Sea ice monitoring method and device based on optical image of stationary satellite

Abstract

The invention discloses a sea ice monitoring method based on a stationary satellite optical image. The sea ice monitoring method comprises the following steps: acquiring GOCI-2 atmospheric corrected data; land mask and sea ice labeling is carried out based on the obtained historical data, a three-classification sea ice training data set is constructed according to a labeling result and a GOCI-2 cloud mask, and labels are seawater, sea ice and cloud mist; generating a sea ice monitoring model based on Swinin-unet, and iteratively training the sea ice monitoring model by using the constructed data set and the category correction cross entropy loss function to obtain an optimal model; and inputting GOCI-2 atmospheric corrected data, using the optimal model to obtain a three-classification result, and integrating sea ice and cloud mist through post-processing to obtain a sea ice monitoring result. The invention further comprises a device for implementing the method. The problem of insufficient sea ice detection and identification capability in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the field of satellite remote sensing technology, and in particular to a sea ice monitoring method and device based on geostationary satellite optical images, a computer product, a memory and a processor. Background Art

[0002] Sea ice is one of the main disasters affecting marine aquaculture and offshore operations, and its range is an important factor in assessing the level of sea ice disasters. The electromagnetic properties of sea ice are different from those of land and water bodies, and can be monitored through reflectivity images obtained by optical satellites or echo images obtained by synthetic aperture radars. Common sea ice monitoring methods include manual annotation, threshold method, machine learning, and deep learning. Among them, although the accuracy of manual annotation is relatively high, it requires corresponding experts to spend a lot of time and energy, and it is difficult to process a large number of sea ice images in a timely manner. The threshold method is easy to use, but it is more sensitive to the solar altitude angle, weather conditions or incident angle, resulting in lower accuracy. Deep learning is a special type of machine learning. In recent years, with the development of deep learning in the automatic recognition and segmentation of satellite images, its accuracy has exceeded the threshold method and machine learning methods, becoming one of the most promising methods for achieving high-efficiency and high-precision sea ice monitoring.

[0003] Existing sea ice monitoring models often use convolution-based deep learning models, which extract texture features that are conducive to semantic segmentation in satellite images by training parameters in multiple convolution kernels at different levels, thereby achieving sea ice monitoring. The commonly used convolutional deep learning model U-Net can achieve sea ice monitoring based on optical images and obtain a rough range of sea ice. However, since the characteristics of clouds and sea ice in optical images are similar, U-Net will mistakenly identify clouds as sea ice to a certain extent, resulting in an overestimation of the monitored sea ice area. In addition, U-Net has a weak ability to identify interglacial waterways in optical images with a resolution of 100 meters.

[0004] Existing sea ice monitoring is generally based on polar orbit synthetic aperture radar images or optical satellite images, the former such as Sentinel-1, GF-3, the latter such as Landsat-8, MOIDS, both of which have low temporal resolution for Bohai Sea. Polar orbit satellites fly around the earth's poles, with a short re-entry period in the polar regions, but a long re-entry period in low-latitude regions. When there are interference factors such as clouds and fog, it is easy to cause data loss, reduce temporal resolution, and make it difficult to meet the needs of long-term observation of Bohai Sea ice. Summary of the invention

[0005] The present application proposes a sea ice monitoring method and device based on geostationary satellite optical images to solve the problem of insufficient sea ice detection and recognition capabilities in the prior art.

[0006] An embodiment of the present application provides a sea ice monitoring method based on geostationary satellite optical images, comprising the following steps. Obtain GOCI-2 atmospheric correction data, perform land mask and sea ice annotation based on the historical data, and construct a three-class sea ice training dataset based on the annotation results and GOCI-2 cloud mask, with labels of sea water, sea ice, and cloud fog; Generate a sea ice monitoring model based on Swin-unet, and use the constructed data set and the category-corrected cross entropy loss function to iteratively train the sea ice monitoring model to obtain the optimal model; The atmospheric-corrected data of GOCI-2 were input, and the three-classification results were obtained using the optimal model. The seawater and cloud labels were then merged through post-processing to finally obtain the sea ice monitoring results.

