Sea ice type identification method and device of synthetic aperture radar image

Abstract

The invention discloses a sea ice type identification method and device based on a synthetic aperture radar image, and belongs to the technical field of satellite remote sensing marine monitoring. The method comprises the steps of 1, preprocessing a dual-polarized SAR image, and extracting input feature data; step 2, constructing a finely labeled sea ice type data set covering a generic north pole scene by combining visual interpretation and external auxiliary data; step 3, inputting the constructed finely labeled sea ice type data set into a sea ice type identification model improved based on a DeepLabv3 + architecture, and performing iterative optimization through a multi-graphics card parallel training mode to generate a high-precision sea ice type identification model; and 4, performing sea ice type identification on the extracted input feature data by using a high-precision sea ice type identification model. According to the method, the incidence angle dependency related to the sea ice type in the original SAR data is more accurately reserved, and the precision of the model is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite remote sensing ocean monitoring, and in particular relates to a method and a device for identifying sea ice types using synthetic aperture radar images. Background Art

[0002] Sea ice is an important part of the Arctic environment and a key indicator and amplifier of global climate change. It plays an indispensable role in the understanding and prediction of global climate change. In addition, accurate and timely high-resolution sea ice information is crucial for human activities in the Arctic, such as offshore operations, ship navigation, fishing activities, and geological exploration. Therefore, sea ice monitoring and mapping are not only an important basis for ensuring the safety and efficiency of these activities, but also have far-reaching significance for protecting the Arctic environment. Currently, remote sensing data is the main source of information on the status of Arctic sea ice.

[0003] Synthetic Aperture Radar (SAR) has become an important technical means for Arctic sea ice mapping because of its high spatial resolution and all-weather imaging capabilities that are not affected by weather and lighting conditions. With the rapid development of SAR technology, a large amount of sea ice data has been accumulated in the Arctic region. Faced with the huge amount of data, the interpretation of ice maps has become time-consuming and inefficient by relying on manual analysis by sea ice experts. Therefore, there is an urgent need to develop robust and automated sea ice type identification technology to support the operational needs of Arctic sea ice mapping.

[0004] Early automatic sea ice type recognition algorithms based on SAR data mostly used traditional machine learning methods, such as random forests, support vector machines, and artificial neural networks. In recent years, with the rapid development of deep learning technology, new opportunities for automatic sea ice type recognition using large amounts of SAR data have gradually emerged. Automated sea ice type recognition algorithms based on deep learning have been applied to Arctic regions such as the Svalbard Islands, the Beaufort Sea, and the Hudson Bay. However, existing methods have not yet been verified in their applicability and interannual robustness in the pan-Arctic range, and have certain limitations. The main drawback is the lack of finely annotated training samples covering the pan-Arctic region, which seriously hinders the accuracy of deep learning algorithms in the pan-Arctic range. In addition, there is significant confusion in the characteristics of different sea ice types in the pan-Arctic region, further limiting the application capabilities of existing deep learning models in the pan-Arctic region. Summary of the invention

[0005] To solve the above technical problems, the present invention provides a method and device for identifying sea ice types in synthetic aperture radar images. Based on manual fine annotation, a sea ice type dataset covering the pan-Arctic region is constructed, and an automatic sea ice type identification algorithm for full-scene SAR images is designed. The algorithm can not only accurately identify various sea ice types in the pan-Arctic region (such as open water, initial ice, one-year ice and multi-year ice), but also has excellent inter-annual robustness, providing a new solution for pan-Arctic sea ice mapping.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying sea ice types using synthetic aperture radar images, the method comprising:

[0008] Step 1: preprocess the dual-polarization SAR image and extract input feature data;

[0009] Step 2: Based on the extracted input feature data, a finely annotated sea ice type dataset covering the pan-Arctic scene is constructed by combining visual interpretation with external auxiliary data;

[0010] Step 3: Input the constructed finely labeled sea ice type dataset into the sea ice type recognition model improved based on the DeepLabv3+ architecture, perform iterative optimization through multi-graphics card parallel training, and generate a high-precision sea ice type recognition model;

[0011] Step 4: Use the high-precision sea ice type recognition model to identify the sea ice type of the dual-polarization SAR image to be identified.

