An AI identification method for sea ice age and type based on polarized light micro-crack detection

Abstract

This invention discloses an AI-based method for identifying sea ice age types based on polarized light microcrack detection, belonging to the field of ice condition detection technology. The method includes: acquiring a polarization image of the target sea ice to construct an original polarization tensor; calculating the total intensity matrix and the linear polarization degree matrix; performing inverse phase normalization on the linear polarization degree matrix to generate a physical space attention mask; extracting a one-dimensional morphological feature vector based on the total intensity matrix; calculating the local grid spatial distribution entropy to extract the local total intensity tensor and the local physical attention mask; inputting the local tensor, mask, and morphological feature vector into a bi-branch heterogeneous network model for forward computation; and outputting an ice age category label. This invention combines the polarization optical characteristics and morphological structure of the ice surface, reducing reflective interference and computational redundancy, and improving the accuracy of sea ice age identification.

Description

Technical Field

[0001] This invention relates to the field of sea ice detection and image data processing technology, specifically to a method for identifying sea ice age type A1 based on polarized light microcrack detection. Background Technology

[0002] With the advancement of polar shipping and resource development, accurate identification of sea ice age has become a fundamental prerequisite for ensuring safe navigation. Current sea ice condition detection methods mainly include satellite radar observation and shipborne optical image analysis. Among these, shipborne optical images, due to their high local spatial resolution, are widely used for real-time identification of ice type in close-range navigation areas.

[0003] However, when conventional optical imaging equipment captures images in polar environments, the sea ice surface is prone to strong specular reflection, leading to frequent localized overexposure in the captured images. This reflective interference masks the subtle physical structural features of the ice surface, such as microcracks and pores formed during its evolution, resulting in the loss of underlying texture information. This makes it difficult for subsequent recognition models to obtain the true ice surface structure data used for classification.

[0004] In terms of image data processing algorithms, existing convolutional neural networks typically perform indiscriminate feature extraction on the entire image. Since polar images contain large areas of uniform, smooth, crack-free ice, global processing consumes significant computational resources on redundant backgrounds. Furthermore, existing recognition schemes do not integrate the optical polarization response properties of sea ice with their physical morphological changes. Lacking underlying physical constraints, the network easily misinterprets reflective spots or shadows on the ice surface as ice age structural features, resulting in high-dimensionality extracted features and a large amount of invalid data. This leads to low accuracy in identifying different ice ages and makes it difficult to output detailed local spatial risk assessment results. Summary of the Invention

[0005] The technical problem solved by this invention is that existing sea ice detection technologies mostly rely on the global visual features of optical images, and do not fully extract the changes in the physical structure of microcracks during the sea ice evolution process; moreover, the high reflectivity of the ice surface in polar scenarios can easily cause the loss of local image features, resulting in low accuracy of ice age identification and difficulty in providing detailed spatial distribution risk assessment parameters.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides an AI-based method for identifying sea ice age and type based on polarized light microcrack detection. This method is applied to a computing system including a polarization acquisition module, a polarization parameter analysis module, a feature extraction and dynamic routing module, and an online inference module. The method steps include:

[0008] The polarization acquisition module acquires polarization images of the target sea ice and constructs the original polarization tensor;

[0009] The polarization parameter analysis module processes the original polarization tensor, calculates the total light intensity matrix, and generates a full-size physical space attention mask.

[0010] The feature extraction and dynamic routing module extracts one-dimensional morphological feature vectors based on the total light intensity matrix, calculates the spatial distribution entropy of the divided local grid, and extracts the local total light intensity tensor and the local physical attention mask.

[0011] The online inference module inputs the local total light intensity tensor, the local physical attention mask, and the one-dimensional morphological feature vector into the network model for forward computation, and outputs the ice age category label.

[0012] In the above scheme, by extracting the parameters of light intensity and linear polarization degree on the ice surface, an attention mask is generated using the polarization information to reduce reflective interference at the data processing front end by utilizing the difference in response to specular reflection and diffuse reflection.

[0013] By introducing spatial distribution entropy to quantify the structural complexity of local regions of an image, and extracting data from densely distributed microcracks in a grid, image computation can be reduced in dimensionality while preserving core physical features.

[0014] Preferably, the polarization acquisition module acquires a polarization image of the target sea ice and constructs the original polarization tensor, including:

[0015] Within the exposure cycle, the pixel values ​​of the target field of view under the preset polarization direction are obtained, and the pixel values ​​of each polarization direction are extracted into a sparse matrix.

[0016] Spatial interpolation is performed on the sparse matrix to generate multiple independent polarization images, and these multiple independent polarization images are stacked to construct the original polarization tensor.

[0017] Preferably, the polarization parameter analysis module processes the original polarization tensor, calculates the total intensity matrix, and generates a full-size physical space attention mask, including:

[0018] The light intensity components of each polarization direction in the original polarization tensor are calculated based on the Stokes vector equation to obtain the total light intensity matrix and the linear polarization degree matrix.

[0019] Perform an inverse normalization operation on the linear polarization degree matrix to generate a full-size physical space attention mask.

[0020] This step utilizes the optical properties of sea ice microcrack regions to enhance light scattering and reduce linear polarization. By using inverse normalization, the low polarization region is mapped to a high-response numerical region in the mask, thereby increasing the weight of the crack region in subsequent calculations.

[0021] Preferably, the feature extraction and dynamic routing module extracts a one-dimensional morphological feature vector based on the total light intensity matrix, including:

[0022] The total light intensity matrix is ​​input into the segmentation network to generate a binary mask for microcracks;

[0023] The number of independent connected components in the binary mask of microcracks is counted by the connected component labeling algorithm, and the area density of microcracks is calculated.

[0024] The number of independent connected domains and the microcrack area density are concatenated to construct a one-dimensional morphological feature vector.

[0025] Preferably, the feature extraction and dynamic routing module calculates the spatial distribution entropy of the divided local grid and extracts the local total light intensity tensor and the local physical attention mask, including:

[0026] The proportion of microcrack pixels within each local grid to the total number of pixels in the corresponding grid is statistically analyzed, and the spatial distribution entropy is obtained by calculating the proportion based on the information entropy formula.

[0027] Extract the spatial coordinate bounding boxes of local grids whose spatial distribution entropy exceeds a preset threshold;

[0028] Based on the spatial coordinate bounding box, the local total intensity tensor and the local physical attention mask are extracted from the total intensity matrix and the full-size physical spatial attention mask.

[0029] Preferably, the network model is a two-branch heterogeneous network model;

[0030] The online inference module inputs the local total light intensity tensor, the local physical attention mask, and the one-dimensional morphological feature vector into the network model for forward computation, including:

[0031] The local total light intensity tensor is input into the convolutional branch network to generate a two-dimensional optical feature map; the local physical attention mask is downsampled to generate an alignment mask; the alignment mask is multiplied element-wise with the two-dimensional optical feature map; and a one-dimensional optical feature vector is output through the pooling layer.

