Sea ice classification method and device based on fusion of SAR complex data amplitude and phase

Abstract

This application relates to a sea ice classification method and apparatus based on amplitude and phase fusion of SAR complex data. The method includes: constructing a SAR single-look complex dataset containing HH / HV dual-polarization real and imaginary parts; after spatial registration, using an adaptive step-size pruning strategy to improve class imbalance and expand the sample size; extracting amplitude and phase features separately; fusing the two types of features based on a two-branch processing framework; inputting the data into a semantic segmentation network; and training the model using a mean-normalized inverse frequency-weighted cross-entropy loss function; inputting the fused features extracted from the measured image using the two-branch framework into the model to obtain pixel-by-pixel sea ice classification results. This method fully exploits the phase information value of SAR complex data, solves the problems of phase entanglement and noise interference, significantly improves the overall accuracy of sea ice classification, and greatly reduces the misclassification rate of similar sea ice categories with low backscattering intensity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing and sea ice classification, and in particular to a sea ice classification method and device based on SAR complex data amplitude and phase fusion. BACKGROUND

[0002] Monitoring the type and distribution of sea ice is of great significance to polar navigation safety and climate change research. Synthetic aperture radar (SAR) has become the core remote sensing data source for sea ice classification due to its all-weather, all-day, and high-resolution imaging advantages. Current SAR-based sea ice classification methods mainly fall into two categories: statistical machine learning and deep learning. Deep learning methods can automatically extract image features, avoiding the complexity of manually designing features in traditional methods, and have better classification accuracy and efficiency, thus becoming the mainstream technology in this field.

[0003] However, existing deep learning sea ice classification techniques still have obvious limitations. Firstly, the information utilization of SAR data is not sufficient. Most methods only use SAR amplitude data as model input, completely ignoring the phase information contained in SAR single-view complex data. However, high-resolution phase information can reflect the microstructure and scattering mechanism of sea ice, and combining it with amplitude information can fully characterize the physical properties of sea ice. Secondly, the SAR sea ice classification data set is scarce and the class distribution is extremely unbalanced. Open water, thick one-year ice, and other majority classes have a high proportion, while young ice, thin one-year ice, and other minority classes have sparse pixels, leading to high misclassification rates when distinguishing similar backscattering intensity sea ice classes based on amplitude data alone. Thirdly, SAR original amplitude data has problems such as speckle noise and outliers, and phase data has problems such as absolute phase randomness and phase angle periodic discontinuity, making it difficult to be directly and effectively utilized by deep learning models.

[0004] In summary, how to fully exploit the complete information of SAR complex data amplitude and phase, address the technical difficulties of phase utilization through targeted processing, and alleviate the problem of data class imbalance to improve sea ice classification accuracy and reduce misclassification rates of similar backscattering intensity classes is a key issue in the field of SAR remote sensing sea ice classification that needs to be addressed. SUMMARY

[0005] Therefore, it is necessary to provide a sea ice classification method and device based on SAR complex data amplitude and phase fusion that can significantly improve sea ice classification accuracy and reduce misclassification rates of similar backscattering intensity classes.

[0006] A sea ice classification method based on SAR complex data amplitude and phase fusion, the method comprising:

[0007] Obtain a SAR single-view complex dataset, which includes multiple sample images and corresponding label images with sea ice category labels. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. Spatial registration processing is performed on the SAR single-view complex dataset. Bilinear interpolation resampling and boundary nearest neighbor filling are performed on the corresponding sample images using the label image as the reference grid to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample images to obtain the processed sample data. A training dataset is constructed based on all the processed sample data. Amplitude and phase features are extracted from the sample data in the training dataset respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor. The feature tensor is then input into a semantic segmentation network to predict the sea ice category pixel by pixel. Based on the prediction results and the corresponding label images, a mean-normalized inverse frequency-weighted cross-entropy loss function is constructed to train and validate the semantic segmentation network, thereby obtaining a sea ice classification model. Acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

[0008] In one embodiment, the sea ice category label corresponds to six types of sea ice: open water, new ice, young ice, thin one-year ice, thick one-year ice, and old ice.

[0009] In one embodiment, the label image is a two-dimensional raster ground truth reference image with pixel-by-pixel sea ice category labeling information. The label image has the same spatial coordinate system as the corresponding sample image and is labeled with a land mask area and an effective sea ice pixel area. The effective sea ice pixel area is labeled with the corresponding sea ice category.

[0010] In one embodiment, when cropping and augmenting the registered sample image using an adaptive step-size cropping strategy: The regions containing a minority of sea ice types in the registered sample image are cropped with a small step size, while the regions containing the majority of sea ice types are cropped with a large step size, to obtain the initial cropped image patch. The initial cropped patches are quality-screened, removing patches with a mask pixel ratio exceeding 30%, and retaining only patches containing at least two valid sea ice categories to obtain valid patches; For each scene, valid patches corresponding to the sample images are selected using random sampling without replacement, with no more than 50 valid patches selected to complete the cropping and augmentation of the sample images.