[0007] In an embodiment of the present application, three types of true value labels, namely sea water, sea ice, and clouds, are generated by image annotation and reading of GOCI-2 cloud masks, where sea water and clouds are subcategories of background labels. Land masking and slicing are performed on remote sensing images and annotated true value images to generate a sea ice training data set for model training.

[0008] In an embodiment of the present application, the sea ice monitoring model of Swin-unet is based on the following three parts: encoder, skip connection and decoder.

[0009] In an embodiment of the present application, a category-corrected cross entropy loss function is introduced, and the model loss is calculated using the dot product result of the prediction result and the category correction matrix.

[0010] In the embodiment of the present application, the encoder structure in the sea ice monitoring model of Swin-unet includes region segmentation, linear embedding, 8 STBs and 3 downsampling. Region segmentation divides the remote sensing image slice into two equal parts in length and width and then superimposes them in the channel direction, so that the original length, width and channel are , , The image slice input becomes This method is better than the commonly used length and width quartering method. The number of channels is increased by linear embedding. Three downsamplings are alternately connected to four groups of STBs, each group of STBs contains two STBs, and the two STBs contain layer normalization, window-based multi-head self-attention modules and multi-layer perceptrons. Among them, the second window-based multi-head self-attention module shifts the window by half the window length.

[0011] In an embodiment of the present application, the decoder in the sea ice monitoring model of Swin-unet includes 4 upsampling, 6 STBs and linear projection. Each group of 2 STBs is used to fuse the scale features input from upsampling and jump connection. The first three upsampling processes expand the feature length and width to twice the original, while halving the number of channels. The last upsampling only doubles the length and width of the feature without changing the number of channels. Finally, the feature is converted into a three-classification result through linear projection, so that its length and width are the same as the original remote sensing image slice, and the number of channels is the same as the number of classifications. The sea water and cloud labels are then merged through post-processing to obtain a single-channel sea ice monitoring result.

[0012] The embodiment of the present application further proposes a sea ice monitoring device based on geostationary satellite optical images, which is used to implement the method described in any embodiment of the present application, including: The acquisition module is used to obtain GOCI-2 atmospheric correction data, annotate land masks and sea ice based on the acquired data, and construct a three-class sea ice training dataset based on the annotation results and GOCI-2 cloud masks, with labels of sea water, sea ice, and cloud fog; The generation module is used to generate a sea ice monitoring model based on Swin-unet, and iteratively train the sea ice monitoring model using the constructed data set and the category-corrected cross entropy loss function to obtain the optimal model; The determination module is used to input the atmospherically corrected data of GOCI-2 and obtain the sea ice monitoring results using the optimal model obtained through training.

[0013] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: To solve the above problems, the present invention proposes a sea ice monitoring method based on geostationary satellite optical images, and uses the Swin-unet model based on the self-attention mechanism to more accurately extract sea ice pixels in the image. Compared with the classic U-net model, Swin-unet is only based on the self-attention mechanism and is a new generation of deep learning model with higher model learning ability and generalization ability. In addition, the present invention also introduces a category-corrected cross entropy loss function to weaken the contribution of differences between subcategories to model training loss, thereby improving the model's ability to identify ice channels. On the other hand, unlike the polar orbit, the geostationary orbit is located at a certain height above the equator. At this height, the geostationary satellite revolves around the earth at an angular velocity of the same magnitude and direction as the earth's rotation, enabling it to capture images of the same location at multiple different time periods in a single day, greatly increasing the temporal resolution and reducing data loss caused by cloud and fog interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 This is a flow chart of an embodiment of a method for monitoring sea ice based on geostationary satellite optical images of the present application; Figure 2(a)-(b) is a deep learning sea ice monitoring model, where Figure 2(a) is a schematic diagram of the Swin-unet model structure, and 2(b) is a schematic diagram of the STBx2 structure; Figure 3 This is a comparative diagram of the sea ice monitoring model in Liaodong Bay; Figure 4 This is a comparative diagram of the sea ice monitoring model in Laizhou Bay and Bohai Bay; Figure 5 This is a schematic diagram of an embodiment of a sea ice monitoring device based on geostationary satellite optical images of the present application. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0016] The embodiment of the present application provides a method and device for sea ice monitoring based on geostationary satellite optical images, including: obtaining geostationary satellite optical atmospheric correction data, extracting cloud masks, manually annotated sea ice true values, constructing a three-classification sea ice training data set, generating a deep learning sea ice monitoring model, training the sea ice monitoring model using a category-corrected cross entropy loss function, and verifying the sea ice monitoring model. The technical solutions provided by the embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0017] Figure 1 This is a flow chart of an embodiment of a sea ice monitoring method based on geostationary satellite optical images of the present application, including the following steps 110-130. Step 110: Obtain GOCI-2 atmospherically corrected data, perform land mask and sea ice annotation based on the acquired historical data, and construct a three-class sea ice training data set based on the annotation results and GOCI-2 cloud mask, with labels of sea water, sea ice, and cloud fog.