[0032] Input a one-dimensional morphological feature vector into a multilayer perceptron branch network to output a one-dimensional hidden layer feature vector.

[0033] The one-dimensional optical feature vector and the one-dimensional hidden layer feature vector are concatenated into a fused feature vector, which is then input into a fully connected classification head to output the ice age category label.

[0034] This step directly multiplies the numerical matrix of the physical mask with the optical feature map output by the convolutional network, achieving numerical-level fusion of physical polarization prior information and image features.

[0035] Preferably, the node weights of the network model are updated during the training phase using a joint loss function;

[0036] The joint loss function includes classification cross-entropy loss and feature distance constraint loss. The feature distance constraint loss is used to calculate the spatial distance between the one-dimensional optical feature vector and the one-dimensional hidden layer feature vector.

[0037] By iterating through the feature distance constraint loss, the feature vectors output by the constraint convolution branch and the multilayer perceptron branch are brought closer to each other in spatial scale, increasing the numerical stability of feature concatenation computation.

[0038] Preferably, the calculation system further includes a state evaluation module; after the online inference module outputs the ice age category label, the method further includes:

[0039] The status assessment module calculates the local risk coefficient based on the ice age category label, microcrack area density, and the mean of the local physical attention mask;

[0040] Based on the spatial coordinate boundary, the local risk coefficients are backfilled into the corresponding coordinate interval of the global risk distribution matrix.

[0041] Preferably, the computing system also includes an early warning output module;

[0042] After backfilling the local risk coefficients to the corresponding coordinate range of the global risk distribution matrix, the method also includes:

[0043] The early warning output module performs spatial integration and mean averaging calculations on the element values ​​within the global risk distribution matrix to obtain the comprehensive navigation safety index.

[0044] The aforementioned risk coefficient backfilling and spatial integration process establishes a quantitative calculation relationship between the local grid classification results and the overall field-of-view safety index.

[0045] Preferably, the method includes the following computational steps: obtaining pixel values ​​in a preset polarization direction and constructing an original polarization tensor through sparse matrix interpolation and stacking;

[0046] The linear polarization degree matrix is ​​solved by algebraic operations based on the Stokes vector equation, and a full-size physical space attention mask is generated by numerical inversion calculation.

[0047] The binary mask generated by the statistical segmentation network is used to calculate the number of connected components and pixel area density, and to generate a one-dimensional morphological feature vector.

[0048] The information entropy of the local microcrack ratio is calculated based on the preset mesh division parameters. Local mesh bounding boxes with entropy greater than the information entropy threshold are extracted, and the local total light intensity tensor and local physical attention mask are extracted.

[0049] A one-dimensional optical feature vector is generated by performing element-wise multiplication on the optical feature map using a downsampled local physical attention mask.

[0050] The ice age category label is obtained by splicing the one-dimensional optical feature vector with the one-dimensional hidden layer feature vector mapped by the one-dimensional morphological feature vector and then performing probabilistic analysis through a fully connected classification head.

[0051] Based on the ice age category label, the preset risk coefficient is queried and backfilled into the global risk distribution matrix. The comprehensive navigation safety index is calculated by performing two-dimensional spatial integration on the element values ​​of the global risk distribution matrix. Based on the threshold comparison results, the warning signal containing coordinates is output.

[0052] A second aspect of the present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the AI ​​identification method for sea ice age and type based on polarized light microcrack detection provided in the first aspect of the present invention.

[0053] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the AI ​​identification method for sea ice age and type based on polarized light microcrack detection provided in the first aspect of the present invention.

[0054] This invention provides a method for identifying sea ice age type (Al) based on polarized light microcrack detection. It has the following advantages:

[0055] 1. This invention calculates the linear polarization degree matrix and performs inverse normalization operation through the polarization parameter analysis module to generate a full-size physical space attention mask. This feature utilizes the optical physical property of enhancing light scattering in sea ice crack regions to reduce linear polarization degree, transforming low polarization regions into high-weight attention values. In the data preprocessing stage, it reduces the interference of high reflectivity of polar ice surfaces and increases the proportion of data features of microcrack regions in the subsequent network model.

[0056] 2. This invention calculates the spatial distribution entropy of the local mesh based on a microcrack binary mask, extracts bounding boxes with distribution entropy higher than a preset threshold to extract the local total light intensity tensor and physical attention mask. This feature quantifies the structural complexity of the local region of the image through information entropy, and selectively filters out large areas of smooth and crack-free redundant background images, thereby achieving dimensionality reduction of the input data and reducing the computational load of the system while retaining the core physical features.

[0057] 3. This invention constructs a dual-branch heterogeneous network model that includes convolutional branches and multilayer perceptron branches. The downsampled physical attention mask is multiplied element-wise with the optical feature map generated by the convolutional branch and then concatenated with the hidden layer vector of the morphological feature map. This feature integrates the polarization optical response constraints of the bottom layer with morphological features such as the number of connected domains and area density at the numerical level, corrects the classification bias that is easily caused by a single visual feature under complex lighting conditions, and improves the accuracy of age recognition. Attached Figure Description

[0058] Figure 1 This is a block diagram of a sea ice microcrack detection and ice age identification system according to an embodiment of the present invention;

[0059] Figure 2 This is a flowchart of a sea ice microcrack detection and ice age identification method according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram illustrating the principle of coaxial acquisition of multi-polarization images according to an embodiment of the present invention.

[0061] Figure 4 This is a flowchart illustrating the polarization parameter analysis and mask construction process according to an embodiment of the present invention.

[0062] Figure 5 This is a flowchart illustrating the topology feature extraction and dynamic routing process according to an embodiment of the present invention.

[0063] Figure 6 This is a flowchart illustrating the offline heterogeneous network collaborative simulation of an embodiment of the present invention.

[0064] Figure 7 This is a flowchart illustrating the output of ice age recognition results and optional risk warnings in an embodiment of the present invention.

[0065] Figure 8 This is a comprehensive data diagram of multimodal feature fusion inference numerical extrapolation and performance verification in an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example:

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] See attached document Figure 1 , Figure 1 This is a block diagram of a sea ice microcrack detection and ice age identification system according to an embodiment of the present invention.

[0070] This invention provides a sea ice microcrack detection and ice age identification system based on polarization parameter analysis and heterogeneous networks, which may include: a polarization acquisition module, a polarization parameter analysis module, a feature extraction and dynamic routing module, a model training module, and an online inference module.

[0071] In an optional implementation, the system also includes a status assessment and early warning output module, which generates early warning information based on the ice age identification results and microcrack risk characteristics.

[0072] The system's overall workflow covers the offline constrained training phase of model parameters and the online forward inference phase for edge devices.

