A deep learning-based sea ice classification and density inversion method

Abstract

The present invention discloses a deep learning-based sea ice classification and density inversion method, which relates to the field of remote sensing image processing technology and aims to solve the problems of the existing technology. The method comprises obtaining raw dual-polarization synthetic aperture radar data and preprocessing it; performing superpixel segmentation on the preprocessed data to obtain multiple superpixels; calculating the posterior probability of the superpixel using a conditional random field model to determine the uncertain superpixel unit; using the uncertain superpixel unit as input, outputting the sea ice and seawater boundary lines within the uncertain superpixel unit based on the Ice-WaterNet network model, and combining them to obtain a sea ice classification result map; and performing sea ice density inversion on the sea ice classification result map to obtain a sea ice density inversion result. The present invention improves the inversion accuracy of sea ice classification and density, and can play a more important role in global climate change monitoring, marine environmental protection, polar resource development and utilization, and other aspects.

Description

Technical Field

[0001] The present invention relates to a sea ice classification and density inversion method based on deep learning, and belongs to the technical field of remote sensing image processing. Background Art

[0002] Sea ice density is not only a key parameter for marine environmental monitoring and climate research, but also serves as an important reference for ensuring the safety of maritime activities and promoting the development and utilization of polar resources. With global climate change and significant shifts in sea ice coverage, sea ice has a significant impact on the global climate system, water balance, and heat transfer, making it a crucial factor in monitoring global climate change. Sea ice also poses significant challenges to maritime activities such as navigation, seabed mining, and polar ocean exploration, and can even trigger catastrophic events. Therefore, accurate and efficient sea ice classification is crucial.

[0003] Synthetic Aperture Radar (SAR) technology offers unique advantages in sea ice monitoring, thanks to its all-day, all-weather observation capabilities, strong penetration, and resistance to cloud cover. The increasing volume of SAR imagery and its increasing spatial and temporal resolution have made SAR a primary data source for sea ice classification.

[0004] SAR-based sea ice classification research not only helps scientists understand the physical properties and distribution patterns of sea ice and its interaction with climate change, but also provides a scientific basis for marine environmental forecasts and sea ice disaster warnings. Specifically:

[0005] On the one hand, sea ice classification helps reveal the inherent connection between sea ice and the climate system. Through detailed sea ice classification, key information such as its type, extent, thickness, and age can be obtained. This allows for analysis of the impact of sea ice changes on the global climate, such as its influence on ocean heat exchange and its regulation of atmospheric temperature. This information is crucial for predicting global climate change trends and developing response strategies.

[0006] Furthermore, sea ice classification information is a crucial reference for ensuring navigational safety. Real-time monitoring of sea ice distribution and types provides accurate route planning and ice avoidance guidance for ships, reducing the risk of accidents such as groundings and collisions caused by sea ice. Furthermore, sea ice classification provides strong support for exploration and resource development activities in polar waters, ensuring smooth progress and the safety of personnel.

[0007] Furthermore, sea ice classification helps assess the impact of sea ice changes on marine ecosystems. Sea ice is a vital component of polar ecosystems, and its changes directly impact the habitats and food chains of polar organisms. Using this information, we can monitor the growth and decline of sea ice and assess its impact on marine ecosystems, providing a basis for decision-making on conservation measures and promoting sustainable development.

[0008] The rapid development of artificial intelligence and deep learning technologies has significantly improved the accuracy and efficiency of sea ice classification. Traditional sea ice classification methods rely primarily on manual visual interpretation and remote sensing image processing, which are subject to significant subjectivity and low efficiency. However, SAR sea ice classification methods based on deep learning can automatically extract sea ice features from SAR imagery, enabling efficient and accurate classification. This not only significantly improves the efficiency and accuracy of sea ice classification but also opens up the possibility of intelligent and automated sea ice monitoring and research.