[0012] In another aspect, the present invention provides a device for identifying sea ice types using synthetic aperture radar images, comprising:

[0013] An input feature data extraction unit is used to preprocess the dual-polarization SAR image and extract input feature data;

[0014] The dataset construction unit is used to construct a finely annotated sea ice type dataset covering the pan-Arctic scene by combining visual interpretation with external auxiliary data;

[0015] The recognition model acquisition unit is used to input the constructed finely labeled sea ice type dataset into the sea ice type recognition model improved based on the DeepLabv3+ architecture, perform iterative optimization through multi-graphics card parallel training, and generate a high-precision sea ice type recognition model;

[0016] The sea ice type identification unit is used to apply a high-precision sea ice type identification model to perform sea ice type identification on the extracted input feature data.

[0017] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory 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 implement the aforementioned method for identifying sea ice types in a synthetic aperture radar image.

[0018] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for identifying sea ice types in a synthetic aperture radar image.

[0019] The beneficial effects of the present invention are:

[0020] The present invention constructs a finely annotated sea ice type dataset covering the pan-Arctic range, including four main sea ice types: seawater, initial ice, one-year ice, and multi-year ice, covering a variety of ice conditions (such as single sea ice type and mixed scenarios of multiple sea ice types). At the same time, the annotation is combined with visual interpretation, sea ice maps, and high-precision sea ice cover products to ensure the accuracy of the annotated data. This dataset fills the gap in finely annotated data in the pan-Arctic region and provides a solid foundation for algorithm training.

[0021] The present invention directly uses the radar incident angle as the input feature, fully considering its influence on co-polarized backscattering and the difference in incident angle dependence between different sea ice types. This avoids the traditional method of uniform incident angle correction processing for all co-polarized backscattering data, more accurately retains the incident angle dependence related to sea ice type in the original SAR data, and improves the accuracy of the model.

[0022] The present invention uses the whole-scene SAR image as input, which provides a larger receptive field for the sea ice type recognition model and can more effectively alleviate the problem of feature confusion between different sea ice types. In addition, compared with the traditional method of using sliced ​​SAR images, the present invention avoids the influence of splicing traces on the classification results.

[0023] Based on the DeepLabv3+ deep learning structure, this paper proposes a pan-Arctic sea ice type recognition model. The model adopts a spatially optimized ConvNeXt backbone network and combines it with the designed EASPP module to fully increase the receptive field and extract multi-scale feature information, which significantly improves the model's ability to express the characteristics of different sea ice types, thereby improving the recognition accuracy of sea ice types in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a method for identifying sea ice types using synthetic aperture radar images according to the present invention;

[0025] Figure 2Schematic diagram of the process of labeling sea ice types;

[0026] Figure 3 This is a schematic diagram of the sea ice type identification model structure;

[0027] Figure 4 Schematic diagram of the ConvNeXt backbone network structure. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0029] Figure 1 A schematic diagram of a flow chart of a method for identifying sea ice types using synthetic aperture radar images provided by the present invention, the method specifically comprising the following steps:

[0030] Step 1: preprocess the dual-polarization SAR image and extract input feature data;

[0031] By preprocessing the dual-polarization SAR image, multi-channel input data is generated, including co-polarization backscatter (HH), cross-polarization backscatter (HV), cross-polarization ratio (HV / HH) and radar incident angle information.

[0032] Specifically, a large number of Sentinel-1 (S1) satellite Extra Wide swath (EW) mode SAR images covering the pan-Arctic region are first collected, and preprocessing steps such as radiation correction, cross-polarization thermal noise removal, land masking and decibelization are performed.

[0033] Radiation correction: Convert the brightness grayscale value of SAR data into radar backscatter coefficient.

[0034] Cross-polarization thermal noise removal: Reduce the impact of thermal noise on cross-polarization SAR images.

[0035] Land Mask: Removes land areas from an image.

[0036] Decibelization: Convert the backscatter coefficient to decibel units to optimize its dynamic range.