[0011] In one embodiment, amplitude feature extraction is performed on the sample data in the training dataset, including: The backscattering intensity data is calculated based on the complex information of the real and imaginary parts of the HH / HV dual polarization in the sample data. A spatial smoothing filter is applied to the backscattering intensity data in linear power form to denoise it while preserving the variation trend of the backscattering intensity. The dynamic range of the denoised backscattering coefficients is truncated within a preset range and normalized to obtain the amplitude characteristics.

[0012] In one embodiment, phase feature extraction is performed on the sample data in the training dataset, including: Based on the complex information of the real and imaginary parts of the HH and HV dual polarizations in the sample data, the phase difference between the HH and HV polarization channels is calculated. The real and imaginary parts of the phase difference are spatially averaged separately, and then the range of values ​​is obtained by solving the arctangent function. The phase angle; The phase angle is decomposed into cosine and sine components and cyclically encoded to obtain the phase feature.

[0013] In one embodiment, the semantic segmentation network is any one of FCN, U-Net, DeepLabV3+, and PSPNet.

[0014] In one embodiment, the mean-normalized inverse frequency-weighted cross-entropy loss function calculates the loss only in the effective sea ice pixel region, ignoring the land mask region. The loss function formula is as follows:

[0015] In the above formula, This indicates the total number of valid pixels in the batch. For the true label of the i-th valid pixel, Let be the predicted probability of the i-th valid pixel. The weights are normalized based on the inverse frequency of the training set categories, and the average weight of all valid sea ice categories is 1.

[0016] This application also provides a sea ice classification device based on SAR complex data amplitude and phase fusion, the device comprising: The dataset acquisition module is used to acquire SAR single-view complex dataset, which includes multiple sample images and label images with sea ice category labels corresponding to each image. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. The training dataset construction module is used to perform spatial registration processing on the SAR single-look complex dataset. It uses the label image as the reference grid to perform bilinear interpolation resampling and boundary nearest neighbor filling on the corresponding sample image to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample image to obtain the processed sample data. The training dataset is constructed based on all the processed sample data. The pixel-by-pixel sea ice category prediction module is used to extract amplitude and phase features from the sample data in the training dataset, respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor, and then the feature tensor is input into the semantic segmentation network to predict the pixel-by-pixel sea ice category. The model training module is used to construct a mean-normalized inverse frequency-weighted cross-entropy loss function based on the prediction results and the corresponding label images, and to train and validate the semantic segmentation network to obtain the sea ice classification model. The sea ice classification module is used to acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, and obtain the pixel-by-pixel sea ice classification result to complete the sea ice classification.

[0017] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above-described sea ice classification method based on SAR complex data amplitude and phase fusion.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described sea ice classification method based on SAR complex data amplitude and phase fusion.

[0019] The aforementioned sea ice classification method and apparatus based on amplitude and phase fusion of SAR complex data acquires a SAR single-view complex dataset including multiple sample images and corresponding label images with sea ice category labels. Using the label images as the reference grid, bilinear interpolation resampling and nearest-neighbor filling are performed on the corresponding sample images in the SAR single-view complex dataset to achieve registration. An adaptive step-size cropping strategy is then used to crop and expand the registered sample images to obtain processed sample data. A training dataset is constructed based on all processed sample data. Amplitude and phase features are extracted from the sample data in the training dataset. A dual-branch feature fusion framework is used to concatenate the amplitude and phase features into a feature tensor. This feature tensor is then input into a semantic segmentation network for pixel-by-pixel sea ice category prediction. Based on the prediction results and the corresponding label images, a mean-normalized inverse frequency-weighted cross-entropy loss function is constructed. The semantic segmentation network is then trained and validated to obtain a sea ice classification model. The fused amplitude and phase features of the SAR single-view complex image are input into the sea ice classification model to obtain pixel-by-pixel sea ice classification results, thus completing the sea ice classification.

[0020] Beneficial effects: This method fully extracts the complete amplitude and phase information from SAR single-look complex data. By using a dual-branch feature fusion framework, it effectively integrates amplitude features with specially processed phase features, overcoming the limitations of using only amplitude information for sea ice classification. It can accurately capture the unique scattering mechanisms of different sea ice types due to differences in dielectric properties and microscopic physical structures, significantly reducing the misclassification rate of similar sea ice types with low backscattering intensity. At the same time, it effectively alleviates the class imbalance problem in sea ice datasets by relying on an adaptive step-size pruning strategy. Combined with targeted training using the mean-normalized inverse frequency-weighted cross-entropy loss function, it makes the training of the semantic segmentation network more targeted, greatly improving the overall accuracy of sea ice classification and the F1 score of each class. Furthermore, the pixel-by-pixel sea ice classification achieved through precise spatial registration with the label image can output high-precision spatial distribution results of sea ice types. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a sea ice classification method based on the fusion of amplitude and phase of SAR complex data in one embodiment. Figure 2 This is a schematic diagram illustrating the category distribution after the adaptive step-size pruning strategy in one embodiment. Figure 3 This is a schematic diagram of the dual-branch processing framework proposed in this method in one embodiment; Figure 4 This is a schematic diagram comparing various F1 scores of the PSPNet model in one embodiment; Figure 5This is a schematic diagram illustrating the visualization results of the PSPNet model in three scenarios in one embodiment; Figure 6 This is a structural block diagram of a sea ice classification device based on SAR complex data amplitude and phase fusion in one embodiment; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] To address the problems in existing technologies, such as SAR sea ice classification relying solely on amplitude information without fully leveraging the value of phase information in complex data, the ease of confusion between similar sea ice categories with low backscattering intensities, and the limitation of classification accuracy due to severe class imbalance in sea ice datasets, this application addresses these issues. Figure 1 As shown, a sea ice classification method based on the fusion of amplitude and phase of SAR complex data is provided, which specifically includes the following steps: Step S100: Obtain the SAR single-view complex dataset. The SAR single-view complex dataset includes multiple sample images and label images with sea ice category labels corresponding to each image. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location.

