Multi-source fusion sea ice density acquisition method, device, equipment and medium

Abstract

This application relates to the field of data processing technology and discloses a multi-source fusion method, device, equipment, and medium for obtaining sea ice density. The method includes: acquiring shipborne radar images and underway observation data; performing optical character recognition (OCR) recognition and preprocessing on the radar images; performing axis analysis and polynomial fitting on the grayscale image to eliminate blind spots; obtaining sea ice distribution through adaptive threshold segmentation; processing underway observation data with a deep convolutional network to fill blind spots; and fusing radar and underway observation data based on area ratio to obtain a fused sea ice density. This application can fuse shipborne ice navigation radar scans and underway observation data, overcoming the limitations of a single data source and extracting a more accurate, stable, and comprehensive sea ice density.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a multi-source fusion sea ice density acquisition method, device, equipment and medium. Background Art

[0002] Sea ice concentration is a quantitative indicator used to assess the distribution of sea ice within a specific area. It is of great significance for polar environmental monitoring, analysis of Arctic shipping lanes, and global climate change research. Currently, observations and research on sea ice concentration rely primarily on ship-borne observations and satellite remote sensing. Ship-borne observations include manual observations and automated imaging observations. Manual observations involve visual observations and recording of sea ice type, thickness, and concentration within a kilometer of the ship using the corresponding standards of the Antarctic Sea Ice Processes and Climate (ASPeCt) or Arctic Ship-Based Sea Ice Observations (ASSIST). Imaging observations utilize shipboard optical imaging systems to acquire optical images and utilize automated recognition algorithms to derive information such as sea ice concentration. Satellite remote sensing utilizes onboard active and passive microwave instruments to collect sea ice information and generate large-scale gridded sea ice concentration products, typically with a resolution of at least three kilometers. Additionally, ship-borne navigation radar systems can also collect sea ice information when operating in ice-bound navigation mode, but these have not yet been widely used to extract sea ice concentration.

[0003] Existing technologies have obvious deficiencies. Underway observations are subject to factors such as visibility and observation angles, and can only cover a limited area around the ship. The observation results are affected by the subjective experience and background knowledge of the observers and have great uncertainty. The sea ice density products based on satellite remote sensing have low resolution, and problems such as the mixing of sea ice of different types, ice thicknesses, and ice ages will reduce the accuracy of density products at the grid scale. Especially in summer, the accuracy of satellite remote sensing products is further reduced by the melting of the sea ice surface. In addition, satellite remote sensing and underway observations lack effective connection in terms of temporal and spatial scales, and they cannot provide cross-scale data support and expanded applications to each other. Although shipborne navigation radars can provide high-resolution sea ice images, they have problems such as central overexposure and remote blind spots. It is difficult to obtain complete and accurate sea ice density when used alone. Summary of the Invention

[0004] The present application provides a multi-source fusion sea ice density acquisition method, device, equipment and medium for fusing shipborne ice zone navigation radar scanning and underway observation data, overcoming the limitations of a single data source and extracting more accurate, stable and comprehensive sea ice density.

[0005] In the first aspect, the present application provides a multi-source fusion method for obtaining sea ice density, which includes: obtaining original radar images and underway observation data collected by a ship-borne ice zone navigation radar, performing OCR text recognition on the original radar image to extract radar parameters, and obtaining an effective radar grayscale image; performing axis density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminating the central overexposed area and the remote blind area, and obtaining a balanced grayscale difference image; inputting the balanced grayscale difference image into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing to obtain vectorized sea ice distribution information and sea ice density in the radar observation area; based on the underway observation data, the deep convolutional network framework with a dual-branch attention mechanism is used to fill the blind area, and the sea ice density in the radar observation area and the underway observation sea ice density are weighted averaged according to the area ratio to obtain a fused sea ice density.

[0006] In a second aspect, the present application provides a multi-source fusion sea ice density acquisition device, the multi-source fusion sea ice density acquisition device comprising:

[0007] An acquisition module is used to obtain the original radar image and underway observation data collected by the ship-borne ice navigation radar, perform OCR text recognition on the original radar image, extract radar parameters, and obtain an effective radar grayscale image;

[0008] A fitting module is used to perform axial density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminate the central overexposed area and the far-end blind area, and obtain a balanced grayscale difference image;

[0009] a segmentation module, configured to input the equalized grayscale difference image into an adaptive inter-class maximum variance method for threshold segmentation and morphological processing, thereby obtaining vectorized sea ice distribution information and sea ice density in the radar observation area;

[0010] An averaging module is used to fill in blind spots based on the underway observation data through a deep convolutional network framework with a dual-branch attention mechanism, and to perform a weighted average of the sea ice density in the radar observation area and the underway observation sea ice density according to the area ratio to obtain a fused sea ice density.

[0011] In a third aspect, a multi-source fusion sea ice density acquisition device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to enable the multi-source fusion sea ice density acquisition device to execute the above-mentioned multi-source fusion sea ice density acquisition method.

[0012] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned multi-source fusion sea ice density acquisition method.

[0013] In the technical solution provided by this application, by acquiring the original radar images and underway observation data collected by the ship-borne ice navigation radar, and using OCR text recognition technology to extract radar parameters, the automatic recognition and extraction of key information such as the radar operating mode, scanning radius and pulse length are realized, avoiding the tediousness and errors of manual parameter input, and improving the degree of automation and efficiency of data processing. Axis density distribution analysis and polynomial fitting are performed on the effective radar grayscale image to solve the problem of radar signal attenuation with distance. By eliminating the central over-exposed area and the far-end blind area, the data quality is ensured, and a balanced grayscale difference image is obtained, which provides a high-quality data foundation for subsequent sea ice identification. The balanced grayscale difference image is input into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing, which realizes the accurate distinction between sea ice and open water. The sea ice distribution information and the sea ice density in the radar observation area are obtained through vectorization processing, which improves the accuracy and availability of the recognition results.

[0014] A deep convolutional network framework with a dual-branch attention mechanism processes underway observations to fill in blind spots, cleverly addressing the problem of missing data in the near-blind zone of the radar. The spatial and channel attention modules in this framework effectively capture the spatial distribution and texture characteristics of sea ice, improving the ability to identify complex ice conditions. A fused sea ice density is obtained by weightedly averaging the sea ice density of the radar observation area and underway observations based on their area ratio. This achieves a rational fusion of multi-source data, fully leveraging the strengths of each data source while overcoming the limitations of a single one. In particular, the two algorithmic features of the deep convolutional network framework with a dual-branch attention mechanism and the adaptive inter-class maximum variance method (AUMV) complement each other functionally. The AUMV provides accurate sea ice distribution information in the radar observation area, providing reference data for the training and validation of the deep convolutional network. The deep convolutional network, on the other hand, processes optical image data to provide supplementary information for radar blind spots. Together, they form a complete description of sea ice distribution. At the same time, the adaptive inter-class maximum variance method based on image segmentation and the deep convolutional network with a dual-branch attention mechanism based on deep learning have been improved and optimized to meet the special needs of sea ice recognition. The former calculates the optimal threshold through the maximum inter-class variance method, adapting to the variability of the grayscale difference between sea ice and water areas; the latter enhances the perception of sea ice characteristics through the attention mechanism, thereby improving the recognition accuracy in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a schematic diagram of an embodiment of a method for obtaining sea ice density by multi-source fusion in an embodiment of the present application;

[0017] Figure 2 This is a schematic diagram of an embodiment of a multi-source fusion sea ice density acquisition device in an embodiment of the present application;

[0018] Figure 3 It is a schematic block diagram of the structure of a multi-source fusion sea ice density acquisition device in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present application provide a multi-source fusion sea ice density acquisition method, device, equipment and medium. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, a method for obtaining sea ice density by multi-source fusion includes:

[0021] Step S101: obtaining original radar images and underway observation data collected by the ship-borne ice navigation radar, performing OCR text recognition on the original radar images to extract radar parameters, and obtaining an effective radar grayscale image;

[0022] Step S102: performing axis density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminating the central overexposed area and the far-end blind area, and obtaining a balanced grayscale difference image;

[0023] Step S103: Input the equalized grayscale difference image into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing to obtain vectorized sea ice distribution information and sea ice density in the radar observation area;

[0024] Step S104: Based on the underway observation data, the blind spots are filled through the deep convolutional network framework with a dual-branch attention mechanism, and the sea ice density in the radar observation area and the underway observation sea ice density are weighted averaged according to the area ratio to obtain the fused sea ice density.

