization entropy characteristics; From the dual-polarization synthetic aperture radar data of the target sea area, the dual-polarization intensity component is extracted to obtain the target intensity characteristics; The target polarization entropy feature and the target intensity feature are fused to obtain the dual polarization feature of the target sea area.

3. The sea ice classification method of claim 2, wherein, Determine the dual-polarization synthetic aperture radar data for the target sea area, including: Acquire raw dual-polarization synthetic aperture radar imagery of the target sea area; Radiometric calibration is performed on the original dual-polarization synthetic aperture radar image of the target sea area to obtain the calibrated image of the target sea area. The calibrated image of the target sea area is filtered to obtain the filtered image of the target sea area. Multi-temporal feature alignment is performed on the filtered images of the target sea area, and the images at different times are registered to the same spatial coordinate system to obtain dual-polarization synthetic aperture radar data of the target sea area.

4. The sea ice classification method of claim 1, wherein, Also includes: Based on the simulated full polarization characteristics of the target sea area, multi-scale enhanced features of the target sea area are generated using a sea ice classification model; Based on the multi-scale enhancement features of the target sea area, superpixel segmentation is performed on the target sea area in the spatial domain to obtain a superpixel mask of the target sea area. Based on the sea ice classification results, multi-scale enhancement features, and superpixel masks of the target sea area, a superpixel conditional random field is constructed and energy minimization inference is performed to obtain the post-processed classification results.

5. The sea ice classification method of claim 4, wherein, Based on the simulated full polarization characteristics of the target sea area, multi-scale enhanced features of the target sea area are generated using a sea ice classification model, including: Based on the simulated full polarization characteristics of the target sea area, the multi-scale feature extraction layer of the sea ice classification model is used to extract local features and context features at each level to obtain the multi-scale features of the target sea area. Based on the multi-scale characteristics of the target sea area, the feature cascade layer of the sea ice classification model is used to perform feature cascade to obtain the multi-scale cascaded features of the target sea area. Based on the multi-scale cascaded features of the target sea area, the spatiotemporal attention mechanism layer of the sea ice classification model is used to extract features and obtain the spatiotemporal attention features of the target sea area. Based on the spatiotemporal attention characteristics of the target sea area, multi-scale enhanced features of the target sea area are generated using the multi-scale enhanced feature generation layer of the sea ice classification model.

6. The sea ice classification method of claim 4, wherein, Based on the multi-scale enhancement features of the target sea area, superpixel segmentation is performed on the target sea area in the spatial domain to obtain a superpixel mask for the target sea area, including: The multi-scale enhancement features of the target sea area are transformed into a uniform color space to obtain a five-dimensional feature vector of the target sea area; the five-dimensional feature vector includes a brightness component, a red-green color component, a yellow-blue color component, a pixel horizontal coordinate, and a pixel vertical coordinate. Based on the five-dimensional feature vector of the target sea area, several cluster centers are initialized and each cluster center is adjusted to the position of minimum gradient. Within the neighborhood of each cluster center, a cluster label is assigned to each pixel based on the weighted distance between each pixel and the five-dimensional feature vector of the cluster center; The cluster centers are updated and iterated repeatedly until convergence is achieved, resulting in a superpixel mask for the target sea area.

7. The sea ice classification method of claim 4, wherein, Based on the sea ice classification results, multi-scale enhancement features, and superpixel masks of the target sea area, a superpixel conditional random field is constructed and energy minimization inference is performed to obtain post-processed classification results, including: Based on the superpixel mask, the pixel-level sea ice classification result is converted into a superpixel-level preliminary classification probability, which is used as a univariate potential energy term. Based on the superpixel mask and the multi-scale enhancement features, the feature similarity between adjacent superpixels is determined as a binary potential term; Construct the energy function of the conditional random field based on the univariate potential energy term and the binary potential energy term; Energy minimization inference is performed on the energy function to obtain the category label of each superpixel; The category labels of each superpixel are mapped back to the pixel level to obtain the post-processing classification result.

