Channel warning method and device based on low earth orbit satellite constellation and storage medium

Abstract

This application discloses a navigation channel early warning method, device, and storage medium based on a low-Earth orbit (LEO) satellite constellation, relating to the field of sea ice monitoring technology. The method includes: a first satellite acquiring navigation channel images; inputting sea ice images contained within the navigation channel images into a pre-trained gradient calculation model to generate corresponding gradient feature maps; classifying the sea ice images to determine their brightness categories; and then calling a first U-Net network, a second U-Net network, and a transition matrix corresponding to the brightness category to determine the sea ice edge lines of the sea ice images. If the first satellite determines that the sea ice poses a threat to a preset navigation channel based on the edge lines, it transmits sea ice-related early warning information to a second satellite in the LEO satellite constellation; the second satellite then transmits the early warning information to vessels within the coverage area. Thus, this application achieves on-orbit identification and real-time broadcasting of navigation channel risks, as well as feature enhancement and targeted processing of sea ice images under different lighting conditions.

Description

Technical Field

[0001] This application relates to the field of sea ice monitoring technology, and in particular to a navigation early warning method, device and storage medium based on a low-Earth orbit satellite constellation. Background Technology

[0002] Currently, for navigation safety early warning in maritime routes (especially polar routes, busy waterways, and coastal areas), existing technologies using low-Earth orbit (LEO) satellites for route monitoring mostly rely on isolated interpretation of single data sources (such as visible light or SAR images). This results in fragmented and incomplete early warning information. For example, they may only identify the presence of sea ice but cannot accurately assess its dynamic threat to specific routes; or they may only detect the position of vessels but cannot determine whether they have deviated from their routes or entered dangerous areas. Furthermore, due to the lack of intelligent processing methods for multi-satellite coordination and multi-temporal data streams from LEO constellations, the timeliness of existing systems is insufficient and cannot meet the modern shipping industry's demand for real-time, accurate, and forward-looking safety early warnings.

[0003] Meanwhile, in the field of sea ice edge detection, existing methods typically rely on a large number of sea ice image samples to train prediction models and Markov chains. For example, related technologies construct training sets by collecting images of multiple types of sea ice to optimize edge recognition algorithms.

[0004] However, these methods often use samples with significant brightness differences and lack effective normalization or feature enhancement processing for factors such as lighting conditions and ice surface reflectivity. Due to uneven brightness distribution and inconsistent feature representation among samples, the edge features between sea ice and the background, as well as between different ice states, are poorly distinguishable, resulting in blurred morphological and textural information. Directly inputting such heterogeneous samples with scattered feature distributions into the model for training not only interferes with the model's learning of key edge features but may also introduce significant noise, reducing the model's generalization ability and detection accuracy under complex ice conditions, ultimately affecting the accuracy and reliability of sea ice edge recognition results.

[0005] There are currently no effective solutions to the technical problems of insufficient timeliness of existing low-orbit satellite early warning systems, which cannot meet the modern shipping industry's demand for real-time, accurate and forward-looking safety early warnings, and the blurring and insufficient distinguishability of edge features due to significant brightness differences in sea ice image samples without effective normalization, which affects the accuracy and reliability of sea ice edge recognition results. Summary of the Invention

[0006] The embodiments of this disclosure provide a navigation early warning method, apparatus, and storage medium based on a low-Earth orbit satellite constellation, to at least solve the technical problems existing in the prior art, such as insufficient timeliness of existing low-Earth orbit satellite early warnings, which cannot meet the requirements of modern shipping for real-time, accurate, and forward-looking safety early warnings, and the fact that sea ice image samples have blurred edge features and insufficient distinguishability due to significant brightness differences and lack of effective normalization, which affects the accuracy and reliability of sea ice edge recognition results.

[0007] According to one aspect of the present disclosure, a navigation channel warning method based on a low-Earth orbit (LEO) satellite constellation is provided, comprising: a first satellite of the LEO satellite constellation acquiring navigation channel images related to a preset navigation channel; the first satellite identifying the edges of sea ice contained in the navigation channel images based on the LEO satellite constellation navigation channel warning method; if the first satellite determines that sea ice poses a threat to the preset navigation channel based on the edges, it transmits sea ice-related warning information to a second satellite of the LEO satellite constellation, wherein the second satellite covers at least a portion of the preset navigation channel; and the second satellite transmitting the warning information to vessels within the coverage area, wherein the operation of identifying the edges of sea ice contained in the sea ice image area based on the LEO satellite constellation navigation channel warning method includes: inputting the sea ice image contained in the navigation channel image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image according to the gradient feature map to determine the brightness category of the sea ice image; and, according to the brightness category of the sea ice image, calling a first U-Net network, a second U-Net network, and a transition matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image.

[0008] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0009] According to another aspect of the present disclosure, a navigation channel early warning device based on a low-Earth orbit (LEO) satellite constellation is also provided, comprising: a navigation channel image acquisition module, used by a first satellite of the LEO satellite constellation to acquire navigation channel images related to a preset navigation channel; a sea ice edge recognition module, used by the first satellite to identify the edge of sea ice contained in the navigation channel image based on a LEO satellite constellation navigation channel early warning method; and an early warning information transmission module, used by the first satellite to transmit sea ice-related early warning information to a second satellite of the LEO satellite constellation when the sea ice is determined to pose a threat to the preset navigation channel based on the edge, wherein the second satellite covers at least a portion of the preset navigation channel; The system includes a navigation warning module for transmitting warning information from the second satellite to vessels within its coverage area. The navigation warning method based on a low-Earth orbit satellite constellation identifies the edges of sea ice contained within a sea ice image region. This includes: inputting the sea ice image contained in the navigation image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image based on the gradient feature map to determine its brightness category; and, based on the brightness category of the sea ice image, calling the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image.

[0010] According to another aspect of the present disclosure, a navigational warning device based on a low-Earth orbit (LEO) satellite constellation is also provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions for processing the following steps: a first satellite of the LEO satellite constellation acquires navigational images related to a preset navigational route; the first satellite identifies the edges of sea ice contained in the navigational images based on a LEO satellite constellation navigational warning method; and, if the first satellite determines that sea ice poses a threat to the preset navigational route based on the edges, it transmits sea ice-related warning information to a second satellite of the LEO satellite constellation, wherein the second satellite covers the preset navigational route. At least a portion; and the second satellite sends early warning information to ships within the coverage area, wherein the navigation early warning method based on a low-Earth orbit satellite constellation, the operation of identifying the edge of sea ice contained in the sea ice image area includes: inputting the sea ice image contained in the navigation image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image according to the gradient feature map to determine the brightness category of the sea ice image; and calling the first U-Net network, the second U-Net network and the transfer matrix corresponding to the brightness category of the sea ice image to determine the sea ice edge line of the sea ice image according to the brightness category of the sea ice image.

[0011] This application first acquires images of the waterway covering the preset waterway from the first satellite of the low-Earth orbit satellite constellation, and identifies the edges of sea ice contained in the waterway images. Based on the identification results, it is determined whether the sea ice edge will pose a threat to the preset waterway. When it is determined that the sea ice edge will pose a threat to the preset waterway, the first satellite prioritizes sending it to the second satellite that can currently or in the near future cover the relevant waterway area via an inter-satellite link. The second satellite then broadcasts the warning information, including the sea ice location, threat level, and suggested routes, directly to all vessels within its coverage area via a dedicated maritime satellite communication link.

[0012] This technical solution accurately assesses the location of sea ice and shipping channels, predicts the degree of threat posed by sea ice to shipping channels, and uses low-Earth orbit satellite constellations for collaborative communication and to predict the degree of threat to shipping channels, thus providing early warnings to passing ships and meeting the modern shipping industry's demand for real-time, accurate, and forward-looking safety warnings.

[0013] Subsequently, sea ice images contained within the channel image are input into a pre-trained gradient calculation model to generate gradient feature maps covering four directions: horizontal, vertical, and two oblique directions (e.g., 45° and 135°), comprehensively capturing edge information in different orientations. Next, the multi-directional gradient feature maps are input into a fully connected multilayer perceptron. This multilayer perceptron learns the correlation and complementarity between gradient features in different directions, uncovering the common representations and differentiated expressions of sea ice edges across multiple dimensions, ultimately outputting a fused feature map that preserves the original spatial dimensions. Then, based on the brightness distribution, gradient response, and texture structure information contained in this fused feature map, the similarity between it and multiple preset category feature centers is calculated, adaptively classifying the current sea ice image into the most matching brightness category. Finally, based on the determined brightness category, the system invokes a first U-Net network for quickly extracting the macroscopic preliminary edge contour, a second U-Net network for outputting feature tensors containing detailed semantic information, and a Markov chain transition matrix for modeling and optimizing the probabilistic dependency between the detailed features of the second U-Net network and the contour output of the first U-Net network. Through probabilistic fusion and optimization of the dual-network results, refined correction, completion, and final determination of the sea ice edge are achieved.