[0025] It is understandable that the execution subject of this application can be a multi-source fusion sea ice density acquisition device, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0026] Specifically, raw radar images and underway observation data collected by the ship's ice navigation radar are acquired. For example, the Chinese research vessel Xuelong 2's navigation radar system includes two radars, one in S-band and one in X-band, installed on the compass deck. It is equipped with a Sigma S6 radar operating system from Rutter, a Canadian company. The radar system operates in three modes: open water, iceberg detection, and ice navigation. During navigation in ice, it uses short and medium pulse scanning, with a typical scanning radius of 0.75 to 8 nautical miles. The acquired radar interface images are then analyzed using optical character recognition (OCR) to identify parameters such as radar operating mode, scanning radius, and pulse length. Images are then classified based on these parameters, eliminating images affected by the STC function and other abnormal images. The STC function linearly adjusts the radar gain to suppress overexposure caused by strong scattered signals near the ship, but this can interfere with subsequent sea ice identification. Images are categorized by pulse length into short-pulse scanning and medium-pulse scanning. Short-pulse scanning images are limited to a scanning radius of 2 to 4 nautical miles, while medium-pulse scanning images are limited to a scanning radius of 3 to 6 nautical miles. The radar scanning image of the circular area is extracted from the classified image, and the blue band component is extracted to obtain the initial grayscale image. The pixel resolution is calculated, and the spatial scale of the initial grayscale image is calibrated to finally obtain the effective radar grayscale image.

[0027] Axial density distribution analysis and polynomial fitting were performed on the effective radar grayscale image. Multiple axes radiating outward from the image center were plotted, and the distribution of pixel values ​​along these axes over distance was calculated. By aggregating the density distribution curves for these axes, the effective observation range for the short-pulse scan image was determined to be 2.4 kilometers, and for the medium-pulse scan image to be 4.0 kilometers. Two problematic areas exist in the observed sea ice images: oversaturated areas near ships and observation blind spots at the far end. Grayscale value saturation analysis was performed on the area within 30 pixels of the image center, with areas exceeding a preset threshold marked as near-end blind spots. The near-end blind spots were precisely quantified based on the radar pulse type and scan radius. The near-end blind spot radii for short-pulse scan radiuses of 2, 3, and 4 nautical miles were 334 meters, 502 meters, and 669 meters, respectively; and for medium-pulse scan radiuses of 3, 4, and 6 nautical miles, the near-end blind spot radii were 502 meters, 669 meters, and 1004 meters, respectively. The effective radar grayscale images are grouped according to radar pulse type and scanning radius. The quadratic polynomial fitting method is used to establish a fitting model. The difference between the actual grayscale value and the fitted value of each pixel is calculated to obtain a balanced grayscale difference image. This processing eliminates the attenuation trend of the radar signal with distance.

[0028] The equalized grayscale difference image was input into the adaptive maximum inter-class variance method for threshold segmentation and morphological processing. First, the effective analysis area was demarcated, defined as a circular region after removing the near-end and far-end blind zones, to obtain the effective analysis area mask. The equalized grayscale difference image and the effective analysis area mask were spatially superimposed and then input into the adaptive maximum inter-class variance method. This method determined the optimal threshold through a traversal approach: the average grayscale value of all non-zero pixels in the masked grayscale difference image was calculated as the initial threshold. Based on this threshold, the image pixels were divided into two classes. The average grayscale value and inter-class variance of the pixels in each class were calculated. By performing a round-trip between the average grayscale values ​​of the two classes, the threshold that maximized the inter-class variance was determined, which was the optimal threshold. The average maximum inter-class variance threshold for the short-pulse scanning image was -19.7, and the average maximum inter-class variance threshold for the medium-pulse scanning image was -20.6. Based on the optimal threshold, the equalized grayscale difference image was binarized to distinguish between sea ice and open water areas. Morphological opening and closing operations are performed on the initial sea ice distribution binary image, and connected domain analysis is applied to eliminate small targets with an area of ​​less than 1000 square meters. The optimized binary image is vectorized, and geometric characteristic parameters such as the outer edge line and area of ​​the sea ice area are extracted. The ratio of the sea ice area to the total area of ​​the effective observation area is calculated to obtain the sea ice density in the radar observation area.

[0029] A deep convolutional network framework with a dual-branch attention mechanism processes sea ice density data from underway observations to fill blind spots. First, a spatial mapping between the radar coordinate system and the geographic coordinate system is established based on the GPS coordinates and heading information provided by the ship's navigation and positioning system. The underway observation data is processed, including density values ​​recorded by visual observations and ice field images collected by shipboard optical equipment. After distortion and geometric correction, the optical images are input into a deep convolutional network framework with a dual-branch attention mechanism. This network consists of four modules: feature extraction, spatial attention, channel attention, and classification. The feature extraction module uses a ResNet-50 backbone network to extract multi-level features. The spatial attention module emphasizes regions with clear boundaries and texture features in the image. The channel attention module enhances channel responses that contribute significantly to sea ice characteristics. Finally, a fully connected layer transforms the feature map into a sea ice distribution probability map. The visual observation density and optical image sea ice density are weighted and combined according to observation quality and reliability to produce a comprehensive underway observation sea ice density. The area ratio of the near-end blind zone to the effective observation area is determined based on the shipborne radar type and scanning radius. The area weight of the cruise observation data is calculated as the ratio of the near-end blind zone area to the total area of ​​the study area, and the area weight of the radar observation data is calculated as the ratio of the annular effective area area to the total area of ​​the study area. The sea ice density of the integrated cruise observation and the sea ice density of the radar observation area are combined according to the data fusion weight.

[0030] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0031] The radar operating mode, scanning radius and pulse length are identified by using the optical character recognition (OCR) method on the original radar image sequence to obtain the radar operating parameters.

[0032] The original radar images are classified according to the radar operating parameters, and images affected by the STC function and other abnormal images are eliminated to obtain valid radar images in the ice area navigation mode;

[0033] Classifying effective radar images into short-pulse scanning images and medium-pulse scanning images according to pulse length, limiting the scanning radius of short-pulse scanning images to 2 to 4 nautical miles, and limiting the scanning radius of medium-pulse scanning images to 3 to 6 nautical miles, to obtain classified radar images;

[0034] Extract the radar scanning image of the circular area from the classified radar image, and extract the blue band component to obtain the initial grayscale image of the radar image;

[0035] The pixel resolution is calculated based on the initial grayscale image and the corresponding radar scanning radius to obtain the standardized pixel resolution;

[0036] The initial grayscale image is spatially scaled according to the standardized pixel resolution to obtain the effective radar grayscale image.