[0073] See attached document Figure 2 , Figure 2 This is a flowchart of a sea ice microcrack detection and ice age identification method according to an embodiment of the present invention.

[0074] This invention provides a method for detecting microcracks in sea ice and identifying ice age based on polarization parameter analysis and heterogeneous networks, comprising the following steps:

[0075] S10: Optical data is acquired from the reflected light field of the target ice layer through the polarization acquisition module. Multiple independent polarization images with preset polarization direction are simultaneously acquired in the target field of view within a single global shutter exposure cycle using an image sensor containing a micro polarizer array, forming the original polarization tensor.

[0076] S20 processes the original polarization tensor through the polarization parameter analysis module, calculates the total light intensity and polarization difference matrix of the target light field based on the Stokes vector equation, calculates the pixel-level linear polarization degree matrix, and performs an inverse phase normalization operation on the linear polarization degree matrix using the optical scattering depolarization characteristics of the ice layer medium to generate a full-size physical space attention mask with the same resolution as the original image.

[0077] S30: The total light intensity matrix is ​​input into the lightweight segmentation network through the feature extraction and dynamic routing module to obtain the microcrack binary mask matrix. The one-dimensional morphological scalar features of the microcrack distribution are statistically analyzed. The microcrack binary mask matrix is ​​divided into local grids. The spatial distribution entropy of each local grid is calculated. The spatial coordinates of the grids with spatial distribution entropy higher than a preset threshold are used as the clipping boundary. The local total light intensity image tensor and local physical mask tensor of the corresponding coordinate region are simultaneously cropped on the total light intensity matrix and the full-size physical space attention mask.

[0078] S40: During the offline constrained training phase, a dual-branch heterogeneous network model is constructed through the model training module. Historical sea ice samples with ice age category labels are used as training samples. One-dimensional morphological scalar features are input into the multilayer perceptron of the first branch for high-dimensional feature mapping. The local total light intensity image tensor is input into the deep convolutional neural network of the second branch. An adaptive average pooling layer is connected after the output of the last convolutional block of the deep convolutional neural network and before the global pooling layer. The local physical mask tensor is spatially downsampled to align the dimension of the network feature map and is multiplied element-wise with the matrix of the current network feature map. The hidden layer feature vectors of the two branches are extracted to construct a joint loss function including classification cross-entropy loss and feature distance constraint loss. The node weight data of the dual-branch heterogeneous network model are iteratively updated.

[0079] S50, in the online forward inference stage, loads the dual-branch heterogeneous network model that has completed weight updates through the online inference module. The polarization acquisition module obtains the original polarization tensor of sea ice with unknown attributes and, after processing by the pre-step, inputs it into the dual-branch heterogeneous network model to perform unidirectional calculation operations. The terminal classifier outputs discrete labels to determine the age and type of sea ice.

[0080] Among them, the sea ice age types include at least two preset ice age categories, and the ice age category labels are determined by historical ice condition records, field measurement data, remote sensing inversion results or manual annotation results in the training samples;

[0081] When the system is further configured with early warning functions, the online inference module also generates risk warning information based on ice age category labels, microcrack area density, and local physical attention response.

[0082] When step S30 outputs multiple highly complex local grids, the online inference module performs forward inference on the local total light intensity tensor, local physical attention mask, and one-dimensional morphological feature vector corresponding to each local grid to obtain multiple local ice age category probability vectors.

[0083] The system performs average fusion, maximum confidence selection, or majority voting on multiple local ice age category probability vectors to obtain the global ice age category label for the target field of view.

[0084] The specific implementation details of the present invention will be described in detail below in conjunction with the above steps.

[0085] See attached document Figure 3 , Figure 3 This is a schematic diagram of the coaxial acquisition principle of multi-polarization images according to an embodiment of the present invention.

[0086] Step S10 is executed by the polarization acquisition module, primarily to acquire the polarization optical field signal of the target sea ice, providing fundamental data support for subsequent polarization physical parameter analysis and network inference. Specifically, step S10 includes the following sub-steps:

[0087] S101 controls the image sensor, which is equipped with a micro-polarizer array inside the polarization acquisition module, to receive the light field reflected from the target ice layer within a single global shutter exposure cycle.

[0088] By setting up a micro-polarizer array, polarization information in multiple directions can be captured simultaneously in a single exposure, thereby reducing the impact of platform displacement caused by wind and waves in polar environments on image quality. Specifically, the surface of the photosensitive pixel layer of this image sensor is photolithographically etched with pixel-level directional polarization gratings, and the micro-polarizer array forms a basic macropixel unit with four adjacent pixels.

[0089] Each macropixel unit has four subpixels covered by polarizing filters with light transmission axis directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees respectively.

[0090] Triggered by a global shutter, the image sensor demodulates the reflected light field of the target ice layer in parallel, outputting a single-frame raw image containing spatially interwoven polarization information.

[0091] The optical signals of four preset polarization directions are collected coplanarly at the same physical time, which effectively reduces the deviation of physical space coordinates caused by the dynamic displacement of the platform due to multiple time-division acquisitions.

[0092] S102, extract polarization pixel data from a single frame of the original image and separate and reconstruct it into an independent sparse matrix.

[0093] Since the polarized pixels in a single frame of the original image are arranged in an alternating pattern, the system traverses the pixel space coordinates of the single frame of the original image and extracts the pixel grayscale values ​​corresponding to the polarization directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees respectively.

[0094] While keeping the original physical space coordinates unchanged, four independent sparse matrices corresponding to different polarization directions are generated.

[0095] In each sparse matrix, at the pixel position where the corresponding angle polarization light value has not been obtained, the system fills the position with a zero value for its grayscale value.

[0096] S103 performs polarization demosaic space interpolation on each sparse matrix to construct the original polarization tensor.

[0097] Spatial interpolation algorithms are used to estimate and fill the neighborhood pixels at the zero-value pixel positions in the sparse matrix, restoring it to the full resolution of the original image. This generates four preset polarization directions with the same spatial resolution under the same target field of view. A two-dimensional image matrix.

[0098] in, The height of the image in pixels. Let be the width in pixels. Define these four two-dimensional image matrices as follows: Represents the polarization image matrix at 0 degrees.

[0099] Represents a 45-degree polarization image matrix. Represents a 90-degree polarization image matrix. This represents a 135-degree polarization image matrix.

[0100] The four two-dimensional image matrices are stacked in an array along the data channel dimension to generate a dimension of... The original polarization tensor in the form of a multidimensional array.

[0101] The original polarization tensor, as the fundamental data source for parameter analysis, is stored in the system's storage space. The specific calculation derivation process of the underlying spatial interpolation algorithm in step S103 can be implemented by those skilled in the art according to conventional digital image processing standards (such as bilinear interpolation or bicubic interpolation algorithms). Its basic operational principles are well-known in the field and will not be elaborated upon here.

