Time sequence sea fog detection method based on multi-channel brightness temperature difference driving of static meteorological satellite

Abstract

The invention belongs to the technical field of computer vision, and particularly relates to a time sequence sea fog detection method based on stationary meteorological satellite multichannel brightness temperature difference driving, which comprises the following steps: acquiring meteorological satellite remote sensing data, and marking sea fog; preprocessing meteorological satellite remote sensing data according to the data source; extracting motion features by using the preprocessed meteorological satellite remote sensing data, and constructing a sea fog motion feature data set; inputting the model training set into a sea fog segmentation neural network model based on time sequence characteristics for training to obtain a training model; and preprocessing data of the test set, inputting the data into the training model for testing, and outputting a prediction result after segmentation, namely a final sea fog detection result. According to the method, the difficulty of motion feature extraction is reduced, the accuracy of sea fog monitoring tasks is improved, the consistency of continuous moment segmentation results is enhanced by comparing learning loss, and better generalization and performance are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite. Background Art

[0002] With the development of artificial intelligence technology and satellite technology in recent years, satellite-based sea fog monitoring is also closely related to people's lives. Therefore, the research on sea fog monitoring based on computer vision has an extremely important impact on my country's aerospace industry and the development of meteorological satellites. It can greatly save the time and resources of meteorological satellite professional researchers, and it is of great significance to conduct in-depth and detailed research on it.

[0003] In the existing solutions, deep learning technology is mainly used to build a neural network that imitates the human brain, and the internal laws and representation levels of sample data are learned through different convolutional layers, pooling layers, etc. Based on a large amount of satellite remote sensing image data and annotated labels as training sets, a semantic segmentation network is used for training, and finally the learning and discrimination of data is achieved. For example, Chinese patent CN110208880A discloses a sea fog detection method based on deep learning and satellite remote sensing technology, including: obtaining satellite remote sensing images, annotating sea fog in the images, and using the images annotated with sea fog as labels for deep learning segmentation network models; preprocessing the training set images to obtain input images that meet preset standards; based on the expanded data set, the deep learning segmentation network model is trained using the back propagation algorithm on the GPU, and after the training is completed, a model that has learned the image characteristics of sea fog is obtained. For example, the Chinese patent CN112287838A discloses an automatic cloud and fog identification method and system based on a stationary meteorological satellite image sequence, which obtains time-continuous stationary meteorological satellite image source data; arranges the stationary meteorological satellite image source data in chronological order to obtain time series image data, and uses a color-based image segmentation method to extract the cloud area and fog area of ​​each image in the stationary meteorological satellite image source data to obtain an initial mask; uses the Farneback optical flow method to generate an optical flow map of the time series image data; calculates the average optical flow of each pixel in the cloud area and fog area based on the optical flow map and the initial mask, and determines the cloud and fog classification threshold; classifies clouds and fog for each connected domain of the stationary meteorological satellite image source data based on the cloud and fog classification threshold to obtain an automatic cloud and fog identification result.

[0004] The existing solutions have the following technical problems:

[0005] (1) Difficulty in extracting motion features: Since the motion rate of sea fog is relatively low, ordinary optical flow algorithms have difficulty capturing the subtle motion characteristics of sea fog when the temporal resolution is relatively high. At the same time, the training of existing solutions only focuses on data at a single moment and ignores the motion information of sea fog, making motion feature extraction difficult.

[0006] (2) Poor consistency of segmentation results: In a time series data with high temporal resolution, the remote sensing images are very similar or even identical; however, there are great differences in the results produced by semantic segmentation, and the inconsistency of the results is very significant.