[0037] Specifically, radar operating parameters are identified from raw radar image sequences using optical character recognition (OCR). OCR, a technology that converts text within an image into machine-editable text, is specifically used in this solution to extract radar operating parameters from radar interface images. For shipborne ice navigation radars, for example, radar interface images typically contain information such as the radar operating mode, scanning radius, and pulse length, often displayed in the edge areas of the interface. The OCR recognition process first preprocesses the interface image, including grayscale conversion, binarization, and noise removal. Then, text regions are detected on the preprocessed image. The detected text regions are fed into a text recognition model, which outputs the text content. This process extracts key parameters from the radar interface image, including the radar operating mode (e.g., ice navigation mode, open water mode, or iceberg detection mode), scanning radius (e.g., 2 nautical miles, 3 nautical miles, or 4 nautical miles), and pulse length (short pulse or medium pulse). After obtaining the radar operating parameters, the raw radar images are classified based on these parameters. Images affected by the STC function and other abnormal images must be eliminated. The STC function is a filtering method built into the radar system. It suppresses overexposure caused by strong scattered signals near the ship by linearly adjusting the radar gain. However, enabling STC can interfere with subsequent sea ice identification. Therefore, it is necessary to identify image features to determine whether the image is affected by STC, such as detecting whether the grayscale value distribution in the central area of ​​the image has obvious radial variation characteristics. In addition, other abnormal images must be eliminated, such as those with abnormal signal strength, missing scan lines, or severely blurred images. Through this screening step, a valid radar image is obtained in ice navigation mode.

[0038] Valid radar images are classified by pulse length. Radar pulse length refers to the duration of the electromagnetic wave pulse emitted by the radar. Different pulse lengths are suitable for different observation conditions. Short pulse scanning is suitable for close-range, high-resolution observations, while medium pulse scanning is suitable for longer-range observations. The scanning radius of short pulse scanning images is limited to 2 to 4 nautical miles, while the scanning radius of medium pulse scanning images is limited to 3 to 6 nautical miles. This classification method ensures that images under different pulse conditions are processed according to their respective characteristics, resulting in classified radar images.

[0039] Extracting the radar scan image of a circular area from the classified radar image is a step tailored to the actual characteristics of radar scans. Radar scans typically appear as a circular area centered on the radar antenna, so this circular area must be extracted from the rectangular interface image. The extraction process first determines the center of the circle (usually the image center) and the radius (determined by the actual size of the radar image within the interface). The image content within this circular area is then captured. After extracting the circular area, the blue band component must be further extracted. Radar images are typically color, and the blue band is highly sensitive for sea ice detection. Extracting the blue band component involves decomposing the RGB color image into red, green, and blue channels. Only the grayscale values ​​of the blue channel are retained to obtain the initial grayscale image of the radar image. Pixel resolution is calculated based on this initial grayscale image and the corresponding radar scan radius. Pixel resolution refers to the actual geographic distance represented by each pixel in the image and is a key parameter that links image data to actual geographic space. The calculation formula is: Pixel resolution = (1852 × Scanning radius in nautical miles) / (Radius pixels in radar image), where 1852 is the number of meters corresponding to one nautical mile, and Radar image radius pixels is the number of pixels from the image center to the edge. This calculation yields a standardized pixel resolution, enabling consistent spatial comparison and analysis of images acquired under different conditions. The initial grayscale image is spatially scaled based on the standardized pixel resolution to produce a valid radar grayscale image. Spatial scale calibration is the process of converting pixel positions in the image into actual geographic coordinates or distances. Specifically, based on the calculated pixel resolution, each pixel position in the image is converted to an actual distance relative to the radar center. This allows subsequent analysis to directly use actual geographic distances, eliminating the need for pixel-to-distance conversion.

[0040] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0041] Draw multiple axes radiating from the center of the image to the outside of the effective radar grayscale image, calculate the distribution of pixel values ​​on each axis in the distance direction, and obtain the axis density distribution curve;

[0042] By aggregating multiple axis density distribution curves, the effective observation range of the radar is determined to be 2.4 kilometers for short pulse scanning images and 4.0 kilometers for medium pulse scanning images.

[0043] The grayscale value saturation analysis is performed on the area within 30 pixels from the center of the effective radar grayscale image. The area exceeding the preset threshold is marked as the near-end blind zone, and the near-end blind zone boundary is obtained;

[0044] The near-end blind zone is accurately quantified according to the radar pulse type and scanning radius, and the near-end blind zone radius under different conditions is determined to obtain the quantified near-end blind zone parameters;

[0045] The effective radar grayscale images are grouped according to radar pulse type and scanning radius, and a fitting model is established using a quadratic polynomial fitting method to obtain a polynomial fitting model.

[0046] The difference between the actual grayscale value of each pixel in the effective radar grayscale image and the fitted value calculated based on the polynomial fitting model is calculated to obtain a balanced grayscale difference image.

[0047] Specifically, multiple axes are drawn from the effective radar grayscale image, radiating outward from the image center. These axes are radially distributed, typically at intervals of 15 or 30 degrees, extending from the image center to the image edge. Once the axes are drawn, the distance-dependent distribution of pixel values ​​along each axis is calculated. This means that along the axis from the image center to the edge, the grayscale value of each pixel is plotted relative to its distance from the center, generating an axis density distribution curve. This density distribution curve reflects the distance-dependent variation of the radar signal and is an important basis for analyzing radar observation range and signal quality. By aggregating multiple axis density distribution curves—averaging the grayscale values ​​of pixels at the same distance along all axes—a more representative distance-grayscale relationship curve can be obtained. Based on this curve, the attenuation of grayscale values ​​at long distances is analyzed to determine the effective observation range. Extensive data analysis has verified that the effective observation range for short-pulse scanning images is 2.4 kilometers, and for medium-pulse scanning images is 4.0 kilometers. The effective observation range refers to the maximum distance at which the radar signal can effectively reflect sea ice conditions. Beyond this distance, the signal quality significantly degrades, making it impossible to accurately identify sea ice. By determining the effective observation distance under different pulse types, the effective observation range of the radar is obtained, which defines the spatial boundary for subsequent data processing.

[0048] Grayscale saturation analysis is performed on the area within 30 pixels of the effective radar grayscale image center to address the common "center overexposure" problem in radar images. Center overexposure refers to the area near the radar antenna where the received scattered signal intensity is excessive, resulting in pixel saturation and an inability to accurately reflect the actual sea ice conditions. Grayscale saturation analysis determines the degree of saturation by calculating the ratio of each pixel's grayscale value to a saturation threshold. Areas where grayscale saturation exceeds a preset threshold are marked as near-end blind zones, and the near-end blind zone boundary is determined. Near-end blind zones are areas near ships that the radar cannot accurately observe and must be filled with other data sources in subsequent processing. The near-end blind zone is precisely quantified based on the radar pulse type and scanning radius, and the near-end blind zone radius is determined for different conditions. For short pulse scanning radii of 2, 3, and 4 nautical miles, the near-end blind zone radius is 334 meters, 502 meters, and 669 meters, respectively. For medium pulse scanning radii of 3, 4, and 6 nautical miles, the near-end blind zone radius is 502 meters, 669 meters, and 1004 meters, respectively. These values, derived from analyzing a large amount of measured data, reflect the size patterns of the near-end blind zone under different scanning conditions. By accurately quantifying the near-end blind zone, we obtain quantitative near-end blind zone parameters, providing a basis for subsequent area division and data fusion.

[0049] The effective radar grayscale images are grouped according to the radar pulse type and scanning radius, and a fitting model is established using the quadratic polynomial fitting method. The purpose of grouping is to perform targeted processing on the image characteristics under different scanning conditions, because radar images under different pulse types and scanning radii have different signal attenuation characteristics. Quadratic polynomial fitting is a commonly used curve fitting method used to describe the attenuation trend of radar signals in the upward direction of distance. The form of the fitting model is , where x is the distance from the pixel to the center of the image, f(x) is the expected grayscale value at that location, is the fitting coefficient. For the case where the short pulse scanning radius is 3 nautical miles, the fitting coefficient is 、 This means that in areas closer to the center, grayscale values ​​decrease with distance (a negative linear term). However, at greater distances, this decreasing trend gradually weakens and eventually turns to an increase due to the presence of the quadratic term. This fitting process yields polynomial fitting models for different scanning conditions. These models describe the ideal distance distribution of pixel grayscale values. The difference between the actual grayscale value of each pixel in the effective radar grayscale image and the fitted value calculated based on the polynomial fitting model is calculated to produce a balanced grayscale difference image. This balanced grayscale difference image eliminates the systematic attenuation of the radar signal with distance, allowing the grayscale values ​​in the image to more directly reflect the reflectivity of sea ice, unaffected by distance. This process is similar to a "detrending" operation, highlighting local variations by subtracting systematic trends.