[0102] See attached document Figure 4 , Figure 4 This is a flowchart of polarization parameter analysis and mask construction according to an embodiment of the present invention.

[0103] Based on the above embodiments, step S20 is specifically executed through the polarization parameter analysis module, which transforms the original polarization tensor obtained from the front end into a feature matrix characterizing the physical state of the target, and constructs a physical space attention mask by combining the depolarization characteristics of sea ice microcracks. In specific implementation, step S20 includes the following sub-steps:

[0104] S201 performs algebraic operations on the four independent polarization image matrices in the original polarization tensor to solve for the Stokes parameter matrix of the target light field.

[0105] In the field of optical physics, the polarization state of a light beam is usually described by the Stokes vector.

[0106] By combining light intensity components with different polarization directions, it is possible to separate the bus polarization information and polarization differences in various directions in natural light.

[0107] The polarization parameter analysis module calculates the total light intensity matrix based on the Stokes vector equation. Horizontal and vertical linear polarization component difference matrix and the difference matrix between the +45° and -45° linear polarization components The specific calculation formula is as follows:

[0108] ;

[0109] ;

[0110] ;

[0111] in, , , , These represent the polarization image matrices extracted in step S10 at 0°, 45°, 90°, and 135°, respectively. The Stokes parameter matrix obtained after this step is calculated. , , All dimensions are .

[0112] S202, based on the obtained Stokes parameter matrix, calculates the pixel-level linear polarization degree matrix.

[0113] The degree of linear polarization reflects the proportion of linearly polarized light components in the total light intensity of the reflected light field.

[0114] The system traverses the pixel space coordinates and calculates the polarization parameters pixel by pixel. The specific formula is as follows:

[0115] ;

[0116] in, This represents the calculated linear polarization degree matrix, whose internal elements are normalized to a range of 0 to 1.

[0117] The minimum positive constant preset by the system is typically set to a range of values ​​of 100%. This is used to prevent division overflow errors caused by the total light intensity (I) being zero during the calculation process, and to ensure the integrity of the algorithm's operating logic.

[0118] The generated linear polarization matrix The spatial resolution remains at .

[0119] S203 utilizes the optical scattering depolarization characteristics of the ice layer medium to perform an inverse phase normalization operation on the linear polarization degree matrix, generating a full-size physical space attention mask with the same resolution as the original image.

[0120] In polar sea ice exploration scenarios, relatively smooth ice layers often produce a high proportion of specular reflection, and the corresponding reflected light has a high degree of linear polarization.

[0121] In areas where there are microcracks inside or on the surface of the ice layer, incident light is scattered and refracted multiple times at the rough interface of the cracks, resulting in an optical depolarization effect, which reduces the degree of linear polarization of the reflected light in that area.

[0122] Based on the above physical principles, the polarization parameter analysis module performs an inverse mapping on the linear polarization degree matrix using the following formula:

[0123] ;

[0124] in, This represents the generated full-size physical space attention mask. Representative dimension is And it is a matrix in which all elements are 1. Through the above inverse normalization operation, the microcrack region that originally exhibited low linear polarization is... The values ​​correspond to higher values ​​in the middle, while the values ​​in the smooth ice surface region with high polarization are suppressed to a lower state.

[0125] Full-size physical space attention mask The tensor dimension is also .

[0126] See attached document Figure 5 , Figure 5 This is a flowchart of topology feature extraction and dynamic routing according to an embodiment of the present invention.

[0127] Based on the above embodiments, step S30 is executed by the feature extraction and dynamic routing module, which is used to extract the macroscopic morphological features of microcracks from the total light intensity matrix, and to filter out local regions with high complexity features based on the information entropy theory, so as to realize the synchronous interception of optical data and physical mask.

[0128] In practice, step S30 includes the following sub-steps:

[0129] S301 uses a lightweight segmentation network to perform feature mapping on the total light intensity matrix, outputs a binary mask of microcracks, and statistically analyzes one-dimensional morphological scalar features.

[0130] Before performing feature mapping, the lightweight segmentation network needs to be built and trained.

[0131] In the specific training phase, historically collected images of total polar sea ice light intensity were used as the sample set, and manually labeled binary images of microcracks were used as the corresponding labels.

[0132] The network outputs a predicted probability map through forward propagation, calculates the error between the predicted result and the label using a binary cross-entropy loss function, and updates the network weights based on gradient descent until the loss function converges. During the inference phase, the feature extraction and dynamic routing module internally deploys a lightweight segmentation network based on a shallow U-Net architecture. This network includes a downsampling encoder consisting of three consecutive convolutional layers with a stride of 2, and an upsampling decoder with a symmetrical structure. Feature fusion is performed between the corresponding layers of the decoder and encoder via skip connections to avoid loss of spatial details during downsampling.

[0133] The total light intensity matrix is ​​solved in step S20. The input data for this network has a tensor dimension of . .

[0134] The network outputs a pixel-level probability map via a Sigmoid activation function at the end, and performs binary classification segmentation based on a preset probability threshold of 0.5, generating a dimension of... Microcrack binary mask .

[0135] exist In the image, areas with a pixel value of 1 represent the location of microcracks, while areas with a pixel value of 0 represent the undamaged background ice surface.

[0136] For the generated binary mask The number of independent connected components in microcracks was counted using a connected component labeling algorithm. and the total pixel area of ​​microcracks .

[0137] Calculate the microcrack area density by combining the total number of pixels The calculation formula is as follows:

[0138] ;

[0139] Number of connected components With area density Channel concatenation is performed to construct a one-dimensional morphological feature vector:

[0140] .

[0141] This eigenvector is used to characterize the degree of structural damage to the target sea ice at a macro-statistical scale.

[0142] For the underlying computational logic of convolutional forward propagation computation and connected component labeling algorithm in lightweight segmentation networks, those skilled in the art can refer to conventional digital image processing specifications for implementation. The basic computational principles are well-known technologies in this field and will not be elaborated here.

[0143] S302 performs local meshing on the microcrack binary mask and calculates the spatial distribution entropy of each mesh region.

[0144] Spatial distribution entropy is an important indicator in information theory used to measure the uncertainty of a system's state. Introducing it into the field of image processing can effectively quantify the degree of disorder in the mixed distribution of targets and background within a local region.

[0145] Binary mask Divide into uniform spatial dimensions A local grid is defined, where (K) is a positive integer, and the grid division parameter is typically 4 or 8.

[0146] The spatial dimension of each local grid is: .