[0014] This application constructs an early warning link based on on-board real-time processing and inter-satellite collaborative distribution, achieving on-orbit identification and immediate broadcasting of navigational risks. Furthermore, by introducing multi-directional gradient feature fusion, it enhances and specifically processes sea ice images under different lighting conditions, significantly improving the adaptability of sea ice edge recognition under varying illumination. Through brightness classification, it adaptively matches the most suitable first U-Net network, second U-Net network, and transition matrix for each image, ensuring optimal matching between model parameters and image characteristics. This achieves high-precision and robust extraction of sea ice edges. This addresses the technical problems of insufficient timeliness of existing low-Earth orbit satellite early warnings, failing to meet the demands of modern shipping for real-time, accurate, and forward-looking safety warnings, and the blurring and insufficient discrimination of edge features due to significant brightness differences in sea ice image samples without effective normalization, thus affecting the accuracy and reliability of sea ice edge recognition results. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings: Figure 1 This is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of this disclosure; Figure 2 This is a flowchart illustrating the airway early warning method based on a low-Earth orbit satellite constellation as described in Embodiment 1 of this disclosure. Figure 3 This is a schematic diagram of the satellite-covered airway according to Embodiment 1 of this disclosure, which describes the airway early warning method based on a low-Earth orbit satellite constellation. Figure 4 This is a schematic flowchart illustrating the generation of the fusion feature map according to the method described in Embodiment 1 of this disclosure; Figure 5 This is a schematic diagram illustrating the classification of methods according to Embodiment 1 of this disclosure; Figure 6 This is a schematic diagram of the first U-Net network structure according to the method described in Embodiment 1 of this disclosure; Figure 7 This is a schematic diagram of the second U-Net network structure according to the method described in Embodiment 1 of this disclosure; Figure 8 This is a schematic diagram of region classification according to the method described in Embodiment 1 of this disclosure; Figure 9 This is a schematic diagram of a low-Earth orbit satellite constellation-based navigation early warning device according to Embodiment 2 of this disclosure; and Figure 10 This is a schematic diagram of a low-Earth orbit satellite constellation-based airway early warning device according to Embodiment 3 of this disclosure. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] Example 1 According to this embodiment, a method embodiment for airway early warning based on a low-Earth orbit satellite constellation is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0019] Figure 1 This is a schematic diagram of the hardware architecture of satellite system 10. (Reference) Figure 1 As shown, satellite system 10 includes an integrated electronic system, which includes a processor, a memory, a bus management module, and a communication interface. The memory is connected to the processor, allowing the processor to access the memory, read program instructions stored in the memory, read data from the memory, or write data to the memory. The bus management module is connected to the processor and also to a bus such as a CAN bus. Thus, the processor can communicate with onboard peripherals connected to the bus through the bus managed by the bus management module. These onboard peripherals include: onboard peripheral 1 (GNSS module), onboard peripheral 2 (fiber optic gyroscope), ... onboard peripheral n (torque flywheel). Furthermore, the processor also communicates with devices such as cameras, star sensors, telemetry and control transponders, and data transmission equipment via the communication interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a satellite system may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0020] It should be noted that, Figure 1 One or more processors and / or other data processing circuits shown herein may generally be referred to as "data processing circuitry". This data processing circuitry may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuitry may be a single, independent processing module, or may be integrated, in whole or in part, into any other element in a computing device. As involved in embodiments of this disclosure, the data processing circuitry serves as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0021] Figure 1 The memory shown can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method for determining the communication frequency band corresponding to the beam in the embodiments of this disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, to implement the above-mentioned method for determining the communication frequency band corresponding to the beam in the application program. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory.

[0022] It should be noted here that, in some optional embodiments, the above... Figure 1 The device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned devices.

[0023] Under the aforementioned operating environment, according to the first aspect of this embodiment, a route early warning method based on a low-Earth orbit satellite constellation is provided. This method comprises... Figure 3 The first satellite shown is implemented, wherein the first satellite includes Figure 1 Satellite system 10 is shown. Figure 2 A flowchart illustrating the method is shown below. (Refer to...) Figure 2 As shown, the method includes: S202: The first satellite of the low-Earth orbit satellite constellation acquires imagery of the route related to the preset route. S204: The first satellite uses a low-Earth orbit satellite constellation-based navigational early warning method to identify the edges of sea ice contained in navigational images; S206: If the first satellite determines, based on the edge of the sea ice, that it poses a threat to the predetermined shipping route, it will transmit sea ice-related early warning information to the second satellite in the low-Earth orbit satellite constellation, wherein the second satellite covers at least a portion of the predetermined shipping route; and S208: The second satellite sends early warning information to ships within the coverage area. The navigation warning method based on a low-Earth orbit satellite constellation includes the following steps for identifying the edges of sea ice contained within a sea ice image area: inputting the sea ice image contained in the navigation image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image according to the gradient feature map to determine its brightness category; and, based on the brightness category of the sea ice image, calling the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image.

[0024] Specifically, such as Figure 3 As shown, a low-Earth orbit satellite constellation consists of multiple satellites working together. In this embodiment, the first satellite and the second satellite are used as examples for illustration.

[0025] Specifically, the first satellite first acquires images of the shipping channel (corresponding to step S202). These images include information on the distribution of sea ice in the preset shipping channel and its surrounding waters.

[0026] Then, combine Figure 4 As shown, the first satellite uses a low-Earth orbit (LEO) satellite constellation-based navigation channel warning method to identify the edges of sea ice contained in navigation channel images. The LEO satellite constellation-based navigation channel warning method involves using the LEO satellite constellation to assess navigation channels in navigation channel images, thereby determining whether a threat exists in the navigation channel and thus requiring a warning (corresponding to step S204).

[0027] Specifically, the first satellite inputs sea ice images (sea ice areas included in the navigation channel images) into a pre-trained gradient calculation model to generate corresponding gradient feature maps. This gradient calculation model is a neural network model or digital image processing model trained on samples and capable of pixel gradient calculations, which can calculate the rate of change of grayscale values ​​for each pixel in the input sea ice image.

[0028] It should be noted that this gradient feature map includes a horizontal gradient map, a vertical gradient map, a first oblique gradient map, and a second oblique gradient map. The horizontal gradient map is a feature map generated by the gradient calculation model after calculating the gray-level differences of pixels in the sea ice image along the horizontal direction. It is used to characterize the brightness changes of the image in the horizontal direction (x-direction) and is sensitive to vertical edges (because vertical edges have drastic changes in the horizontal direction). The vertical gradient map is a feature map generated by the gradient calculation model after calculating the gray-level differences of pixels in the sea ice image along the vertical direction. It is used to characterize the brightness changes of the image in the vertical direction (y-direction) and is sensitive to horizontal edges (because horizontal edges have drastic changes in the vertical direction). The first and second oblique gradient maps are feature maps generated by the gradient calculation model after calculating the gray-level differences of pixels in the sea ice image in two mutually perpendicular oblique directions (such as 45° and 135°), respectively. They are used to characterize the oblique sea ice edge information in the image.

[0029] Furthermore, the first satellite concatenates the horizontal gradient map, vertical gradient map, first oblique gradient map, and second oblique gradient map along the channel dimension to form an input feature tensor containing multi-directional edge features (its dimension can be represented as H×W×C, where H is the number of height pixels of the sea ice image, W is the number of width pixels, and C is the number of channels of the gradient feature map; in this embodiment, C=4). This input feature tensor is then input into a pre-trained multilayer perceptron (MLP). This MLP is a fully connected neural network model that learns the relationships between gradient features in different directions (such as the cross-response between horizontal and vertical gradients, and the fusion of complementary information between oblique gradients and other directional gradients) to uncover the common features and differentiated expressions of sea ice edges in multiple dimensions. Then, through dimension mapping operations, a fused feature map that maintains spatial dimensional consistency with the input gradient feature map (i.e., the H×W dimension remains unchanged) is output.

[0030] Next, the first satellite classifies the input sea ice image according to the fused feature map, determining its brightness category. Specifically, based on information such as brightness distribution, gradient response, and texture structure in the sea ice image corresponding to the fused feature map, the similarity between the fused feature map and n preset category feature centers is calculated, and the current sea ice image is assigned to the brightness category with the highest similarity. This brightness category represents a set of sea ice images clustered in the feature space and possessing similar imaging characteristics (including but not limited to overall illumination conditions, ice surface reflectivity, and visual appearance). For example, based on the physical state and appearance of sea ice, common categories may include, but are not limited to: Open water: typically exhibits the lowest brightness characteristics, appearing as dark areas in images; New ice / thin ice: It is thinner and has lower brightness characteristics, appearing as a dark gray color; One-year ice: has a moderate level of brightness characteristics, and the surface brightness may be enhanced due to snow cover; Multi-year ice: Due to its dense structure, salt precipitation, and increased bubbles, it usually exhibits the highest brightness characteristics; Melting pool: Brightness characteristics depend on water depth and may appear as localized bright areas; Snow-covered areas: Due to the high reflectivity of snow, these areas appear as characteristic bright areas.