[0007] Satellite-based sea fog detection is already a common technical means and has many mature applications. However, time series information plays an important role in sea fog monitoring tasks. It contains physical information about the formation of sea fog and other characteristics. Existing sea fog monitoring methods or models cannot extract the subtle motion characteristics of sea fog well. In addition, the consistency of segmentation results in time series tasks also needs to be optimized. Therefore, how to extract time series information is the key to further improve the accuracy of sea fog monitoring. Summary of the invention

[0008] The present invention proposes a time-series sea fog detection method based on multi-channel brightness temperature difference drive of geostationary meteorological satellites, which uses multi-channel brightness temperature difference drive to extract information in the time series to detect sea fog, aiming to better utilize time series information to further improve the monitoring accuracy. Specifically, the present invention uses the brightness temperature difference data of the infrared channel at a single moment to establish a quantitative relationship with the water vapor content, increase the feature distinction between clouds and fog, and then use the optical flow algorithm to extract motion information in the brightness temperature difference data at adjacent moments, further strengthen the cloud and fog distinction features, and use contrast loss in the training model to construct positive and negative pairs of samples, increase the differences between samples, optimize the consistency of segmentation results, and finally realize the use of sea fog time series information, further improving the accuracy of sea fog monitoring.

[0009] The present invention introduces different motion properties of clouds and fog that change over time, which can obtain more accurate identification results, effectively improve the accuracy of cloud and fog identification, and has high universality. There are obvious differences in the motion characteristics of different types of clouds and sea fog. Taking continuous time sea fog samples as training sets and introducing motion features in the same sequence can improve the accuracy of sea fog detection.

[0010] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0011] The present invention provides a time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite, comprising the following steps:

[0012] Step S1: Obtain meteorological satellite remote sensing data and mark sea fog;

[0013] Step S2: preprocessing the meteorological satellite remote sensing data according to the data source;

[0014] Step S3: constructing a sea fog movement feature dataset using the preprocessed meteorological satellite remote sensing data;

[0015] Step S4: inputting the model training set into the sea fog segmentation neural network model based on time series characteristics for training to obtain a training model;

[0016] Step S5: pre-process the test set data, input it into the training model for testing, and output the segmented prediction result, i.e., the final sea fog detection result.

[0017] The preprocessing includes: (1) unification of spatial resolution; (2) unification of data format; and (3) dimensional processing.

[0018] Among them, the unification of spatial resolution refers to the normalization of the area represented by a pixel in different channels; the unification of data format refers to the unification of the data format into a data format available to the deep learning framework; the dimensionality processing refers to generating a three-dimensional three-channel pseudo-color image by channel transformation and numerical processing of high-channel data.

[0019] The specific implementation process of step S3 is as follows:

[0020] S3.1: Annotate and group the pre-processed meteorological satellite remote sensing data according to the data source time to generate multiple time series data;

[0021] S3.2: Based on the time series data at each moment, the brightness temperature difference data is generated by using the channel data for difference;

[0022] S3.3: Use the brightness temperature difference data and the optical flow algorithm to generate a sea fog motion feature dataset.

[0023] Furthermore, in S3.1, the preprocessed meteorological satellite remote sensing data are grouped according to the time label, the data in the same time series are grouped into one group, and the time resolution of the time series is normalized to ensure that the time resolution of each time series is equal and each group is greater than or equal to 20 meteorological satellite remote sensing data.

[0024] Furthermore, in S3.3, in the same time series, the optical flow algorithm is used to extract motion features from the generated brightness temperature difference data. The data processed by the optical flow algorithm is used to describe the motion information of image pixels in a time series in the field of video processing or computer vision. By analyzing the pixel motion between consecutive frames, the movement speed and direction of the subject in the image are determined, and a sea fog motion feature dataset is generated. The dataset is then integrated into the pseudo-color image through matrix splicing to generate a three-dimensional four-channel sea fog motion feature dataset as a model training set.

[0025] Among them, the sea fog segmentation neural network model based on time series characteristics uses MiTB3 as the backbone network, adopts Mix Transformer as the encoder, and designs contrastive learning loss to optimize the consistency of time series data segmentation results, and finally outputs the segmentation results.

[0026] Furthermore, the specific calculation formula of the contrastive learning loss is as follows:

[0027]

[0028] Among them, x represents the query fog area characteristics, x + With x - represents the positive and negative sample pairs, τ represents the temperature hyperparameter, N represents the number of sample pairs, and Represents the nth positive and negative sample pair.