[0147] The feature extraction and dynamic routing module traverses all grids and, for the (i)th grid, counts the number of microcrack pixels with a pixel value of 1 within that grid. And calculate its percentage of the total pixels in a single grid. The proportion (p_{i}) is calculated using the following formula:

[0148] ;

[0149] The spatial distribution entropy of this local grid is obtained by quantizing the proportional data using information entropy theory. The calculation formula is as follows:

[0150] ;

[0151] in, Characterizing the first The complexity of the distribution of microcracks mixed with the smooth ice surface in a local grid; the larger the value, the more complex the morphological features in the local area.

[0152] This is a preset, extremely small positive number, and its value range is set to 1. The space between these two points is used to avoid numerical overflow caused by the independent variable being zero in logarithmic operations, thus ensuring the completeness of the information entropy algorithm's computational logic.

[0153] S303, based on the calculation results of spatial distribution entropy, performs synchronous dynamic interception of the total light intensity matrix and the physical space attention mask.

[0154] Set spatial distribution entropy threshold The feature extraction and dynamic routing module compares the spatial distribution entropy of each local grid cell one by one. With the set threshold When a local mesh is detected to satisfy... When the condition is met, extract the spatial coordinate bounding box system of the high-complexity grid on the two-dimensional image plane. .

[0155] To avoid the logical dead zone problem where the algorithm flow is interrupted because the spatial distribution entropy of all grids is below the threshold, if no grid that meets the condition is detected after traversal, the spatial distribution entropy is directly selected. The local grid with the largest value is selected as the target for interception.

[0156] Based on this spatial coordinate bounding box system, the feature extraction and dynamic routing modules respectively operate within the total light intensity matrix. and the full-size physical space attention mask generated in step S20 Perform synchronous tensor slice truncation operation.

[0157] When multiple spatial coordinate bounding boxes exist, multiple local total light intensity tensors are obtained respectively. and multiple corresponding local physical attention masks This forms a local tensor set. After the above slicing process, the high-complexity local total light intensity tensor is extracted. and the corresponding local physical attention mask Both of these tensor dimensions remain unchanged. .

[0158] Extracting the generated local tensor and Combining one-dimensional morphological feature vectors Together, they serve as the foundational data for offline heterogeneous network collaborative simulation, which is then fed into the next processing stage.

[0159] See attached document Figure 6 , Figure 6 This is a flowchart of an offline heterogeneous network collaborative simulation according to an embodiment of the present invention.

[0160] Based on the above embodiments, step S40 is executed through the heterogeneous network inference module, which is used to fuse local optical tensors, physical space masks and one-dimensional morphological features to output a probability assessment result characterizing the target sea ice age category.

[0161] In practice, step S40 includes the following sub-steps:

[0162] Sea ice of different ages varies in crystal structure, salt bubble content, surface roughness, crack density, and optical scattering depolarization characteristics.

[0163] Depending on the formation time and structural evolution of sea ice, the number of connected domains, area density, and local spatial distribution complexity of microcracks on the surface and inside exhibit different statistical characteristics.

[0164] Therefore, this invention inputs the physical spatial attention mask obtained by inverting the linear polarization degree, the one-dimensional morphological features of microcracks, and the local total light intensity image features into a dual-branch heterogeneous network to achieve the identification of the target sea ice age category.

[0165] S401, Constructing a heterogeneous neural network model and performing the offline training phase. To achieve collaborative processing of multimodal data, a heterogeneous neural network model is deployed within the heterogeneous network inference module.

[0166] The model structurally includes a convolutional branch network for extracting two-dimensional spatial features, a multilayer perceptron branch network for processing one-dimensional vectors, and a fully connected classification head for feature concatenation.

[0167] Before conducting actual business simulations, the model needs to be trained using historical samples.

[0168] Specifically, the local total light intensity tensor, local physical attention mask, and one-dimensional morphological feature vector from the same source in the historical detection data are extracted as training sample sets, and the ice age category label corresponding to the training sample is converted into a discrete vector in one-hot encoding format as supervised training label.

[0169] The ice age category is determined by historical ice condition maps, on-site borehole measurement results, remote sensing ice age inversion results, or manual expert annotation results.

[0170] During the forward propagation process of training, the model predicts the ice age category probability distribution based on the sample input and output, and uses the cross-entropy loss function to calculate the error between the predicted probability distribution and the true label.

[0171] The real labels are the ice age category labels corresponding to the training samples, and are converted into discrete vectors with a unique encoding format.

[0172] The gradient of the loss function with respect to the weights of each layer of the network is calculated using the backpropagation algorithm, and the network parameters are updated using the optimizer until the loss value converges to a preset interval, typically set to . between.

[0173] For the specific calculation process of backpropagation algorithm and optimizer parameter update during model training, those skilled in the art can refer to the conventional deep learning specifications. The basic operation principle is a well-known technology in this field and will not be elaborated here.

[0174] The convolutional branch network of the heterogeneous neural network model includes at least three convolutional blocks, each convolutional block including a convolutional layer, a batch normalization layer, a non-linear activation layer and a downsampling layer;

[0175] Multilayer perceptron branch networks include at least two fully connected layers to increase the number of connected components. and area density The one-dimensional morphological feature vectors are mapped to a high-dimensional hidden layer feature space;

[0176] The fully connected classification head outputs an ice age category probability vector based on the fused feature vector. The joint loss function includes classification cross-entropy loss. and feature distance constraint loss The joint loss function is expressed as:

[0177] ;

[0178] in, These are the preset weighting coefficients.

[0179] The feature distance constraint loss The one-dimensional optical hidden layer feature vector used to constrain the output of the convolution branch and the one-dimensional morphological hidden layer feature vector of the multilayer perceptron branch are close to each other in the same ice age category samples.

[0180] Before calculating the feature distance constraint loss, the system uses a linear mapping layer to uniformly map the one-dimensional optical hidden layer feature vector output by the convolution branch and the one-dimensional morphological hidden layer feature vector output by the multilayer perceptron branch to the feature space of the same dimension. Then, it calculates the Euclidean distance or cosine distance between the two to enhance the class consistency between different modal features.

[0181] S402 extracts two-dimensional spatial features through a convolutional branch network and performs feature intervention by combining a physical spatial attention mask.

[0182] During the deduction phase, the high-complexity local total light intensity tensor obtained in step S30 will be used... Input convolutional branch network.

[0183] This branch network consists of multiple stacked convolutional layers, batch normalization layers, and activation function layers. It downsamples and extracts features from the input tensor, and outputs a dimension of [dimensionality missing] after the last convolutional block and before the global pooling layer. Two-dimensional optical feature map .

[0184] in, Represents the number of feature channels, and These represent the height and width of the downsampled feature map, respectively. Simultaneously, to introduce the optical depolarization constraint of ice microcracks, the heterogeneous network inference module receives the local physical attention mask extracted in step S30. Its initial tensor dimension is .

[0185] Using the adaptive average pooling algorithm Perform spatial dimension scaling operation to generate a size of Alignment mask To ensure that it is consistent with the feature map Maintain consistency in spatial resolution.