[0037] After the above processing, four-channel data are generated for the input of the sea ice type identification model, including HH polarization backscatter, HV polarization backscatter, cross-polarization ratio and radar incident angle characteristics.

[0038] Step 2: Build a finely labeled pan-Arctic sea ice type dataset;

[0039] Based on the extraction of input feature data, a finely annotated sea ice type dataset covering typical scenes in the pan-Arctic is constructed by combining visual interpretation with external auxiliary data.

[0040] Specifically, we first selected SAR images from different months that covered various ice conditions, including single-type sea ice scenes (such as initial ice, one-year ice, or multi-year ice) and mixed-type scenes (such as seawater mixed with sea ice, and alternating between two or more types of sea ice). Then, we manually interpreted the selected SAR images and verified them with the ice map provided by the United States National Ice Center (USNIC) to generate accurate and detailed annotation results. During the annotation process, the pixels of each SAR image were refined into four main types: seawater, initial ice, one-year ice, and multi-year ice. The annotation process is as follows: Figure 2 Finally, the high-precision Arctic sea ice cover product is used to mask the seawater in the annotated SAR images, and a sea ice type dataset covering typical pan-Arctic scenes is constructed.

[0041] Ultimately, the dataset covers typical sea ice scenes in different regions and months, providing diverse and reliable training samples for model training.

[0042] Step 3: training the sea ice type recognition model;

[0043] The constructed finely annotated dataset is input into the improved sea ice type recognition model based on the DeepLabv3+ architecture, and is iteratively optimized multiple times through multi-graphics card parallel training to finally generate a high-precision sea ice type recognition model.

[0044] Specifically, Figure 3 The schematic diagram of the structure of the sea ice type recognition model in the embodiment of the present invention is shown. The model is targetedly improved based on the DeepLabv3+ deep learning model architecture, in which the black rectangular dashed box marks the improved part of the DeepLabv3+ network structure in this embodiment. First, the model takes the whole scene SAR image as input, including four-channel features of HH polarization backscatter, HV polarization backscatter, cross-polarization ratio and radar incident angle. Secondly, the model adopts an encoder-decoder structure, in which the encoder is responsible for extracting the features of the input SAR image, and the decoder is responsible for mapping these features back to the original resolution to generate pixel-level sea ice type recognition results.

[0045] In this embodiment, the DeepLabv3+ structure is improved as follows: In the encoder, the spatially optimized ConvNeXt backbone network and the newly designed EASPP module are used to replace the Xception backbone network and the ASPP module in the original DeepLabv3+ structure respectively. The ConvNeXt backbone network is used to extract basic features, and its structure is as follows: Figure 4As shown, k represents the convolution kernel size, s represents the step size, and p represents the edge padding operation. The ConvNeXt network consists of five stages, namely the initial stage and the first to fourth stages. Each stage consists of multiple ConvNeXt convolution blocks that integrate depthwise separable convolution, standard convolution, layer normalization, and Gaussian error linear unit (GELU) activation functions. The second, third, and fourth stages all contain downsampling layers to extract features of different scales. At the same time, in order to improve the resolution of the feature map, the initial stage uses 2 2 standard convolutions instead of the original 4 The shallow features are extracted by the initial, first stage and downsampling layers of ConvNeXt and then passed to the decoder to preserve high-resolution detail information. Then, the features extracted by the ConvNeXt backbone network are further integrated by the newly proposed EASPP module. The EASPP module is responsible for extracting and fusing multi-scale features to obtain deep features, providing rich classification information for the decoder. This module consists of five branches: a global pooling layer responsible for capturing global information, three expansion rates of 6, 12, and 18, and a 3D image processing layer with different image processing functions. 3 dilated convolutions and 1 3 3. By cascading these branches in a regular manner with increasing dilation rates, the EASPP module significantly increases the receptive field and thus enhances the ability to distinguish sea ice types.