[0031] Finally, the first satellite, based on the brightness category of the sea ice image, invokes the corresponding first U-Net network, second U-Net network, and transition matrix to determine the sea ice edge line of the image. Specifically, both the first and second U-Net networks are semantic segmentation networks based on the U-Net architecture, and they process the same input sea ice image in parallel. In practical applications, the first U-Net network can be configured to focus on quickly extracting preliminary, explicit edge structures. The second U-Net network can be configured to characterize the probability value of each pixel in the sea ice image belonging to a snow-covered area, a sea ice melting area, or a normal area. The transition matrix (Markov chain transition matrix) is a probability transition matrix learned in advance through training. This matrix establishes the probabilistic dependency and optimization relationship between the rich feature tensor (context and detail states) output by the second U-Net network and the contour map (macro-boundary states) output by the first U-Net network. That is, it uses the detail and context features provided by the second network to refine, complete, and optimize the macro-contour generated by the first network.

[0032] Furthermore, after identifying the edge of the sea ice contained in the image of the shipping channel, the first satellite compares the location of the sea ice edge with the location of the shipping channel to further determine whether the sea ice edge poses a threat to the predetermined shipping channel. If the sea ice edge poses a threat to the predetermined shipping channel, it means that ships traveling along the predetermined shipping channel may hit the ice when they reach the sea ice location.

[0033] When the first satellite determines that the edge of the sea ice poses a threat to the preset shipping lane, it sends the warning information to the second satellite, which can currently or in the near future cover the relevant shipping lane area, via an inter-satellite link (corresponding to step S206). This solves the problem of limited overhead time and insufficient continuous coverage capability of a single low-orbit satellite for a specific area. Through real-time information relay within the constellation, it ensures that the warning information can be transmitted to the most suitable downlink node as soon as possible.

[0034] Furthermore, after receiving the warning information from the first satellite, the second satellite, since its instantaneous coverage area includes at least a portion of the threatened pre-defined shipping lane, can directly broadcast the warning information, including sea ice location, threat level, and suggested routes, to all vessels within the pre-defined shipping lane within its coverage area via a dedicated maritime satellite communication link (such as VDES, satellite AIS, etc.) (corresponding to step S208). This notifies vessels to avoid the pre-defined shipping lane where sea ice exists, preventing them from encountering sea ice while sailing in that lane. This significantly reduces the delay from detection to warning, avoiding potential delays introduced by information relay through ground stations.

[0035] As described in the background section, existing methods for waterway monitoring using low-Earth orbit satellites largely rely on isolated interpretation of single data sources (such as visible light or SAR images), leading to fragmented and incomplete early warning information. Furthermore, current sea ice edge detection methods typically depend on training prediction models and Markov chains with a large number of sea ice image samples. For example, related technologies construct training sets by collecting multiple types of sea ice images to optimize edge recognition algorithms. However, the samples used in these methods often exhibit significant brightness differences and lack effective normalization or feature enhancement processing for factors such as lighting conditions and ice surface reflectivity. Due to uneven brightness distribution and inconsistent feature representation among samples, the edge feature differentiation between sea ice and the background, as well as between different ice states, is insufficient, resulting in blurred morphological and textural information. Directly inputting such heterogeneous and haphazardly distributed samples into the model for training not only interferes with the model's learning of key edge features but may also introduce significant noise, reducing the model's generalization ability and detection accuracy under complex ice conditions, ultimately affecting the accuracy and reliability of sea ice edge recognition results.

[0036] In view of this, this application first acquires channel images covering the preset channel from the first satellite of the low-Earth orbit satellite constellation, and identifies the edges of sea ice contained in the channel images. Based on the identification results, it is determined whether the sea ice edge will pose a threat to the preset channel. When it is determined that the sea ice edge will pose a threat to the preset channel, the first satellite prioritizes sending it to the second satellite that can currently or in the near future cover the relevant channel area via an inter-satellite link. The second satellite then directly broadcasts the warning information, including the sea ice location, threat level, and suggested routes, to all vessels within its coverage area via a dedicated maritime satellite communication link.

[0037] This technical solution accurately assesses the location of sea ice and shipping channels, predicts the degree of threat posed by sea ice to shipping channels, and uses low-Earth orbit satellite constellations for collaborative communication and to predict the degree of threat to shipping channels, thus providing early warnings to passing ships and meeting the modern shipping industry's demand for real-time, accurate, and forward-looking safety warnings.

[0038] Subsequently, sea ice images contained within the channel image are input into a pre-trained gradient calculation model to generate gradient feature maps covering four directions: horizontal, vertical, and two oblique directions (e.g., 45° and 135°), comprehensively capturing edge information in different orientations. Next, the multi-directional gradient feature maps are input into a fully connected multilayer perceptron. This multilayer perceptron learns the correlation and complementarity between gradient features in different directions, uncovering the common representations and differentiated expressions of sea ice edges across multiple dimensions, ultimately outputting a fused feature map that preserves the original spatial dimensions. Then, based on the brightness distribution, gradient response, and texture structure information contained in this fused feature map, the similarity between it and multiple preset category feature centers is calculated, adaptively classifying the current sea ice image into the most matching brightness category. Finally, based on the determined brightness category, the system invokes a first U-Net network for quickly extracting the macroscopic preliminary edge contour, a second U-Net network for outputting a feature tensor containing detailed semantic information, and a Markov chain transition matrix for modeling and optimizing the probabilistic dependency between the detailed features of the second U-Net network and the contour output of the first U-Net network. Through probabilistic fusion and optimization of the dual-network results, refined correction, completion, and final determination of the sea ice edge are achieved.

[0039] This application constructs an early warning link based on on-board real-time processing and inter-satellite collaborative distribution, achieving on-orbit identification and immediate broadcasting of navigational risks. Furthermore, by introducing multi-directional gradient feature fusion, it enhances and specifically processes sea ice images under different lighting conditions, significantly improving the adaptability of sea ice edge recognition under varying illumination. Through brightness classification, it adaptively matches the most suitable first U-Net network, second U-Net network, and transition matrix for each image, ensuring optimal matching between model parameters and image characteristics. This achieves high-precision and robust extraction of sea ice edges. This addresses the technical problems of insufficient timeliness of existing low-Earth orbit satellite early warnings, failing to meet the demands of modern shipping for real-time, accurate, and forward-looking safety warnings, and the blurring and insufficient discrimination of edge features due to significant brightness differences in sea ice image samples without effective normalization, thus affecting the accuracy and reliability of sea ice edge recognition results.

[0040] Optionally, the operation of determining that sea ice poses a threat to the shipping channel based on the edge includes: determining the nearest first distance between the edge and the preset shipping channel; and determining that sea ice poses a threat to the preset shipping channel if the first distance is less than a preset distance threshold.

[0041] Specifically, the first satellite uses image analysis algorithms to calculate the shortest spatial distance between each edge pixel on the identified sea ice edge line and the boundary of a preset shipping channel (represented in the image as one or more predefined shipping channel regions or centerlines). By traversing and calculating all edge points, the minimum value is selected as the first distance. This first distance reflects the degree to which the sea ice is closest to the shipping channel and is the primary quantitative indicator for assessing whether the sea ice poses a threat to the preset shipping channel.

[0042] It should be noted that image analysis algorithms refer to a collection of digital image processing and geometric calculation methods that run on the first satellite and are used to extract and quantify spatial relationships from images. It is a general term for program steps known to those skilled in the art for calculating the minimum distance between a specific target in an image and a preset geographical area.

[0043] Furthermore, the first satellite uses a pre-set safety boundary parameter as a preset distance threshold. This preset distance threshold is determined comprehensively based on factors such as navigation safety regulations, ship handling characteristics, sea ice drift speed, and error tolerance. Then, the first satellite compares the calculated first distance with the preset distance threshold in real time. If the first distance is less than the preset distance threshold, it indicates that the sea ice has intruded or is sufficiently close to the safety boundary of the shipping channel, and the first satellite automatically determines that the sea ice poses a threat to the preset shipping channel.

[0044] By employing the above methods, visual edge information is transformed into clear risk decisions, achieving a crucial leap from observing sea ice to assessing its danger. By introducing quantified, preset distance thresholds, the first satellite on board can identify urgent risks requiring immediate warnings in real time, automatically, and consistently. This directly supports the generation and triggering of subsequent warning information, ensuring the timeliness and reliability of the entire warning process.

[0045] Optionally, the operation of the second satellite to send warning information to ships within the coverage area includes: the second satellite determining the direction of movement of the ship based on the ship's trajectory; and sending the warning information to the ship if it is determined based on the direction of movement that the ship is moving towards the sea ice.