[0186] Then, align the mask. Extended through tensor broadcast mechanism and with two-dimensional optical feature maps Perform element-wise Hadamard product operation, the specific formula is as follows:

[0187] ;

[0188] in, This represents a two-dimensional feature tensor constrained by physical rules. This represents element-wise multiplication.

[0189] Through this algebraic operation, the system assigns higher response weights to the features of the microcrack region and suppresses redundant background features in the smooth ice surface region.

[0190] S403 performs one-dimensional morphological feature mapping and multi-dimensional feature cascade fusion to output the target sea ice physical state assessment results.

[0191] In multimodal data processing, since the local two-dimensional optical matrix and the one-dimensional macroscopic morphological scalar are in completely different feature distribution spaces, direct splicing can easily lead to the curse of dimensionality and feature submersion. Therefore, it is necessary to perform heterogeneous mapping and dimensionality reduction through independent network branches.

[0192] Specifically, the one-dimensional morphological feature vector (V_{\mathrm{morph}}) constructed in step S30 is input into the multilayer perceptron branch network.

[0193] This branch network contains two fully connected layers to map low-dimensional morphological scalars to a high-dimensional hidden feature space, with an output dimension of... One-dimensional feature vector .

[0194] Simultaneously, the two-dimensional feature tensor generated in step S402 is processed. Perform a global average pooling operation to reduce its dimensionality to a length of One-dimensional feature vector ,in .

[0195] The heterogeneous network inference module will, in terms of data channel dimension and To splice the components together, the length of the structure is... fused feature vector .

[0196] This feature fusion mechanism preserves the local microscopic optical features of damage while introducing macroscopic connected domain and area distribution state information.

[0197] Ultimately, the feature vectors will be fused. The input is a fully connected classification head consisting of a linear layer and a Softmax activation function, calculated as follows:

[0198] ;

[0199] in, This represents the output probability vector for the ice age category. and These represent the weight matrix and bias term of the fully connected classification head, respectively.

[0200] Specifically, The dimension is , The dimension is ,in This represents the total number of preset ice age categories.

[0201] vector Each element in the table corresponds to the predicted probability of the target sea ice under different ice age categories. The system selects the label corresponding to the highest probability value as the specific business meaning output, thereby realizing the identification of the ice age category of the target sea ice.

[0202] See attached document Figure 7 , Figure 7 This is a flowchart of the ice age recognition result output and optional risk warning according to an embodiment of the present invention.

[0203] Based on the above embodiments, step S50 is executed by the online inference module, which is used to parse the ice age category probability vector output by the heterogeneous network into an ice age category label, and in optional embodiments, combines microcrack morphological features and physical space attention response to generate risk warning information.

[0204] In practice, step S50 includes the following sub-steps:

[0205] S501, analyze the ice age category probability vector to determine the predicted ice age category of the target sea ice.

[0206] The online inference module receives the ice age category probability vector output in step S40. .

[0207] In classification and decision theory, converting the soft probabilities output by a network into classification labels requires setting clear judgment criteria.

[0208] The system uses the maximum probability determination rule to parse the vector and extracts the category index value corresponding to the probability element with the largest value, which is used as the specific ice age category label for the local sea ice region.

[0209] Assuming the total number of preset ice age categories is The corresponding set of category labels is Obtain the current local tensor to predict the ice age category. The calculation formula is as follows:

[0210] ;

[0211] in, Representing vectors The Middle The probability values ​​corresponding to each category.

[0212] The online inference module outputs the corresponding ice age category label. When the system is configured with a risk warning function, the status assessment and warning output module further analyzes the ice age category label and microcrack area density. and local physical attention mask Calculate the local risk coefficient from the mean value. .

[0213] The risk mapping table is pre-established using historical geological statistics. It internally records the corresponding mapping relationship between various ice age categories, microcrack area density ranges, and quantified risk coefficients. As ice age stability decreases or microcrack area density increases, the mapped risk coefficients show a monotonically increasing trend. For example, when the ice age category corresponds to a stable ice layer and the microcrack area density is below the first threshold, the local risk coefficient takes a low-risk range value.

[0214] When the ice age category corresponds to nascent ice or a thin ice layer and the microcrack area density is higher than the second threshold, the local risk coefficient is taken as the high-risk range value.

[0215] Local risk coefficient The value range is set between 0 and 1.

[0216] S502, in the optional risk warning mode, constructs a global risk distribution matrix and performs spatial coordinate backfilling of local risk coefficients.

[0217] To quantify the overall safety status of the target's field of view, the system initializes a tensor with dimension [missing information]. A two-dimensional matrix whose internal elements are all zero is defined as the global risk distribution matrix. .

[0218] Based on the spatial coordinate bounding box system of the high-complexity grid on the two-dimensional image plane recorded in step S30 ( The status assessment and early warning output module will extract the local risk coefficient. Assign to The positions of all pixels within this coordinate range.

[0219] If multiple high-complexity grids that meet the conditions are detected simultaneously in step S30, these grids are traversed, and each grid is input into the bi-branch heterogeneous network for forward reasoning to obtain the corresponding ice age category label or local risk coefficient. Then, they are backfilled independently according to their respective spatial coordinate intervals.

[0220] For low-complexity smooth ice surface regions not hit by the dynamic interception mechanism in step S30, their... The corresponding position in the code retains the initial value of zero.

[0221] By performing spatial coordinate mapping operations, a global risk distribution matrix containing global topological risk information of the target's field of view is generated. .

[0222] S503, in the optional risk warning mode, calculates the comprehensive navigation safety index based on the global risk distribution matrix and triggers navigation alarm actions in combination with preset thresholds.

[0223] In practical operational scenarios such as polar exploration, it is necessary to measure the safety of the entire exploration area. By accumulating and averaging the local risk coefficients in a two-dimensional spatial domain, it is possible to effectively quantify the total area proportion of risky regions within the field of view and the severity of the risk.

[0224] The status assessment and early warning output module outputs the global risk distribution matrix. The comprehensive navigation safety index for the current complete field of view is obtained by spatial integration and mean averaging of all element values ​​within the field of view. The specific calculation formula is as follows:

[0225] ;

[0226] Where (x) and (y) represent the horizontal and vertical spatial coordinates of the matrix pixels, respectively. This indicates that the global risk distribution matrix is ​​in the coordinate system (…). Risk values ​​at the location.

[0227] The value ranges from 0 to 1. The closer the value is to 1, the more robust and stable the overall structure of the target sea ice area is.

[0228] Set a security index threshold .

[0229] The status assessment and early warning output module will calculate in real time... With safety index threshold Perform numerical comparison.

[0230] Among them, the safety index threshold The safety level is determined based on the pressure resistance of the operating vehicle, its sailing speed, historical risk sample statistics, or manually preset safety levels.