[0046] In the decoder, the deep features are first upsampled by a factor of four to match the resolution of the shallow features. Then, the deep and shallow features are concatenated and combined through 3 3 standard convolutions generate fusion features. After the 1 convolution layer adaptively selects important features, it performs 4 times upsampling to restore the resolution to the same level as the input SAR image. Finally, the softmax function is used to map the features into probability distributions to determine the highest probability value of the sea ice type (seawater, initial ice, one-year ice, multi-year ice) to which each pixel belongs, thereby obtaining the sea ice type recognition result at the pixel level.

[0047] In addition, this embodiment adopts the cross entropy loss function, combined with the AdamW optimizer and multi-GPU parallel training strategy, to accelerate the model training process and smooth parameter optimization. At the same time, the training effect is evaluated by indicators such as overall accuracy, intersection-over-union ratio and F1 score. After multiple training and verification, the network parameters are updated by minimizing the loss function and back-propagating the gradient until the accuracy of the model output reaches or exceeds the ideal value and is saved as the final sea ice type recognition model.

[0048] Step 4: Apply the model to identify sea ice types;

[0049] The trained sea ice type recognition model is used to perform pixel-level classification on the preprocessed dual-polarization SAR images to generate sea ice type recognition results, including four main types: sea water, initial ice, one-year ice and multi-year ice.

[0050] In summary, the present invention provides a method for identifying sea ice types from synthetic aperture radar images. Through manual visual interpretation combined with external data assistance, a finely annotated sea ice type dataset covering typical pan-Arctic scenes is constructed, covering four sea ice types: seawater, initial ice, one-year ice, and multi-year ice. The method uses the whole-scene SAR image as input, and comprehensively considers the polarization characteristics (co-polarization, cross-polarization backscattering, and cross-polarization ratio) and radar incident angle information. At the same time, the DeepLabv3+ structure is improved using the spatially optimized ConvNeXt network and the designed EASPP module, and the network is trained in combination with the constructed dataset, so that the deep learning model can fully learn and distinguish the characteristics of different sea ice types. After multiple iterative training, the obtained sea ice type recognition model has a high-precision classification function, which can effectively distinguish SAR images in mixed scenes of multiple sea ice types, realizes the automatic and accurate classification of sea ice types in the pan-Arctic region, and has interannual robustness, providing a new solution for Arctic sea ice mapping and business applications.

[0051] In another aspect, the present invention provides a device for identifying sea ice types using synthetic aperture radar images, comprising:

[0052] An input feature data extraction unit is used to preprocess the dual-polarization SAR image and extract input feature data;

[0053] The dataset construction unit is used to construct a finely annotated sea ice type dataset covering the pan-Arctic scene by combining visual interpretation with external auxiliary data;

[0054] The recognition model acquisition unit is used to input the constructed finely labeled sea ice type dataset into the sea ice type recognition model improved based on the DeepLabv3+ architecture, perform iterative optimization through multi-card parallel training, and generate a high-precision sea ice type recognition model;

[0055] The sea ice type identification unit is used to apply a high-precision sea ice type identification model to perform sea ice type identification on the extracted input feature data.

[0056] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory 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 implement the aforementioned method for identifying sea ice types in a synthetic aperture radar image.

[0057] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for identifying sea ice types in a synthetic aperture radar image.

[0058] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying sea ice types using synthetic aperture radar images, characterized in that: The method comprises: Step 1: preprocess the dual-polarization SAR image and extract input feature data; Step 2: Based on the extracted input feature data, a finely annotated sea ice type dataset covering the pan-Arctic scene is constructed by combining visual interpretation with external auxiliary data; Step 3: Input the constructed finely labeled sea ice type dataset into the sea ice type recognition model improved based on the DeepLabv3+ architecture, perform iterative optimization through multi-graphics card parallel training, and generate a high-precision sea ice type recognition model; Step 4: Use the high-precision sea ice type recognition model to identify the sea ice type of the dual-polarization SAR image to be identified.

2. The method for identifying sea ice types using synthetic aperture radar images according to claim 1, characterized in that: The step 1 includes preprocessing the dual-polarization SAR image to generate multi-channel input data, including co-polarization backscatter, cross-polarization backscatter, cross-polarization ratio and radar incident angle information.

3. The method for identifying sea ice types using synthetic aperture radar images according to claim 2, characterized in that: The preprocessing of the dual-polarization SAR image includes radiation correction, cross-polarization thermal noise removal, land masking and decibelization.

