Ocean near-surface sub-mesoscale process flow field inversion method and device

Abstract

The invention provides an ocean near-surface sub-mesoscale process flow field inversion method and device. The method provided by the invention comprises the following steps: screening an ocean near-surface sub-mesoscale representative sea area; obtaining the structure image information of the frontal surface and the filament of the phytoplankton in the marine sub-mesoscale process; acquiring displacement velocity field information at the same position; the structure image information and the displacement velocity field information are matched, and a first training sample set is constructed; correcting the first training sample set based on the second data set and the third data set to obtain a second training sample set; training the initial model by using the second training sample set to obtain a flow field inversion model; obtaining a plurality of sub-mesoscale sea surface water color images of the sea area to be detected; and processing the sub-mesoscale sea surface water color image by using the flow field inversion model to obtain a displacement velocity field. The invention provides an ocean near-surface sub-mesoscale process flow field inversion method and device. The method and device are used for solving the problem that a current ocean flow field inversion algorithm is not high in sub-mesoscale process inversion precision.

Description

Technical Field

[0001] This application relates to the technical field of ocean dynamic process identification, and particularly to a method and device for inverting the flow field of near-surface submesoscale processes in the ocean. Background Art

[0002] Submesoscale processes play a crucial role in ocean energy dissipation and are the key link connecting large-scale balanced dynamics and small-scale unbalanced turbulence. Among them, the development of observation technologies is very important for studying submesoscale motions. The new generation of altimetry satellite data developed in the past, such as surface water and ocean topography, has improved the spatial resolution of ocean dynamic field data to about 2 kilometers, bringing new opportunities for the study of ocean submesoscale processes. However, its revisit cycle of about 20 days cannot capture the rapid changes in submesoscale motions on the hourly to daily time scales. Therefore, traditional satellite-based altimetry data is difficult to capture the fine-scale structure of submesoscale processes and cannot meet the needs of more in-depth research on submesoscale motions. Summary of the Invention

[0003] In view of this, this application provides a method and device for inverting the flow field of near-surface submesoscale processes in the ocean to solve the problem of low inversion accuracy of the current ocean flow field inversion algorithm for submesoscale processes.

[0004] Specifically, this application is implemented through the following technical solutions:

[0005] The first aspect of this application provides a method for inverting the flow field of near-surface submesoscale processes in the ocean, and the method includes:

[0006] Screen representative sea areas of near-surface submesoscale processes from ocean regions;

[0007] Use a geostationary ocean color imaging satellite to obtain structural image information of phytoplankton fronts and filaments of submesoscale processes in the representative sea areas;

[0008] Use a first database to obtain displacement velocity field information of ocean submesoscale processes at the same location;

[0009] Match the structural image information and the displacement velocity field information to construct a first training sample set;

[0010] Correct the first training sample set based on a second data set and a third data set to obtain a second training sample set; wherein, the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of ocean large-scale processes;

[0011] Use the second training sample set to train an initial model to obtain a flow field inversion model;

[0012] Use the geostationary ocean color imaging satellite to observe the sea area to be detected, and obtain multiple sub-meso-scale sea surface color images of the sea area to be detected;

[0013] Use the flow field inversion model to process the sub-meso-scale sea surface color image, and obtain the displacement velocity field of the surface sub-meso-scale process in the sea area to be detected.

[0014] The second aspect of the present application provides a device for inverting the flow field of the ocean near-surface sub-meso-scale process. The device includes an acquisition module, a construction module, and a processing module; wherein,

[0015] The acquisition module is used to screen the representative sea areas of the ocean near-surface sub-meso-scale from the ocean area;

[0016] The acquisition module is further used to use the geostationary ocean color imaging satellite to obtain the structural image information of the phytoplankton front and filaments of the sub-meso-scale process in the representative sea area;

[0017] The acquisition module is further used to use the first database to obtain the displacement velocity field information of the ocean sub-meso-scale process at the same location;

[0018] The construction module is used to match the structural image information and the displacement velocity field information, and construct a first training sample set;

[0019] The construction module is further used to correct the first training sample set based on the second data set and the third data set to obtain a second training sample set; wherein, the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of the ocean large-scale process;

[0020] The construction module is further used to use the second training sample set to train the initial model to obtain a flow field inversion model;

[0021] The processing module is used to use the geostationary ocean color imaging satellite to observe the sea area to be detected, and obtain multiple sub-meso-scale sea surface color images of the sea area to be detected;

[0022] The processing module is further used to use the flow field inversion model to process the sub-meso-scale sea surface color image, and obtain the displacement velocity field of the surface sub-meso-scale process in the sea area to be detected.

[0023] The method and device for inverting the flow field of the ocean near-surface submesoscale process provided by this application, in order to improve the model's ability to invert the flow field of the submesoscale process, first screen representative sea areas, select representative sea areas where both the identification target and the interference background environment exist simultaneously, to improve the pertinence and representativeness of the samples; secondly, based on the samples composed of ocean data obtained from the representative sea areas, the structural image information of phytoplankton fronts and filaments and the displacement velocity field information at the same position are obtained. Secondly, the first training sample set composed of representative sea area data is corrected using the second data set containing sea surface height data and the third data set containing geostrophic flow field displacement velocity field data of the ocean large-scale process. On the one hand, the fineness of the information increases, enabling the model to learn more details of the ocean submesoscale process; on the other hand, the large-scale information of the second and third data set samples corrects the submesoscale information of the first training sample set in the same dimension, enabling the model to learn more accurate feature information, thereby improving the fineness and accuracy of the model prediction, realizing the multi-scale information fusion from the submesoscale to the ocean large-scale, and helping to deeply understand the behavioral characteristics of the ocean submesoscale process in different scale environments. Further, the initial model is trained using the second training sample set constructed based on multi-source data, greatly optimizing the training data of the model. The corrected training sample set combines the advantages of multiple data, reduces the errors and uncertainties that may exist in a single data source, and improves the accuracy of the relationship between the submesoscale sea surface color image and the displacement velocity field learned by the flow field inversion model. Using the trained flow field inversion model to process the submesoscale sea surface color image of the sea area to be detected observed by the geostationary ocean color imaging satellite, the displacement velocity field of the surface submesoscale process of this sea area can be obtained. This process realizes the rapid and accurate conversion from satellite observation images to flow field information, providing strong data support for oceanographic research, ocean environment monitoring, ocean resource management, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flowchart of the first embodiment of the method for inverting the flow field of the ocean near-surface submesoscale process provided by this application; Figure 2 It is a comparison diagram of the effects of the flow field inversion model and the traditional model at the same time shown in the exemplary embodiment of this application; Figure 3 It is a schematic structural diagram of the first embodiment of the device for inverting the flow field of the ocean near-surface submesoscale process provided by this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application.