[0024] Step S110: Spatial registration processing is performed on the SAR single-look complex dataset. The corresponding sample images are resampled by bilinear interpolation and boundary nearest neighbor filling with the label image as the reference grid to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample images to obtain the processed sample data. The training dataset is constructed based on all the processed sample data.

[0025] Step S120: Extract amplitude and phase features from the sample data in the training dataset respectively. Using a dual-branch feature fusion framework, concatenate the amplitude and phase features into a feature tensor. Then, input the feature tensor into the semantic segmentation network to predict the sea ice category pixel by pixel.

[0026] Step S130: Based on the prediction results and the corresponding label images, construct the mean-normalized inverse frequency-weighted cross-entropy loss function, train and validate the semantic segmentation network, and obtain the sea ice classification model.

[0027] Step S140: Obtain the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

[0028] In this method, a complex dataset specifically designed for SAR sea ice classification is first constructed, preserving the amplitude and phase information of the SAR single-look complex data. This fully mines and utilizes the complete complex information in the SAR complex data for sea ice classification. Secondly, the phase information in the complex data undergoes targeted processing. By calculating inter-channel phase difference, spatial averaging, and cyclic encoding, the microscopic physical features of sea ice contained in the phase are effectively extracted, compensating for the information limitations and recognition constraints of using only amplitude information for sea ice classification. Finally, to address the severe class imbalance problem in the sea ice dataset, an adaptive step-size pruning strategy combined with a mean-normalized inverse frequency-weighted cross-entropy loss function is used for effective processing. Simultaneously, training and validation experiments are conducted based on various classic semantic segmentation networks, comprehensively verifying the effectiveness and practicality of SAR complex information in improving sea ice classification accuracy and reducing the misclassification rate of easily confused categories.

[0029] In step S100, the SAR single-look complex dataset is constructed based on the ASID dataset, a benchmark dataset in the field of deep learning sea ice mapping. Using the same observation data and coverage area of ​​the Arctic surrounding waters from January 2018 to December 2021 as the ASID dataset, 41 valid scenes are selected to construct a NetCDF format SAR single-look complex dataset. Each valid scene corresponds to one set of dual-polarization Sentinel-1 ultrawide (EW) synthetic aperture radar (SAR) observation data after NERSC noise correction. A valid scene refers to SLC format data corresponding to the GRD format data in the ASID dataset. The GRD format is SAR amplitude data after multi-look processing and radiometric correction, while the SLC format is SAR single-look complex data that retains complete real and imaginary part information. Both are different format products from the same satellite observation period and the same coverage area; the correspondence between formats is a fundamental condition for a data unit to become a valid scene.

[0030] In this embodiment, the SAR single-view complex dataset includes multiple sample images and corresponding label images with sea ice category labels. The label image is the SOD label map in the ASID dataset that matches the sample images. Based on the SOD (Stage Of Development) label, sea ice types are divided into six categories: open water, new ice, young ice, thin one-year ice, thick one-year ice, and old ice. The pixels of each sample image completely retain the real and imaginary complex information of HH / HV dual polarization to capture the complete complex information in the SAR complex data. Each sample image is associated with auxiliary variables such as incident angle and geographical location to achieve spatial matching between the sample image and the label image.

[0031] Specifically, the label image is a two-dimensional raster ground truth reference image with pixel-by-pixel sea ice category annotation information. The label image has the same spatial coordinate system as the corresponding sample image and is labeled with land mask area and effective sea ice pixel area. The effective sea ice pixel area is labeled with the corresponding sea ice category.

[0032] In step S110, in order to achieve the goal of pixel-by-pixel learning, the sample images in the SAR observation data, i.e. the SAR single-view complex dataset, are registered to the corresponding label grid on a scene-by-scene basis.

[0033] In this embodiment, the SOD label image is used as the reference grid. The corresponding sample images in the dataset are resampled using bilinear interpolation to ensure the size of the sample images perfectly matches the label images. Nearest neighbor padding is used to fill in the boundary regions of the sample images. The resampled sample images are then saved as a representation of the SOD label images. Figure 1 The y / x coordinates are determined, while the SOD label image remains in its original state without interpolation, thus achieving accurate spatial registration between the sample image and the label image.

[0034] Furthermore, to address the severe class imbalance problem in the registered dataset, where most classes such as open water and thick one-year ice account for over 92% of the pixels, while the minority classes such as young ice, thin one-year ice, and old ice have extremely sparse pixel distributions, as shown in Table 1, an adaptive step-size cropping strategy is adopted to crop and expand the registered sample images.