[0018] GOCI-2 stands for Geostationary Ocean Color Imager Ⅱ. It is carried on the geostationary satellite GK-2B launched by South Korea in February 2020. It has a spatial resolution of 250m and can image 10 scenes from 7:00 to 16:00 Beijing time every day, with a temporal resolution of 1 hour.

[0019] Constructing a remote sensing image sea ice training data set includes obtaining GOCI-2 images after atmospheric correction, randomly selecting training data according to the image date, annotating sea ice with reference to true color synthetic images, reading GOCI-2 cloud masks, generating three types of true value labels labeled as sea water, sea ice, and clouds, performing land masking and slicing on remote sensing images and annotated true value images, and generating a sea ice training data set for model training. Since the goal of the present invention is to monitor sea ice in remote sensing images, and sea water labels and cloud labels are actually background categories, that is, sea water and clouds are subcategories of the background, and when evaluating the sea ice monitoring model, only the model's ability to distinguish between sea ice and background is evaluated. Therefore, this data set is a three-category sea ice training data set, the purpose of which is to stimulate the learning potential of the deep learning model for different features in the same category.

[0020] Step 120: Generate a sea ice monitoring model based on Swin-unet, and iteratively train the sea ice monitoring model using the constructed data set and the category-corrected cross entropy loss function to obtain the optimal model.

[0021] Swin-unet refers to a deep learning model based on the self-attention mechanism and containing a symmetric encoding-decoding structure. A deep learning sea ice monitoring model based on Swin-unet is generated. The structure of Swin-unet is shown in Figure 2. Its generation is based on the following three parts: encoder, jump connection and decoder.

[0022] The three-category sea ice training data set obtained in step 110 and the deep learning sea ice monitoring model are used for training. The training requires multiple iterations to adjust the model weights to the best value so that the model converges. In each iteration, the difference between the output result of the calculation model and the true value, that is, the model loss, needs to be compared and transmitted back to the model to provide a basis for parameter adjustment. The function for calculating the model loss is called the loss function. In order to reduce the misclassification of ice channels in hundred-meter resolution images by existing models, the present invention introduces a category-corrected cross entropy loss function while using the three-category data set. The formula is as follows:

[0023] in To correct the cross entropy loss, is a true value, To predict the results, Correction matrix for categories.

[0024] The difference from the traditional cross entropy loss function is that the model loss is calculated by multiplying the prediction result with the category correction matrix, thereby reducing the falsely high loss caused by the difference between subcategories. The expression is:

[0025] In the formula is the correction matrix, is the correction factor, and its value range is .

[0026] Correction Matrix The dimension of is generally 2, and the size of each dimension is the same as the number of subcategories. The values ​​on the diagonal are all 1, which means that the loss generated by each subcategory itself is retained when calculating the loss function. Can exist in other locations ,when When it is 0, it means that the category correction matrix is ​​not used, and the category correction cross entropy function is consistent with the traditional cross entropy function; The larger the value, the greater the degree of correction, and the more tolerant the loss function is to the misclassification between subclasses. When it is 1, it means that the loss function treats subcategories as the same category at calculation time, which may cause the model training to fail to converge. The position in the matrix represents the subcategory that needs to be corrected, such as Located in the first row and third column, it represents the situation where the loss function will correct the model to classify the first category as the third category. Therefore, in the correction matrix They are located in the third column of the first row and the first column of the third row respectively, which means that the loss function will correct the misclassification between the first category and the third category. In the sea ice monitoring model, the first category and the third category represent sea water and clouds, respectively, both of which are subcategories of the background. The value of is determined by multiple experiments and is 0.7 in this scenario.