[0026] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit the present application. The singular forms "a", "the", and "said" used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0027] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to a determination".

[0028] Specific embodiments are given below to introduce the technical solutions of the present application in detail.

[0029] Figure 1 This is a flowchart of the first embodiment of the method for inverting the flow field of ocean near-surface submesoscale processes provided by the present application. Please refer to Figure 1 , the method provided in this embodiment may include:

[0030] S101. Screen representative sea areas of ocean near-surface submesoscale from the ocean area.

[0031] Specifically, the submesoscale process is a kind of movement in the ocean, and its horizontal scale is between 0.1 - 10 km. This scale range makes the submesoscale process occupy a unique spatial position in the ocean, different from both larger-scale ocean circulations and smaller-scale turbulent motions. Vertically, it changes at a speed of 100 m / day. This relatively fast vertical change speed plays a key role in the vertical exchange of matter and energy in the ocean. It can promote the transport of nutrients between different water layers, affect the growth and distribution of marine organisms such as phytoplankton, and thus have a chain reaction on the entire marine ecosystem.

[0032] Furthermore, the submesoscale process is a key link connecting the large-scale balanced dynamics and the small-scale unbalanced turbulence. At larger scales, it is influenced by balanced dynamics such as ocean circulation. At the same time, through its own movements and interactions, it transfers energy and matter to smaller-scale turbulent motions, promoting the cascade transfer of ocean energy. Phytoplankton in the submesoscale process often form filamentous structures, and the diurnal variation of horizontal transport is significant. Such distribution and movement changes of phytoplankton not only affect their own survival and reproduction but also have an impact on the entire ocean food web.

[0033] Traditional submesoscale flow field inversion models cannot capture the rapid changes of submesoscale motions on the hourly to daily time scales, making it difficult to obtain complete information on the fine-scale structure of submesoscale processes. When dealing with complex filamentous structures (such as submesoscale vortices), it is difficult to accurately match, resulting in large velocity errors and angular errors, which affect the accurate identification and analysis of submesoscale processes.

[0034] Given these deficiencies of traditional submesoscale flow field inversion models, it is necessary to construct special samples. By screening representative submesoscale sea areas near the ocean surface, obtaining the structural image information of phytoplankton fronts and filaments in this sea area and the displacement velocity field information at the same location, and matching the two to construct the first training sample set, the physical relationship between the distribution of substances (phytoplankton) and the flow field in the ocean can be accurately captured. Using the second dataset (sea surface height data) and the third dataset (geostrophic flow field displacement velocity field data of large-scale ocean processes) to correct the first training sample set to obtain the second training sample set, realizing the multi-scale information fusion from submesoscale to large-scale ocean. The special samples constructed in this way can provide more comprehensive and accurate data for model training, helping the model learn more detailed submesoscale process details and more accurate feature information, thereby improving the model's ability to invert the submesoscale process flow field, making up for the deficiencies of traditional models, and meeting the needs of more in-depth research on submesoscale motions.

[0035] Specifically, representative submesoscale sea areas near the ocean surface are screened from ocean regions, including:

[0036] (1) Obtain the motion conditions of the submesoscale process in each ocean region.

[0037] Specifically, the motion conditions of the submesoscale process in ocean regions include rich mesoscale vortices, submesoscale vortices, phytoplankton fronts, filaments and other features. These motion conditions can provide more flow field change information in the study of submesoscale ocean process flow field inversion, enabling the model to fully learn the physical properties of submesoscale motions, thereby improving the analytical ability of submesoscale flow fields.

[0038] Furthermore, the movement of sub - mesoscale processes in each ocean region can be obtained through satellites, such as geostationary ocean color satellites and altimetry satellites.

[0039] (2)Obtain the characteristic differences between phytoplankton and filaments and the environment in each ocean region.

[0040] Specifically, after obtaining the movement of sub - mesoscale processes in each ocean region, the geostationary ocean color satellite (GOCI - II with a spatial resolution of 250m and 10 observations per day) can be used to obtain the ocean surface chlorophyll a (Chl - a) concentration data. The Chl - a concentration is an important indicator of phytoplankton biomass. By analyzing the distribution and changes of Chl - a concentration in different ocean regions, the abundance differences of phytoplankton can be understood. At the same time, ocean color images can show the aggregation patterns of phytoplankton, such as the areas where algal blooms form and the positions of phytoplankton fronts. These information helps to preliminarily judge the distribution characteristics of phytoplankton in different regions and its relationship with the environment. According to factors such as ocean geography, climate, and ecology, each ocean region is classified, such as the open ocean area, coastal area, upwelling area, tropical sea area, temperate sea area, frigid sea area, etc. By comparing the characteristic differences between phytoplankton and filaments in different types of regions, the relationship between them and the regional environmental characteristics (such as ocean circulation patterns, water temperature distribution, nutrient supply mechanisms, etc.) is analyzed.

[0041] Furthermore, based on the above - mentioned analysis results, a comprehensive evaluation model is constructed to incorporate the characteristics of phytoplankton and filaments and environmental factors into a unified framework. The model can be a statistical model (such as multiple regression model, structural equation model, etc.) or a mechanism model based on physical - ecological processes, which is used to quantitatively predict the changes of phytoplankton and filament characteristics under different environmental conditions. Through model simulation, the characteristic differences between phytoplankton and filaments and the environment can be further verified.

[0042] Furthermore, obtaining the characteristic differences between phytoplankton and filaments and the environment in each ocean region can also include the following steps:

[0043] 2.1. Obtain multiple ocean color images of each ocean region, and each ocean color image includes phytoplankton, filaments, and the ocean region background environment at the same time.

[0044] Specifically, the water color image can display information such as the aggregation pattern of phytoplankton, the area where algal blooms form, and the position of the phytoplankton front. Each water color image contains both phytoplankton and filaments, as well as the background environment of the ocean area. By obtaining multiple water color images of different ocean areas, the relationship between phytoplankton, filaments and the background environment in different areas can be observed comprehensively and multi-dimensionally. The water color images of each ocean area can be obtained by a geostationary ocean color imager satellite.