[0029] Furthermore, in step S4, the sea fog segmentation loss is calculated using binary cross entropy loss and hinge loss:

[0030]

[0031] Among them, y and Represent the true labels and the results predicted by the model respectively.

[0032] The time-series sea fog detection system based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite provided by the present invention is used to implement the time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite, and the system comprises:

[0033] Input module: used to obtain sufficient amount of meteorological satellite remote sensing data that meets the time resolution requirements;

[0034] Data preprocessing module: annotate sea fog in meteorological satellite remote sensing data, unify spatial resolution, unify data format, process dimensions, and generate pseudo-color images through channel processing;

[0035] Sea fog motion feature extraction module: mainly used to construct a sea fog motion feature dataset using pre-processed meteorological satellite remote sensing data;

[0036] Sea fog segmentation neural network model module based on time series characteristics: used to input the model training set into the sea fog segmentation neural network model based on time series characteristics for training to obtain a training model;

[0037] Output module: Use the test set data and the trained model to output the predicted data.

[0038] The beneficial effects of the present invention are:

[0039] (1) The difficulty of motion feature extraction is reduced and the accuracy of sea fog monitoring tasks is improved: The present invention can better extract the motion information of sea fog and further improve the recognition accuracy in sea fog monitoring tasks.

[0040] (2) High consistency of segmentation results: The present invention uses contrastive learning loss to enhance the consistency of segmentation results at consecutive moments, thus achieving better generalization and performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention provides a flow chart of a time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite.

[0042] Figure 2 This is a flow chart of the optical flow algorithm in the present invention.

[0043] Figure 3 This is the structural diagram of the neural network model for sea fog segmentation based on temporal characteristics.

[0044] Figure 4 The present invention provides a structural composition block diagram of a time-series sea fog detection system based on multi-channel brightness temperature difference drive of a geostationary meteorological satellite. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0046] In the present invention, it mainly involves a deep learning method for extracting time information based on time series images. Its main technologies include ① brightness temperature difference, that is, the difference operation of different channel data of satellite remote sensing images. At the physical level, the influence of factors such as surface temperature and aerosol can be ignored, and a quantitative relationship between brightness temperature difference and water vapor content is established. The water vapor content between clouds and fogs is very different, so as to increase the feature representation between clouds and fogs. Since the fog area in the infrared band is close to the background, the edge of the fog area can be blurred well in the background, while increasing the feature distinction between clouds and fog, the ambiguous features between the fog edge and the inside of the cloud are reduced. ② Optical flow algorithm is the motion mode or change mode of pixel points in the image in time. It is used to describe the motion information of image pixel points in a time sequence in the field of video processing or computer vision. By analyzing the pixel motion between consecutive frames, the motion speed and direction of the subject in the image are determined. The brightness temperature difference data at different times is used as the input of the optical flow algorithm, which can extract the motion information of the target in the remote sensing image, increase the difference between clouds and fog, and construct a training set as input. ③ Contrastive learning is an effective method for learning two different categories. It achieves better segmentation results by learning the differences between samples of different categories and the similarities between samples of the same category. It constructs positive and negative pairs of samples in the samples to increase the differences between samples. Using contrastive learning loss in the training of a time series can reduce the differences between foggy areas, thereby achieving the effect of optimizing the consistency of time series data segmentation results.

[0047] In a first aspect, the present invention provides a time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite.

[0048] like Figure 1 As shown, the present invention provides a time-series sea fog detection method based on multi-channel brightness temperature difference driven by geostationary meteorological satellites, which uses optical flow algorithm and contrast learning to extract the motion characteristics of sea fog, completes sea fog detection based on meteorological satellite remote sensing data, and further improves the accuracy of sea fog detection. The specific implementation process is as follows:

[0049] Step S1: data acquisition and annotation;

[0050] Obtain sufficient meteorological satellite remote sensing data that meets the temporal resolution requirements, and mark the sea fog in the meteorological satellite remote sensing data;

[0051] Step S2: data preprocessing;

[0052] The acquired meteorological satellite remote sensing data is preprocessed according to the data source, mainly including: unification of spatial resolution, unification of data format (processing into a data format available for deep learning framework), dimensionality processing and generating pseudo-color images through channel processing.