[0035] Table 1. Variable Descriptions for SAR Single-Look Complex Dataset

[0036] In this embodiment, when cropping and expanding the registered sample image using an adaptive step-size cropping strategy, small-step cropping is applied to regions containing a few types of sea ice, while large-step cropping is applied to regions dominated by most types of sea ice, resulting in initial cropped patches. These initial cropped patches undergo quality screening, removing patches with a mask pixel ratio exceeding 30%, and retaining only patches containing at least two valid sea ice categories, thus obtaining valid patches. For each scene's sample image, valid patches are selected using random sampling without replacement, with no more than 50 valid patches selected to complete the sample image cropping and expansion process.

[0037] Specifically, the sea ice category codes on the SOD label map are first scanned and identified to accurately identify areas in the sample image that contain a minority of sea ice types, such as young ice, thin one-year ice, and old ice, as well as areas dominated by the majority of sea ice types, such as open water and thick one-year ice. For the identified sample image areas containing a minority of sea ice types, a small step size of 128 pixels is used for cropping to maximize the generation of candidate slices for this type of area. For the sample image areas dominated by the majority of sea ice types, a large step size of 256 pixels is used for cropping to reasonably control the total number of samples generated for this type of area. Based on the above differentiated step size cropping method, the initial cropped patches of the sample image are obtained.

[0038] Furthermore, all initial cropped patches undergo quality screening, removing patches with a mask pixel ratio exceeding 30%, and retaining only patches containing at least two valid sea ice categories to obtain valid patches. Finally, for each scene's sample images, valid patches are selected using random sampling without replacement, with no more than 50 valid patches chosen to complete the sample cropping and expansion for a single scene. This process expands the dataset to 1318 patches of size 256×256, significantly improving the class balance of the dataset and ensuring sufficient sample representation for a few sea ice categories.

[0039] like Figure 2The diagram illustrates the class distribution after applying the adaptive step-size pruning strategy. The red bars represent the class distribution of the training set after pruning and expansion, while the blue bars represent the class distribution of the original data. A comparison clearly shows a significant improvement in class balance: the pixel proportion of open water decreased from 59.9% to 47.6%, indicating a marked reduction in its dominance, while the pixel proportions of various rare sea ice categories increased substantially. Specifically, the pixel proportion of thin one-year ice increased from 0.3% to 2.5%, an increase of over eight times; the pixel proportion of old ice increased from 2.4% to 9.9%, an increase of nearly four times; and the pixel proportions of other minority categories such as young ice and new ice also increased to varying degrees. This data rebalancing mechanism effectively compensates for the sparse minority class samples in the original data, ensuring sufficient sample representation for all types of sea ice, providing reliable data support for the effective training of subsequent deep learning models.

[0040] In this embodiment, in order to prevent spatial information leakage, all processed sample data are divided into training set, validation set and test set according to the scenario, with a division ratio of 70%:20%:10% to prevent spatial information leakage, and finally a training dataset that meets the training requirements of the model is constructed.

[0041] In step S120, in order to systematically evaluate the contribution of complex information to sea ice classification, complex features were extracted based on the constructed dataset, and amplitude features and phase features were obtained respectively.

[0042] In this embodiment, amplitude feature extraction is performed on the sample data in the training dataset, including: calculating backscattering intensity data based on the complex information of the real and imaginary parts of the HH / HV dual polarization in the sample data; applying a spatial smoothing filter to denoise the backscattering intensity data in the form of linear power, preserving the variation trend of the backscattering intensity; and normalizing the dynamic range of the denoised backscattering coefficients within a preset range to obtain the amplitude feature.

[0043] Specifically, amplitude characteristics are key features for characterizing the surface roughness and dielectric properties of sea ice. However, speckle noise and outlier interference in the original SAR sample data can affect the effectiveness of the characteristics. Therefore, this method first calculates the backscattering intensity data of the corresponding polarization based on the complex information of the real (I) and imaginary (Q) parts of the HH / HV dual polarizations according to the following formula, which is expressed as:

[0044] Furthermore, the backscattering intensity data in the form of linear power is then adaptively processed.

[0045] In this embodiment, unlike the conventional method of directly converting the original intensity data into the logarithmic domain, a spatial smoothing filter with a kernel size of 7×7 is first applied to the linear power data. This operation can effectively suppress speckle noise while preserving the original variation trend of backscattering intensity to the greatest extent, avoiding the loss of feature information. At the same time, in order to match the physical distribution characteristics of sea ice, physical constraint-based normalization processing is performed on the denoised backscattering intensity data. According to the statistical characteristics of the constructed SAR complex dataset, the dynamic range of the backscattering coefficient is truncated within a preset range of [-45, 5] dB. This adaptive truncation method can effectively eliminate outlier interference caused by instrument noise, specular reflection, and other factors, significantly enhance the feature contrast between the low scattering region (flat ice) and the high scattering region (rough ice), and finally obtain high-quality amplitude features that can be directly used for model training.