[0045] 2.2. Identify the first boundaries of the phytoplankton and filaments from each of the water color images.

[0046] Specifically, through image recognition technology, such as object detection algorithms based on deep learning, a model capable of accurately identifying phytoplankton and filaments can be trained by learning a large number of known sample images of phytoplankton and filaments. This model can analyze features such as the color, texture, and shape of pixels in the water color image, separate the phytoplankton and filaments from the background environment, and then determine their boundaries to obtain the first boundaries. For another example, a convolutional neural network (CNN) can also be used. It extracts image features through multiple convolutional layers and pooling layers, and then determines whether a pixel belongs to phytoplankton and filaments through a classifier, and finally outlines their boundaries to obtain the first boundaries.

[0047] 2.3. Identify the second boundaries of the background environment from each of the water color images.

[0048] Specifically, after identifying the first boundaries of the phytoplankton and filaments in the water color image, the part of the water color image other than the phytoplankton and filaments is the background environment of the ocean area. Using an edge detection algorithm in image processing, such as the Canny edge detection algorithm, which calculates the gradient intensity and direction of pixels in the image to find the places where the gray value changes significantly in the image, thereby determining the boundaries of the background environment to obtain the second boundaries. It is also possible to use a threshold segmentation method. According to the gray scale characteristics of the image, a suitable threshold is set to distinguish the background environment from the phytoplankton and filaments, obtain the boundaries of the background environment, and then determine these boundaries as the second boundaries of the background environment.

[0049] 2.4. Calculate the overlapping area after the intersection of the first boundaries and the second boundaries in each of the water color images.

[0050] Specifically, after obtaining the first boundary of phytoplankton and filaments and the second boundary of the background environment in each water color image, the area calculation method in graphics is used to represent the boundaries as polygons, and the overlapping area is obtained by calculating the area of the intersection of the two polygons. For example, the vector cross product method can be used to calculate the area of a polygon. For two intersecting polygons (i.e., the polygons formed by the first boundary and the second boundary), find their intersection region and then calculate the area of the intersection region. This overlapping area to a certain extent reflects the overlapping degree between phytoplankton and filaments and the background environment, providing a data basis for calculating the feature differences in the subsequent steps.

[0051] 2.5. Calculate the weighted sum of the overlapping areas of all water color images to obtain the total overlapping area, and use the total overlapping area as the feature difference.

[0052] Specifically, since the quality and reliability of different water color images are different, in order to calculate the feature differences more accurately, it is necessary to perform weighted processing on the overlapping area of each water color image. The size of the weight is positively correlated with the recognition accuracy of the boundary. Accurately identifying the boundaries of phytoplankton, filaments and the background environment is crucial for calculating the overlapping area after their intersection. If the boundary recognition is inaccurate, the calculation of the overlapping area will deviate, resulting in unreliable calculation results of the feature differences. For example, if there is a misjudgment when identifying the first boundary of phytoplankton and filaments, and parts that do not belong to phytoplankton are included within the boundary, it will make the calculation of the overlapping area too large, thus affecting the judgment of the feature differences. Therefore, the higher the boundary recognition accuracy, the greater the weight of the corresponding water color image when calculating the feature differences.

[0053] Furthermore, the weight can also be determined by identifying indicators such as the brightness, contrast, exposure, and sharpness of the water color image to evaluate whether the water color image is clearly recognizable. These indicators reflect the quality of the water color image. The better the quality of the water color image, the higher the accuracy of boundary recognition. Compare these indicators with the standard indicators, calculate the ratio between them, and use this ratio as the weight corresponding to the water color image. The specific weight determination indicators are set according to actual needs and are not restricted in this embodiment.

[0054] Furthermore, after obtaining the weight of each water color image, multiply the overlapping area of each water color image by its corresponding weight and then sum them up to obtain the total overlapping area. The total overlapping area comprehensively considers the information of multiple water color images and can more comprehensively and accurately reflect the feature differences between phytoplankton, filaments and the environment, providing a key basis for screening representative sea areas of near-surface sub-mesoscale in the ocean.

[0055] (3)Screen each of the ocean regions according to the feature differences to obtain the first screening result of the representative sea areas.

[0056] Specifically, the characteristic differences refer to the characteristic differences between phytoplankton and filaments in the sea area and the environment. After determining the characteristic differences of each sea area, the sea areas with characteristic differences greater than the characteristic difference threshold are retained. The representative sea areas in the first screening result include the sea areas with characteristic differences greater than the characteristic difference threshold in each sea area. It should be noted that the characteristic difference threshold is set according to actual needs and is not limited in this embodiment. It can be understood that a large characteristic difference means that the difference between phytoplankton and filaments and the environment is more significant, the characteristics of submesoscale processes in these sea areas are more obvious, and the information contained in the sea areas with large characteristic differences is more diverse. When training the model, the data of these sea areas can provide more valuable learning materials for the model, enabling the model to capture the relationship between more complex and subtle submesoscale processes and environmental factors. Retaining the sea areas with large characteristic differences can avoid overfitting of the model to a specific local area. The marine environment is complex and diverse, and the submesoscale processes in different regions are affected by a variety of factors. If only the sea areas with small characteristic differences are selected for training, the model may only learn the characteristics of submesoscale processes in some similar environments and have poor generalization ability when facing other different environments. However, the sea areas with large characteristic differences cover a wider range of marine environmental conditions. Selecting these sea areas for training can enable the model to learn the commonalities and differences of submesoscale processes in various marine environments, enhance its adaptability to different marine environments, and improve the reliability of the model in performing flow field inversion in different sea areas.

[0057] (4) Determine the distribution characteristics of each representative sea area in the first screening result of the representative sea areas, select the first screening result of the representative sea areas according to the distribution characteristics to obtain the second screening result, and determine the representative sea areas according to the second screening result.

[0058] Specifically, in this embodiment, key attention is paid to the sea areas where active submesoscale vortices, frontal activities, and the distribution of phytoplankton show obvious submesoscale characteristics, which can provide obvious characteristic information and enable the model to better learn the relationship between frontal activities and the displacement velocity field of phytoplankton and submesoscale processes.