[0053] Specifically, the original format of meteorological satellite remote sensing data is a three-dimensional array. Taking the Fengyun-4A satellite as an example, it contains 14 channels of raw data, which are marked as satellite data of the same size of 1 channel. Spatial resolution unification refers to the normalization of the area represented by a pixel in different channels; data format unification refers to the unification of the data format into a data format that can be used by the deep learning framework; dimensional processing refers to the generation of a three-dimensional three-channel pseudo-color image by channel transformation and numerical processing of high-channel data. Preprocessing of meteorological satellite remote sensing data ensures the uniformity, integrity and accuracy of the data in the subsequent data training process.

[0054] Step S3: constructing a sea fog movement feature dataset using the preprocessed meteorological satellite remote sensing data;

[0055] S3.1: Annotate and group the pre-processed meteorological satellite remote sensing data according to the data source time to generate multiple time series data;

[0056] Specifically, the preprocessed meteorological satellite remote sensing data are grouped according to time labels, the data in the same time series are grouped together, and the time resolution of the time series is normalized to ensure that the time resolution of each time series is equal and each group is greater than or equal to 20 meteorological satellite remote sensing data.

[0057] S3.2: Based on the time series data at each moment, the brightness temperature difference data is generated by using the channel data for difference;

[0058] Specifically, based on the meteorological satellite remote sensing data at each individual moment, taking the Fengyun-4A satellite as an example, Figure 2 As shown in the figure, the brightness temperature difference data is generated by subtracting the channel data of the infrared channel 11 microns and 12 microns (channels 12 and 13). This method can approximately ignore the influence of factors such as surface temperature and aerosols, thereby establishing a quantitative relationship with the water vapor content, blurring the edge of the sea fog area to the background, increasing the distinction between cloud and fog features and reducing the ambiguous features between the fog edge and the cloud interior, and ensuring constant brightness, satisfying the conditions for calculating the motion vector using the optical flow algorithm.

[0059] S3.3: Using the brightness temperature difference data and processing it with the optical flow algorithm, the final sea fog motion feature dataset is generated;

[0060] Specifically, in the same time series, the optical flow algorithm is used to extract motion features from the generated brightness temperature difference data. The data processed by the optical flow algorithm is used to describe the motion information of image pixels in a time series in the field of video processing or computer vision. By analyzing the pixel motion between consecutive frames, the movement speed and direction of the subject in the image are determined to generate the final sea fog motion feature dataset. The features between clouds and fog are obviously different and are integrated into the pseudo-color image through matrix splicing to generate a three-dimensional four-channel sea fog motion feature dataset as a model training set.

[0061] Different from the traditional optical flow algorithm, the present invention increases the difference between sea fog and clouds through the brightness temperature difference, and can better meet the data source requirements of the optical flow algorithm. While hiding the fog in the background, it further solves the ambiguity problem of the similarity between the overall cloud and the edge of the fog, and extracts the motion characteristics of the sea fog well.

[0062] Step S4: inputting the model training set into the sea fog segmentation neural network model based on time series characteristics for training, and obtaining the training model after sufficient training;

[0063] The sea fog segmentation neural network model based on time series characteristics adopted in the present invention uses MiTB3 as the backbone network, Mix Transformer (MiT) as the encoder, and designs contrastive learning loss to optimize the consistency of time series data segmentation results, and finally outputs the segmentation results.

[0064] In the present invention, the main function of the sea fog segmentation neural network model based on time series characteristics is to extract and decode the overall remote sensing data features, thereby realizing the detection of sea fog.