[0027] Step 130: Input the GOCI-2 atmospheric corrected data, use the optimal model to obtain the three-classification results, and integrate the sea water and cloud labels after post-processing to obtain the sea ice monitoring results.

[0028] The example of this application provides a method for calculating the training loss of a deep learning model. The present invention reduces the false high training loss caused by the difference in factor categories during the model training process by introducing a category correction matrix into the cross entropy loss function, and further utilizes the three-category sea ice training data set to reduce the misclassification caused by ice channels. In this application, ice channels refer to navigable water surfaces formed after the sea ice is broken. Their width in images with a resolution of 100 meters is generally several to a dozen pixels, and they belong to the background category when monitoring sea ice.

[0029] Figure 2(a)-(b) is a schematic diagram of the deep learning sea ice monitoring model, which includes an encoder, skip connection, and decoder, which are explained as follows.

[0030] Encoder. The encoder is used to extract features of different scales from remote sensing image slices. Its structure includes region segmentation, linear embedding, 8 STBs (Swin-transformer blocks) and 3 downsampling. Region segmentation divides the remote sensing image slice into two equal parts in length and width and then superimposes them in the channel direction, making the original length, width and channel equal to , , The image slice input becomes After many experiments, it was found that the accuracy of the sea ice monitoring model constructed by dividing the length and width into two equal parts is better than the commonly used method of dividing the length and width into four equal parts. Then, through linear embedding, the number of channels is increased to , The value of is usually 96. Then it passes through 4 groups of STBs and three downsamplings. Each group of STBs consists of 2 similar STBs, both of which contain layer normalization, window-based multi-head self-attention modules (W-MSA or SW-MSA) and multi-layer perceptrons (such as Figure 2b ). The difference is that the second window-based multi-head self-attention module will shift the window by half the window length to achieve information exchange between different windows. The downsampling process will reduce the length and width to half of the original, while doubling the number of channels, so that STB can extract features of different scales.

[0031] Skip connection. The structure of skip connection is simple and it is used to transfer features of different scales extracted by the encoder to the decoder.

[0032] Decoder. The decoder is used to fuse the features of different scales extracted by the encoder, restore them to the scale of remote sensing images, and output sea ice monitoring results. The decoder contains 4 upsampling, 6 STBs and linear projections. The structure of STB is the same as that of the encoder. Each group of 2 is used to fuse the scale features input from upsampling and jump connection. The first three upsampling processes expand the feature length and width to twice the original, while halving the number of channels. The last upsampling is special. It only doubles the length and width of the feature without changing the number of channels. That is, the feature size of its output is Finally, the features are converted into three-classification results through linear projection, so that their length and width are the same as the original remote sensing image slices, and the number of channels is the same as the number of classifications, which is 3. Then, the single-channel sea ice monitoring results are obtained through post-processing.

[0033] This application example introduces the deep learning model Swin-unet to generate a sea ice monitoring model based on geostationary satellite optical images, reducing the model's misclassification of clouds and fog. The model input is the image slices after the land mask of the 12 available bands contained in the GOCI-2 image. The model parameters are iteratively trained using the three-classification true value labels. The classification results of seawater, sea ice, and clouds and fog based on the GOCI-2 image are obtained through the best model. Finally, the seawater and clouds and fog are processed as background categories to obtain the sea ice monitoring results based on geostationary satellite optical images.