[0046] Specifically, the second satellite first acquires AIS messages from vessels within its coverage area in real time through its onboard Automatic Identification System (AIS) receiver or by fusing other onboard sensing data. From these messages, it parses the vessels' dynamic information, including but not limited to their position, speed above ground, and heading above ground. By processing and analyzing this dynamic information (e.g., calculating the vector changes of their recent positions), the satellite determines the vessels' current direction of movement.

[0047] Then, the first satellite performs spatial correlation analysis between the vessel's direction of movement, current location, and the sea ice threat location included in the warning information. If the analysis indicates that the vessel's direction of movement vector is pointing in a direction that may lead it into the sea ice threat area (i.e., the vessel is currently or is about to sail towards this risk area), the vessel is determined to be a high-risk target, and a warning information containing details such as the specific location and extent of the sea ice is immediately sent to that specific vessel. Conversely, if the vessel's direction of movement indicates that it is moving away from or sailing parallel to the sea ice area within a safe range, the warning is not triggered or its transmission priority is reduced.

[0048] By using the above methods, we can ensure that early warning information is only sent to vessels that are truly facing imminent risks, and avoid irrelevant information interfering with other vessels.

[0049] Optionally, the operation of inputting the sea ice image into a pre-trained gradient calculation model to generate the corresponding gradient feature map includes: inputting the sea ice image into the pre-trained gradient calculation model; and the gradient calculation model performing convolution operations on the sea ice image using a horizontal kernel, a vertical kernel, a first oblique kernel, and a second oblique kernel, respectively, and outputting the corresponding horizontal gradient map, vertical gradient map, first oblique gradient map, and second oblique gradient map.

[0050] Specifically, sea ice images are input into a pre-trained gradient calculation model. The model then performs convolution operations on the sample sea ice images using horizontal kernels, vertical kernels, 45° oblique kernels, and 135° oblique kernels, thereby outputting corresponding horizontal gradient maps, vertical gradient maps, first oblique gradient map (45 degrees), and second oblique gradient map (135 degrees), achieving comprehensive capture of sea ice edges with different orientations.

[0051] It should be noted that the gradient calculation model has four pre-defined convolutional kernels (horizontal kernel, vertical kernel, first diagonal kernel, and second diagonal kernel). Each kernel is a fixed-size matrix (e.g., a 3×3 matrix) with edge detection capabilities, and the matrix element values ​​are designed based on gradient operators, such as the kernel parameters of the Sobel and Prewitt operators. The weight parameters of each kernel have been optimized through training to ensure sensitivity to brightness changes in the corresponding direction. In the application phase, the gradient calculation model controls each kernel to perform sliding convolution on the pixel matrix of the sea ice image with a pre-defined stride (e.g., stride of 1). The kernel performs element-wise multiplication and summation with the corresponding local pixel region in the image to obtain the gradient value at that location (the gradient value represents the intensity of the pixel grayscale change, and the sign of the gradient value represents the direction of change). By traversing all pixel regions of the image, four single-channel feature maps with the same spatial dimension (H×W) as the input sea ice image are output: the horizontal gradient map, the vertical gradient map, the first diagonal gradient map, and the second diagonal gradient map.

[0052] Using the above method, this gradient calculation model generates four orientation-sensitive gradient feature maps in parallel from an input sea ice image. This achieves preliminary, directional extraction of multi-directional edge information from the sea ice image, providing comprehensive and physically meaningful low-level feature inputs for subsequent feature fusion.

[0053] Optionally, a horizontal kernel is used to detect the vertical edges of the sea ice image and generate a horizontal gradient map. The corresponding computation matrix is: ; in, , , , , , The value is a preset value.

[0054] Specifically, the horizontal kernel is a 3×3 convolutional kernel used in the gradient calculation model to perform horizontal gradient operations. Its corresponding computation matrix is ​​a pre-configured fixed-parameter matrix, which is the core computational carrier for realizing vertical edge detection and generating horizontal gradient maps. This computational matrix structure is designed so that the middle column has zero values, while the left and right columns are assigned opposite numerical weights. This effectively suppresses interference from vertical (between vertical and horizontal pixels) grayscale changes during convolution operations, focusing instead on detecting horizontal (between horizontal and horizontal pixels) grayscale differences. For example, the horizontal kernel of the Sobel operator is: ; It's important to clarify that edge direction refers to the direction of the edge line itself (tangential direction), such as a vertical edge, which is an edge line running vertically. Gradient direction, on the other hand, refers to the direction of the greatest rate of change in image brightness, which is perpendicular to the edge line. In sea ice images, at the boundary between seawater and sea ice, the difference in reflectivity creates a brightness contrast, resulting in an edge. For example, when seawater appears dark and sea ice appears bright, or vice versa, a significant brightness jump occurs at the boundary. Therefore, for a vertically oriented edge line (i.e., the edge direction is vertical), its brightness changes most rapidly in the horizontal direction, hence its gradient direction is horizontal.

[0055] Meanwhile, this invention utilizes Sobel to determine kernels (horizontal kernels, vertical kernels, and oblique kernels), wherein the standard kernel for the horizontal kernel is fixed. , , It is a negative number. , , It is a positive number. Furthermore, , , , , , It is a constant pre-set based on gradient detection requirements (such as response sensitivity to grayscale changes in sea ice images and noise suppression capability). Its value logic can refer to the kernel parameter design rules of classic edge detection operators such as Sobel and Prewitt to ensure accurate capture of grayscale changes in the horizontal direction.

[0056] When the gradient calculation model performs operations, the calculation matrix of the horizontal kernel performs sliding convolution on the pixel matrix of the sea ice image with a preset stride (e.g., stride of 1). In an embodiment of the present invention, for any 3×3 local pixel region in the image, the pixel value of the region is multiplied with the corresponding element of the horizontal kernel calculation matrix, and then all multiplication results are summed to obtain the gradient value of the local region (i.e., the quantized value of the horizontal grayscale change). For example, if the pixel value of a certain local pixel region is: ; The corresponding gradient value is × + × + × + × + × + × (The elements in the middle column have no computational contribution as their values ​​are 0). By traversing all local pixel regions of the sea ice image and arranging the gradient values ​​of each region according to their spatial location, a horizontal gradient map can be generated. In this gradient map, the larger the gradient value, the more drastic the horizontal grayscale change in the sea ice image, indicating that this location represents the vertical edge region of the sea ice.

[0057] The above methods enable targeted detection of vertical edge regions in sea ice images. The generated horizontal gradient map can clearly quantify the degree of grayscale change in the horizontal direction of the image, while effectively eliminating the interference of vertical pixel changes.

[0058] Simultaneously, the vertical kernel is used to detect the horizontal edges of the sea ice image and generate a vertical gradient map. The corresponding calculation matrix is: ; in, , , , , , The value is a preset value.

[0059] Specifically, the vertical kernel is a 3×3 convolutional kernel used in the gradient calculation model to perform gradient operations in the vertical direction. Its corresponding computation matrix is ​​a pre-configured fixed-parameter matrix, which is the core computational carrier for realizing horizontal edge detection and generating vertical gradient maps. This computational matrix structure is designed so that the middle row has zero values, while the upper and lower rows are assigned opposite numerical weights. This effectively suppresses horizontal (between left and right pixels) grayscale variation interference during convolution operations, focusing instead on detecting vertical (between upper and lower pixels) grayscale differences. For example, the vertical kernel of the Sobel operator is: ; It should be noted that for horizontally oriented edge lines (i.e., edges with a horizontal direction), their brightness changes most rapidly in the vertical direction; therefore, their gradient direction is vertical. Meanwhile, , , It is a negative number. , , It is a positive number. Furthermore, , , , , , It is a constant pre-set based on gradient detection requirements (such as response sensitivity to grayscale changes in sea ice images and noise suppression capability). Its value logic can refer to the kernel parameter design rules of classic edge detection operators such as Sobel and Prewitt to ensure accurate capture of grayscale changes in the vertical direction.

[0060] During the gradient calculation model's computation, the vertical kernel matrix performs sliding convolution on the pixel matrix of the sea ice image with a preset stride (e.g., stride 1): for any 3×3 local pixel region in the image, the pixel value of that region is multiplied by the corresponding element of the vertical kernel calculation matrix, and then all multiplication results are summed to obtain the gradient value of that local region (i.e., the quantized value of the vertical grayscale change). For example, if the pixel value of a certain local pixel region is: ; The corresponding gradient value is × + × + × + × + × + × (The middle row elements have no computational contribution because their values ​​are 0). By traversing all local pixel regions of the sea ice image and arranging the gradient values ​​of each region according to their spatial location, a vertical gradient map can be generated. In this gradient map, the larger the gradient value, the more drastic the vertical grayscale change in the sea ice image, which corresponds to the horizontal edge region of the sea ice.

[0061] The above methods enable targeted detection of horizontal edge regions in sea ice images. The generated vertical gradient map can clearly quantify the degree of grayscale change in the vertical direction of the image, while effectively eliminating interference from pixel changes in the horizontal direction.