[0065] like Figure 3 As shown, the specific implementation process is as follows:

[0066] S4.1: Mix Transformer (MiT) is used as the encoder. MiT takes meteorological satellite remote sensing data integrated with prior physical information as input and outputs spatial feature maps of different resolutions, namely multi-level and multi-scale feature maps similar to CNN. These feature maps contain detailed information from high-resolution to semantic information at low resolution. The performance of semantic segmentation tasks is improved by combining different levels of resolution.

[0067] S4.2: Use contrastive learning loss to construct more positive pairs between foggy areas and foggy areas of consecutive time points in the same time series by constructing positive and negative pairs, improve the similarity of foggy areas, and optimize the consistency of timing problems. The specific calculation formula of contrastive learning loss is as follows:

[0068]

[0069] Among them, x represents the query fog area characteristics, x + With x - represents the positive and negative sample pairs, τ represents the temperature hyperparameter, N represents the number of sample pairs, and represents the nth positive and negative sample pair; the final contrastive learning loss is calculated for all images in the batch (batch-size) during the training process. The present invention optimizes the consistency of time series data segmentation results by introducing contrastive learning loss in the sea fog monitoring task.

[0070] S4.3: In order to calculate the sea fog segmentation loss, binary cross entropy (BCE) loss and hinge loss are introduced. The sea fog segmentation loss is calculated by binary cross entropy (BCE) loss and hinge loss. The specific calculation formula is as follows:

[0071]

[0072]

[0073] Among them, y and Represent the true labels and the results predicted by the model respectively.

[0074] S4.4: The sea fog segmentation neural network model based on time series characteristics generates a trained model through training data sets, data annotation and GPU training.

[0075] The MLP layer and decoder restore the low-dimensional vector to high-dimensional image data through a series of deconvolution layer operations, thereby outputting the deep learning network sea fog monitoring results.

[0076] Step S5: preprocess the test set data according to the data preprocessing method of step S2, and then input it into the training model for testing, and output the segmented prediction result, that is, the final sea fog detection result, in the form of a labeled file of the sea fog in the specified area.

[0077] The present invention analyzes the physical characteristics of sea fog and utilizes brightness temperature difference data to better extract the subtle movement characteristics of sea fog and achieve a more ideal sea fog differentiation effect.

[0078] In the second aspect, the present invention provides a time-series sea fog detection system based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite, which is used to implement a time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite provided by the present invention.

[0079] like Figure 4 As shown, the present invention provides a time-series sea fog detection system based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite, which specifically includes the following modules:

[0080] Input module, data preprocessing module, sea fog motion feature extraction module, sea fog segmentation neural network model module based on time series characteristics and output module.

[0081] The functions and roles of each module are as follows:

[0082] Input module: used to obtain sufficient amount of meteorological satellite remote sensing data that meets the time resolution requirements;

[0083] Data preprocessing module: annotate sea fog in meteorological satellite remote sensing data, unify spatial resolution, unify data format, process dimensions, and generate pseudo-color images through channel processing;

[0084] Sea fog motion feature extraction module: mainly used to construct a sea fog motion feature data set using pre-processed meteorological satellite remote sensing data; specifically, the pre-processed meteorological satellite remote sensing data is annotated and grouped according to the data source time to generate multiple time series data; based on the time series data at each moment, the channel data is used to generate brightness temperature difference data; the brightness temperature difference data is processed by the optical flow algorithm to generate the final sea fog motion feature data set;

[0085] Sea fog segmentation neural network model module based on time series characteristics: mainly used to input the model training set into the sea fog segmentation neural network model based on time series characteristics for training, and obtain the training model after sufficient training; the sea fog segmentation neural network model based on time series characteristics adopted by the present invention uses MiTB3 as the backbone network, Mix Transformer (MiT) as the encoder, and designs contrastive learning loss to optimize the consistency of time series data segmentation results, and finally outputs the segmentation results;

[0086] Output module: Use the test set data and the trained model to output the predicted data.