[0231] When the judgment At that time, the system generates a structured early warning signal containing the current spatial coordinates, safety index data, and risk distribution matrix.

[0232] The warning signal is then sent to an external navigation control terminal via a communication interface to assist in route replanning and risk avoidance actions.

[0233] For the underlying encoding method and network communication transmission protocol of the warning signal, those skilled in the art can refer to conventional industrial bus standards or wireless communication specifications for implementation. The communication protocol architecture is a well-known technology in this field and will not be elaborated here.

[0234] To enable those skilled in the art to gain a deeper understanding of the technical essence of the present invention, the forward inference calculation process of the above system and method is illustrated below with a specific business application scenario and a series of measured values.

[0235] Meanwhile, corresponding experimental verification data based on standard test datasets are provided to demonstrate the progressiveness of the technical solution of this invention.

[0236] See attached document Figure 8 , Figure 8 This is a comprehensive data graph of multimodal feature fusion inference numerical extrapolation and performance verification according to an embodiment of the present invention.

[0237] In a feasible scenario for scientific research navigation of a polar icebreaker, the polarization acquisition module mounted on the bow of the icebreaker acquires the original polarization tensor with a field of view dimension of 1024×1024 in front of the target ice layer within a single global shutter cycle.

[0238] In the polarization parameter analysis stage, taking a representative ice surface pixel with spatial coordinates (500, 500) in the target field of view as an example, the image sensor acquires the grayscale scalar values ​​of this pixel in four preset directions as follows: , , , 90.

[0239] The system performs algebraic operations based on the Stokes vector equation to obtain the total light intensity of the pixel. Difference between horizontal and vertical linear polarization components and the difference in diagonal polarization components In the system preset constants Calculate the linear polarization degree of the pixel under the given conditions. 0.5385.

[0240] Further calculations using the depolarization characteristic inversion formula yielded the numerical value of this coordinate point within the physical space attention mask. This value reflects a moderate degree of rough scattering on the ice surface at this location.

[0241] In the topology feature extraction and dynamic routing stages, the system will use dimensions of The total light intensity matrix is ​​fed into a lightweight segmentation network.

[0242] The number of connected components with microcracks within the field of view was identified using a connected component labeling algorithm. Total pixel area of ​​microcracks The area density was calculated. Therefore, a one-dimensional morphological feature vector representing the macroscopic damage in this field of view is constructed. .

[0243] System settings for mesh generation parameters The field of view is uniformly divided into 16 dimensions. The local mesh at coordinate index (2,3) was found to contain 13,107 microcrack pixels, accounting for 65,536 of the total pixels in the mesh. The system has a preset overflow prevention constant. The spatial distribution entropy of the grid is calculated based on the information entropy theory. .

[0244] The system sets the spatial distribution entropy threshold. Since 0.7219 > 0.6, the system determines that the grid region has a high-complexity structure, and then simultaneously extracts the corresponding local total light intensity tensor. With local physical attention mask Its tensor dimension is 1.

[0245] During the offline heterogeneous network collaborative simulation phase, the captured local total light intensity tensor After forward propagation through the convolutional branch network, the number of downsampled output channels is... ,high ,width Two-dimensional optical feature map .

[0246] Synchronously, local physical attention mask Scaling by adaptive average pooling Alignment mask And injected through element-wise Hadamard product The feature tensor after intervention is obtained. .

[0247] After flattening it globally, it generates a length of One-dimensional feature vector Meanwhile, the multilayer perceptron branch pairs Perform spatial mapping and output length. One-dimensional hidden layer feature vector

[0248] The two are cascaded to form a fused feature vector of length 4224. The data is then sent to the fully connected classification header. The system has three preset ice age categories (…). The categories are: Category 1 represents newly formed ice, Category 2 represents one-year ice, and Category 3 represents multi-year ice.

[0249] The classification head calculates the output probability vector. .

[0250] Based on the maximum probability rule, the system determines the predicted ice age category of this local area to be Category 2 (one-year ice).

[0251] During the risk assessment phase, based on the conditions of ice age category 2 and microcrack area density of 0.08, the system queries a pre-set risk mapping table and assigns a risk coefficient to the local mesh. Assume that a total of 10 elements above the entropy threshold are detected within the field of view of this exposure. The grid is defined as follows: the local risk coefficient of the mapped grid is 0.65 after deduction, and the other 6 grids are regarded as smooth safe regions (risk coefficient is 0).

[0252] The system will backfill the above coefficients to Global risk distribution matrix middle.

[0253] Perform two-dimensional spatial integration and mean averaging calculations over the entire field of view: the sum of risk values ​​for all grids is... Dividing by the total number of pixels, 1,048,576, yields a global average risk of 0.40625.

[0254] Comprehensive navigation safety index .

[0255] The system presets safety index thresholds for icebreakers. If the comparison result is 0.59375 < 0.75, the system triggers a navigation alarm and outputs a risk avoidance and replanning signal containing coordinates and risk intensity to the ship's control bus.

[0256] To further verify the effectiveness of the technical solution provided by this invention and the improvement effect on technical indicators, Figure 8 The paper simultaneously presents the receiver operating characteristic (ROC) curves and performance comparison results of the system of the present invention and conventional single-modal networks on a standard ice condition dataset.

[0257] The historical sample set used in the experiment contained 8,000 sets of measured optical and polarization tensors from the same source, divided into training and testing sets in an 8:2 ratio. The ground truth labels were jointly calibrated using in-situ ice core drilling samples combined with high-resolution satellite microwave scattering data.

[0258] To achieve controlled variable verification, three comparative models were constructed in this experiment; Model A is a conventional single-modal depth residual network (calculated only from the total light intensity image);

[0259] Model B is a cascaded network (inputting total light intensity image and microcrack morphology features); Model C is the heterogeneous neural network of this invention (integrating total light intensity image, physical space attention mask and one-dimensional morphology features).

[0260] The hyperparameters of the model training were kept consistent, with an initial learning rate of 0.001, a batch size of 32, and an iteration period of 100 epochs.

[0261] After evaluation using forward inference on the test set, Model A achieved an overall classification accuracy of 82.4%, but it exhibited severe misclassifications in the optical feature boundary region between newly formed ice and one-year-old ice, resulting in a recall rate of only 79.1%.

[0262] After introducing macroscopic morphological constraints on microcracks, Model B improved its accuracy to 87.6%.

[0263] The proposed model C, due to its deep integration of a physical rule mask based on Stokes polarization analysis, effectively overcomes the masking of shallow crack features by the high reflectivity of the surface, achieving an overall classification accuracy of 95.8%, with precision and recall rates of 95.2% and 96.1%, respectively.

[0264] like Figure 8As shown in the ROC validation curve, after recording the network training states with and without the feature distance constraint loss term.