[0062] Additionally, the first oblique kernel is used to detect the upper left to lower right edge of the sea ice image, generating the first oblique gradient map. The corresponding calculation matrix is: ; in, , , , , , The value is a preset value.

[0063] Specifically, the first diagonal kernel is a 3×3 convolutional kernel used in the gradient calculation model to perform gradient operations from the bottom left to the top right (anti-diagonal). Its corresponding computation matrix is ​​a pre-configured fixed-parameter matrix, which is the core computational carrier for realizing edge detection from the top left to the bottom right (diagonal) and generating the first diagonal gradient map. This computational matrix structure is designed so that the values ​​in the diagonal direction are zero, while the two sides of the diagonal direction are assigned opposite numerical weights. This effectively suppresses grayscale variation interference in the diagonal direction during convolution operations, focusing instead on detecting grayscale differences in the anti-diagonal direction. For example, the first diagonal kernel of the Sobel operator is: ; It should be noted that for edge lines running diagonally (i.e., from the upper left to the lower right), the brightness changes most rapidly along the anti-diagonal direction; therefore, its gradient direction is along the anti-diagonal direction. Also, , , It is a negative number. , , It is a positive number. Furthermore, , , , , , It is a constant pre-set based on gradient detection requirements (such as response sensitivity to grayscale changes in sea ice images and noise suppression capability). Its value logic can refer to the kernel parameter design rules of classic edge detection operators such as Sobel and Prewitt to ensure accurate capture of grayscale changes in the anti-diagonal direction.

[0064] During the gradient calculation model's computation, the first oblique kernel matrix performs sliding convolution on the pixel matrix of the sea ice image with a preset stride (e.g., stride of 1): for any 3×3 local pixel region in the image, the pixel value of that region is multiplied by the corresponding element of the first oblique kernel calculation matrix, and then all multiplication results are summed to obtain the gradient value of that local region (i.e., the quantized value of the grayscale change from the lower left to the upper right). For example, if the pixel value of a certain local pixel region is: ; The corresponding gradient value is × + × + × + × + × + × (Diagonal elements have no computational contribution as their value is 0). By traversing all local pixel regions of the sea ice image and arranging the gradient values ​​of each region according to their spatial location, the first oblique gradient map can be generated. In this gradient map, the larger the gradient value, the more drastic the grayscale change in the sea ice image from the lower left to the upper right, i.e., this location is the upper left to lower right edge region of the sea ice.

[0065] The above method enables targeted detection of the sea ice image from the upper left to the lower right edge region. The generated first diagonal gradient map can clearly quantify the degree of grayscale change in the anti-diagonal direction of the image, while effectively eliminating the interference of pixel changes in the diagonal direction.

[0066] In addition, the second oblique kernel is used to detect the lower left to upper right edge of the sea ice image, generating a second oblique gradient map, and the corresponding calculation matrix is: ; in, , , , , , The value is a preset value.

[0067] Specifically, the second diagonal kernel is a 3×3 convolutional kernel used in the gradient calculation model to perform gradient operations from the top left to the bottom right (diagonal). Its corresponding computation matrix is ​​a pre-configured fixed-parameter matrix, which is the core computational carrier for realizing edge detection from the bottom left to the top right (anti-diagonal) and generating the second diagonal gradient map. This computational matrix structure is designed so that the values ​​in the anti-diagonal direction are zero, while the left and right sides of the anti-diagonal direction are assigned opposite numerical weights. This effectively suppresses gray-level variation interference in the anti-diagonal direction during convolution operations, focusing instead on detecting gray-level differences in the diagonal direction. For example, the second diagonal kernel of the Sobel operator is: ; It should be noted that for edge lines with an anti-diagonal orientation (i.e., edge direction from lower left to upper right), the brightness changes most rapidly along the diagonal direction; therefore, its gradient direction is diagonal. Also, , , It is a negative number. , , It is a positive number. Furthermore, , , , , , It is a constant pre-set based on gradient detection requirements (such as response sensitivity to grayscale changes in sea ice images and noise suppression capability). Its value logic can refer to the kernel parameter design rules of classic edge detection operators such as Sobel and Prewitt to ensure accurate capture of grayscale changes in the diagonal direction.

[0068] During the gradient calculation model's computation, the second oblique kernel matrix performs sliding convolution on the pixel matrix of the sea ice image with a preset stride (e.g., stride of 1): for any 3×3 local pixel region in the image, the pixel value of that region is multiplied by the corresponding element of the second oblique kernel calculation matrix, and then all multiplication results are summed to obtain the gradient value of that local region (i.e., the quantized value of the grayscale change from the top left to the bottom right). For example, if the pixel value of a certain local pixel region is: ; The corresponding gradient value is × + × + × + × + × + × (Anti-diagonal elements have no computational contribution as their value is 0). By traversing all local pixel regions of the sea ice image and arranging the gradient values ​​of each region according to their spatial location, a second oblique gradient map can be generated. In this gradient map, the larger the gradient value, the more drastic the grayscale change in the sea ice image from the upper left to the lower right, i.e., this location is the lower left to upper right edge region of the sea ice.

[0069] The above method enables targeted detection of the sea ice image from the lower left to the upper right edge region. The generated second diagonal gradient map can clearly quantify the degree of grayscale change in the diagonal direction of the image, while effectively eliminating interference from pixel changes in the opposite diagonal direction.

[0070] Optionally, the first U-Net network and the second U-Net network are trained as follows: multiple sample sea ice images and edge information annotated in each sample sea ice image are acquired; each sample sea ice image is input into a preset gradient calculation model to generate corresponding gradient feature maps; the gradient feature maps of each sample sea ice image are input into a multilayer perceptron to obtain fused feature maps corresponding to each sample sea ice image; the fused feature maps of all sample sea ice images are clustered to generate multiple brightness categories, where each brightness category corresponds to a first U-Net network and a second U-Net network; and the first U-Net network and the second U-Net network corresponding to the brightness category are trained using sample sea ice images belonging to the same brightness category and the edge information corresponding to the sample sea ice images.

[0071] Specifically, the training process of the first U-Net network and the second U-Net network is based on the adaptive training logic of "sample preprocessing - feature extraction - category division - classification training" to ensure that each network can accurately adapt to the sea ice image features corresponding to the brightness category.

[0072] Specifically, combined Figure 4 As shown, firstly, sea ice images covering different lighting conditions and ice states (such as open water, new ice, multi-year ice, etc.) are collected as a sample set. Simultaneously, pixel-level edge annotation is performed on each sample sea ice image. That is, through manual annotation or high-precision detection equipment calibration, it is determined whether each pixel in the sample image belongs to the sea ice edge region, forming an edge annotation map consistent with the spatial dimension of the sample image (this annotation map serves as supervision data for training and is used for subsequent network loss calculation).

[0073] Next, each sample sea ice image is input into a preset gradient calculation model (with the same structure and parameters as the gradient calculation model used in the application stage). Convolution operations are performed using the model's built-in horizontal kernel, vertical kernel, and diagonal kernel to generate a multi-directional gradient feature map (containing gradient maps in the horizontal, vertical, and two diagonal directions) for each sample sea ice image, providing basic gradient information for subsequent feature extraction.

[0074] Furthermore, the multi-directional gradient feature maps corresponding to each sea ice image sample are concatenated into an input feature tensor along the channel dimension and then input into a pre-trained multilayer perceptron (with the same structure and parameters as the multilayer perceptron used in the application stage). The multilayer perceptron outputs a fused feature map with the same spatial dimension as the sample sea ice image through deep fusion operations on the multi-directional gradient features. This fused feature map integrates the omnidirectional edge information and brightness-related features of the sample.

[0075] Subsequently, reference Figure 5 As shown, based on the brightness-related features such as brightness distribution and gradient response contained in the fused feature maps, a pre-defined clustering algorithm (such as the K-Means algorithm) is used to group all sample sea ice images: sample sea ice images with similar feature distributions (i.e., similar brightness characteristics and edge feature patterns) are grouped together to form a brightness category. Simultaneously, a dedicated first U-Net network and a second U-Net network are configured for each brightness category to ensure that each network is only adapted to the sample features of its corresponding category. This generates n brightness categories. The clustering operation can classify sea ice images with fused regions and sea ice images with snow-covered regions. Therefore, it is not necessary to specifically collect sea ice images with fused regions and sea ice images with snow-covered regions.

[0076] Finally, for each brightness category: the fused feature map of all samples in that category (or the superimposed features of the fused feature map and the original sea ice image) is used as the network input; the edge annotation map corresponding to the sample is used as the supervision label; The network parameters are optimized using the backpropagation algorithm: For the first U-Net network, the difference between the preliminary edge map output by the network and the labeled edge map is used as the loss function to train it to learn the explicit edge structure mapping relationship of sea ice images of this category. For the second U-Net network, the difference between the region probability map output by the network and the corresponding region label is used as the loss function to train it to learn the detailed region feature mapping relationship of sea ice images of this category; until the network loss value converges to a preset threshold, the training of the first U-Net network and the second U-Net network for the corresponding brightness category is completed.