[0059] Furthermore, analyze the distribution characteristics of each representative sea area in the first screening result of representative sea areas, such as geographical location distribution, marine area type distribution, and the influence distribution of ocean circulation systems, etc. When determining the geographical location distribution of each representative sea area, determine the latitude position of each representative sea area on the earth and clarify whether it is in the tropics, temperate zone or frigid zone. The sea areas in different latitude zones have significant differences in light, temperature and climate conditions, and these factors will greatly affect the marine ecosystem and marine dynamic processes; when determining the marine area type distribution of each representative sea area, check which type the representative sea area belongs to, such as the oceanic region, the coastal area, the upwelling area, etc. Because there are differences in the ocean circulation patterns, water temperature distributions, and nutrient supply mechanisms in different types of regions. For example, the coastal area is greatly affected by the land, rich in nutrients, with rapid growth of phytoplankton, and the submesoscale processes may be more complex; the upwelling area will bring deep-sea nutrients to the surface, promoting the massive reproduction of phytoplankton, and its submesoscale flow field also has unique characteristics; when determining the influence distribution of ocean circulation systems on each representative sea area, it is necessary to study which ocean circulation system the representative sea area is affected by. Large ocean circulations such as the Kuroshio and the Gulf Stream will have a profound impact on the water temperature, salinity and flow field of the surrounding sea areas, and thus affect the submesoscale processes. The material transport and energy transfer methods in the submesoscale processes in the sea areas affected by different circulation systems are different, and the distribution and movement of phytoplankton will also vary.

[0060] Furthermore, to enable the model to learn the flow field characteristics of submesoscale processes in various marine environments, select areas as evenly as possible in the first screening result of representative sea areas. Cover sea areas in different oceans, such as the Pacific Ocean, the Atlantic Ocean, the Indian Ocean, etc., to ensure that the model can learn the characteristics of submesoscale processes in different ocean environments; select sea areas in different latitude zones, including the tropics, temperate zones and frigid zones, to make the model adapt to submesoscale processes under different temperature and light conditions; select areas affected by different ocean circulation systems, such as sea areas affected by warm currents and cold currents, to let the model understand the action mechanisms of different circulations on submesoscale processes, and focus on sea areas with active submesoscale vortices, frontal activities and obvious submesoscale characteristics in the distribution of phytoplankton. These sea areas can provide obvious characteristic information, which helps the model better learn the relationship between frontal activities and the displacement velocity field between phytoplankton and submesoscale processes. For example, in sea areas where submesoscale vortices are active, phytoplankton will show a specific distribution pattern under the action of the vortices. By studying these sea areas, we can more deeply understand the influence of submesoscale flow fields on the distribution of phytoplankton.

[0061] Determine the areas in the second screening result obtained through the above screening as representative sea areas. In subsequent research, using the data on submesoscale processes in these representative sea areas to train the model can better simulate the displacement velocity field of submesoscale processes, improve the accuracy and reliability of the flow field inversion model, and provide strong data support for oceanographic research, ocean environmental monitoring, ocean resource management, etc.

[0062] S102. Use a geostationary ocean color imaging satellite to obtain the structural image information of phytoplankton fronts and filaments in the submesoscale processes of the representative sea areas.

[0063] Specifically, the structural image information of phytoplankton fronts and filaments includes the shape characteristics, texture characteristics, color characteristics, and boundary characteristics of phytoplankton and filaments, etc. The structural image information of phytoplankton fronts and filaments can reflect the changes in ocean color and the distribution characteristics of phytoplankton, thereby reflecting submesoscale processes. A geostationary ocean color imaging satellite is a type of satellite used to observe the ocean, with relatively high spatial resolution. For example, the spatial resolution can reach 250 m, and it can clearly observe phenomena such as submesoscale vortices on the order of several kilometers in the ocean, especially in areas such as phytoplankton fronts, filaments, and the edges of mesoscale vortices. At the same time, it has a high temporal resolution. For example, it can conduct 8 observations per day and can capture the changes in ocean processes on the hourly scale. It is located in a geostationary orbit and can continuously observe specific areas.

[0064] S103. Use the first database to obtain the displacement velocity field information of the ocean submesoscale process at the same location.

[0065] Specifically, the first database is the MITgcm LLC4320 simulation product obtained from the data portal of the National Aeronautics and Space Administration of the United States. Based on the MITgcm with LLC (Latitude - Longitude Cap, latitude - longitude polar cap) grid configuration, the nominal horizontal resolution is 1 / 48° (about 2 km), and the temporal resolution is 1 hour. The LLC4320 simulation includes 90 vertical layers. The vertical resolution near the sea surface is about 1 m, and it is about 30 m at 500 m. The resolution of the upper ocean process is optimized. This first dataset provides key variables such as sea level anomaly, temperature, salinity, and velocity per hour in the native grid.

[0066] Furthermore, search for the displacement velocity field information of the ocean submesoscale process at the same location from the first dataset. The displacement velocity field contains information such as the velocity, displacement, vortices, and spatio - temporal changes of ocean water body movement, which can ensure the continuity of the studied submesoscale process in time and space and accurately track its evolution process at a specific location.

[0067] S104. Match the structural image information and the displacement velocity field information to construct a first training sample set.

[0068] Specifically, matching the structural image information and the displacement velocity field information includes:

[0069] (1) Extract the structural features of the phytoplankton front and the filaments in the structural image information.

[0070] Specifically, using image segmentation and feature extraction techniques, identify and extract the structural features of the phytoplankton front and filaments from the structural image, and convert them into a structured data format for subsequent analysis, such as vector representations of edge contours, texture features, etc.

[0071] (2) Perform spatial grid processing on the displacement velocity field information, and extract the displacement features and velocity features of each grid cell of the displacement velocity field information.

[0072] Specifically, for the displacement velocity field information obtained from the first database, perform spatial grid processing on it. Each grid cell corresponds to a certain area (such as corresponding to an image pixel or pixel block), and extract the displacement features and velocity features of each grid cell. The implementation process of feature extraction can refer to the description in related technologies and will not be elaborated here.

[0073] (3) Process the structural image information and the displacement velocity field information according to the spatial scale and time series, and match the displacement velocity field information with the structural features of the phytoplankton front and the structural features of the filaments.

[0074] Specifically, using GIS technology to match based on geographical coordinates from the spatial scale, obtain the structural image information and the displacement velocity field information that are in the same geographical reference system, have consistent resolutions, and establish a corresponding relationship pixel by pixel or region by region. Perform time calibration on the structural image information and the displacement velocity field information from the time series to obtain the structural image information and the displacement velocity field information that are continuous and synchronized in time. Then, by analyzing the correlation between the structural image information and the displacement velocity field information after spatial scale and time series processing, the matching of the structural image information and the displacement velocity field information is realized.