[0009] However, due to rising temperatures during the melting period, sea ice begins to melt, gradually transforming from a solid to a liquid or semi-solid state. This process is accompanied by a decrease in ice thickness, the formation of meltwater on the ice surface, and the destruction of the ice structure. These changes make the sea ice diverse and complex in both appearance and internal structure, making the sea ice characteristics observed by remote sensing complex and difficult to describe using unified characteristics and models. In addition, the dynamic changes of sea ice are a major difficulty in classification. During the melting process, sea ice may be affected by various external forces such as wind, waves, and currents, causing it to drift, break up, and accumulate. These dynamic changes not only change the distribution and morphology of sea ice but also increase the difficulty of sea ice classification.

[0010] In summary, the SAR sea ice classification during the melt period is affected by many factors, including the complexity of the physical changes of sea ice, the interference of environmental factors, and the complexity of data processing and analysis. Summary of the Invention

[0011] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a sea ice classification and density inversion method based on deep learning, so as to improve the inversion accuracy of sea ice classification and density, and play a more important role in global climate change monitoring, marine environmental protection, polar resource development and utilization, etc.

[0012] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0013] The present invention provides a sea ice classification and density inversion method based on deep learning, comprising:

[0014] Obtain raw dual-polarization synthetic aperture radar data;

[0015] Preprocessing the original dual-polarization synthetic aperture radar data to obtain preprocessed dual-polarization synthetic aperture radar data;

[0016] Perform superpixel segmentation on the preprocessed dual-polarization synthetic aperture radar data to obtain multiple superpixels;

[0017] The conditional random field model is used to calculate the posterior probability of superpixels, and the superpixels whose posterior probability is within the set interval are taken as uncertain superpixel units;

[0018] Taking the uncertain superpixel unit as input, the sea ice and sea water boundary line within the uncertain superpixel unit is obtained based on the output of the pre-trained Ice-WaterNet network model;

[0019] The sea ice and seawater boundaries within all uncertain superpixel units are combined to obtain the sea ice classification result map;

[0020] Perform sea ice density inversion based on the sea ice classification result map to obtain the sea ice density inversion result;

[0021] Among them, the Ice-WaterNet network model is constructed based on the U-Net network model. The construction method includes connecting each convolutional layer in the decoder of the U-Net network model to an uncertainty measurement module. The uncertainty measurement module is used to divide the uncertain superpixel unit into an uncertain area and a certain area, and calculate the uncertainty of the features in the uncertain area to give high weights to features with large uncertainty.

[0022] Furthermore, the preprocessing of the original dual-polarization synthetic aperture radar data to obtain preprocessed dual-polarization synthetic aperture radar data includes:

[0023] Read noise parameters and radiation calibration parameters from raw dual-polarization synthetic aperture radar data;

[0024] The remote sensing image pixel brightness value of the original dual-polarization synthetic aperture radar data is converted into the backscattering coefficient according to the noise parameter and the radiation calibration parameter. The expression is as follows:

[0025] ;

[0026] in, represents the backscattering coefficient, Represents the brightness value of the remote sensing image pixel, represents the noise parameter, represents the radiation calibration parameter;

[0027] The backscatter coefficient is converted into the backscatter coefficient of the reference incident angle, which is the preprocessed dual-polarization synthetic aperture radar data. Its expression is as follows:

[0028] ;

[0029] in, represents the corrected backscatter coefficient, i.e., the preprocessed dual-polarization synthetic aperture radar data, represents the backscattering coefficient, represents the reference incident angle, represents the slope coefficient, represents the local angle of incidence.

[0030] Furthermore, the superpixel segmentation performed on the preprocessed dual-polarization synthetic aperture radar data adopts a meanshift segmentation method.

[0031] Furthermore, the expression of the conditional random field model is as follows:

[0032] ;

[0033] in, represents the posterior probability, represents the normalization coefficient, Indicates the The unit potential energy function of superpixels, Indicates the superpixels and The potential energy function of superpixels; Indicates the The class of superpixels, Indicates the superpixel classes, where the classes are classified as sea ice or sea water; and Both represent superpixel sets and do not contain each other; Indicates the The backscatter coefficient of superpixels, represents the model weight parameter, represents the model input features, , Indicates the superpixels and The cross-correlation function of superpixels, represents the Dilick function, when hour, The value is 1, otherwise it is 0.