[0077] By using the above methods, the first U-Net network and the second U-Net network corresponding to each brightness category can accurately adapt to the feature patterns of sea ice images of that category, avoiding the problem of insufficient generalization ability caused by a single network adapting to all brightness samples, and ensuring the edge detection accuracy in the subsequent inference stage.

[0078] Optionally, the operation of calling the first U-Net network, the second U-Net network, and the transfer matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image includes: calling the first U-Net network corresponding to the brightness category to generate a sea ice prediction contour map corresponding to the sea ice image; calling the second U-Net network corresponding to the brightness category to generate a prediction tensor corresponding to the sea ice image; comparing the sea ice prediction contour map and the prediction tensor to determine the normal sea ice region, seawater region, snow cover region, and melt region; and determining the sea ice edge line according to the transfer matrix.

[0079] Specifically, refer to Figure 6 As shown, after determining the brightness category of the sea ice image, the sea ice image is input into the first U-Net network corresponding to that brightness category, and the first U-Net network is used to generate a sea ice prediction contour map corresponding to the sea ice image. However, because there is melting and snow accumulation at the sea ice edges, the prediction contour map indicates a rough image of the sea ice edge contour and cannot accurately determine the true edge contour of the sea ice.

[0080] It should be noted that the first U-Net network includes an encoding network, skip-layer connection module 1, skip-layer connection module 2, and a decoding network. The encoding network includes an encoding module, which contains convolutional layers and max-pooling layers. After inputting the sea ice image into the encoding network, multiple encoded feature maps at different levels are obtained. Then, skip-layer connection modules 1 and 2 are used to transmit the encoded feature maps output by the encoding module to decoding modules 1 and 2 in the decoding network, respectively. Figure 6 Both the skip-layer connection module 1 and skip-layer connection module 2 shown include fully connected layers. In the decoding network, decoding module 1 performs convolutional fusion and upsampling on the encoded feature map from skip-layer connection module 1 through convolutional and upsampling layers to output a first decoded feature map. Decoding module 2 then fuses the first decoded feature map with the encoded feature map from skip-layer connection module 2 through a clustering layer, and then performs convolutional and upsampling processing to finally output a second decoded feature map with the same size as the original image, which serves as the sea ice prediction contour map.

[0081] While inputting the sea ice image into the first U-Net network corresponding to the brightness category, the sea ice image is also input into the second U-Net network corresponding to the brightness category. The second U-Net network generates a prediction tensor (feature tensor) corresponding to the sea ice image. The prediction tensor includes at least one prediction matrix, and each prediction matrix corresponds to the cause of different types of misidentified edges (snow accumulation and sea ice melting), to normal sea ice, and to seawater.

[0082] It should be noted that the reference Figure 7 As shown, the overall structure of the second U-Net network is similar to that of the first U-Net network, also including an encoding network, skip-layer connection modules, and a decoding network. However, the design goals and output results of the second U-Net network differ from those of the first U-Net network. Specifically, the working principle of the second U-Net network is as follows: First, the encoding network (on the left) downsamples the input sea ice image and extracts multi-level feature maps through convolutional layers and max pooling layers. Then, the multi-level feature maps are passed to the corresponding layers of the decoding network through skip-layer connection modules 1 and 2 (the fully connected layer structure in the figure). In the decoding network, decoding module 1 performs convolutional fusion and upsampling on the feature map from skip-layer connection module 1, and passes the result to decoding module 2. Decoding module 2 then fuses the result with the feature map from skip-layer connection module 2 through a pooling layer, and outputs a feature map after convolution and upsampling. Finally, the feature map is input into a softmax classifier, and after pixel-level multi-class probability calculation, the predicted tensors for different regions of the sea ice image are finally output.

[0083] The elements of the prediction matrix indicate the probability that a corresponding location in the sea ice image exists at the misidentified edge corresponding to the prediction matrix, as well as the probability of normal sea ice. However, because sea ice edges may have melting and snow accumulation, the probability values ​​of this prediction tensor only indicate the approximate location of snow accumulation and melting, and do not accurately reflect the actual edge location of the sea ice.

[0084] For example, Table 1 shows a partial prediction matrix for snow-covered areas: Table 1 This represents the probability value of the corresponding pixel in the image within the prediction matrix for the snow-covered region of the j-th brightness category. In the prediction matrix corresponding to the snow-covered region, all probability values ​​except those for the snow-covered region are 0.

[0085] For example, Table 2 shows some probability values ​​for seawater areas: Table 2 Table 3 shows some probability values ​​for snow-covered areas: Table 3 For example, Table 4 shows a partial prediction matrix of the melting region: Table 4 This represents the probability value of the corresponding pixel in the image within the prediction matrix for the melted region of the j-th brightness category. In the prediction matrix corresponding to the melted region, all probability values ​​except those for the melted region are 0.

[0086] For example, Table 5 shows some probability values ​​for sea ice areas: Table 5 For example, Table 6 shows some probability values ​​for the melting region: Table 6 Furthermore, such as Figure 8 As shown, the predicted sea ice profile map and the predicted tensor are compared to determine the normal sea ice region, seawater region, snow cover region, and melt region (the probability value of the sea ice region can be used). To identify, the probability value of a seawater area can be used (To identify), and to determine the sea ice edge line based on the transition matrix. Specifically, first, the transition matrices A1~A1 of the Markov chain are constructed. n The transition matrices A1~A n Corresponding to brightness category 1 to brightness category n.

[0087] Among them, parameters This represents the probability of transitioning from state "edge pixels moving inward" to state "edge pixels moving inward" in category j (j=1~n); parameter This represents the probability of transitioning from the state "edge pixel moves inward" to the state "edge pixel remains in its current position" in category j. parameter This represents the probability of transitioning from state "edge pixels moving inward" to state "edge pixels moving outward" in category j. parameter This represents the probability of transitioning from the state "edge pixel remains in its current position" to the state "edge pixel moves inward" in category j. parameter This represents the probability of transitioning from the state "edge pixel keeps its current position" to the state "edge pixel keeps its current direction" in category j. parameter This represents the probability of transitioning from the state "edge pixel remains in its current position" to the state "edge pixel moves outward" in category j. parameter This represents the probability of transitioning from state "edge pixels moving outward" to state "edge pixels moving inward" in category j. parameter This represents the probability of transitioning from the state "edge pixels move outward" to the state "edge pixels remain in their current position" in category j; and parameter This represents the probability of transitioning from state "edge pixels move outward" to state "edge pixels move outward" in category j.

[0088] After constructing the transition matrix of the Markov chain, the parameter values ​​of the transition matrix are generated based on the state of the transition direction of the sample edge pixels. Here, the samples are the pixels that historically determined the edges.

[0089] This sample is a classified sea ice image, and the sea ice images of category j are shown in Table 7: Table 7 Among them, the edge line of the sea ice in the melting area is a pixel. ~ A line that forms a continuous line. It is the starting point. It is the endpoint. (Regarding the sample) ~ The pre-transition state (i.e., the first state) and post-transition state (i.e., the second state) corresponding to the sample are determined respectively.

[0090] Specifically, for each sample ( ), based on the sample Compared with the previous sample The position of the pixel and Determine and sample The corresponding state before the transition. For example, when the pixel... The location is Within (how to understand "within" will be explained below), Pre-transition state For edge pixels, move inward; when pixel Location and The same, this sample Pre-transition state Maintain the current position for edge pixels; when pixel The location is In addition, Pre-transition state This involves pushing edge pixels outwards.

[0091] In addition, for each sample ( ), based on the sample With the next sample pixels and Determine and sample The corresponding state after the transition. For example, when the pixel... The location is In addition, samples post-transition state For edge pixels, move inward; when pixel Location and The same, this sample post-transition state Maintain the current position for edge pixels; when pixel The location is Within, the sample post-transition state This involves pushing edge pixels outwards.

[0092] For example, for samples According to this sample Compared with the previous sample pixels and Determine and sample Corresponding state before transition and based on the sample With the next sample pixels and Determine and sample Corresponding post-transition state .

[0093] And so on, for the samples According to this sample Compared with the previous sample pixels and Determine the relationship with this sample Corresponding state before transition and based on the sample With the next sample pixels and Determine and sample Corresponding post-transition state .

[0094] Therefore, based on the sample Pre-transition state and the state after the transition Determine the relationship with each sample The corresponding state transition information is shown in Table 8: Table 8 Therefore, the parameters of the transition matrix A can be determined according to the following formula: in This represents the number of samples whose state before the transition was "edge pixels moving inward"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed inward" and whose state after the transfer is "edge pixels are pushed inward".