[0034] A total of 124 Bohai GOCI-2 images that were not used in the preparation of the sea ice dataset and the training of the sea ice monitoring model were used to verify the best deep learning sea ice monitoring model in step 3, and the overall accuracy index was obtained. The sea ice monitoring accuracy index IoU (Intersection over union) was compared with the existing sea ice monitoring model (based on U-Net). The IoU of the sea ice monitoring model of the present invention was 81.3%, and the IoU of the existing model was 76.7%, an increase of 4.6%. The comparison results are shown in the following example: Figure 3 and Figure 4 shown. Figure 3 This is a comparative diagram of the sea ice monitoring model in Liaodong Bay. Figure 4 The following is a comparative diagram of the sea ice monitoring model in Laizhou Bay and Bohai Bay. In the two figures, (a) is the true color composite image of GOCI-2, (b) is the annotated true value, (c) is the sea ice monitoring result of the existing model, and (d) is the sea ice monitoring result of the present invention. Figure 3~4 In the circled area in sub-figure (c), compared with the existing models, the sea ice monitoring model of the present invention reduces the misclassification caused by clouds and / or ice channels.

[0035] Figure 5 This is a schematic diagram of an embodiment of a sea ice monitoring device based on geostationary satellite optical images of the present application. The device is used to implement the method described in any embodiment of the present application, including: The acquisition module 41 is used to obtain the GOCI-2 atmospheric corrected data, perform land mask and sea ice annotation based on the acquired data, and construct a three-class sea ice training data set based on the annotation results and the GOCI-2 cloud mask, with the labels being sea water, sea ice, and cloud fog.

[0036] The generation module 42 is used to generate a sea ice monitoring model based on Swin-unet, and iteratively train the sea ice monitoring model using the constructed data set and the category-corrected cross entropy loss function to obtain the optimal model.

[0037] The determination module 43 is used to input the atmospheric corrected data of GOCI-2 and obtain the sea ice monitoring results using the optimal model obtained through training.

[0038] The method and device of the present application have the following advantages compared with the prior art: (I) Application of Swin-unet in sea ice monitoring using geostationary satellite optical images. Swin-unet is a new generation of deep learning model based on the self-attention mechanism. Compared with the U-Net model based on convolutional neural network, it was launched later, has stronger parameter learning ability, higher model generalization ability, and is more suitable for data sets composed of large amounts of data. Swin-unet is applied to sea ice monitoring using geostationary satellite optical images. Compared with existing models, it reduces misclassification caused by clouds and fog and improves the accuracy of sea ice monitoring.

[0039] (ii) Introducing the category-corrected cross entropy loss function. Through investigation, it was found that the GOCI-2 cloud mask not only contains cloud and fog information, but also includes sea ice and ice-interglacial waterway information. In order to rationally utilize the ice-interglacial waterway information in the GOCI-2 cloud mask, the present invention first removes the sea ice information in the cloud mask, and uses the clouds and ice-interglacial waterways as cloud and fog labels during the training of the sea ice monitoring model. Then, the category-corrected cross entropy loss function is introduced to reduce the loss between the seawater label and the cloud and fog label during model training, so that the model treats the two categories as approximate categories. Finally, based on the accuracy test, the use of the category-corrected cross entropy loss function improves the overall accuracy of the sea ice monitoring model, while improving the model's ability to recognize ice-interglacial waterways without affecting the model's ability to recognize sea ice.

[0040] It should be noted that the execution subject of each step of the method provided in the embodiment of the present application can be the same device, or the method can be executed by different devices. For example, the execution subject of step 110 and step 120 can be device 1, and the execution subject of step 130 can be device 2; for another example, the execution subject of step 110 can be device 1, and the execution subject of steps 120 and 130 can be device 2; and so on.

[0041] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] Therefore, the present application also proposes a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in any embodiment of the present application is implemented.

[0043] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0044] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0045] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0046] Therefore, an embodiment of the present application also proposes a computer product, including a computer program, which, when executed by a processor, implements the sea ice monitoring method based on geostationary satellite optical images as described in any embodiment of the present application.

[0047] Furthermore, the present application also proposes an electronic device (or computing device), comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any embodiment of the present application when executing the computer program.