[0034] Furthermore, the Ice-WaterNet network model is an encoder-decoder structure, wherein the number of convolutional layers in the encoder and decoder are both 5, the encoder is used to calculate channel and spatial features, and the decoder is used to use multi-scale context information to extract different types of floating ice, and each convolutional layer in the decoder is provided with an attention module, which is used to use the difference between HH and HV channels to distinguish sea ice and seawater to determine the buffer area.

[0035] Furthermore, the data processing process of the uncertainty measurement module includes:

[0036] For each uncertain superpixel unit, the buffer area corresponding to the maximum output probability during model deconvolution is recorded as the uncertain area;

[0037] Using a rectangular model as a filtering window, continuously changing the perceptual field of view of the filtering window, obtaining the sea ice boundary line within the uncertain region to redivide the uncertain region into a deterministic region and an uncertain region, until the size of the filtering window reaches or exceeds the size of the minimum bounding rectangle of the uncertain superpixel unit, wherein the aspect ratio of the rectangular model is the aspect ratio of the minimum bounding rectangle of the uncertain superpixel unit;

[0038] Calculate the uncertainty of the features in the uncertainty region of the uncertain superpixel unit, and obtain the uncertainty values ​​of the features in the uncertainty region of all uncertain superpixel units;

[0039] The features in the uncertainty region of all uncertain superpixel units are sorted according to their uncertainty values, and features with large uncertainty values ​​are given high weights.

[0040] Furthermore, the uncertainty measurement module calculates the uncertainty of the feature within the uncertainty region in the uncertain superpixel unit using the following expression:

[0041] ;

[0042] in, represents the sea ice forecast uncertainty, represents the uncertainty of seawater prediction, represents the uncertainty characteristics of sea ice, represents the uncertainty characteristics of seawater, Indicates the offset.

[0043] Furthermore, the pre-training method of the Ice-WaterNet network model includes:

[0044] Obtaining an ice condition map dataset, wherein the ice condition map dataset is annotated with a reference true value;

[0045] The Ice-WaterNet network model is trained with the ice map dataset as input and the sea ice and sea water boundary lines as output. During the training process, the model parameters of the Ice-WaterNet network model are adjusted using a loss function. When the loss function value reaches the convergence condition, the pre-trained Ice-WaterNet network model is obtained.

[0046] Furthermore, the loss function adopts a binary cross entropy function, which is expressed as:

[0047] ;

[0048] in, represents the binary cross entropy function, Represents the first The output of the convolutional layer, Indicates the number of convolutional layers of the Ice-WaterNet network model, Indicates the true value of the benchmark.

[0049] Furthermore, the sea ice density inversion is performed based on the sea ice classification result map to obtain the sea ice density inversion result, including:

[0050] Downsampling the sea ice classification result according to a preset resolution to obtain a downsampled sea ice classification result;

[0051] The downsampled sea ice classification results are projected according to the projection grid to obtain the percentage of sea ice pixels in each sea ice classification result projection grid, that is, the sea ice density of each sea ice classification result projection grid;

[0052] The sea ice density of each sea ice classification result projection grid is combined to obtain the sea ice density inversion result.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This paper uses a conditional random field model to extract uncertain sea ice objects at the superpixel level. Then, by building an Ice-WaterNet deep learning model and adopting an iterative strategy to optimize the segmentation edges of superpixels, the accuracy of sea ice classification and density inversion is improved. This method can play a more important role in global climate change monitoring, marine environmental protection, and polar resource development and utilization.

[0055] The present invention can effectively improve the inversion accuracy of sea ice classification and density during the melting period through uncertainty superpixel extraction and uncertainty measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of a flow chart of a method for sea ice classification and density inversion based on deep learning in one embodiment of the present invention;

[0057] Figure 2 Schematic diagram of the processing flow of the Ice-WaterNet model of the sea ice classification and density inversion method based on deep learning in one embodiment of the present invention;

[0058] Figure 3 Schematic diagram of comparison of the intersection-over-union ratio of CRF, U-net, and Ice-WaterNet for a deep learning-based sea ice classification and density inversion method in one embodiment of the present invention;

[0059] Figure 4 Schematic diagram showing the comparison of the accuracy of CRF, U-net, and Ice-WaterNet for sea ice classification and density inversion methods based on deep learning in one embodiment of the present invention;