[0095] in This represents the number of samples whose state before the transition was "edge pixels being pushed inward"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed inward" and whose state after the transfer is "edge pixels remain in their current position".

[0096] in This represents the number of samples whose state before the transition was "edge pixels being pushed inward"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed inward" and whose state after the transfer is "edge pixels are pushed outward".

[0097] in This represents the number of samples whose state before the transition is "edge pixels remain in their current positions"; The number of samples whose state before the transfer was "edge pixels remain in their current position" and whose state after the transfer is "edge pixels move inward".

[0098] in This represents the number of samples whose state before the transition is "edge pixels remain in their current positions"; This refers to the number of samples whose state before the transfer is "edge pixels remain in their current position" and whose state after the transfer is "edge pixels remain in their current position".

[0099] in This represents the number of samples whose state before the transition is "edge pixels remain in their current positions"; The number of samples whose state before the transfer was "edge pixels remain in their current position" and whose state after the transfer is "edge pixels are pushed outward".

[0100] in This represents the number of samples whose state before the transition was "edge pixels being pushed outwards"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed outward" and whose state after the transfer is "edge pixels are pushed inward".

[0101] in This represents the number of samples whose state before the transition was "edge pixels being pushed outwards"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed outward" and whose state after the transfer is "edge pixels remain in their current position".

[0102] in This represents the number of samples whose state before the transition was "edge pixels being pushed outwards"; This refers to the number of samples whose state before the transfer was "edge pixels are pushed outwards" and whose state after the transfer is "edge pixels are pushed outwards".

[0103] Then, the true sea ice edge line is determined in the melting / snow accumulation area based on the transfer matrix.

[0104] Specifically, in embodiments of the present invention, the starting pixel is determined. The specific steps for achieving the initial state include: First, determine the center point of the sea ice area, and then measure the distance from the center point to the center point. Draw a circle with radius , if Outside the circle, then The initial state is that the edge pixels are pushed outwards. If Inside the circle, then The initial state is that the edge pixels are pushed inward. If Remain still (i.e., and (If they are the same pixel), then The initial state is that the edge pixels maintain their current positions.

[0105] Next, based on the transition matrices A1~A of the Markov chain... n For the current pixel The determination is made based on the progression of the state. Specifically, it involves obtaining and... The current first state (edge ​​pixels moving outward) corresponds to three state transition probabilities (i.e., the probability of transitioning from "edge pixels moving outward" to "edge pixels moving inward", "edge pixels remaining in place", and "edge pixels moving outward"). By comparing these three probability values, the state corresponding to the highest transition probability is determined as the current pixel. The second state and the next edge pixel The first state, and based on this, determine the next edge pixel. The position. For example, if the transition probability value of "edge pixels moving inward" is the highest, then The second state is determined as "edge pixels moving inward", and the corresponding determination is made. The location.

[0106] Furthermore, As the new current pixel, repeat the above process: from the center point to the current pixel. Draw a circle with a radius of 1, and determine the current pixel based on the state transition probabilities defined by the pre-constructed Markov chain transition matrix. The second state and the next pixel The first state. This process iterates in this manner until the preset endpoint is reached. The position is determined to obtain an ordered sequence of edge pixels. ~ Finally, the sequence of edge pixels is connected sequentially to form a sea ice edge line that passes through the snow-covered area or the sea ice melting area.

[0107] It should be noted that when the distance from the center point to the current pixel is... Draw a circle with radius to determine the next pixel. At that time, first use the current pixel point Divide the area into 9 pixels centered on the current pixel, excluding the current pixel. Next, use the other 8 pixels as the next pixel. The initial candidate pixels are determined. Then, the next pixel is determined based on the state transition probabilities defined by the pre-constructed Markov chain transition matrix. The transition state (movement direction), if the next pixel point The first condition is "edge pixels are moved inwards," which selects the pixel outside the circle from the initial candidate pixels as the next pixel. A set of candidate pixels for a given direction. This set may contain multiple pixels; therefore, pixels matching the given direction are excluded from this set. and Pixels in the same direction are used to obtain a set of valid candidate pixels. Finally, according to a preset selection rule (in this embodiment of the invention, the selection rule is random sampling), any one pixel is selected from the set of valid candidate pixels as... .

[0108] It is worth noting that during the iterative determination of the edge line based on the Markov chain transition matrix, three extreme migration scenarios requiring constraints may occur: For example, if an edge pixel is determined to be "moving outward" based on its state transition probability, and subsequent pixels are continuously determined to move in the same direction based on the highest probability, the edge line will continuously expand towards the seawater, potentially eventually crossing the actual boundary. To avoid such situations leading to the infinite spread of the edge line, this application sets a restriction region in the snow accumulation and melting areas. Once the calculated position of an edge pixel exceeds the range of this restriction region, its migration direction is forcibly changed to "moving inward," thereby constraining the edge line within a reasonable physical boundary.

[0109] The restricted area is determined as follows: Based on the edges of the initially identified melted or snow-covered areas, morphological dilation is applied. For an edge with an initial width of W pixels, n dilation operations are performed using a k×k square structuring element. The width of the dilated edge can be expressed as: W 膨胀后 =W+2(k-1)n The area covered by this expansion is defined as the boundary of the region, which is used to physically constrain extreme shift behavior during Markov chain inference.

[0110] Similarly, if edge pixels continuously "push inward," the edge line may shrink into the sea ice. Therefore, when the calculated position of the edge line exceeds the limit area, it will be forced to "push outward" to maintain the reasonable position of the edge line.

[0111] Furthermore, if an edge pixel is continuously judged as "remaining in its current position" for more than the preset number of iterations, the system will force it to be judged as "edge pixel moving inward" or "edge pixel moving outward" based on its surrounding context, in order to ensure the continuity and convergence of the edge line determination.

[0112] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0113] Therefore, according to this embodiment, this application achieves on-orbit identification and real-time broadcasting of navigational risks by constructing an early warning link based on on-board real-time processing and inter-satellite collaborative distribution. Furthermore, by introducing multi-directional gradient feature fusion, it achieves feature enhancement and targeted processing of sea ice images under different lighting conditions, significantly improving the adaptability of sea ice edge recognition under different lighting conditions. It can also adaptively match the most suitable first U-Net network, second U-Net network, and transition matrix for each image through brightness classification, ensuring optimal matching between model parameters and image characteristics, thereby achieving high-precision and robust extraction of sea ice edges. This addresses the technical problems existing in the prior art, such as insufficient timeliness of existing low-orbit satellite early warnings, which cannot meet the modern shipping industry's demand for real-time, accurate, and forward-looking safety early warnings, and the technical issues of blurred edge features and insufficient distinguishability caused by significant brightness differences in sea ice image samples without effective normalization, affecting the accuracy and reliability of sea ice edge recognition results.

[0114] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0116] Example 2 Figure 9 A low-Earth orbit satellite constellation-based navigational early warning device according to this embodiment is shown, which corresponds to the method described according to Embodiment 1. (Reference) Figure 9As shown, the device includes: a channel image acquisition module 910, used by a first satellite of a low-Earth orbit (LEO) satellite constellation to acquire channel images related to a preset channel; a sea ice edge recognition module 920, used by the first satellite to identify the edge of sea ice contained in the channel image based on a channel warning method of the LEO satellite constellation; a warning information transmission module 930, used by the first satellite to transmit sea ice-related warning information to a second satellite of the LEO satellite constellation, wherein the second satellite covers at least a portion of the preset channel, when the first satellite determines that sea ice poses a threat to the preset channel based on the edge; and a channel warning module 940, used for... The second satellite transmits early warning information to ships within its coverage area. The navigation warning method based on a low-Earth orbit satellite constellation includes the following steps for identifying the edges of sea ice contained within a sea ice image region: inputting the sea ice image contained in the navigation image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image according to the gradient feature map to determine its brightness category; and, based on the brightness category of the sea ice image, calling the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image.

[0117] It should be noted that the low-Earth orbit satellite constellation-based airway early warning device provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0118] Therefore, according to this embodiment, this application achieves on-orbit identification and real-time broadcasting of navigational risks by constructing an early warning link based on on-board real-time processing and inter-satellite collaborative distribution. Furthermore, by introducing multi-directional gradient feature fusion, it achieves feature enhancement and targeted processing of sea ice images under different lighting conditions, significantly improving the adaptability of sea ice edge recognition under different lighting conditions. It can also adaptively match the most suitable first U-Net network, second U-Net network, and transition matrix for each image through brightness classification, ensuring optimal matching between model parameters and image characteristics, thereby achieving high-precision and robust extraction of sea ice edges. This addresses the technical problems existing in the prior art, such as insufficient timeliness of existing low-orbit satellite early warnings, which cannot meet the modern shipping industry's demand for real-time, accurate, and forward-looking safety early warnings, and the technical issues of blurred edge features and insufficient distinguishability caused by significant brightness differences in sea ice image samples without effective normalization, affecting the accuracy and reliability of sea ice edge recognition results.