[0087] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A time-series sea fog detection method based on multi-channel brightness temperature difference driven by geostationary meteorological satellites, characterized in that: The following steps are involved: Step S1: Obtain meteorological satellite remote sensing data and mark sea fog; Step S2: preprocessing the meteorological satellite remote sensing data according to the data source; Step S3: constructing a sea fog movement feature dataset using the preprocessed meteorological satellite remote sensing data; Step S4: inputting the model training set into the sea fog segmentation neural network model based on time series characteristics for training to obtain a training model; Step S5: pre-process the test set data, input it into the training model for testing, and output the segmented prediction result, i.e., the final sea fog detection result.

2. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 1 is characterized in that: The preprocessing includes: (1) unification of spatial resolution; (2) unification of data format; and (3) dimensional processing.

3. The time-series sea fog detection method based on multi-channel brightness temperature difference driving of geostationary meteorological satellite according to claim 2 is characterized in that: The unification of spatial resolution refers to the normalization of the area represented by a pixel in different channels; the unification of data format refers to the unification of the data format into a data format available to the deep learning framework; the dimensionality processing refers to generating a three-dimensional three-channel pseudo-color image by channel transformation and numerical processing of high-channel data.

4. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 1 is characterized in that: The specific implementation process of step S3 is as follows: S3.1: Annotate and group the pre-processed meteorological satellite remote sensing data according to the data source time to generate multiple time series data; S3.2: Based on the time series data at each moment, the brightness temperature difference data is generated by using the channel data for difference; S3.3: Use the brightness temperature difference data and the optical flow algorithm to generate a sea fog motion feature dataset.

5. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 4 is characterized in that: In S3.1, the preprocessed meteorological satellite remote sensing data are grouped according to time labels, the data in the same time series are grouped into one group, and the time resolution of the time series is normalized to ensure that the time resolution of each time series is equal and each group is greater than or equal to 20 meteorological satellite remote sensing data.

6. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 4 is characterized in that: In S3.3, in the same time series, the optical flow algorithm is used to extract motion features from the generated brightness temperature difference data. The data processed by the optical flow algorithm is used to describe the motion information of image pixels in a time series in the field of video processing or computer vision. By analyzing the pixel motion between consecutive frames, the movement speed and direction of the subject in the image are determined, and a sea fog motion feature dataset is generated. It is then integrated into the pseudo-color image through matrix splicing to generate a three-dimensional four-channel sea fog motion feature dataset as a model training set.

7. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 1 is characterized in that: The sea fog segmentation neural network model based on time series characteristics uses MiTB3 as the backbone network, adopts MixTransformer as the encoder, and designs contrastive learning loss to optimize the consistency of time series data segmentation results, and finally outputs the segmentation results.

8. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 7 is characterized in that: The specific calculation formula of the contrastive learning loss is as follows: Among them, x represents the query fog area characteristics, x + With x - represents the positive and negative sample pairs, τ represents the temperature hyperparameter, N represents the number of sample pairs, and Represents the nth positive and negative sample pair.

9. The method for detecting sea fog based on the time series of multi-channel brightness temperature difference driven by geostationary meteorological satellite according to claim 8 is characterized in that: In step S4, the sea fog segmentation loss is calculated using binary cross entropy loss and hinge loss: Among them, y and Represent the true labels and the results predicted by the model respectively.

10. A time-series sea fog detection system based on multi-channel brightness temperature difference driven by geostationary meteorological satellites, characterized in that: The system is used to implement the time-series sea fog detection method based on multi-channel brightness temperature difference driving of a geostationary meteorological satellite as described in any one of claims 1 to 9, and comprises: Input module: used to obtain sufficient amount of meteorological satellite remote sensing data that meets the time resolution requirements; Data preprocessing module: annotate sea fog in meteorological satellite remote sensing data, unify spatial resolution, unify data format, process dimensions, and generate pseudo-color images through channel processing; Sea fog motion feature extraction module: mainly used to construct a sea fog motion feature dataset using pre-processed meteorological satellite remote sensing data; Sea fog segmentation neural network model module based on time series characteristics: used to input the model training set into the sea fog segmentation neural network model based on time series characteristics for training to obtain a training model; Output module: Use the test set data and the trained model to output the predicted data.

Images