[0048] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium. Computer-readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0049] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, the elements defined by the sentence "comprises..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0050] Those skilled in the art will appreciate that, unless expressly stated, the singular forms "a", "an", "said" and "the" used may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, modules and / or groups thereof. It should be understood that when a module is connected to another module, it may be directly connected to the other modules, or there may be intermediate modules. In addition, the connection here may include a wireless connection. The term "and / or" used in this application includes all or any unit and all combinations of one or more associated listed items.

[0051] Those skilled in the art will understand that unless otherwise defined, all terms (including technical, technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0052] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A sea ice monitoring method based on geostationary satellite optical images, characterized in that: The following steps are involved: Obtain GOCI-2 atmospherically corrected data; Based on the historical data obtained, land masks and sea ice annotations are performed, and a three-class sea ice training dataset is constructed based on the annotation results and GOCI-2 cloud masks, with labels of sea water, sea ice, and clouds; Generate a sea ice monitoring model based on Swin-unet, and use the constructed data set and the category-corrected cross entropy loss function to iteratively train the sea ice monitoring model to obtain the optimal model; The atmospherically corrected data of GOCI-2 were input and the sea ice monitoring results were obtained using the optimal model.

2. The sea ice monitoring method based on geostationary satellite optical images according to claim 1, characterized in that: By annotating images and reading GOCI-2 cloud masks, three types of true value labels are generated, namely sea water, sea ice, and clouds. Sea water and clouds are subcategories of background labels. Land masks and slices are performed on remote sensing images and annotated true value images to generate a sea ice training dataset for model training.

3. The sea ice monitoring method based on geostationary satellite optical images according to claim 1, characterized in that: Swin-unet's sea ice monitoring model is based on the following three parts: encoder, skip connection and decoder.

4. The sea ice monitoring method based on geostationary satellite optical images according to claim 1, characterized in that: The category-corrected cross entropy loss function is introduced, and the model loss is calculated using the dot product result of the prediction result and the category-corrected matrix.

5. The sea ice monitoring method based on geostationary satellite optical images according to claim 1, characterized in that: The encoder structure in Swin-unet's sea ice monitoring model includes region segmentation, linear embedding, 8 STBs, and 3 downsampling; Regional segmentation divides the remote sensing image slices into two equal parts in length and width and then superimposes them in the channel direction, making the original length, width and channel , , The image slice input becomes size; Increase the number of channels through linear embedding; Three downsamplings are alternately connected to four groups of STBs, each group of STBs contains two STBs, and both STBs contain layer normalization, window-based multi-head self-attention modules and multi-layer perceptrons. Among them, the second window-based multi-head self-attention module shifts the window by half the window length.

6. The sea ice monitoring method based on geostationary satellite optical images according to claim 1, characterized in that: The decoder in Swin-unet's sea ice monitoring model contains 4 upsampling, 6 STBs and linear projection; Every group of 2 STBs is used to fuse the scale features input from upsampling and jump connections; the first three upsampling processes expand the feature length and width to twice the original, and halve the number of channels. The last upsampling only doubles the length and width of the feature without changing the number of channels. Finally, the features are converted into three-classification results through linear projection, so that their length and width are the same as the original remote sensing image slices, and the number of channels is the same as the number of classifications. After post-processing, the sea water and cloud labels are integrated to obtain the sea ice monitoring results of a single channel.

7. A sea ice monitoring device based on geostationary satellite optical images, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The acquisition module is used to obtain GOCI-2 atmospheric correction data, annotate land masks and sea ice based on the acquired data, and construct a three-class sea ice training dataset based on the annotation results and GOCI-2 cloud masks, with labels of sea water, sea ice, and cloud fog; The generation module is used to generate a sea ice monitoring model based on Swin-unet, and iteratively train the sea ice monitoring model using the constructed data set and the category-corrected cross entropy loss function to obtain the optimal model; The determination module is used to input the atmospherically corrected data of GOCI-2 and obtain the sea ice monitoring results using the optimal model obtained through training.

8. A computer product, comprising a computer program, characterized in that When the computer program is executed by a processor, the sea ice monitoring method based on geostationary satellite optical images as described in 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 program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Images