[0046] Specifically, the dynamic range of the backscattering coefficient [-45, 5] dB is an empirical threshold set by combining the sensor's system noise and the microwave scattering characteristics of sea ice itself. After removing outliers, [-45, 5] dB can cover most of the effective scattering range of sea ice in polar regions. This range is adjustable. For more refined adjustments, histogram statistics can be performed on the entire dataset, and adaptive truncation can be performed using the 1% and 99th percentiles.

[0047] In this embodiment, phase features can capture the fine structural details of sea ice that are difficult to identify based on amplitude information alone, and are an important supplement to amplitude features. The single-channel absolute phase has randomness due to path delay and cannot be directly used for sea ice classification. Therefore, the phase difference (IPD) between the HH and HV polarization channels, which can reveal the coherent structure of the scattering target, is selected for phase feature extraction.

[0048] Specifically, phase feature extraction is performed on the sample data in the training dataset, including: calculating the phase difference between the HH and HV polarization channels based on the complex information of the real and imaginary parts of the HH and HV dual polarizations in the sample data; spatially averaging the real and imaginary parts of the phase difference; and then solving the problem using the arctangent function to obtain the value range. The phase angle is decomposed into cosine and sine components and cyclically encoded to obtain the phase feature.

[0049] Furthermore, based on the real parts of the HH and HV bipolarizations in the sample data... virtual part The complex representation of bipolarization is calculated as follows:

[0050] In the above formula, Then, using the conjugate multiplication formula:

[0051] in The phase difference between the HH and HV polarization channels can be calculated using the formula that represents the conjugate of the HV polarization complex number. .

[0052] Furthermore, to avoid feature instability caused by directly calculating the phase angle from noisy pixels, this method does not directly solve for the phase angle of the pixel. Instead, it first performs spatial averaging on the real and imaginary parts of the IPD calculation result to effectively estimate the expected value of local coherence and ensure the stability of the phase representation. Subsequently, based on the spatially averaged real and imaginary parts, the phase angle is calculated using the arctangent function. The calculated range of values ​​is: The phase angle is then determined. Next, to address the periodic discontinuity of this phase angle, a cyclic encoding strategy is employed. Trigonometric functions are used to perform nonlinear mapping on the phase angle feature map pixel by pixel, calculating its continuous cosine component feature map. Characteristic map of sinusoidal components This allows for the construction of a continuous and mathematically consistent feature space for subsequent neural network training, ultimately yielding phase features that can effectively characterize the microstructure of sea ice.

[0053] In this embodiment, a dual-branch feature fusion framework is used to concatenate amplitude and phase features into a feature tensor. The amplitude and phase features are stacked into a tensor of size N by the feature concatenation module. H The tensor W, where dimension N depends on the specific configuration (N=2 for amplitude-only configuration, N=4 for combined configuration). Finally, the fused tensor is fed into a semantic segmentation network to generate the final pixel-by-pixel sea ice classification map.

[0054] Furthermore, when concatenating amplitude and phase features, the dual-polarization amplitude and phase data are packaged into a 4-channel tensor during the initial data preparation stage. For these 4 channels, UNet and DeepLabV3+ use convolutional layers to stack the 4 channels and directly concatenate and fuse them at the input. Specifically, FCN and PSPNet perform the concatenation and fusion after feature extraction.

[0055] Specifically, the semantic segmentation network can be any one of FCN, U-Net, DeepLabV3+, or PSPNet. These models cover a variety of architectural paradigms, such as encoder-decoder, pyramid pooling, and sparse convolution.

[0056] In step S130, to address the severe class imbalance problem in sea ice distribution, a mean-normalized inverse frequency-weighted cross-entropy loss is constructed, with the loss function formula as follows:

[0057] In the above formula, This indicates the total number of valid pixels (non-mask pixels) in the batch. For the true label of the i-th valid pixel, Let be the predicted probability of the i-th valid pixel. The weights are normalized based on the inverse frequency of the training set categories, and the average weight of all valid sea ice categories is 1.

[0058] Specifically, to ensure numerical stability, the weights were normalized so that the average weight of all valid categories (C=6) was 1, expressed as:

[0059] In the above formula, It is the total number of pixels in category c, and the denominator is... This represents the sum of the inverse frequencies of all valid categories. This normalization step ensures that the magnitude of the weighted loss remains consistent with the scale of the original loss.

[0060] In this embodiment, the model is optimized using the AdamW optimizer, with an initial learning rate of... The weight decays to Performance evaluation uses Overall Accuracy (OA) and Macro-F1 score, both of which are calculated only on valid pixels (excluding masks). OA represents the proportion of correctly classified pixels out of the total number of valid pixels:

[0061] In the above formula, C=6 represents the total number of categories. This represents the total number of valid pixels. The Macro-F1 score is defined as the arithmetic mean of the F1 scores for each class, ensuring that all classes have equal weight regardless of their sample size.

[0062] In the above formula, and These represent the precision and recall of category c, respectively.

[0063] In this embodiment, the core architecture used for feature extraction and model training in this method is a dual-branch processing framework. This framework is specifically designed for the amplitude and phase information characteristics of SAR single-look complex data, realizing the separation, extraction, targeted processing and effective fusion of the two types of features. It is a key design to improve the accuracy of sea ice classification.