[0075] Specifically, in terms of spatial scale, using Geographic Information System (GIS) technology, the structural image and the displacement velocity field are precisely matched based on geographical coordinates. Ensure that each pixel or region in the image can accurately correspond to the corresponding position in the velocity field. This is the basis for establishing the connection between the two. For example, through means such as satellite positioning information and map projection transformation, align the structural image information and the displacement velocity field information in the same geographical reference system for subsequent correlation analysis. If the resolutions of the structural image and the displacement velocity field are inconsistent, resolution coordination processing is required. Interpolation algorithms such as bilinear interpolation and nearest neighbor interpolation can be used to interpolate the velocity field data to the same or similar resolution as the image, so that the two match in spatial scale, facilitating the establishment of the correspondence between the structure and the velocity pixel by pixel or region by region.

[0076] Furthermore, in terms of time series, according to the satellite observation time and the time stamps of the velocity field data, perform time calibration operations to make the two reflect the ocean state at the same or similar times. For data with inconsistent time resolutions, methods such as time interpolation or averaging can be used to construct a continuous and synchronized data set in time, and match and analyze a series of continuous structural images with the corresponding displacement velocity field time series. Observe the evolution process of the phytoplankton front and filamentous structures over time, as well as the corresponding changes in the velocity field, and analyze the dynamic correlation between the two in the time dimension. For example, by calculating the correlation between the change rate of the structural characteristics and the change in the velocity field, reveal the time lag effect and periodic change law of the interaction between the structure and the flow field in the ocean submesoscale process.

[0077] Furthermore, based on the extracted structural characteristics, displacement characteristics, and velocity characteristics of the phytoplankton front and the filamentous structures, use statistical analysis or machine learning methods to construct a mathematical model between the structural image information and the displacement velocity field information. For example, in one embodiment, regression analysis is used to establish a quantitative relationship model between the length of the phytoplankton front and the magnitude of the local flow field velocity; a neural network model is used to learn the complex mapping relationship between the filamentous texture structure and the vorticity of the surrounding velocity field. By analyzing the linear or non-linear correlation between the structural image information and the displacement velocity field information, match the displacement velocity field information with the structural image information.

[0078] S105. Correct the first training sample set based on the second data set and the third data set to obtain a second training sample set; wherein, the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of the large-scale ocean process.

[0079] Specifically, although the first training sample set can provide detailed information on submesoscale processes, its description of large-scale ocean processes is relatively limited. The sea surface height data in the second dataset and the geostrophic flow field displacement velocity field data of large-scale ocean processes in the third dataset can provide background information on large-scale ocean circulation. Since submesoscale ocean processes occur and evolve under the background of large-scale ocean circulation, by using the second and third datasets to correct the first training sample set, the first training sample set can contain more comprehensive ocean dynamic information, forming a complete information chain from large scale to submesoscale, which helps the model to more accurately understand the behavioral characteristics of submesoscale processes in different scale environments and improve the accuracy and reliability of submesoscale process prediction.

[0080] Further, the correction of the first training sample set based on the second and third datasets includes:

[0081] (1) Construct a conversion function based on the geostrophic balance relationship, and the conversion function is used to calculate the geostrophic flow field displacement velocity field of the second dataset.

[0082] Specifically, in the ocean, there is a geostrophic balance relationship between geostrophic flow and sea surface height anomaly, and the conversion function can be constructed through the following function: ; where is the Coriolis parameter; is the geostrophic flow velocity; is the acceleration due to gravity; is the sea surface height anomaly; is the horizontal direction.

[0083] For the second dataset, substituting its sea surface height data into this conversion function can obtain the geostrophic flow field displacement velocity field corresponding to each data in the second dataset, which can be directly compared with the displacement velocity field in the first training sample set within the same geographical area and time range, and the consistency and differences between the two can be judged based on the geostrophic balance relationship.

[0084] (2) Within the same location and time range, use the geostrophic flow field displacement velocity fields of the second and third datasets to compare with the displacement velocity field of the first training sample set to obtain the displacement velocity field difference.

[0085] Specifically, the first training sample set has a high spatial and temporal resolution, while the third data set has a lower resolution, which reflects a larger-scale ocean dynamic field. To make them corresponding, spatial filtering can be performed on the displacement velocity field in the first training sample set. For example, a low-pass filter can be used to extract its large-scale geostrophic current component to match the data in the third data set in terms of scale, and then the displacement velocity field difference between the first training sample set and the third data set can be obtained.

[0086] (3) Correct the first training sample set according to the second data set and the displacement velocity field difference between the third data set and the first training sample set.

[0087] Specifically, combining the above description, analyze the displacement velocity field difference between the second data set and the first training sample set and the displacement velocity field difference between the third data set and the first training sample set. According to the displacement velocity field difference, correct the data in the first training sample set. In specific implementation, the data in the first training sample set and its corresponding displacement velocity field difference can be weighted and summed, and the data after weighted summation is used to replace the original data in the first training sample set.

[0088] S106. Use the second training sample set to train the initial model to obtain a flow field inversion model.

[0089] Specifically, the training process of the flow field inversion model includes:

[0090] (1) Input the second training sample set into the initial model;

[0091] (2) The initial model learns the relationship between the submesoscale sea surface water color image and the displacement velocity field from the second training sample set.

[0092] Specifically, it can be understood that the second training sample set is a high-quality data set after correction, which contains submesoscale sea surface water color images and corresponding displacement velocity field information. These data come from multiple data sources, integrating ocean information of different scales and types, providing rich and accurate learning materials for the initial model. The sea surface water color image can reflect the distribution of phytoplankton, filaments, etc. on the ocean surface, while the displacement velocity field describes the motion state of ocean water masses. There is a complex physical connection between the two. For example, the aggregation and diffusion of phytoplankton are often affected by the ocean current field, and the change of the ocean current field may also interact with the growth and distribution of phytoplankton. By inputting the second training sample set into the initial model, the model can access the information of the interconnection of various factors in the real ocean environment, and thus learn the potential internal relationship pattern between the submesoscale sea surface water color image and the displacement velocity field.

[0093] (3) Determine the loss function. With the goal of minimizing the loss function, iterate the parameters of the flow field inversion model so that the difference between the predicted value and the true value of the displacement velocity field is less than the threshold.