[0060] Figure 5 Schematic diagram showing a comparison of the recall rates of CRF, U-net, and Ice-WaterNet for a deep learning-based sea ice classification and density inversion method in an embodiment of the present invention;

[0061] Figure 6 Schematic diagram of the comparison of F1 scores of CRF, U-net, and Ice-WaterNet for a deep learning-based sea ice classification and density inversion method in an embodiment of the present invention;

[0062] Figure 7 A schematic diagram of sea ice density inversion results of Ice-WaterNet, a deep learning-based sea ice classification and density inversion method, in one embodiment of the present invention;

[0063] Figure 8 Schematic diagram of ASI's sea ice density inversion results. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0065] like Figure 1 As shown, an embodiment of the present invention provides a sea ice classification and density inversion method based on deep learning, comprising the following steps:

[0066] Original dual-polarization synthetic aperture radar data is obtained. In this embodiment, the original dual-polarization synthetic aperture radar data is original dual-polarization Setninel-1 SAR data.

[0067] Preprocess the raw dual-polarization synthetic aperture radar data, including radiometric calibration and incident angle correction. Specifically:

[0068] Radiometric calibration: Read the noise parameters and radiometric calibration parameters from the annotation file of the original dual-polarization Setninel-1 SAR data, and convert the remote sensing image pixel brightness values ​​of the original dual-polarization Setninel-1 SAR data into backscatter coefficients. The expression is as follows:

[0069] ;

[0070] in, represents the backscattering coefficient, Represents the brightness value of the remote sensing image pixel, represents the noise parameter, Represents the radiation calibration parameters.

[0071] Incident angle correction: The backscatter coefficient is converted to the backscatter coefficient of the reference incident angle, which is the preprocessed dual-polarization synthetic aperture radar data. Its expression is as follows:

[0072] ;

[0073] in, represents the corrected backscatter coefficient, i.e., the preprocessed dual-polarization synthetic aperture radar data, represents the backscattering coefficient, represents the reference incident angle, represents the slope coefficient, represents the local angle of incidence.

[0074] In this embodiment, The value is 2. The value is 26°.

[0075] Perform superpixel segmentation on the pre-processed dual-polarization synthetic aperture radar data (i.e., the corrected backscatter coefficients), specifically:

[0076] Firstly, the preprocessed dual-polarization synthetic aperture radar data are divided into multiple superpixels using the meanshift segmentation method.

[0077] Then, the superpixel-level conditional random field (CRF) model is used to calculate the posterior probability of the superpixel. The expression of the conditional random field model is as follows:

[0078] ;

[0079] in, represents the posterior probability, represents the normalization coefficient, Indicates the The unit potential energy function of superpixels, Indicates the superpixels and The potential energy function of superpixels; Indicates the The class of superpixels, Indicates the superpixel classes, where the classes are classified as sea ice or sea water; and Both represent superpixel sets and do not contain each other; Indicates the The backscatter coefficient of superpixels, represents the model weight parameter, represents the model input features, , Indicates the superpixels and The cross-correlation function of superpixels, represents the Dilick function, when hour, The value is 1, otherwise it is 0.

[0080] The superpixel segmentation process uses the minimum energy function method and outputs the final posterior probability , marking superpixels whose posterior probabilities are within a set interval as uncertain superpixel units, effectively reducing the amount of computation while ensuring the accuracy of sea ice and water classification. In this embodiment, the interval is set to 0.15-0.85.

[0081] Build the Ice-WaterNet network model:

[0082] The Ice-WaterNet network model is built based on the U-Net network model. The Ice-WaterNet network model has an encoder-decoder structure, in which the number of convolutional layers in both the encoder and decoder is 5. The encoder is used to calculate channel and spatial features, and the decoder is used to use multi-scale contextual information to extract different types of floating ice. Each convolutional layer in the decoder is also equipped with an attention module, which explicitly describes long-range global information. It can use the difference between the HH and HV channels to distinguish sea ice and seawater to determine the buffer area, solving the ice and water classification ambiguity problem caused by frost flowers on open waters and ice channels caused by wind and waves, thereby improving the accuracy of ice and water classification.