[0064] In step S140, the SAR single-look complex image to be measured is acquired. Based on the above dual-branch processing framework, the same feature extraction and fusion operation as the training data is performed on the measured image. The fused features are then input into the sea ice classification model that has been trained and validated. After the model performs pixel-by-pixel inference calculation, the sea ice classification result is output, thus completing the sea ice classification.

[0065] Specifically, the dual-branch processing framework is divided into three core stages: complex data processing, feature fusion, and semantic segmentation. Taking the original SAR single-look complex image as input, it is fed into two independent feature extraction branches for dedicated processing. The amplitude extraction branch sequentially completes intensity calculation, speckle removal and noise reduction, logarithmic scaling transformation and normalization operations to extract amplitude features that can accurately characterize the surface roughness and dielectric properties of sea ice. The phase extraction branch sequentially completes inter-channel cross product, complex domain spatial filtering and cyclic encoding operations to extract phase features that can capture the fine structural details of sea ice, effectively making up for the information shortcomings of amplitude features in distinguishing easily confused sea ice categories. Subsequently, the amplitude and phase features extracted from the two branches are stacked and fused through the feature splicing module to generate a feature tensor of size N×H×W. Finally, the fused feature tensor is input into a sea ice classification model composed of classic semantic segmentation networks such as FCN, U-Net, DeepLabV3+, and PSPNet. After model inference, a pixel-by-pixel sea ice classification map is output, achieving accurate identification of six types of sea ice: open water, new ice, young ice, thin one-year ice, thick one-year ice, and old ice.

[0066] In this embodiment, after acquiring the SAR single-look complex image to be measured, it is also necessary to perform routine SAR image preprocessing operations, including orbit correction, radiometric calibration, and noise removal.

[0067] like Figure 3 The diagram shown illustrates the dual-branch processing framework of this method, illustrating the complete process from inputting SAR single-look complex data, through amplitude and phase dual-branch feature extraction, to feature fusion, semantic segmentation, and finally outputting a sea ice classification map.

[0068] In this paper, the effectiveness of the proposed method is also demonstrated experimentally. To verify the effectiveness of the proposed dual-branch framework in correctly handling phase information, two different input configurations were set up for comparison. The baseline configuration uses only amplitude intensity as input, denoted as "Amp.", while the proposed framework combines amplitude and phase features, denoted as "Comb.". The training of all models followed the procedure in step S130. The subsequent analysis provides a comprehensive comparison of the four deep learning architectures, systematically presents quantitative evaluation metrics, delves into the specific improvements for each ice type, and discusses the qualitative visualization interpretation of the classification results.

[0069] Table 2 presents a quantitative performance comparison of four representative semantic segmentation models under different input configurations. Experimental results show that introducing complex information consistently and significantly improves OA and Macro-F1 scores across all tested architectures. Specifically, FCN achieves the most significant gain in classification accuracy, with an OA improvement of 10%. The PSPNet network model not only achieves the highest overall performance (both metrics reaching 72%), but also records the largest increase in Macro-F1 score, improving by 10%. Similarly, U-Net and DeepLabV3+ also show significant enhancements, with Macro-F1 scores improving by 8%, further demonstrating that properly processed phase information is crucial for recognizing sea ice features. In summary, incorporating complex information improves OA by more than 5% and Macro-F1 by more than 8% for all models. These findings confirm that relying solely on amplitude information constitutes a bottleneck in sea ice classification, and that this limitation can be effectively alleviated by rationally utilizing SAR complex data.

[0070] Table 2 Performance Comparison Analysis of Four Semantic Segmentation Models

[0071] To explore the specific sources of performance improvement, the F1 scores of the PSPNet model on the validation set for each category were further analyzed, such as... Figure 4 As shown, the red bars represent the proposed combined method, and the blue bars represent the amplitude-only benchmark. While the OW classification maintains high accuracy and robustness, substantial breakthroughs were observed across all sea ice categories, particularly Yi and Thin FYI. In the amplitude-only benchmark configuration (Amp.), distinguishing different sea ice development stages is extremely challenging, with F1 scores generally around 60% for each sea ice category. This is attributed to the fact that different sea ice types (such as Yi and Thin FYI) tend to exhibit similar backscattering intensities, leading to significant confusion in the amplitude domain. However, the introduction of phase information effectively mitigates this confusion. Figure 4 As shown, the F1 score of YI was significantly improved by 12.1%, while Thin FYI and OI also achieved significant gains of 11.7% and 11.6%, respectively. Although the amplitude responses of these sea ice types overlap, they exhibit distinctly different surface roughness and dielectric properties, which manifests in the data as differences in phase correlation and inter-channel phase difference. The proposed framework effectively utilizes these phase cues to optimize the classification boundary, demonstrating that phase information provides key discriminative features for distinguishing complex sea ice types.