[0094] Specifically, the loss function can be determined through the following steps:

[0095] (1) For each pixel position in the submesoscale sea surface water color image, obtain its predicted displacement velocity vector and true displacement velocity vector respectively.

[0096] Specifically, when processing the submesoscale sea surface water color image, each pixel position may correspond to different flow field characteristics. By obtaining the predicted displacement velocity vector and the true value of each pixel position respectively, the detailed distribution of the flow field in the image space can be accurately described. The displacement velocity vector contains two important physical quantities, namely displacement and velocity. Displacement describes the position change of the water body within a certain time interval, while velocity represents the speed and direction of the displacement. In the ocean flow field, these two quantities are interrelated and jointly determine the dynamic characteristics of the flow field.

[0097] (2) Calculate the square of the displacement and the square of the velocity in the predicted displacement velocity vector and the true displacement velocity vector.

[0098] (3) Calculate the difference between the square of the displacement and the square of the velocity, and take the square root of the difference to obtain the loss function.

[0099] Specifically, the loss function can be expressed by the following formula: ; where, is the displacement vector in the true displacement velocity vector; is the velocity vector in the true displacement velocity vector; is the displacement vector in the predicted displacement velocity vector; is the velocity vector in the predicted displacement velocity vector.

[0100] Further, the structure of the flow field inversion model is introduced as follows:

[0101] The flow field inversion model includes: a feature extraction module, a convolution module, an iterative update module, and an output flow upsampling module;

[0102] The feature extraction module is used to extract features from the sub-mesoscale sea surface water color image to obtain a dense feature map. The resolution of the dense feature map is low resolution, and the dense feature map includes a feature map of the temperature gradient and a feature map of the texture feature of the sea area to be detected;

[0103] The convolution module includes a plurality of convolution kernels with different sizes, and the plurality of convolution kernels are used to calculate the correlation of the dense feature map;

[0104] The iterative update module is used to learn and predict the movement trends of the filaments and the phytoplankton front according to the update gate and reset gate mechanisms, combining the current input of the dense feature map and the correlation volume, to obtain the flow field prediction value;

[0105] The output flow upsampling module is used to process the flow field prediction value to restore the subtle displacement changes of the filaments and the phytoplankton front on the ocean surface, to obtain the predicted displacement velocity field of the ocean sub-mesoscale process.

[0106] S107. Use the geostationary ocean color imaging satellite to observe the sea area to be detected, and obtain multiple sub-mesoscale sea surface water color images of the sea area to be detected.

[0107] S108. Use the flow field inversion model to process the sub-mesoscale sea surface water color image to obtain the displacement velocity field of the surface sub-mesoscale process of the sea area to be detected.

[0108] Specifically, using the flow field inversion model to process the sub-mesoscale sea surface water color image includes:

[0109] (1) Extract features from the sub-mesoscale sea surface water color image to obtain a dense feature map.

[0110] Specifically, combined with the above description, use the feature extraction module in the flow field inversion model to extract features from the sub-mesoscale sea surface water color image, which is implemented by a convolutional neural network and consists of three levels of downsampling. This feature extraction module can extract complex feature information (such as temperature gradient and texture features, etc.) from two sets of consecutive sub-mesoscale sea surface water color images, and map the input image (256 pixels × 256 pixels) to a low-resolution dense feature map with an initial image resolution of 1 / 8. For the filament structure and front area in the ocean, this layer can capture their unique temperature gradient and texture features, convert these features into vector representations that can be processed by the model, and provide a basis for accurately identifying and analyzing the movement and changes of filaments and fronts.

[0111] (2) Calculate the correlation between the dense feature maps.

[0112] Specifically, a four - layer correlation pyramid structure is used. Through hierarchical pooling operations, while maintaining high - resolution image information, the visual similarity between two sets of consecutive input dense feature maps is effectively calculated. For filaments and fronts, their corresponding positions in consecutive images can be accurately located. Even when they undergo complex changes such as distortion and rotation, their similarity features can be accurately identified, thereby providing a key basis for accurately calculating the displacement velocity and ensuring that the model can effectively handle the dynamic changes of filaments and fronts. For the implementation process of calculating the correlation between dense feature maps, please refer to the description in the related technology and will not be elaborated here.

[0113] (3) Combine the dense feature maps and the correlation to learn and predict the movement trends of the filaments and the phytoplankton fronts, and obtain the flow - field prediction value.

[0114] Specifically, taking the dense feature maps and their correlation as inputs, the flow - field inversion model can learn the complex relationship between the movement trends of filaments and phytoplankton fronts and marine environmental factors. Through learning a large number of training samples, the flow - field inversion model gradually establishes a mapping model between the feature maps and the movement trends. Based on the learned mapping model, when a new sub - mesoscale sea - surface water - color image is input, the flow - field inversion model can predict the movement trends of filaments and phytoplankton fronts according to the extracted dense feature maps and their correlation, and then obtain the flow - field prediction value. This prediction value not only contains the velocity magnitude and direction information of the ocean water body in this area, but also reflects the influence of the movement of filaments and fronts on the flow field and the driving effect of the flow field on them.

[0115] (4) Process the flow - field prediction value to obtain the predicted displacement velocity field.

[0116] Specifically, the flow - field prediction value is a vector field containing various information and needs to be further processed to obtain the predicted displacement velocity field for analysis. The velocity vectors in the flow - field prediction value can be decomposed according to the spatial coordinates to obtain the velocity components in each direction, and then these components are combined into a displacement velocity field through a certain algorithm so that it can accurately represent the displacement changes of the ocean water body at the sub - mesoscale. At the same time, the velocity field may also be smoothed to remove the influence of noise and outliers, improving the accuracy and reliability of the displacement velocity field.

[0117] (5) Use the predicted displacement velocity field to calculate the sub - mesoscale kinetic - energy spectrum, Chl - a concentration spectrum, and their flux changes.