[0083] The improvement of the Ice-WaterNet network model over the U-Net network model is that each convolutional layer in the decoder of the U-Net network model is connected to an uncertainty measurement module. The uncertainty measurement module is used to divide the uncertain superpixel unit into uncertain areas and certain areas, and calculate the uncertainty value of the features in the uncertain area to give high weights to features with large uncertainty.

[0084] The channel and spatial features of any superpixel unit in the encoder are calculated and then fused in the encoder through a cross-attention mechanism. With each iteration, the convolutional features of each convolutional layer are updated. The uncertainty measurement module divides the uncertain superpixel unit into uncertain and certain regions. The uncertainty of the features in the uncertain region is calculated, and features with greater uncertainty are given greater weight.

[0085] The data processing process of the uncertainty measurement module includes:

[0086] For any superpixel, the initial edge of the superpixel is denoted as T0, and the edge at the Kth erosion is denoted as Tk. Tk and the edge at T0 serve as a buffer zone. The erosion process uses a rectangular model (with an aspect ratio equal to the aspect ratio of the superpixel's minimum bounding rectangle). Using the U-Net spatial attention mechanism, the buffer zone corresponding to the maximum output probability during U-Net deconvolution is marked as the U-Net uncertainty region. The loop terminates when the filter window exceeds or reaches the size of the superpixel's minimum bounding rectangle. In other words, the U-Net here determines the potential boundary between sea ice and seawater within the superpixel by changing the window's field of view, thereby dividing the uncertain superpixel into a certain region and an uncertain region.

[0087] The uncertainty of the features in the uncertainty region of the uncertain superpixel unit is calculated, and the uncertainty values ​​of the features in the uncertainty region of all uncertain superpixel units are obtained.

[0088] The features in the uncertainty region of all uncertain superpixel units are sorted according to their uncertainty values, and features with large uncertainty values ​​are given high weights.

[0089] The uncertainty measurement module calculates the uncertainty of the features within the uncertainty region in the uncertain superpixel unit as follows:

[0090] ;

[0091] in, represents the uncertainty in sea ice forecasts, represents the uncertainty of seawater prediction, represents the uncertainty characteristics of sea ice, represents the uncertainty characteristics of seawater, Indicates the offset.

[0092] In this embodiment, the offset is 0.5. The uncertainty measurement module reduces the uncertainty of superpixels at different scales during the layer-by-layer processing of the convolution layer and prevents feature loss, making the final output sea ice classification result more accurate.

[0093] Pre-train the constructed Ice-WaterNet network model:

[0094] Obtain an ice condition map dataset, which is annotated with benchmark true values.

[0095] The Ice-WaterNet network model is trained with the ice map dataset as input and the sea ice and sea water boundary lines as output. During the training process, the model parameters of the Ice-WaterNet network model are adjusted using a loss function. The loss function uses a binary cross entropy function, which is expressed as:

[0096] ;

[0097] in, represents the binary cross entropy function, Represents the first The output of the convolutional layer, Indicates the number of convolutional layers of the Ice-WaterNet network model, Indicates the true value of the benchmark.

[0098] When the loss function value reaches the convergence condition, the pre-trained Ice-WaterNet network model is obtained.

[0099] Taking the uncertain superpixel unit as input, the sea ice and sea water boundary line within the uncertain superpixel unit is obtained based on the output of the pre-trained Ice-WaterNet network model. The sea ice and sea water boundary lines within all uncertain superpixel units are combined to obtain the sea ice classification result map.

[0100] Sea ice concentration retrieval:

[0101] The sea ice classification result is downsampled according to the preset resolution to obtain the downsampled sea ice classification result.

[0102] The downsampled sea ice classification results are projected according to the projection grid to obtain the percentage of sea ice pixels in each sea ice classification result projection grid, that is, the sea ice density of each sea ice classification result projection grid. The sea ice density inversion result is obtained by combining the sea ice density of each sea ice classification result projection grid.