4. The method for identifying sea ice types using synthetic aperture radar images according to claim 1, characterized in that: The step 2 comprises: selecting SAR images of different months and covering various ice conditions, including single-type sea ice scenes and mixed-type scenes, wherein the single-type sea ice scenes include initial ice, one-year ice or multi-year ice, and the mixed-type scenes include a mixture of seawater and sea ice, and alternation of two or more types of sea ice; generating fine-marking results after manually visually interpreting the selected SAR images and verifying them; using high-precision Arctic sea ice cover products to perform seawater masking on the fine-marked SAR images, and constructing a fine-marked sea ice type dataset covering pan-Arctic scenes.

5. The method for identifying sea ice types using synthetic aperture radar images according to claim 1, characterized in that: The step 3 comprises: The whole-scene SAR image in the constructed finely annotated sea ice type dataset is input into the sea ice type recognition model, including four-channel features of co-polarization backscatter, cross-polarization backscatter, cross-polarization ratio and radar incident angle; The sea ice type recognition model uses an encoder to extract shallow features and deep features of the input full-scene SAR image, and uses a decoder to fuse the shallow features and the deep features and map them back to the original resolution to generate a pixel-level sea ice type recognition result; The cross entropy loss function is adopted, combined with the AdamW optimizer and multi-GPU parallel training strategy to accelerate the model training process and smooth parameter optimization. The training effect is evaluated by overall accuracy, intersection-over-union ratio and F1 score. The network parameters are updated by minimizing the loss function and back-propagating the gradient until the accuracy of the sea ice type recognition model output reaches or exceeds the ideal value and is saved as the final high-precision sea ice type recognition model.

6. The method for identifying sea ice types using synthetic aperture radar images according to claim 5, characterized in that: The encoder includes a spatially optimized ConvNeXt backbone network and an EASPP module; The ConvNeXt backbone network is used to extract shallow features, including an initial stage and stages 1 to 4, each stage is composed of multiple ConvNeXt convolution blocks, and stages 2, 3 and 4 also include a downsampling layer; The EASPP module is used to extract and fuse shallow features to obtain multi-scale deep features. It includes five branches: a global pooling layer responsible for capturing global information, a 3-layer deep pooling layer with three expansion rates of 6, 12, and 18, and a 4-layer deep pooling layer with three expansion rates of 6, 12, and 18. 3 dilated convolutions and 1 3 Standard convolution of 3; The decoder connects the deep features and shallow features and passes 3 3 standard convolution generates fusion features, and the subsequent 1 After the convolution layer adaptively selects important features, it performs 4-fold upsampling and uses the softmax function to map the adaptively selected important features into probability distribution to determine the highest probability value of the sea ice type to which each pixel belongs.

7. The method for identifying sea ice types using synthetic aperture radar images according to claim 1, characterized in that: The step 4 includes using a high-precision sea ice type recognition model to perform pixel-level classification on the pre-processed dual-polarization SAR image to be identified, and generating sea ice type recognition results, including four types: seawater, initial ice, one-year ice and multi-year ice.

8. A device for identifying sea ice types using synthetic aperture radar images, characterized in that: include: An input feature data extraction unit is used to preprocess the dual-polarization SAR image and extract input feature data; The dataset construction unit is used to construct a finely annotated sea ice type dataset covering the pan-Arctic scene based on the extracted input feature data and by combining visual interpretation with external auxiliary data; The recognition model acquisition unit is used to input the constructed finely labeled sea ice type dataset into the sea ice type recognition model improved based on the DeepLabv3+ architecture, perform iterative optimization through multi-card parallel training, and generate a high-precision sea ice type recognition model; The sea ice type recognition unit is used to apply a high-precision sea ice type recognition model to perform sea ice type recognition on the dual-polarization SAR image to be recognized.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the method for identifying sea ice types in synthetic aperture radar images as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, and when the instructions are executed by a processor, the processor can implement a method for identifying sea ice types in a synthetic aperture radar image as described in any one of claims 1 to 7.

Images