[0050] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0051] The radar scanning range is divided into an effective analysis area, which is defined as the circular area after removing the near-end blind area and the far-end blind area, and the effective analysis area mask is obtained;

[0052] The equalized grayscale difference image is spatially superimposed with the effective analysis area mask, and the adaptive inter-class maximum variance method is input. The threshold is initialized to the global average of the grayscale value. By traversing the potential threshold and calculating the inter-class variance, the optimal threshold is obtained according to the maximum inter-class method.

[0053] The equalized grayscale difference image is binarized according to the optimal threshold to distinguish the sea ice area from the open water area, and the initial sea ice distribution binary image is obtained;

[0054] Morphological opening and closing operations were performed on the initial binary image of sea ice distribution, and connected domain analysis was applied to remove small objects with an area of ​​less than 1000 square meters to obtain the optimized binary image of sea ice distribution.

[0055] The optimized sea ice distribution binary image is vectorized to extract the outer edge line and area of ​​the sea ice area to obtain vectorized sea ice distribution information.

[0056] The ratio of the sea ice area in the vectorized sea ice distribution information to the total area of ​​the effective observation area is calculated to obtain the sea ice density in the radar observation area.

[0057] Specifically, the radar scanning range is divided into an effective analysis area, which is defined as a circular area after excluding the near-end blind zone and the far-end blind zone. The near-end blind zone refers to the overexposed area near the ship due to the strong radar signal, while the far-end blind zone refers to the area beyond the effective observation range of the radar. The effective analysis area can be mathematically represented as a circular ring: ,in is the image center coordinate, is the near-end blind zone radius, is the pixel radius corresponding to the effective observation distance. Using this spatial definition, a binary mask image is generated, where the pixel value in the effective analysis area is set to 1 and the pixel value in other areas is set to 0, thus obtaining the effective analysis area mask. Spatially superimposing the equalized grayscale difference image with the effective analysis area mask involves multiplying the two images pixel by pixel. Specifically, areas with a median value of 1 in the mask retain the original equalized grayscale difference value, while areas with a median value of 0 in the mask have their corresponding equalized grayscale difference value set to 0. This operation ensures that subsequent analysis is performed only on pixels within the effective analysis area, avoiding contamination of the results by near- and far-end blind zones. The superimposed image is then input into the adaptive maximum inter-class variance method for threshold segmentation. The adaptive maximum inter-class variance method, also known as the OTSU method, is an algorithm that automatically determines the threshold for image binarization. This algorithm maximizes the inter-class variance to determine the optimal threshold that maximizes the discrimination between the two binarized classes (foreground and background).

[0058] The initialization process of the OTSU method is to calculate the average grayscale value of all non-zero pixels in the image and use it as the initial threshold. In the masked equalized grayscale difference image, only the pixels in the effective analysis area participate in this calculation. For example, if there are 10,000 pixels in the effective analysis area and the sum of their grayscale values ​​is 500,000, then the initial threshold is 500,000 / 10,000=50. Next, the traversal process begins: according to the current threshold u, the image pixels are divided into two categories (the set of pixels less than or equal to the threshold) The set of pixels whose sum is greater than the threshold ), calculate the average gray value of the two types of pixels respectively and , and then calculate the between-class variance Then continue in and All possible thresholds are traversed within a preset accuracy range, and the inter-class variance is calculated. The threshold that maximizes the inter-class variance is output as the optimal threshold. The average maximum inter-class variance threshold for short-pulse scan images is -19.7, and the average maximum inter-class variance threshold for medium-pulse scan images is -20.6. The equalized grayscale difference image is binarized based on the optimal threshold to distinguish between sea ice and open water areas. Binarization is the process of converting a grayscale image into a single value (usually 0 and 1). In this scheme, pixels in the equalized grayscale difference image with a value greater than the optimal threshold are considered to be sea ice areas and assigned a value of 1; pixels with a value less than or equal to the optimal threshold are considered to be open water areas and assigned a value of 0. This process produces an initial binary image of the sea ice distribution, which intuitively shows the spatial distribution of sea ice within the radar observation range.

[0059] Morphological opening and closing operations are performed on the initial binary sea ice distribution image to optimize its quality. Morphological operations are image processing methods based on mathematical morphology theory and are commonly used to remove noise and disconnected areas. Opening, which involves erosion followed by dilation, can remove small isolated points and noise; closing, which involves dilation followed by erosion, can fill small holes and gaps within the target area. The combination of these two operations effectively improves the quality of the binary image, smoothing the boundaries of sea ice areas and reducing noise interference.

[0060] Connected domain analysis is applied to eliminate small targets with an area of ​​less than 1000 square meters, further improving the reliability of sea ice identification. Connected domain analysis refers to finding all interconnected pixel regions in a binary image. Each region is called a connected domain. For each connected domain, its area (that is, the number of pixels contained) is calculated and multiplied by the square of the pixel resolution to obtain the actual area. For example, if a connected domain contains 100 pixels and the pixel resolution is 10 meters, its actual area is Square meters. For connected domains with an actual area of ​​less than 1,000 square meters, they are removed from the sea ice area because these tiny targets may be noise or other non-sea ice targets rather than real sea ice. Through this processing, an optimized binary image of sea ice distribution is obtained. Vectorization of the optimized binary image of sea ice distribution is the process of converting the region boundaries in the binary image into vector data. The main steps of vectorization include: boundary tracking, polygon fitting and attribute calculation. Boundary tracking is to extract the pixel sequence of the region boundary from the binary image; polygon fitting is to simplify the boundary pixel sequence into a polygon composed of a small number of vertices; attribute calculation is to calculate the geometric features of the region based on the polygon, such as area, perimeter, etc. Through vectorization processing, the geometric feature parameters such as the outer edge line and area of ​​the sea ice area are obtained. These vector data are more suitable for subsequent spatial analysis and visualization.

[0061] The ratio of the sea ice area in the vectorized sea ice distribution information to the total area of ​​the effective observation area is calculated to obtain the sea ice density in the radar observation area. The sea ice area refers to the sum of the areas of all sea ice areas after vectorization, and the total area of ​​the effective observation area is the area of ​​the entire circular effective analysis area, that is, ,in is the near-end blind zone radius, The formula for calculating sea ice density is: sea ice density = sea ice area / total area of ​​effective observation area. This ratio reflects the degree of sea ice coverage within the effective observation area and is an important indicator for describing the density of sea ice distribution.

[0062] In a specific embodiment, the process of performing the step of spatially superimposing the equalized grayscale difference image and the effective analysis area mask may specifically include the following steps:

[0063] Multiply the equalized grayscale difference image by the effective analysis area mask to obtain the grayscale difference image after mask processing;

[0064] Calculate the average gray value of all non-zero pixels in the gray difference image after mask processing as the initial threshold;

[0065] According to the initial threshold, the grayscale difference image pixels after mask processing are divided into two categories: a set of pixels less than or equal to the initial threshold and a set of pixels greater than the initial threshold, and the image binary classification result is obtained;

[0066] Calculate the average grayscale value of the two types of pixel sets in the image binary classification results respectively to obtain the average grayscale value within the class;

[0067] Determine the set of potential thresholds based on the average grayscale value within the class combined with the error accuracy, traverse all possible thresholds and calculate the corresponding inter-class means;

[0068] The corresponding threshold is determined according to the maximum inter-class variance, and the updated threshold is output as the optimal threshold.