[0118] Specifically, the predicted displacement velocity field provides basic data for calculating the submesoscale kinetic energy spectrum, Chl-a concentration spectrum, and their flux variations. The submesoscale kinetic energy spectrum can be obtained by performing spectral analysis on the displacement velocity field, which reflects the distribution of ocean kinetic energy at different wavenumbers (or spatial scales) and helps study the transfer and dissipation mechanisms of ocean energy within the submesoscale range. The Chl-a concentration spectrum is calculated based on the Chl-a concentration information in the sea surface water color image. Combining with the displacement velocity field, the distribution characteristics of the Chl-a concentration at the submesoscale and its interaction relationship with the flow field can be analyzed. Further, by calculating the flux variations, the transfer rates and directions of substances (such as phytoplankton-related substances represented by Chl-a) and energy during the submesoscale process can be understood, further revealing the laws of material cycling and energy flow in the ocean ecosystem and providing important quantitative indicators for in-depth understanding of the physical, chemical, and biological interactions in the ocean submesoscale process.

[0119] Furthermore, Figure 2 For the comparison diagram of the effects of the flow field inversion model and the traditional model at the same time shown in the exemplary embodiments of the present application, please refer to Figure 2 , Figure 2 In row (a) of [reference], the velocity field data of the flow field inversion model and the traditional models (the maximum correlation coefficient algorithm and the physical constraint convergence algorithm) are shown. The ground truth data is the accurate velocity field information as a reference; the velocity field of the flow field inversion model is obtained by using the flow field inversion model proposed in the present application; the velocity field obtained by the maximum correlation coefficient algorithm represents the traditional maximum cross-correlation algorithm; the velocity field obtained by the physical constraint convergence algorithm is based on the optimized algorithm of decorrelation scale analysis. Figure 2 In row (b) of [reference], the average velocity error (EPE) data of the flow field inversion model and the traditional models (the maximum correlation coefficient algorithm and the physical constraint convergence algorithm) are shown. Figure 2 In row (c) of [reference], the average angular error (AEE) data of the flow field inversion model and the traditional models (the maximum correlation coefficient algorithm and the physical constraint convergence algorithm) are shown. In terms of the endpoint error, the average velocity error of the flow field inversion model is 0.14 m / s, the average velocity error of the maximum correlation coefficient algorithm is 0.40 m / s, and the average velocity error of the physical constraint convergence algorithm is 0.39 m / s; in terms of the average angular error, the average angular error of the flow field inversion model is 5.58°, the average angular error of the maximum correlation coefficient algorithm is 12.41°, and the average angular error of the physical constraint convergence algorithm is 11.91°. It can be clearly seen from these data that the flow field inversion model provided by the present application performs better in error control and can simulate the velocity field more accurately.

[0120] The method provided in this embodiment screens representative sea areas, matches and analyzes the structural image information of phytoplankton fronts and filaments with the displacement velocity field information at the same location, and can accurately capture the physical relationship between the material distribution (phytoplankton) and the flow field in the ocean. Using the ocean data of the representative sea areas to form the first training sample set, and correcting it with the second data set and the third data set to obtain the second training sample set, realizing the multi-scale information fusion from the sub-mesoscale to the large ocean scale, which helps to deeply understand the behavioral characteristics of the ocean sub-mesoscale process in different scale environments. Further, training the initial model with the second training sample set constructed based on multi-source data greatly optimizes the training data of the model. The corrected training sample set combines the advantages of multiple data sources, reduces the errors and uncertainties that may exist in a single data source, and improves the accuracy of the relationship between the sub-mesoscale sea surface color image and the displacement velocity field learned by the flow field inversion model. Using the trained flow field inversion model to process the sub-mesoscale sea surface color image of the sea area to be detected observed by the geostationary ocean color imaging satellite can obtain the displacement velocity field of the sub-mesoscale process on the sea surface of this sea area. This process realizes the rapid and accurate conversion from satellite observation images to flow field information, providing strong data support for oceanography research, ocean environment monitoring, ocean resource management, etc. In addition, by calculating and analyzing the sub-mesoscale kinetic energy spectrum, Chl-a concentration spectrum and their flux changes, the laws of the transfer and transformation of ocean energy and matter at the sub-mesoscale can be revealed, providing an important basis for the development of the basic theory of ocean science.

[0121] Corresponding to the foregoing embodiment of a method for inverting the flow field of the ocean near-surface sub-mesoscale process, the present application also provides an embodiment of an apparatus for inverting the flow field of the ocean near-surface sub-mesoscale process.

[0122] Figure 3 It is a schematic structural diagram of the first embodiment of the apparatus for inverting the flow field of the ocean near-surface sub-mesoscale process provided by the present application. Please refer to Figure 3 , the apparatus provided in this embodiment includes an acquisition module 310, a construction module 320 and a processing module 330; wherein,

[0123] The acquisition module 310 is used to screen the representative sea areas of the ocean near-surface sub-mesoscale from the ocean area;

[0124] The acquisition module 310 is further used to obtain the structural image information of phytoplankton fronts and filaments of the ocean sub-mesoscale process by using the geostationary ocean color imaging satellite;

[0125] The acquisition module 310 is further used to obtain the displacement velocity field information of the ocean sub-mesoscale process at the same location by using the first database;

[0126] The construction module 320 is configured to match the structural image information and the displacement velocity field information to construct a first training sample set;

[0127] The construction module 320 is further configured to correct the first training sample set based on a second data set and a third data set to obtain a second training sample set; wherein, the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of large-scale ocean processes;

[0128] The construction module 320 is further configured to train an initial model using the second training sample set to obtain a flow field inversion model;

[0129] The processing module 330 is configured to use the geostationary ocean color imaging satellite to observe the sea area to be detected, and obtain multiple sub-meso-scale sea surface color images of the sea area to be detected;

[0130] The processing module 330 is further configured to process the sub-meso-scale sea surface color image using the flow field inversion model to obtain the displacement velocity field of the surface sub-meso-scale process of the sea area to be detected.

[0131] The device of this embodiment can be used to execute Figure 1 the steps of the method embodiment shown. The specific implementation principle and process are similar, and will not be elaborated here.

[0132] For the implementation process of the functions and roles of each unit in the above device, please refer to the implementation process of the corresponding steps in the above method for details, which will not be elaborated here.

[0133] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0134] The above are only the preferred embodiments of this application, and are not intended to limit this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A method for inverting the flow field of sub-mesoscale processes near the ocean surface, characterized in that: The method comprises: Screening of sub-mesoscale representative sea areas near the ocean surface from the ocean area; Using a geostationary ocean color imaging satellite to obtain structural image information of phytoplankton fronts and filaments representing sub-mesoscale processes in the ocean; The first database is used to obtain the displacement velocity field information of the ocean submesoscale process at the same location; Matching the structural image information and the displacement velocity field information to construct a first training sample set; Correcting the first training sample set based on a second data set and a third data set to obtain a second training sample set; wherein the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of large-scale ocean processes; Using the second training sample set to train the initial model to obtain a flow field inversion model; Utilizing the geostationary ocean color imaging satellite to observe the sea area to be detected, and obtaining a plurality of sub-mesoscale sea surface water color images of the sea area to be detected; The sub-mesoscale sea surface water color image is processed using the flow field inversion model to obtain the displacement velocity field of the surface sub-mesoscale process of the sea area to be detected.