8. A model training method, comprising: include: Determine the full polarization and dual polarization characteristics of the sample sea area; The full polarization characteristics of the sample sea area are used to characterize the multidimensional features of the full polarization synthetic aperture radar data of the sample sea area. The dual-polarization characteristics of the sample sea area are used to characterize the multidimensional features of the dual-polarization synthetic aperture radar data of the sample sea area. Based on the full polarization characteristics of the sample sea area, a teacher feature map and a teacher probability map are generated using a pre-trained teacher network, and the teacher feature map is used as the true value of the full polarization characteristics of the sample sea area. The pre-trained teacher network is a neural network model trained using the full polarization features of the sample sea area and the real labels; the teacher feature map is an intermediate feature map extracted by the pre-trained teacher network; the teacher probability map is a probability distribution map of sea ice categories in the sample sea area output by the pre-trained teacher network; and the real labels indicate the distribution of sea ice categories in the sample sea area. Based on the bipolarity characteristics of the sample sea area, the generator of the student network is used to generate simulated full polarization characteristics of the sample sea area as a student feature map. Based on the simulated full polarization features and the true values ​​of the full polarization features of the sample sea area, a discriminator of the student network is used to generate a corresponding discrimination score; the discrimination score includes the simulated discrimination score corresponding to the simulated full polarization features of the sample sea area and the true discrimination score corresponding to the true values ​​of the full polarization features of the sample sea area. Based on the simulated full polarization characteristics of the sample sea area, a student prediction map is generated using the classifier of the student network; the student prediction map is a probability distribution map of sea ice categories in the sample sea area output by the classifier of the student network. Based on the differences between the student prediction map and the teacher probability map, the differences between the student feature map and the teacher feature map, the differences between the student prediction map and the true label, and the discrimination score, the generator, discriminator, and classifier of the student network are jointly trained, and the trained generator and classifier are used as the feature enhancement model and the sea ice classification model, respectively.

9. The model training method of claim 8, wherein, Based on the differences between the student prediction map and the teacher probability map, the differences between the student feature map and the teacher feature map, the differences between the student prediction map and the true label, and the discrimination score, the generator, discriminator, and classifier of the student network are jointly trained. The trained generator and classifier are used as the feature enhancement model and the sea ice classification model, respectively, including: Based on the difference between the student prediction map and the teacher probability map, the Kohlbek-Leibler divergence is used to determine the distillation loss of the response layer; Based on the difference between the student feature map and the teacher feature map, the feature layer consistency loss is determined using mean squared error constraints; Based on the difference between the student's predicted image and the true label, the cross-entropy method is used to determine the supervised loss for the classification task. The adversarial generation loss is determined based on the difference between the real and simulated discriminant scores and the gradient penalty term; The total loss is calculated based on the response layer distillation loss, the feature layer consistency loss, the classification task supervision loss, and the adversarial generation loss. The generator, discriminator, and classifier of the student network are jointly trained based on the total loss, and the trained generator and classifier are used as the feature enhancement model and sea ice classification model, respectively.

10. The model training method of claim 9, wherein, Based on the difference between the real and simulated discriminant scores and the gradient penalty term, the adversarial generation loss is determined, including: Linear interpolation is performed on the true values ​​and simulated full polarization characteristics of the sample sea area to obtain interpolated samples; The interpolated sample is input into the discriminator of the student network to obtain the discriminant score of the interpolated sample; Determine the gradient norm of the discrimination score of the interpolated sample relative to the interpolated sample, and construct a gradient penalty term based on the difference between the gradient norm and one; The adversarial generation loss is determined based on the difference between the real and simulated discriminant scores and the gradient penalty term.

11. The model training method of claim 8, wherein, Determine the full polarization and dual polarization characteristics of the sample sea area, including: Determine the full polarization synthetic aperture radar data and dual polarization synthetic aperture radar data for the sample sea area; Based on the fully polarimetric synthetic aperture radar data of the sample sea area, Pauli decomposition is used to determine the Pauli decomposition components, and the fully polarimetric characteristics of the sample sea area are obtained. Based on the dual-polarization synthetic aperture radar data of the sample sea area, the polarization entropy components are determined by eigenvalue decomposition to obtain the sample polarization entropy characteristics. From the dual-polarization synthetic aperture radar data of the sample sea area, the dual-polarization intensity component is extracted to obtain the sample intensity characteristics; The polarization entropy feature and the intensity feature of the sample are fused to obtain the dual polarization feature of the sample sea area.

12. The model training method according to claim 11, characterized in that, Based on the fully polarimetric synthetic aperture radar data of the sample sea area, Pauli decomposition is used to determine the Pauli decomposition components, thereby obtaining the fully polarimetric characteristics of the sample sea area, including: Pauli decomposition was performed on the polarization scattering matrix in the fully polarimetric synthetic aperture radar data of the sample sea area to obtain odd scattering components, even scattering components and volume scattering components. The power of the odd-order scattering components is determined as the scattering characteristic of the sample surface; The power of the even-order scattering components is determined as the dihedral scattering characteristic of the sample; The power of the volume scattering component is determined as the volume scattering characteristic of the sample; The scattering characteristics of the sample surface, the scattering characteristics of the sample dihedral angle, and the scattering characteristics of the sample volume are fused to obtain the full polarization characteristics of the sample sea area.