[0069] Specifically, the equalized grayscale difference image is multiplied by the valid analysis area mask to produce a masked grayscale difference image. This step multiplies the two images pixel by pixel, eliminating the influence of invalid areas on the threshold calculation. Specifically, each pixel value in the equalized grayscale difference image is multiplied by the mask value (0 or 1) at the corresponding position. When the mask value is 1, the original grayscale value is retained; when the mask value is 0, the result is 0, indicating that the pixel is not included in the subsequent calculation. This spatial superposition operation ensures that only pixels within the valid analysis area participate in the threshold calculation, improving the algorithm's relevance and accuracy. The average grayscale value of all non-zero pixels in the masked grayscale difference image is calculated as the initial threshold. Non-zero pixels are pixels within the valid area where the mask value is 1. The calculation formula is: Initial threshold = Sum of all non-zero pixel grayscale values ​​ / Number of non-zero pixels. For example, if there are 10,000 pixels in the valid area and the total grayscale value is 150,000, then the initial threshold is 150,000 / 10,000 = 15. This initial threshold serves as the starting point of the traversal process and provides the basis for subsequent image classification.

[0070] Based on the initial threshold, the masked grayscale difference image pixels are divided into two categories: a set of pixels less than or equal to the initial threshold, and a set of pixels greater than the initial threshold, resulting in a binary image classification result. In the equalized grayscale difference image, the grayscale value reflects the degree of deviation of the pixel from the fitted trend. Generally, larger positive values ​​indicate sea ice areas (strong radar reflection), while smaller negative values ​​indicate open water areas (weak radar reflection). Using the threshold to classify pixels into two categories effectively distinguishes between sea ice and water areas. Continuing with the above example, if the initial threshold is 15, pixels with grayscale values ​​less than or equal to 15 are classified as the first category (likely corresponding to water areas), and pixels with grayscale values ​​greater than 15 are classified as the second category (likely corresponding to sea ice).

[0071] Calculate the average grayscale value of each pixel set in each of the two categories in the binary image classification results to obtain the intra-class average grayscale value. This step averages the grayscale values ​​of all pixels within each category to obtain representative values ​​for each category. For example, if the first category contains 6,000 pixels with a total grayscale value of 30,000, the average grayscale value of the first category is 30,000 / 6,000 = 5; if the second category contains 4,000 pixels with a total grayscale value of 120,000, the average grayscale value of the second category is 120,000 / 4,000 = 30. These two averages represent the grayscale characteristics of the two categories and are used in subsequent threshold update calculations.

[0072] The range of all potential thresholds is determined based on the average grayscale value within the class and the desired error precision. Continuing with the previous example, if the preset threshold precision is 0.1, all possible thresholds include all values ​​between 5.0 and 30.0 in steps of 0.1, totaling 251. All possible thresholds are iterated over and the corresponding inter-class variance is calculated. This updating method is based on the idea that the optimal threshold should be located midway between the features of the two classes, maximizing the inter-class variance between the two classes. At this threshold, the classification results show the greatest difference between the two classes, which helps improve classification accuracy.

[0073] The preset threshold precision is a small positive number, such as 0.1 or 0.01. The higher the preset precision, the larger the possible threshold range. Continuing with the above example, assuming that the between-class variance reaches a maximum value of 300 when the threshold is 17.3, 17.3 is output as the optimal threshold.

[0074] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0075] According to the GPS coordinates and heading information provided by the ship's navigation and positioning system, a spatial mapping relationship between the radar coordinate system and the geographic coordinate system is established to obtain spatial registration parameters;

[0076] The data source of sea ice density data observed during underway is determined. If the density value is recorded by manual visual observation, it is considered to be an overall value evenly distributed within the observation range. If the ice field image is collected by ship-borne optical equipment, distortion correction and geometric correction are performed to obtain a corrected optical image.

[0077] The corrected optical image is input into a deep convolutional network framework with a dual-branch attention mechanism. The deep convolutional network framework consists of four modules: feature extraction, spatial attention, channel attention, and classification. The sea ice density in the optical image is obtained.

[0078] The artificial visual observation density and optical image sea ice density are weighted and combined according to the observation quality and credibility to obtain the comprehensive underway observation sea ice density.

[0079] The area ratio of the near-end blind area to the effective observation area is determined based on the shipborne radar type and scanning radius, and the area weight of the underway observation data and the area weight of the radar observation data are calculated to obtain the data fusion weight.

[0080] According to the data fusion weight, the weighted average of the comprehensive cruise observation sea ice density and the radar observation area sea ice density is calculated to obtain the fused sea ice density.

[0081] Specifically, a spatial mapping relationship is established between the radar coordinate system and the geographic coordinate system based on the GPS coordinates and heading information provided by the ship's navigation and positioning system. The ship's navigation and positioning system typically provides information such as the ship's current latitude and longitude, heading angle, and speed. The radar coordinate system is a local coordinate system centered on the radar antenna. Establishing this spatial mapping relationship involves determining the coordinate origin, aligning the coordinate axes, and unifying the scale. Specifically, a point (x_r, y_r) in the radar coordinate system is transformed to a point (x_g, y_g) in the geographic coordinate system through rotation and translation. The transformation involves a rotation based on the ship's heading angle θ and a translation based on the GPS position. This spatial registration ensures spatial consistency between radar and underway observation data, laying the foundation for subsequent data fusion. Data source determination for underway observation sea ice density data allows for appropriate processing methods to be applied to different types of observation data. Underway observation data primarily comes in two forms: density values ​​recorded by visual observations and ice field images collected by shipboard optical equipment. Concentration values ​​recorded by visual observations are typically estimated and recorded by observers within a radius of approximately 1 km around the ship, based on the corresponding standards of the Antarctic Sea Ice Processes and Climate (ASPeCt) or the Arctic Ship-Based Sea Ice Observations (ASSIST). This observation method produces an overall concentration value, such as 0.3 (indicating that sea ice covers 30% of the observed area). Due to the lack of detailed spatial distribution information, this concentration value is assumed to be uniformly distributed throughout the observed area. Images of ice fields collected by shipboard optical equipment require distortion correction and geometric correction. Distortion correction eliminates image distortion caused by the camera lens, typically using camera parameters (focal length, principal point coordinates, distortion coefficients, etc.). Geometric correction involves projecting the image onto a horizontal plane based on the camera's mounting position and attitude to align it with the geographic coordinate system. These corrections produce an optical image that reflects the true distribution of sea ice.

[0082] The rectified optical image is fed into a deep convolutional network framework with a dual-branch attention mechanism, which utilizes deep learning techniques to automatically identify and calculate sea ice density. This framework is a neural network architecture specifically designed for sea ice identification and consists of four modules: feature extraction, spatial attention, channel attention, and classification. The feature extraction module, using a ResNet-50 network as its backbone, performs multi-level feature extraction on the input image to generate a feature map. The spatial attention module focuses on spatial location information in the feature map, emphasizing regions with distinct boundaries and texture features, which helps distinguish the boundary between sea ice and open water. The channel attention module focuses on channel information in the feature map, enhancing the responses of channels that contribute significantly to sea ice characteristics and suppressing the influence of irrelevant channels. Finally, the classification module transforms the processed feature map into a sea ice distribution probability map through a fully connected layer and calculates the overall sea ice density. This deep learning framework automatically extracts sea ice features and calculates sea ice density from optical images, overcoming the subjectivity and inefficiency of traditional manual identification. The visual observation density and optical imagery sea ice density are weighted and combined according to observation quality and reliability to produce the integrated underway observation sea ice density. This step integrates observational data from different sources to obtain a more accurate near-field sea ice density. Observation quality and reliability are assessed based on factors such as observation conditions, equipment status, and data integrity. For example, when visibility is good and equipment is functioning properly, observation quality is high; otherwise, it is low. The weighted combination formula is: Integrated underway observation sea ice density = α × visual observation density + β × optical imagery sea ice density, where α and β are weighting coefficients, and α + β = 1. For example, if the visual observation quality and reliability are higher, α might be 0.7 and β might be 0.3; if the optical imagery quality is better, α might be 0.3 and β might be 0.7. This weighted combination leverages the advantages of multiple observation methods and improves the reliability of near-field sea ice density data.