[0119] Example 3 Figure 10 A low-Earth orbit satellite constellation-based navigational early warning device according to this embodiment is shown, which corresponds to the method according to Embodiment 1. (Reference) Figure 10As shown, the device includes: a processor 1010; and a memory 1020, connected to the processor 1010, for providing the processor 1010 with instructions to process the following steps: a first satellite of a low-Earth orbit (LEO) satellite constellation acquires channel images related to a preset channel; the first satellite identifies the edges of sea ice contained in the channel images based on a channel warning method of the LEO satellite constellation; if the first satellite determines that sea ice poses a threat to the preset channel based on the edges, it transmits sea ice-related warning information to a second satellite of the LEO satellite constellation, wherein the second satellite covers at least a portion of the preset channel; and The second satellite sends early warning information to ships within its coverage area. The navigation warning method based on a low-Earth orbit satellite constellation includes the following steps: inputting the sea ice image contained within the navigation image into a pre-trained gradient calculation model to generate a corresponding gradient feature map; classifying the sea ice image according to the gradient feature map to determine its brightness category; and, based on the brightness category of the sea ice image, calling the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category to determine the sea ice edge line of the sea ice image.

[0120] It should be noted that the low-Earth orbit satellite constellation-based airway early warning device provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0121] Therefore, according to this embodiment, this application achieves on-orbit identification and real-time broadcasting of navigational risks by constructing an early warning link based on on-board real-time processing and inter-satellite collaborative distribution. Furthermore, by introducing multi-directional gradient feature fusion, it achieves feature enhancement and targeted processing of sea ice images under different lighting conditions, significantly improving the adaptability of sea ice edge recognition under different lighting conditions. It can also adaptively match the most suitable first U-Net network, second U-Net network, and transition matrix for each image through brightness classification, ensuring optimal matching between model parameters and image characteristics, thereby achieving high-precision and robust extraction of sea ice edges. This addresses the technical problems existing in the prior art, such as insufficient timeliness of existing low-orbit satellite early warnings, which cannot meet the modern shipping industry's demand for real-time, accurate, and forward-looking safety early warnings, and the technical issues of blurred edge features and insufficient distinguishability caused by significant brightness differences in sea ice image samples without effective normalization, affecting the accuracy and reliability of sea ice edge recognition results.

[0122] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0123] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0128] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for early warning of navigation routes for a low-Earth orbit satellite constellation, characterized in that, include: The first satellite of the low-Earth orbit satellite constellation acquires imagery of the flight path related to the preset flight path. The first satellite, based on the low-Earth orbit satellite constellation's navigation warning method, identifies the edges of sea ice contained in the navigation image; If the first satellite determines that the sea ice poses a threat to the preset shipping route based on the edge, it will send early warning information related to the sea ice to the second satellite of the low-Earth orbit satellite constellation, wherein the second satellite covers at least a portion of the preset shipping route; as well as The second satellite transmits the warning information to ships within its coverage area, and the operation of identifying the edge of sea ice contained in the sea ice image area using the low-Earth orbit satellite constellation-based navigation warning method includes: The sea ice image contained in the channel image is input into a pre-trained gradient calculation model to generate the corresponding gradient feature map; Based on the gradient feature map, the sea ice images are classified to determine the brightness category of the sea ice images; and Based on the brightness category of the sea ice image, the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category are invoked to determine the sea ice edge line of the sea ice image.

2. The method according to claim 1, characterized in that, The operation of determining that the sea ice poses a threat to the shipping channel based on the edge includes: Determine the first distance closest to the edge from the preset channel; and If the first distance is less than a preset distance threshold, it is determined that the sea ice poses a threat to the preset shipping route.

3. The method according to claim 1, characterized in that, The operation of the second satellite transmitting the early warning information to ships within its coverage area includes: The second satellite determines the direction of movement of the vessel based on its trajectory; and If the vessel is determined to be moving toward the sea ice based on the direction of movement, the warning information will be sent to the vessel.

4. The method according to claim 1, characterized in that, The operation of inputting the sea ice image into a pre-trained gradient calculation model to generate the corresponding gradient feature map includes: The sea ice image is input into a pre-trained gradient calculation model; and The gradient calculation model uses a horizontal kernel, a vertical kernel, a first oblique kernel, and a second oblique kernel to perform convolution operations on the sea ice image, respectively, and outputs the corresponding horizontal gradient map, vertical gradient map, first oblique gradient map, and second oblique gradient map.

5. The method according to claim 4, characterized in that, The horizontal kernel is used to detect the vertical edges of the sea ice image and generate the horizontal gradient map. The corresponding calculation matrix is: ； in, , , , , , For a preset value, and in which The vertical kernel is used to detect the horizontal edges of the sea ice image and generate the vertical gradient map. The corresponding calculation matrix is: ； in, , , , , , For a preset value, and in which The first oblique kernel is used to detect the upper left to lower right edge of the sea ice image, generating the first oblique gradient map, and the corresponding calculation matrix is: ； in, , , , , , For a preset value, and in which The second oblique kernel is used to detect the lower left to upper right edge of the sea ice image, generating the second oblique gradient map. The corresponding calculation matrix is: ； in, , , , , , The value is a preset value.

6. The method according to claim 1, characterized in that, The first U-Net network and the second U-Net network are trained in the following manner: Acquire multiple sample sea ice images and the edge information marked in each sample sea ice image; Each sample of sea ice image is input into a preset gradient calculation model to generate a corresponding gradient feature map. The gradient feature maps of each sample sea ice image are input into the multilayer perceptron to obtain the fused feature maps corresponding to each sample sea ice image; Clustering is performed on the fused feature maps of all sea ice samples to generate multiple brightness categories, where each brightness category corresponds to a first U-Net network and a second U-Net network; as well as The first U-Net network and the second U-Net network corresponding to the brightness category are trained using sample sea ice images belonging to the same brightness category and the edge information corresponding to the sample sea ice images.

7. The method according to claim 1, characterized in that, The operation of determining the sea ice edge line of the sea ice image by invoking the first U-Net network, the second U-Net network, and the transfer matrix corresponding to the brightness category includes: The first U-Net network corresponding to the brightness category is invoked to generate a sea ice prediction contour map corresponding to the sea ice image; The second U-Net network corresponding to the brightness category is invoked to generate the prediction tensor corresponding to the sea ice image; The predicted sea ice profile map and the predicted tensor are compared to determine the normal sea ice region, seawater region, snow cover region, and meltwater region; and The sea ice edge line is determined based on the aforementioned transition matrix.

8. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 7 is performed by a processor.

9. A navigational early warning device based on a low-Earth orbit satellite constellation, characterized in that, include: The airway image acquisition module is used to acquire airway images related to preset airways by the first satellite of the low-Earth orbit satellite constellation. The sea ice edge recognition module is used by the first satellite in the low-Earth orbit satellite constellation-based navigation early warning method to identify the edges of sea ice contained in the navigation image; The early warning information transmission module is used to transmit early warning information related to the sea ice to the second satellite of the low-Earth orbit satellite constellation when the first satellite determines that the sea ice poses a threat to the preset shipping route based on the edge of the sea ice, wherein the second satellite covers at least a part of the preset shipping route; as well as A navigation warning module is used by the second satellite to transmit the warning information to ships within its coverage area, and wherein the navigation warning method based on a low-Earth orbit satellite constellation includes the operation of identifying the edge of sea ice contained in the sea ice image area, comprising: The sea ice image contained in the channel image is input into a pre-trained gradient calculation model to generate the corresponding gradient feature map; Based on the gradient feature map, the sea ice images are classified to determine the brightness category of the sea ice images; and Based on the brightness category of the sea ice image, the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category are invoked to determine the sea ice edge line of the sea ice image.

10. A navigational early warning device based on a low-Earth orbit satellite constellation, characterized in that, include: processor; as well as A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: The first satellite of the low-Earth orbit satellite constellation acquires imagery of the flight path related to the preset flight path. The first satellite, based on the low-Earth orbit satellite constellation's navigation warning method, identifies the edges of sea ice contained in the navigation image; If the first satellite determines that the sea ice poses a threat to the preset shipping route based on the edge, it will send early warning information related to the sea ice to the second satellite of the low-Earth orbit satellite constellation, wherein the second satellite covers at least a portion of the preset shipping route; as well as The second satellite transmits the warning information to ships within its coverage area, and the operation of identifying the edge of sea ice contained in the sea ice image area using the low-Earth orbit satellite constellation-based navigation warning method includes: The sea ice image contained in the channel image is input into a pre-trained gradient calculation model to generate the corresponding gradient feature map; Based on the gradient feature map, the sea ice images are classified to determine the brightness category of the sea ice images; and Based on the brightness category of the sea ice image, the first U-Net network, the second U-Net network, and the transition matrix corresponding to the brightness category are invoked to determine the sea ice edge line of the sea ice image.

Images