13. The model training method according to claim 11, characterized in that, Determine the full polarization synthetic aperture radar (SAP) data and dual polarization synthetic aperture radar (DAP) data for the sample sea area, including: Acquire the original polarimetric image of the sample sea area; the original polarimetric image of the sample sea area is the original fully polarimetric synthetic aperture radar image or the original dual polarimetric synthetic aperture radar image of the sample sea area. Radiometric calibration is performed on the original polarization image of the sample sea area to obtain the calibrated image of the sample sea area. The calibrated image of the sample sea area is filtered to obtain the filtered image of the sample sea area. Multi-temporal feature alignment is performed on the filtered images of the sample sea area, and the images at different times are registered to the same spatial coordinate system to obtain the polarization data of the sample sea area; the polarization data of the sample sea area is the fully polarimetric synthetic aperture radar data or the dual polarimetric synthetic aperture radar data of the sample sea area.

14. The model training method according to claim 8, characterized in that, Also includes: The full polarization characteristics of the sample sea area are input into the teacher network to obtain the teacher prediction map; the teacher prediction map is the probability distribution map of sea ice categories of the sample sea area output by the teacher network. The teacher network is trained based on the difference between the teacher's predicted graph and the real label to obtain a pre-trained teacher network.

15. The model training method according to claim 8, characterized in that, The generator of the student network includes an encoder and a decoder constructed using a progressive growth strategy. The downsampling module of the encoder and the upsampling module of the decoder are structurally symmetrical, and skip connections are used between the same resolution levels of the encoder and the decoder. Based on the dual-polarization characteristics of the sample sea area, a simulated full-polarization characteristic of the sample sea area is generated using the generator of the student network, including: Based on the dual-polarization characteristics of the sample sea area, the encoder's downsampling modules are used to generate layer-by-layer encoded feature maps of the sample sea area. Based on the coded feature maps of each level of the sample sea area, the decoder generates the decoded feature maps of each level of the sample sea area layer by layer using each upsampling module. The decoded feature map output by the last upsampling module of the decoder is determined as the simulated full polarization feature of the sample sea area.

16. The model training method according to claim 15, characterized in that, The discriminator of the student network includes a multi-scale downsampling layer, a mini-batch standard deviation layer, and a global average pooling layer. The multi-scale downsampling layer includes multiple downsampling residual blocks that grow synchronously with the generator of the student network. Based on the simulated full polarization characteristics and ground truth values ​​of the full polarization characteristics of the sample sea area, the discriminator of the student network generates corresponding discrimination scores, including: Based on the feature map to be discriminated, multi-scale discriminative features are extracted using the multi-scale downsampling layer to obtain the first feature map; the feature map to be discriminated is the simulated full polarization feature or the true value of the full polarization feature of the sample sea area; Based on the feature map to be discriminated in the current batch, the second-order statistic is determined using the small-batch standard deviation layer and spliced ​​to the first feature map to obtain the second feature map; Based on the second feature map, the confidence score of the feature map to be discriminated is determined using the global average pooling layer to obtain the discrimination score.

17. The model training method according to claim 8, characterized in that, Based on the simulated full polarization characteristics of the sample sea area, a student prediction map is generated using the classifier of the student network, including: Based on the simulated full polarization characteristics of the sample sea area, the multi-scale feature extraction layer of the classifier is used to extract local features and context features at each level to obtain the multi-scale features of the sample sea area. Based on the multi-scale characteristics of the sample sea area, the feature cascade layer of the classifier is used to perform feature cascade to obtain the multi-scale cascaded features of the sample sea area. Based on the multi-scale cascaded features of the sample sea area, the spatiotemporal attention mechanism layer of the classifier is used to extract features and obtain the spatiotemporal attention features of the sample sea area. Based on the spatiotemporal attention characteristics of the sample sea area, the probability distribution map of sea ice categories in the sample sea area is generated using the classification probability generation layer of the classifier, and used as the student prediction map.

18. An electronic device comprising: One or more processors; Storage device for storing one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1 to 17.

19. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 17.

20. A computer program product comprising a computer program stored on at least one of a readable storage medium and an electronic device, the computer program implementing the method according to any one of claims 1 to 17 when executed by a processor.

Images