[0083] Based on the shipborne radar type and scanning radius, the ratio of the near-end blind zone to the effective observation area is determined, and the area weights of the underway observation data and the radar observation data are calculated. This step determines the contribution of different data sources to the final fusion. For short-pulse scanning radii of 2, 3, and 4 nautical miles, the underway observation weights are 2.0%, 4.6%, and 8.2%, respectively, while the radar observation weights are 98.0%, 95.4%, and 91.8%, respectively. For medium-pulse scanning radii of 3, 4, and 6 nautical miles, the underway observation weights are 1.6%, 2.8%, and 6.3%, respectively, while the radar observation weights are 98.4%, 97.2%, and 93.7%, respectively. These weights are calculated based on the ratio of the near-end blind zone to the effective observation area, reflecting the relative importance of different data sources in terms of spatial coverage.

[0084] The fused sea ice density is calculated by weighting the combined underway observations and the radar observation area's sea ice density according to the data fusion weights. This is the final step in multi-source fusion, integrating sea ice density data from different sources and regions into a comprehensive sea ice density covering the entire study area. The weighted average formula is: fused sea ice density = combined underway observations × underway observation weight + radar observation area's sea ice density × radar observation weight. This area-based weighting ensures that the contributions of near-field (radar blind spots) and far-field (radar effective observation areas) to the final results are consistent with their actual spatial extents, preventing data from any one region from excessively influencing the overall results.

[0085] In a specific embodiment, the process of calculating the area weight of the cruise observation data and the area weight of the radar observation data may specifically include the following steps:

[0086] The effective observation range is determined based on the radar pulse type. The effective observation range for short pulse scanning is 2.4 kilometers, and the effective observation range for medium pulse scanning is 4.0 kilometers. The effective observation boundary of the radar is obtained.

[0087] According to the radar effective observation boundary and the near-end blind zone boundary, the effective observation area is calculated to obtain the annular effective area area;

[0088] Calculate the near-end blind zone area, and determine the near-end blind zone area when the short pulse scanning radius is 2, 3, and 4 nautical miles, and the near-end blind zone area when the medium pulse scanning radius is 3, 4, and 6 nautical miles, respectively, to obtain the standardized near-end blind zone area;

[0089] Calculate the total study area as the sum of the proximal blind area and the annular effective area to obtain the total study area area;

[0090] The area weight of the cruise observation data is calculated as the ratio of the proximal blind area to the total area of ​​the study area to obtain the cruise observation weight;

[0091] The area weight of the radar observation data was calculated as the ratio of the area of ​​the annular effective region to the total area of ​​the study area to obtain the radar observation weight.

[0092] Specifically, the effective observation distance is determined according to the radar pulse type, which is an empirical value obtained based on the analysis of a large amount of measured data. Short pulse scanning is a radar that emits electromagnetic wave pulses with a shorter duration, which is suitable for close-range high-resolution observations, and its effective observation distance is 2.4 kilometers; medium pulse scanning emits electromagnetic wave pulses with a longer duration, which is suitable for longer-distance observations, and its effective observation distance is 4.0 kilometers. The effective observation distance refers to the maximum distance at which the radar signal can effectively reflect the sea ice situation. Beyond this distance, the signal quality is significantly reduced and the sea ice cannot be accurately identified. By determining the effective observation distance under different pulse types, the radar effective observation boundary is obtained, which defines the effective range of the radar data. According to the radar effective observation boundary and the proximal blind zone boundary, the area of ​​the effective observation area is calculated to obtain the area of ​​the annular effective area. The proximal blind zone refers to the overexposed area near the radar antenna due to excessive signal strength, which cannot accurately observe sea ice; the effective observation boundary is the maximum range that the radar signal can effectively observe. The effective observation area is the annular area between these two boundaries, and its area calculation formula is: the area of ​​the annular effective area ,in is the effective observation distance, For example, when the short pulse scanning radius is 3 nautical miles, the effective observation distance is 2.4 kilometers, and the near-end blind zone radius is 502 meters, the annular effective area is This circular area is where the radar can effectively observe sea ice and is also the basic area for calculating the sea ice density in the radar observation area.

[0093] Calculate the near-end blind zone area, determine the near-end blind zone area when the short pulse scanning radius is 2, 3, and 4 nautical miles, and the near-end blind zone area when the medium pulse scanning radius is 3, 4, and 6 nautical miles, and obtain the standardized near-end blind zone area. The near-end blind zone area refers to the area of ​​the circular area with the radar antenna as the center and the radius as the near-end blind zone radius. The calculation formula is: near-end blind zone area ,in When the short pulse scanning radius is 2, 3, and 4 nautical miles, the near-end blind zone radius is 334 meters, 502 meters, and 669 meters respectively, and the corresponding near-end blind zone areas are square meters, ≈791404 square meters, square meters; when the medium pulse scanning radius is 3, 4, and 6 nautical miles, the near-end blind zone radius is 502 meters, 669 meters, and 1004 meters respectively, and the corresponding near-end blind zone areas are square meters, square meters, ≈ 3166196 square meters. These near-end blind spots are areas that cannot be effectively observed by radar and need to be filled by comprehensive underway observation data.

[0094] Calculate the total study area as the sum of the near-end blind zone area and the annular effective area area to obtain the total study area area. The total study area refers to the entire circular area from the center of the ship to the effective observation boundary of the radar, including the near-end blind zone and the annular effective observation area. The calculation formula is: Total study area area = near-end blind zone area + annular effective area area ,in is the effective observation distance. For example, when the short pulse scanning radius is 3 nautical miles, the total area of ​​the study area is ≈ 18,086,511 square meters. This total area defines the scope of the entire study area and is the basis for calculating the data fusion weight. The area weight of the comprehensive cruise observation data is calculated as the ratio of the proximal blind zone area to the total area of ​​the study area, and the comprehensive cruise observation weight is obtained. Comprehensive cruise observation is mainly used to fill the data of the proximal blind zone, so its weight should be equivalent to the proportion of the proximal blind zone in the total study area. The calculation formula is: Comprehensive cruise observation weight = proximal blind zone area / total area of ​​the study area For example, for a short pulse scanning radius of 3 nautical miles, the weight of the comprehensive underway observation is , or 4.6%. For short-pulse scanning radii of 2, 3, and 4 nautical miles, the weights for under-the-ship observations were 2.0%, 4.6%, and 8.2%, respectively. For medium-pulse scanning radii of 3, 4, and 6 nautical miles, the weights for under-the-ship observations were 1.6%, 2.8%, and 6.3%, respectively. These weights reflect the relative importance of the near-end blind zone within the entire study area under different scanning conditions.

[0095] The area weight of the radar observation data is calculated as the ratio of the area of ​​the annular effective area to the total area of ​​the study area to obtain the radar observation weight. Radar observation is mainly used to obtain data in the annular effective area, so its weight should be equivalent to the proportion of the annular effective area in the total study area. The calculation formula is: Radar observation weight = area of ​​the annular effective area / total area of ​​the study area For example, for a short-pulse scan radius of 3 nautical miles, the radar observation weight is 1 - 0.046 = 0.954, or 95.4%. For short-pulse scan radii of 2, 3, and 4 nautical miles, the radar observation weights are 98.0%, 95.4%, and 91.8%, respectively. For medium-pulse scan radii of 3, 4, and 6 nautical miles, the radar observation weights are 98.4%, 97.2%, and 93.7%, respectively. These weights reflect the relative importance of the annular effective area within the entire study area under different scanning conditions.