[0103] like Figure 2 As shown in the figure, taking the Sentinel-1 SAR data of the Fram Strait region in the Arctic on August 1, 2023 as an example, for the SAR data that has been radiometrically calibrated and ensemble corrected, the initial segmentation result is obtained by CRF superpixel segmentation, and the superpixels with a posterior probability between 0.15 and 0.85 are marked as uncertain superpixels. For any uncertain superpixel unit, the uncertain superpixel is segmented based on the multi-layer convolution and iterative strategy of Ice-WaterNet. In the loop process, it can be found that as the filter window increases, more sea ice objects are observed, so we only need to focus on Figure 2The sea ice object in region 1 is sufficient. Once the sea ice object in region 1 is accurately extracted, the remaining pixels in region 1 and all pixels in region 2 are marked as open water. The segmentation edge corresponding to the minimum uncertainty is taken as the final uncertain superpixel segmentation result.

[0104] Figures 3 to 6 This is the accuracy verification result of the ice-water classification result based on Ice-WaterNet of the present invention. Among the four evaluation indicators, the Ice-WaterNet model of the present invention performs best, indicating that the algorithm of the present invention has the best performance in the segmentation accuracy of the ice-water boundary and the suppression of the misclassification of seawater targets; on the other hand, Ice-WaterNet has a smaller fluctuation range among the four evaluation indicators, indicating that the model has better stability. Specifically, the intersection-over-union ratio reflects the similarity between the classification result and the true value, among which the CRF and U-Net results are similar, while the classification result of Ice-WaterNet is closer to the true value. Although the accuracy of U-Net is higher than that of CRF, the recall rate is lower than that of the CRF result, indicating that there are more false alarms in the U-Net model, that is, the uncertainty of sea ice is greater. The CRF result in the F1 score is also better than the U-Net result, indicating that CRF can better achieve a balance between accuracy and recall. In addition, Ice-WaterNet performs best in the F1 score, which shows that the method of the present invention performs best in ice-water classification.

[0105] Figure 7 and Figure 8 This is the sea ice density inversion result based on Ice-WaterNet and ASI density products. Overall, the two density results are highly consistent, but the Ice-WaterNet result of our method has higher resolution and displays more detailed sea ice information. Furthermore, in low-density areas such as the sea ice edge and polynyas, our method produces richer texture features.

[0106] In summary, the deep learning-based sea ice classification and density inversion method provided by the present invention improves the inversion accuracy of sea ice classification and density, and can play a more important role in global climate change monitoring, marine environmental protection, polar resource development and utilization, etc.

[0107] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A deep learning-based sea ice classification and density inversion method, characterized by: include: Obtain raw dual-polarization synthetic aperture radar data; Preprocessing the original dual-polarization synthetic aperture radar data to obtain preprocessed dual-polarization synthetic aperture radar data; Perform superpixel segmentation on the preprocessed dual-polarization synthetic aperture radar data to obtain multiple superpixels; The conditional random field model is used to calculate the posterior probability of superpixels, and the superpixels whose posterior probability is within the set interval are taken as uncertain superpixel units; Taking the uncertain superpixel unit as input, the sea ice and sea water boundary line within the uncertain superpixel unit is obtained based on the output of the pre-trained Ice-WaterNet network model; The sea ice and seawater boundaries within all uncertain superpixel units are combined to obtain the sea ice classification result map; Perform sea ice density inversion based on the sea ice classification result map to obtain the sea ice density inversion result; The Ice-WaterNet network model is constructed based on the U-Net network model. The construction method includes connecting each convolutional layer in the decoder of the U-Net network model to an uncertainty measurement module, wherein the uncertainty measurement module is used to divide the uncertain superpixel unit into an uncertain region and a certain region, and calculate the uncertainty of the features in the uncertain region to assign a high weight to the features with large uncertainty; The data processing process of the uncertainty measurement module includes: For each uncertain superpixel unit, the buffer area corresponding to the maximum output probability during model deconvolution is recorded as the uncertain area; Using a rectangular model as a filtering window, continuously changing the perceptual field of view of the filtering window, obtaining the sea ice boundary line within the uncertain region to redivide the uncertain region into a deterministic region and an uncertain region, until the size of the filtering window reaches or exceeds the size of the minimum bounding rectangle of the uncertain superpixel unit, wherein the aspect ratio of the rectangular model is the aspect ratio of the minimum bounding rectangle of the uncertain superpixel unit; The uncertainty of the features in the uncertainty region of the uncertain superpixel unit is calculated, and the uncertainty value of the features in the uncertainty region of all uncertain superpixel units is obtained. The expression is as follows: ； in, represents the uncertainty in sea ice forecasts, represents the uncertainty of seawater prediction, represents the uncertainty characteristics of sea ice, represents the uncertainty characteristics of seawater, Indicates the offset; The features in the uncertainty region of all uncertain superpixel units are sorted according to their uncertainty values, and features with large uncertainty values ​​are given high weights.

2. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The preprocessing of the original dual-polarization synthetic aperture radar data to obtain preprocessed dual-polarization synthetic aperture radar data includes: Read noise parameters and radiation calibration parameters from raw dual-polarization synthetic aperture radar data; The remote sensing image pixel brightness value of the original dual-polarization synthetic aperture radar data is converted into the backscattering coefficient according to the noise parameter and the radiation calibration parameter. The expression is as follows: ； in, represents the backscattering coefficient, Represents the brightness value of the remote sensing image pixel, represents the noise parameter, represents the radiation calibration parameter; The backscatter coefficient is converted into the backscatter coefficient of the reference incident angle, which is the preprocessed dual-polarization synthetic aperture radar data. Its expression is as follows: ； in, represents the corrected backscatter coefficient, i.e., the preprocessed dual-polarization synthetic aperture radar data, represents the backscattering coefficient, represents the reference incident angle, represents the slope coefficient, represents the local angle of incidence.

3. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The superpixel segmentation of the preprocessed dual-polarization synthetic aperture radar data adopts the meanshift segmentation method.

4. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The expression of the conditional random field model is as follows: ； in, represents the posterior probability, represents the normalization coefficient, Indicates the The unit potential energy function of superpixels, Indicates the superpixels and The potential energy function of superpixels; Indicates the The class of superpixels, Indicates the superpixel classes, where the classes are classified as sea ice or sea water; and Both represent superpixel sets and do not contain each other; Indicates the The backscatter coefficient of superpixels, represents the model weight parameter, represents the model input features, , Indicates the superpixels and The cross-correlation function of superpixels, represents the Dilick function, when hour, The value is 1, otherwise it is 0.

5. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The Ice-WaterNet network model is an encoder-decoder structure, in which the number of convolutional layers in the encoder and decoder are both 5. The encoder is used to calculate channel and spatial features, and the decoder is used to extract different types of floating ice using multi-scale context information. Each convolutional layer in the decoder is equipped with an attention module, which is used to distinguish sea ice from sea water by using the difference between the HH and HV channels to determine the buffer area.

6. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The pre-training method of the Ice-WaterNet network model includes: Obtaining an ice condition map dataset, wherein the ice condition map dataset is annotated with a reference true value; The Ice-WaterNet network model is trained with the ice map dataset as input and the sea ice and sea water boundary lines as output. During the training process, the model parameters of the Ice-WaterNet network model are adjusted using a loss function. When the loss function value reaches the convergence condition, the pre-trained Ice-WaterNet network model is obtained.

7. The deep learning-based sea ice classification and density inversion method according to claim 6, characterized in that: The loss function adopts the binary cross entropy function, which is expressed as: ； in, represents the binary cross entropy function, Represents the first The output of the convolutional layer, Indicates the number of convolutional layers of the Ice-WaterNet network model, Indicates the true value of the benchmark.

8. The deep learning-based sea ice classification and density inversion method according to claim 1, characterized in that: The sea ice density inversion is performed based on the sea ice classification result map to obtain the sea ice density inversion result, including: Downsampling the sea ice classification result according to a preset resolution to obtain a downsampled sea ice classification result; The downsampled sea ice classification results are projected according to the projection grid to obtain the percentage of sea ice pixels in each sea ice classification result projection grid, that is, the sea ice density of each sea ice classification result projection grid; The sea ice density of each sea ice classification result projection grid is combined to obtain the sea ice density inversion result.

Images