[0072] To visually verify the effectiveness of IPD, Figure 5The classification results of the PSPNet model are shown, achieving the highest improvement in evaluation metrics. From left to right, the columns correspond to SAR (HH) images, ground truth (GT) ASID, predictions from the Amp. model (amplitude only), and predictions from the Comb. model (incorporating complex information), respectively. In the first scenario (first row), the model faces severe inter-class confusion. As shown in the GT, the left region is dominated by NI. However, because NI is similar to the backscattering characteristics of thicker ice types in the amplitude domain, the Amp. model incorrectly classifies almost the entire region as Thick FYI. By introducing phase information, the Comb. model successfully corrects this confusion and accurately identifies the NI region, producing results consistent with the GT. The challenge in the second example (second row) lies in distinguishing adjacent developmental stages: NI and YI. In the Amp. prediction, these two categories are merged into a single category (YI) because their amplitude textures are almost indistinguishable. In contrast, the Comb. model uses phase features to capture subtle surface differences, thus successfully delineating the boundary between the NI and YI regions. Finally, the third row focuses on OI identification, which typically exhibits complex textures. Although the baseline model struggles to maintain spatial continuity, leading to severe undersegmentation of the OI region in the prediction results, the introduction of IPD suppresses segmentation noise. This restores the structural integrity of the OI, consistent with the significant improvement in the F1 score of the OI observed in the quantitative analysis.

[0073] The aforementioned sea ice classification method based on the fusion of amplitude and phase data from SAR complex data effectively overcomes the long-standing technical challenges of phase entanglement and noise interference in traditional methods by implementing targeted processing of phase information through complex domain spatial filtering and cyclic coding, thus fully preserving the structural characteristics and physical integrity of the phase information. Experimental results fully demonstrate that the introduction of inter-channel phase difference (IPD) has achieved a key breakthrough in distinguishing between open water (OW) and thin annual ice (Thin FYI). Due to their similar backscattering intensities, these two types of sea ice are prone to severe classification confusion in benchmark models that rely solely on amplitude information. However, phase features can accurately capture the dielectric properties and microscale roughness differences unique to different sea ice types that cannot be reflected in the amplitude domain, thereby achieving effective differentiation of easily confused sea ice categories and significantly improving classification accuracy.

[0074] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0075] In one embodiment, such as Figure 6 As shown, a sea ice classification device based on SAR complex data amplitude and phase fusion is provided, including: a dataset acquisition module 200, a training dataset construction module 210, a pixel-by-pixel sea ice category prediction module 220, a model training module 230, and a sea ice classification module 240, wherein: The dataset acquisition module 200 is used to acquire a SAR single-view complex dataset, which includes multiple sample images and label images with sea ice category labels corresponding to each image. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. The training dataset construction module 210 is used to perform spatial registration processing on the SAR single-look complex dataset. It uses the label image as the reference grid to perform bilinear interpolation resampling and boundary nearest neighbor filling on the corresponding sample image to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample image to obtain the processed sample data. The training dataset is constructed based on all the processed sample data. The pixel-by-pixel sea ice category prediction module 220 is used to extract amplitude features and phase features from the sample data in the training dataset, respectively, and use a dual-branch feature fusion framework to concatenate the amplitude features and phase features into a feature tensor. The feature tensor is then input into a semantic segmentation network to predict the pixel-by-pixel sea ice category. The model training module 230 is used to construct a mean-normalized inverse frequency-weighted cross-entropy loss function based on the prediction results and the corresponding label images, and to train and validate the semantic segmentation network to obtain the sea ice classification model. The sea ice classification module 240 is used to acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

[0076] Specific limitations regarding the sea ice classification device based on SAR complex data amplitude and phase fusion can be found in the limitations of the sea ice classification method based on SAR complex data amplitude and phase fusion mentioned above, and will not be repeated here. Each module in the aforementioned sea ice classification device based on SAR complex data amplitude and phase fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0077] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a sea ice classification method based on SAR complex data amplitude and phase fusion. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0078] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0079] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Obtain a SAR single-view complex dataset, which includes multiple sample images and corresponding label images with sea ice category labels. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. Spatial registration processing is performed on the SAR single-view complex dataset. Bilinear interpolation resampling and boundary nearest neighbor filling are performed on the corresponding sample images using the label image as the reference grid to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample images to obtain the processed sample data. A training dataset is constructed based on all the processed sample data. Amplitude and phase features are extracted from the sample data in the training dataset respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor. The feature tensor is then input into a semantic segmentation network to predict the sea ice category pixel by pixel. Based on the prediction results and the corresponding label images, a mean-normalized inverse frequency-weighted cross-entropy loss function is constructed to train and validate the semantic segmentation network, thereby obtaining a sea ice classification model. Acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