[0096] The above describes the method for obtaining sea ice density by multi-source fusion in the embodiment of the present application. The following describes the device for obtaining sea ice density by multi-source fusion in the embodiment of the present application. Figure 2 In one embodiment of the present application, a device for acquiring sea ice density through multi-source fusion includes:

[0097] An acquisition module 201 is configured to acquire original radar images and sea ice density data collected by a ship-borne ice navigation radar, perform OCR text recognition on the original radar images, extract radar parameters, and obtain an effective radar grayscale image.

[0098] The fitting module 202 is used to perform axis density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminate the central overexposed area and the remote blind area, and obtain a balanced grayscale difference image;

[0099] The segmentation module 203 is used to input the equalized grayscale difference image into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing to obtain vectorized sea ice distribution information and sea ice density in the radar observation area;

[0100] The averaging module 204 is used to fill the blind spots based on the sea ice density data observed during underway observations through a deep convolutional network framework with a dual-branch attention mechanism, and to perform a weighted average of the sea ice density in the radar observation area and the comprehensive underway observation sea ice density according to the area ratio to obtain a fused sea ice density.

[0101] Through the collaborative efforts of the aforementioned components, the system automatically identifies and extracts key information such as the radar's operating mode, scanning radius, and pulse length by acquiring raw radar images and underway sea ice density data collected by the ship-borne ice navigation radar and extracting radar parameters using optical character recognition (OCR) text recognition technology. This avoids the tedious and error-prone manual parameter input and improves the automation and efficiency of data processing. Axis density distribution analysis and polynomial fitting of the effective radar grayscale image address the problem of radar signal attenuation with distance. Data quality is ensured by eliminating central overexposed areas and far-end blind spots, resulting in a balanced grayscale difference image, providing a high-quality data foundation for subsequent sea ice identification. The balanced grayscale difference image is then input into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing, enabling precise differentiation between sea ice and open water. Vectorization is then used to obtain sea ice distribution information and sea ice density in the radar observation area, improving the accuracy and usability of the identification results. A deep convolutional network framework with a dual-branch attention mechanism processes underway sea ice density data to fill blind spots, cleverly addressing the problem of missing data in the near-blind zone of the radar. The spatial attention and channel attention modules in this framework effectively capture the spatial distribution and texture characteristics of sea ice, improving the ability to identify complex ice conditions. A fused sea ice density is obtained by weightedly averaging the sea ice density in the radar observation area and the integrated underway observations based on their area ratio. This achieves a rational fusion of multi-source data, fully leveraging the strengths of each data source while overcoming the limitations of a single one. In particular, the deep convolutional network framework with a dual-branch attention mechanism and the adaptive inter-class maximum variance method (AUMV) complement each other functionally. The AUMV provides accurate sea ice distribution information in the radar observation area, providing reference data for the training and validation of the deep convolutional network. The deep convolutional network, on the other hand, processes optical image data to provide supplementary information for radar blind spots. Together, they form a complete description of sea ice distribution. At the same time, the adaptive inter-class maximum variance method based on image segmentation and the deep convolutional network with a dual-branch attention mechanism based on deep learning have been improved and optimized to meet the special needs of sea ice recognition. The former calculates the optimal threshold through the maximum inter-class variance, adapting to the variability of the grayscale difference between sea ice and water; the latter enhances the perception of sea ice characteristics through the attention mechanism, thereby improving the recognition accuracy in complex backgrounds.

[0102] above Figure 2 The multi-source fusion sea ice density acquisition device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The multi-source fusion sea ice density acquisition device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0103] Figure 3This is a schematic diagram of the structure of a multi-source fusion sea ice concentration acquisition device provided by an embodiment of the present invention. This multi-source fusion sea ice concentration acquisition device 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing applications 333 or data 332. Memory 320 and storage media 330 may be either transient or persistent storage. The program stored in storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions and operations within multi-source fusion sea ice concentration acquisition device 300. Furthermore, processor 310 may be configured to communicate with storage medium 330, allowing multi-source fusion sea ice concentration acquisition device 300 to execute the series of instructions and operations stored in storage medium 330 to implement the steps of the multi-source fusion sea ice concentration acquisition method described above.

[0104] The multi-source fusion sea ice density acquisition device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the multi-source fusion sea ice density acquisition device shown does not constitute a limitation on the multi-source fusion sea ice density acquisition device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0105] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the multi-source fusion sea ice density acquisition method.

[0106] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0107] If the integrated unit is implemented in the form of 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, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a multi-source fusion sea ice density acquisition device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-source fusion method for obtaining sea ice density, characterized in that: The method comprises: Obtaining original radar images collected by the ship-borne ice navigation radar and sea ice density data observed manually during navigation, performing OCR text recognition on the original radar images to extract radar parameters and obtain an effective radar grayscale image; Performing axis density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminating the central overexposed area and the far-end blind area, and obtaining a balanced grayscale difference image; Inputting the equalized grayscale difference image into an adaptive inter-class maximum variance method for threshold segmentation and morphological processing to obtain vectorized sea ice distribution information and sea ice density in the radar observation area; Based on the sea ice density data observed manually while underway, the blind spots are filled through the deep convolutional network framework of the dual-branch attention mechanism. The sea ice density in the radar observation area and the sea ice density data observed manually while underway are weighted averaged according to the area ratio to obtain the fused sea ice density, including: establishing a spatial mapping relationship between the radar coordinate system and the geographic coordinate system according to the GPS coordinates and heading information provided by the ship's navigation and positioning system to obtain spatial registration parameters; judging the data source of the sea ice density data observed manually while underway, and if the density value is recorded by manual visual observation, it is regarded as an overall value uniformly distributed within the observation range; if it is an ice field image collected by ship-borne optical equipment, distortion correction and geometric correction are performed to obtain a corrected optical image; the corrected A deep convolutional network framework with a dual-branch attention mechanism for optical image input is constructed. The deep convolutional network framework consists of four modules: feature extraction, spatial attention, channel attention, and classification. The sea ice density of the optical image is obtained. The artificial visual observation density and the optical image sea ice density are weightedly combined according to the observation quality and credibility to obtain the comprehensive underway artificial observation sea ice density data. The area ratio of the proximal blind area to the effective observation area is determined according to the shipborne radar type and scanning radius, and the area weights of the artificial observation data and the radar observation data are calculated to obtain the data fusion weight. According to the data fusion weight, the comprehensive underway artificial observation sea ice density data and the sea ice density in the radar observation area are weightedly averaged to obtain the fused sea ice density.

2. The multi-source fusion sea ice density acquisition method according to claim 1 is characterized in that: The method of obtaining an original radar image collected by a ship-borne ice navigation radar and sea ice density data manually observed during navigation, performing OCR text recognition on the original radar image to extract radar parameters, and obtaining an effective radar grayscale image includes: Identifying the radar operating mode, scanning radius, and pulse length of the original radar image sequence using an OCR optical text recognition method to obtain radar operating parameters; Classifying the original radar image according to the radar operating parameters, eliminating images affected by the STC function and other abnormal images, and obtaining a valid radar image in the ice area navigation mode; Classifying the effective radar image into a short pulse scan image and a medium pulse scan image according to pulse length, limiting the scan radius of the short pulse scan image to 2 to 4 nautical miles, and limiting the scan radius of the medium pulse scan image to 3 to 6 nautical miles, to obtain a classified radar image; Extracting a radar scanning image of a circular area from the classified radar image, and extracting a blue band component to obtain an initial grayscale image of the radar image; Calculating pixel resolution based on the initial grayscale image and the corresponding radar scanning radius to obtain a standardized pixel resolution; The initial grayscale image is spatially scaled according to the standardized pixel resolution to obtain the effective radar grayscale image.