[0265] Experimental data show that the feature distance constraint loss introduced in this invention Afterwards, the intra-class aggregation degree of heterogeneous modes in the hidden layer space is significantly improved. The area under the curve (AUC) value of the method proposed in this invention is 0.975, which is better than 0.912 after removing the feature distance constraint.

[0266] Furthermore, thanks to the dynamic routing interception mechanism based on spatial distribution entropy, which eliminates more than 30% of redundant smooth ice surface data, the system of this invention takes only 45 milliseconds for a single frame complete inference on the edge computing terminal, fully meeting the timeliness requirements for dynamic ice condition detection during polar navigation.

[0267] This confirms that the present invention, while ensuring recognition accuracy, possesses outstanding algorithm completeness and engineering deployment feasibility.

[0268] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A sea ice age type identification method based on polarized light microcrack detection, characterized in that, The method, applicable to a computing system including a polarization acquisition module, a polarization parameter analysis module, a feature extraction and dynamic routing module, and an online inference module, comprises only one independent claim and includes the following steps: The polarization acquisition module acquires the polarization image of the target sea ice and constructs the original polarization tensor; The polarization parameter analysis module processes the original polarization tensor, calculates the total light intensity matrix, and generates a full-size physical space attention mask. The feature extraction and dynamic routing module extracts one-dimensional morphological feature vectors based on the total light intensity matrix, calculates the spatial distribution entropy of the divided local grid, and extracts the local total light intensity tensor and the local physical attention mask. The online inference module inputs the local total light intensity tensor, the local physical attention mask, and the one-dimensional morphological feature vector into the network model for forward computation, and outputs the ice age category label.

2. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 1, characterized in that, The polarization acquisition module acquires the polarization image of the target sea ice and constructs the original polarization tensor, including: Within the exposure cycle, the pixel values ​​of the target field of view under the preset polarization direction are obtained, and the pixel values ​​of each polarization direction are extracted into a sparse matrix. Spatial interpolation is performed on the sparse matrix to generate multiple independent polarization images, and the multiple independent polarization images are stacked to construct the original polarization tensor.

3. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 2, characterized in that, The polarization parameter analysis module processes the original polarization tensor, calculates the total intensity matrix, and generates a full-size physical space attention mask, including: The light intensity components of each polarization direction in the original polarization tensor are calculated based on the Stokes vector equation to obtain the total light intensity matrix and the linear polarization degree matrix. Perform an inverse normalization operation on the linear polarization degree matrix to generate the full-size physical space attention mask.

4. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 3, characterized in that, The feature extraction and dynamic routing module extracts a one-dimensional morphological feature vector based on the total light intensity matrix, including: The total light intensity matrix is ​​input into the segmentation network to generate a binary mask with microcracks. The number of independent connected components in the binary mask of microcracks is counted by a connected component labeling algorithm, and the area density of microcracks is calculated. The number of independent connected regions and the microcrack area density are concatenated to construct the one-dimensional morphological feature vector.

5. The method according to claim 4, characterized in that, The feature extraction and dynamic routing module calculates the spatial distribution entropy of the divided local grid and extracts the local total light intensity tensor and the local physical attention mask, including: The proportion of microcrack pixels within each local grid to the total number of pixels in the corresponding grid is statistically analyzed, and the spatial distribution entropy is obtained by calculating the proportion based on the information entropy formula. Extract the spatial coordinate bounding boxes of local grids whose spatial distribution entropy exceeds a preset threshold; Based on the spatial coordinate bounding box, the local total light intensity tensor and the local physical attention mask are extracted on the total light intensity matrix and the full-size physical spatial attention mask.

6. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 1, characterized in that, The network model is a two-branch heterogeneous network model; The online inference module inputs the local total light intensity tensor, the local physical attention mask, and the one-dimensional morphological feature vector into the network model for forward computation, including: The local total light intensity tensor is input into a convolutional branch network to generate a two-dimensional optical feature map. The local physical attention mask is downsampled to generate an alignment mask. The alignment mask is then multiplied element-wise with the two-dimensional optical feature map, and a one-dimensional optical feature vector is output through a pooling layer. The one-dimensional morphological feature vector is input into a multilayer perceptron branch network, which outputs a one-dimensional hidden layer feature vector. The one-dimensional optical feature vector and the one-dimensional hidden layer feature vector are concatenated to form a fused feature vector, which is then input into a fully connected classification head to output the ice age category label.

7. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 6, characterized in that, The node weights of the network model are updated during the training phase using a joint loss function. The joint loss function includes classification cross-entropy loss and feature distance constraint loss, wherein the feature distance constraint loss is used to calculate the spatial distance between the one-dimensional optical feature vector and the one-dimensional hidden layer feature vector.

8. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 7, characterized in that, The computing system also includes a state evaluation module; After the online inference module outputs the ice age category label, the method further includes: The state assessment module calculates the local risk coefficient based on the ice age category label, microcrack area density, and the mean of the local physical attention mask. Based on the spatial coordinate bounding box, the local risk coefficient is backfilled into the corresponding coordinate interval of the global risk distribution matrix.

9. The AI-based method for identifying sea ice age and type based on polarized light microcrack detection according to claim 8, characterized in that, The computing system also includes an early warning output module; After the local risk coefficients are backfilled to the corresponding coordinate interval of the global risk distribution matrix, the method further includes: The early warning output module performs spatial integration and mean averaging calculations on the element values ​​within the global risk distribution matrix to obtain the comprehensive navigation safety index.

10. The method according to claim 9, characterized in that, The complete steps of the sea ice microcrack detection and numerical simulation calculation process included in the method are as follows: Obtain pixel values ​​in the preset polarization direction and construct the original polarization tensor through sparse matrix interpolation and stacking; The linear polarization degree matrix is ​​solved by algebraic operations based on the Stokes vector equation, and a full-size physical space attention mask is generated by numerical inversion calculation. The binary mask generated by the statistical segmentation network is used to calculate the number of connected components and pixel area density, thereby generating the one-dimensional morphological feature vector. The information entropy of the local microcrack ratio is calculated based on the preset mesh division parameters. Local mesh bounding boxes with information entropy greater than the threshold are extracted, and the local total light intensity tensor and the local physical attention mask are extracted. The one-dimensional optical feature vector is generated by performing element-wise multiplication on the optical feature map using the downsampled local physical attention mask. The one-dimensional optical feature vector is concatenated with the one-dimensional hidden layer feature vector mapped from the one-dimensional morphological feature vector, and the ice age category label is obtained by probabilistic parsing through a fully connected classification head. The preset risk coefficient is queried based on the ice age category label and backfilled into the global risk distribution matrix. The comprehensive navigation safety index is obtained by performing two-dimensional spatial integration on the element values ​​of the global risk distribution matrix. Based on the threshold comparison results, a warning signal containing coordinates is output.

Images