[0080] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain a SAR single-view complex dataset, which includes multiple sample images and corresponding label images with sea ice category labels. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. Spatial registration processing is performed on the SAR single-view complex dataset. Bilinear interpolation resampling and boundary nearest neighbor filling are performed on the corresponding sample images using the label image as the reference grid to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample images to obtain the processed sample data. A training dataset is constructed based on all the processed sample data. Amplitude and phase features are extracted from the sample data in the training dataset respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor. The feature tensor is then input into a semantic segmentation network to predict the sea ice category pixel by pixel. Based on the prediction results and the corresponding label images, a mean-normalized inverse frequency-weighted cross-entropy loss function is constructed to train and validate the semantic segmentation network, thereby obtaining a sea ice classification model. Acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A sea ice classification method based on amplitude and phase fusion of SAR complex data, characterized in that, The method includes: Obtain a SAR single-view complex dataset, which includes multiple sample images and corresponding label images with sea ice category labels. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. Spatial registration processing is performed on the SAR single-view complex dataset. Bilinear interpolation resampling and boundary nearest neighbor filling are performed on the corresponding sample images using the label image as the reference grid to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample images to obtain the processed sample data. A training dataset is constructed based on all the processed sample data. Amplitude and phase features are extracted from the sample data in the training dataset respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor. The feature tensor is then input into a semantic segmentation network to predict the sea ice category pixel by pixel. Based on the prediction results and the corresponding label images, a mean-normalized inverse frequency-weighted cross-entropy loss function is constructed to train and validate the semantic segmentation network, thereby obtaining a sea ice classification model. Acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, obtain the pixel-by-pixel sea ice classification result, and complete the sea ice classification.

2. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, The sea ice category labels correspond to six types of sea ice: open water, new ice, young ice, thin one-year ice, thick one-year ice, and old ice.

3. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 2, characterized in that, The label image is a two-dimensional raster ground truth reference image with pixel-by-pixel sea ice category labeling information. The label image has the same spatial coordinate system as the corresponding sample image and is labeled with land mask area and effective sea ice pixel area. The effective sea ice pixel area is labeled with the corresponding sea ice category.

4. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, When cropping and augmenting registered sample images using an adaptive step-size cropping strategy: The regions containing a minority of sea ice types in the registered sample image are cropped with a small step size, while the regions containing the majority of sea ice types are cropped with a large step size, to obtain the initial cropped image patch. The initial cropped patches are quality-screened, removing patches with a mask pixel ratio exceeding 30%, and retaining only patches containing at least two valid sea ice categories to obtain valid patches; For each scene, valid patches corresponding to the sample images are selected using random sampling without replacement, with no more than 50 valid patches selected to complete the cropping and augmentation of the sample images.

5. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, The amplitude feature is extracted from the sample data in the training dataset, including: The backscattering intensity data is calculated based on the complex information of the real and imaginary parts of the HH / HV dual polarization in the sample data. A spatial smoothing filter is applied to the backscattering intensity data in linear power form to denoise it while preserving the variation trend of the backscattering intensity. The dynamic range of the denoised backscattering coefficients is truncated within a preset range and normalized to obtain the amplitude characteristics.

6. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, Phase feature extraction is performed on the sample data in the training dataset, including: Based on the complex information of the real and imaginary parts of the HH and HV dual polarizations in the sample data, the phase difference between the HH and HV polarization channels is calculated. The real and imaginary parts of the phase difference are spatially averaged separately, and then the range of values ​​is obtained by solving the arctangent function. The phase angle; The phase angle is decomposed into cosine and sine components and cyclically encoded to obtain the phase feature.

7. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, The semantic segmentation network can be any one of FCN, U-Net, DeepLabV3+, and PSPNet.

8. The sea ice classification method based on SAR complex data amplitude and phase fusion according to claim 1, characterized in that, The mean-normalized inverse frequency-weighted cross-entropy loss function calculates the loss only in the valid sea ice pixel region, ignoring the land mask region. The loss function formula is as follows: In the above formula, This indicates the total number of valid pixels in the batch. For the true label of the i-th valid pixel, Let be the predicted probability of the i-th valid pixel. The weights are normalized based on the inverse frequency of the training set categories, and the average weight of all valid sea ice categories is 1.

9. A sea ice classification device based on SAR complex data amplitude and phase fusion, characterized in that, The device includes: The dataset acquisition module is used to acquire SAR single-view complex dataset, which includes multiple sample images and label images with sea ice category labels corresponding to each image. The pixels of each sample image are complex information of the real and imaginary parts of HH / HV dual polarization, and are associated with auxiliary variables such as incident angle and geographical location. The training dataset construction module is used to perform spatial registration processing on the SAR single-look complex dataset. It uses the label image as the reference grid to perform bilinear interpolation resampling and boundary nearest neighbor filling on the corresponding sample image to achieve registration. Then, an adaptive step-size cropping strategy is used to crop and expand the registered sample image to obtain the processed sample data. The training dataset is constructed based on all the processed sample data. The pixel-by-pixel sea ice category prediction module is used to extract amplitude and phase features from the sample data in the training dataset, respectively. Using a dual-branch feature fusion framework, the amplitude and phase features are concatenated into a feature tensor, and then the feature tensor is input into the semantic segmentation network to predict the pixel-by-pixel sea ice category. The model training module is used to construct a mean-normalized inverse frequency-weighted cross-entropy loss function based on the prediction results and the corresponding label images, and to train and validate the semantic segmentation network to obtain the sea ice classification model. The sea ice classification module is used to acquire the SAR single-view complex image to be measured, input the amplitude and phase fusion features of the SAR single-view complex image into the sea ice classification model, and obtain the pixel-by-pixel sea ice classification result to complete the sea ice classification.

Images