2. The method according to claim 1, characterized in that The matching of the structural image information and the displacement velocity field information comprises: extracting structural features of the phytoplankton front and the filaments in the structural image information; Performing spatial gridding processing on the displacement velocity field information to extract displacement characteristics and velocity characteristics of each grid unit of the displacement velocity field information; The structural image information and the displacement velocity field information are processed according to the spatial scale and time series, and the displacement velocity field information is matched with the structural features of the phytoplankton front and the structural features of the filaments.

3. The method according to claim 1, characterized in that The correcting the first training sample set based on the second data set and the third data set comprises: constructing a conversion function based on a geostrophic balance relationship, wherein the conversion function is used to calculate a geostrophic flow field displacement velocity field of the second data set; In the same position and time range, the displacement velocity field of the geostrophic flow field of the second data set and the third data set is compared with the displacement velocity field of the first training sample set to obtain a displacement velocity field difference; The first training sample set is corrected according to the difference in displacement velocity field between the second data set and the third data set and the first training sample set.

4. The method according to claim 1, characterized in that The method of screening representative sub-mesoscale ocean areas near the surface of the ocean from the ocean area includes: Obtain the motion of submesoscale processes in various ocean regions; Obtain the characteristic differences between phytoplankton and filaments and the environment in each ocean area; Screening the various ocean areas according to the characteristic differences to obtain a first screening result of representative ocean areas; Determine the distribution characteristics of each representative sea area in the first screening result of the representative sea areas, select the first screening result of the representative sea areas according to the distribution characteristics to obtain a second screening result, and determine the representative sea area according to the second screening result.

5. The method according to claim 4, characterized in that The obtaining of characteristic differences between phytoplankton and filaments and the environment in each ocean area includes: Acquire multiple water color images of each ocean area, each of the water color images includes phytoplankton, filaments, and the background environment of the ocean area; Identifying the first boundary of the phytoplankton and the filament from each of the water color images; identifying a second boundary of the background environment from each of the water color images; Calculating the overlapping area of ​​the first boundary and the second boundary in each of the water color images; The weighted sum of the overlapping areas of all water color images is calculated to obtain the total overlapping area, and the total overlapping area is used as the characteristic difference.

6. The method according to claim 1, characterized in that The flow field inversion model includes: a feature extraction module, a convolution module, an iterative update module and an output flow upsampling module; The feature extraction module is used to extract features from the sub-mesoscale sea surface water color image to obtain a dense feature map, the resolution of the dense feature map is low resolution, and the dense feature map includes a feature map of the temperature gradient and a feature map of the texture feature of the sea area to be detected; The convolution module includes a plurality of convolution kernels, the plurality of convolution kernels have different sizes, and the plurality of convolution kernels are used to calculate the correlation of the dense feature map; The iterative update module is used to learn and predict the movement trends of the filaments and the phytoplankton front according to the update gate and reset gate mechanism, combined with the current input of the dense feature map and the correlation volume, to obtain a flow field prediction value; The output flow upsampling module is used to process the flow field prediction value, restore the slight displacement changes of the filaments and the phytoplankton front on the ocean surface, and obtain the predicted displacement velocity field of the ocean sub-mesoscale process.

7. The method according to claim 1, characterized in that The training process of the flow field inversion model includes: Inputting the second training sample set into the initial model; The initial model learns the relationship between the sub-mesoscale sea surface water color image and the displacement velocity field from the second training sample set; A loss function is determined, and with the goal of minimizing the loss function, the flow field inversion model parameters are iterated to make the difference between the predicted value of the displacement velocity field and the true value of the displacement velocity field smaller than a threshold.

8. The method according to claim 7, characterized in that Determining the loss function comprises: For each pixel position in the sub-mesoscale sea surface water color image, respectively obtaining a displacement velocity vector prediction value and a displacement velocity vector true value thereof; Calculating the square of displacement and the square of velocity in the displacement velocity vector prediction value and the displacement velocity vector true value; The difference between the square of the displacement and the square of the velocity is calculated, and the square root of the difference is taken to obtain the loss function.

9. The method according to claim 1, characterized in that: The method of processing the sub-mesoscale sea surface water color image by using the flow field inversion model comprises: Extracting features from the sub-mesoscale sea surface water color image to obtain a dense feature map; Calculating correlations between the dense feature maps; Combining the dense feature map with the correlation, learning and predicting the movement trends of the filaments and the phytoplankton front, and obtaining a flow field prediction value; Processing the flow field prediction value to obtain a predicted displacement velocity field; The predicted displacement velocity field is used to calculate the sub-mesoscale kinetic energy spectrum, Chl-a concentration spectrum and flux variation thereof.

10. An ocean near-surface sub-mesoscale process flow field inversion device, characterized in that: The device comprises an acquisition module, a construction module and a processing module; wherein, The acquisition module is used to screen the sub-mesoscale representative sea areas near the ocean surface from the ocean area; The acquisition module is also used to acquire structural image information of phytoplankton fronts and filaments representing sub-mesoscale processes in the ocean using a geostationary ocean color imaging satellite; The acquisition module is further used to acquire the displacement velocity field information of the ocean submesoscale process at the same position using the first database; The construction module is used to match the structural image information and the displacement velocity field information to construct a first training sample set; The construction module is further used to correct the first training sample set based on a second data set and a third data set to obtain a second training sample set; wherein the second data set includes sea surface height data, and the third data set includes geostrophic flow field displacement velocity field data of large-scale ocean processes; The construction module is further used to train the initial model using the second training sample set to obtain a flow field inversion model; The processing module is used to observe the sea area to be detected by using the geostationary ocean color imaging satellite to obtain multiple sub-mesoscale sea surface water color images of the sea area to be detected; The processing module is also used to process the sub-mesoscale sea surface water color image using the flow field inversion model to obtain the displacement velocity field of the surface sub-mesoscale process of the sea area to be detected.

Images