3. The multi-source fusion sea ice density acquisition method according to claim 1 is characterized in that: The performing of axis density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminating the central overexposed area and the remote blind area, and obtaining a balanced grayscale difference image includes: Drawing a plurality of axes radiating outward from the center of the image for the effective radar grayscale image, calculating the distribution of pixel values ​​on each axis in the distance direction, and obtaining an axis density distribution curve; By aggregating multiple axis density distribution curves, it is determined that the effective observation distance of the short pulse scanning image is 2.4 kilometers, and the effective observation distance of the medium pulse scanning image is 4.0 kilometers, thereby obtaining the effective observation range of the radar; Performing a grayscale value saturation analysis on an area within 30 pixels from the center of the effective radar grayscale image, marking an area exceeding a preset threshold as a near-end blind zone, and obtaining a near-end blind zone boundary; The near-end blind zone is accurately quantified according to the radar pulse type and the scanning radius, and the near-end blind zone radius under different conditions is determined to obtain a quantified near-end blind zone parameter; The effective radar grayscale images are grouped according to radar pulse type and scanning radius, and a fitting model is established using a quadratic polynomial fitting method to obtain a polynomial fitting model. The polynomial fitting model is used to describe the attenuation trend of the radar signal in the upward direction of the distance. The form of the polynomial fitting model is: , where x is the distance from the pixel to the center of the image, f(x) is the expected grayscale value at the corresponding position, 、 and is the fitting coefficient; The difference between the actual grayscale value of each pixel in the effective radar grayscale image and the fitting value calculated based on the polynomial fitting model is calculated to obtain the equalized grayscale difference image.

4. The multi-source fusion sea ice density acquisition method according to claim 1 is characterized in that: The step of inputting the equalized grayscale difference image into the adaptive inter-class maximum variance method for threshold segmentation and morphological processing to obtain vectorized sea ice distribution information and sea ice density in the radar observation area includes: The radar scanning range is divided into an effective analysis area, which is defined as the circular area after removing the near-end blind area and the far-end blind area, and the effective analysis area mask is obtained; Spatially superimposing the equalized grayscale difference image and the effective analysis area mask, inputting the adaptive inter-class maximum variance method, initializing the iterative threshold to the global average of the grayscale value, and iteratively updating the threshold until convergence to obtain the optimal threshold; performing binarization processing on the equalized grayscale difference image according to the optimal threshold value to distinguish sea ice areas from open water areas, thereby obtaining an initial sea ice distribution binary image; performing morphological opening and closing operations on the initial sea ice distribution binary image, and applying connected domain analysis to eliminate small objects with an area of ​​less than 1000 square meters, to obtain an optimized sea ice distribution binary image; performing vectorization processing on the optimized sea ice distribution binary image, extracting the outer edge line and area of ​​the sea ice area, and obtaining the vectorized sea ice distribution information; The ratio of the sea ice area in the vectorized sea ice distribution information to the total area of ​​the effective analysis area is calculated to obtain the sea ice density in the radar observation area.

5. The multi-source fusion sea ice density acquisition method according to claim 4 is characterized in that: The equalized grayscale difference image is spatially superimposed with the effective analysis area mask, an adaptive inter-class maximum variance method is input, an iterative threshold is initialized to a global average of grayscale values, and the threshold is iteratively updated until convergence to obtain an optimal threshold, including: multiplying the equalized grayscale difference image by the effective analysis area mask to obtain a grayscale difference image after mask processing; Calculating the average grayscale value of all non-zero pixels in the grayscale difference image after mask processing as the initial iteration threshold to obtain an iteration starting threshold; Dividing the grayscale difference image pixels after the mask processing into two categories according to the iteration starting threshold: a set of pixels less than or equal to the iteration starting threshold and a set of pixels greater than the iteration starting threshold, to obtain an image binary classification result; Calculate the average grayscale value of the two types of pixel sets in the binary classification result of the image respectively to obtain the average grayscale value within the class; Calculate the new threshold value based on the average grayscale value within the class as the arithmetic average of the average grayscale values ​​of the two classes, and obtain the iteratively updated threshold value; The above classification, calculation and updating process is repeated until the absolute value of the difference between the thresholds before and after the iterative update is less than the preset convergence threshold, and the last updated threshold is output as the optimal threshold.

6. The multi-source fusion sea ice density acquisition method according to claim 1 is characterized in that: The method of determining the area ratio of the near-end blind area to the effective observation area according to the shipborne radar type and scanning radius, calculating the area weight of the manual observation data and the area weight of the radar observation data, and obtaining the data fusion weight includes: The effective observation range is determined based on the radar pulse type. The effective observation range for short pulse scanning is 2.4 kilometers, and the effective observation range for medium pulse scanning is 4.0 kilometers. The effective observation boundary of the radar is obtained. Calculating the effective observation area according to the radar effective observation boundary and the near-end blind zone boundary to obtain the annular effective area; Calculate the near-end blind zone area, and determine the near-end blind zone area when the short pulse scanning radius is 2, 3, and 4 nautical miles, and the near-end blind zone area when the medium pulse scanning radius is 3, 4, and 6 nautical miles, respectively, to obtain the standardized near-end blind zone area; Calculate the total study area as the sum of the proximal blind area and the annular effective area to obtain the total study area area; The area weight of the manual observation data was calculated as the ratio of the proximal blind area to the total area of ​​the study area to obtain the manual observation weight; The area weight of the radar observation data was calculated as the ratio of the area of ​​the annular effective region to the total area of ​​the study area to obtain the radar observation weight.

7. A multi-source fusion sea ice density acquisition device, characterized in that: A method for obtaining sea ice density by multi-source fusion according to any one of claims 1 to 6, wherein the multi-source fusion sea ice density obtaining device comprises: An acquisition module is used to obtain the original radar image collected by the ship-borne ice navigation radar and the sea ice density data observed manually during navigation, perform OCR text recognition on the original radar image to extract radar parameters, and obtain an effective radar grayscale image; A fitting module is used to perform axial density distribution analysis and polynomial fitting on the effective radar grayscale image, eliminate the central overexposed area and the far-end blind area, and obtain a balanced grayscale difference image; a segmentation module, configured to input the equalized grayscale difference image into an adaptive inter-class maximum variance method for threshold segmentation and morphological processing, thereby obtaining vectorized sea ice distribution information and sea ice density in the radar observation area; The averaging module is used to fill the blind spots based on the sea ice density data of the artificial observation underway through the deep convolutional network framework of the dual-branch attention mechanism, and perform weighted averaging on the sea ice density of the radar observation area and the sea ice density of the artificial observation underway according to the area ratio to obtain the fused sea ice density, including: establishing a spatial mapping relationship between the radar coordinate system and the geographic coordinate system according to the GPS coordinates and heading information provided by the ship's navigation and positioning system to obtain spatial registration parameters; judging the data source of the sea ice density data of the artificial observation underway, if the density value is recorded by manual visual observation, it is regarded as an overall value uniformly distributed within the observation range; if it is an ice field image collected by ship-borne optical equipment, distortion correction and geometric correction are performed to obtain a corrected optical image; the correction The positive optical image is input into a deep convolutional network framework with a dual-branch attention mechanism. The deep convolutional network framework consists of four modules: feature extraction, spatial attention, channel attention, and classification, to obtain the sea ice density of the optical image; the artificial visual observation density and the optical image sea ice density are weightedly combined according to the observation quality and credibility to obtain the comprehensive underway artificial observation sea ice density data; the area ratio of the proximal blind area to the effective observation area is determined according to the type and scanning radius of the shipborne radar, and the area weights of the artificial observation data and the radar observation data are calculated to obtain the data fusion weight; according to the data fusion weight, the comprehensive underway artificial observation sea ice density data and the sea ice density in the radar observation area are weightedly averaged to obtain the fused sea ice density.

8. A multi-source fusion sea ice density acquisition device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the multi-source fusion sea ice density acquisition method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to perform the multi-source fusion sea ice density acquisition method according to any one of claims 1 to 6.

Images