CN115345238B

CN115345238B - Method and device for generating seawater transparency fusion data

Info

Publication number: CN115345238B
Application number: CN202210985459.6A
Authority: CN
Inventors: 周旋; 李自强; 宋帅; 周江涛; 安玉柱; 赵亚明; 薛彦广; 张阳; 王舸
Original assignee: 61741 Unit Of Pla
Current assignee: 61741 Unit Of Pla
Priority date: 2022-08-17
Filing date: 2022-08-17
Publication date: 2023-04-07
Anticipated expiration: 2042-08-17
Also published as: CN115345238A

Abstract

The application provides a method and a device for generating seawater transparency fusion data, comprising the following steps: inputting historical seawater transparency fusion data m days before the target area and satellite remote sensing seawater transparency data m +1 days into the estimation data to generate an antagonistic network, and obtaining estimated seawater transparency fusion data m +1 days; and inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into the fusion data to generate an antagonistic network, thereby obtaining the target seawater transparency fusion data of the m +1 th day. Therefore, a time-space change rule for fighting against the network excavation seawater transparency can be generated through the estimated data, estimated seawater transparency fusion data are generated, and missing data reconstruction is achieved; and then an antagonistic network is generated through the fusion data to realize the fusion of the seawater transparency fusion data and the high-precision satellite remote sensing seawater transparency, and finally the full-coverage high-precision single-day seawater transparency fusion data is generated.

Description

Method and device for generating seawater transparency fusion data

Technical Field

The application relates to the technical field of ocean exploration, in particular to a method and a device for generating seawater transparency fusion data.

Background

The sea water transparency refers to the maximum visible depth of the transparent scale when it is immersed in sea water, and is closely related to the physical properties, chemical composition, suspended matter in sea water and dynamic state of sea. The method for obtaining the seawater transparency with full coverage, high space-time resolution and high precision has important significance for researching the physical and chemical properties of the seawater, monitoring the marine ecology, producing the marine fishery and guaranteeing the marine military activities.

At present, in the prior art, seawater transparency data are often obtained through a satellite remote sensing and data interpolation processing technology. However, due to factors such as cloud occlusion, solar flare pollution and satellite orbit clearance, the seawater transparency data acquired in this way is insufficient in terms of spatial coverage, detail characteristics and the like, resulting in poor data quality.

Disclosure of Invention

In view of this, an object of the present application is to provide a method and a device for generating seawater transparency fusion data, wherein historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days before the target area are input into pre-trained estimated data to generate an antagonistic network, so as to obtain estimated seawater transparency fusion data m +1 days generated by an estimated data generator; inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into pre-trained fusion data to generate an antagonistic network, and obtaining target seawater transparency fusion data of the m +1 th day of the target area generated by the fusion data generator. Therefore, a time-space change rule for fighting against the network excavation seawater transparency can be generated through the estimated data, estimated seawater transparency fusion data are generated, and missing data reconstruction is achieved; and then, fusion of seawater transparency fusion data and high-precision satellite remote sensing seawater transparency is realized through a countermeasure network generated by fusing data, and finally, full-coverage and high-precision single-day seawater transparency fusion data is generated.

The embodiment of the application provides a method for generating seawater transparency fusion data, which comprises the following steps:

acquiring historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days after the target area; m is a positive integer;

inputting the historical seawater transparency fusion data into pre-trained pre-estimation data to generate pre-estimation data generators in an antagonistic network, and obtaining pre-estimation seawater transparency fusion data of the m +1 th day of the target area; the predicted data generation countermeasure network comprises a predicted data generator and a predicted data discriminator, and is obtained by mutual countermeasure training of the predicted data generator and the predicted data discriminator;

inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into fusion data which is trained in advance to generate fusion data generators in an countermeasure network, and obtaining target seawater transparency fusion data of the m +1 th day of the target area; the fusion data generation countermeasure network comprises the fusion data generator, a global data discriminator and a satellite data discriminator; the fusion data generation countermeasure network is obtained by the mutual countermeasure of the fusion data generator and the global data arbiter and the mutual countermeasure training of the fusion data generator and the satellite data arbiter.

Further, the pre-estimation data generation countermeasure network and the fusion data generation countermeasure network are trained by the following means:

inputting the historical seawater transparency data of the target area into initial estimation data to generate an initial estimation data generator in an countermeasure network, and obtaining first historical estimation seawater transparency data estimated by the initial estimation data generator;

inputting the first historical pre-estimated seawater transparency data and the historical seawater transparency data into initial pre-estimated data to generate an initial pre-estimated data discriminator in an countermeasure network, and obtaining a first discrimination result of the initial pre-estimated data discriminator on the first historical pre-estimated seawater transparency data;

generating an confrontation network for the initial estimation data based on the historical seawater transparency data, the first historical estimation seawater transparency data and the first judgment result, and pre-training the confrontation network to obtain an estimation data generation confrontation network after pre-training is completed and second historical estimation seawater transparency data estimated by the initial estimation data generator when pre-training is completed;

inputting the second historical pre-estimated seawater transparency data and the historical satellite remote sensing seawater transparency data of the target area into initial fusion data to generate an initial fusion data generator in a countermeasure network, and obtaining first historical seawater transparency fusion data generated by the initial fusion data generator;

performing Poisson fusion on the historical satellite remote sensing seawater transparency data and the second historical estimated seawater transparency data to obtain historical seawater transparency Poisson fusion data;

inputting the historical seawater transparency Poisson fusion data and the first historical seawater transparency fusion data into initial fusion data to generate an initial global data discriminator in a countermeasure network, and obtaining a second discrimination result of the initial global data discriminator on the first historical seawater transparency fusion data;

inputting the historical satellite remote sensing seawater transparency data and the first historical seawater transparency fusion data into initial fusion data to generate an initial satellite data discriminator in a countermeasure network, and obtaining a third discrimination result of the initial satellite data discriminator on the first historical seawater transparency fusion data;

generating an confrontation network for the initial fusion data based on the historical seawater transparency Poisson fusion data, the historical satellite remote sensing seawater transparency data, the first historical seawater transparency fusion data, the second judgment result and the third judgment result, and pre-training the confrontation network to obtain a pre-trained confrontation network generated fusion data and second historical seawater transparency fusion data generated by the initial fusion data generator when pre-training is completed;

and respectively re-inputting the pre-trained estimated data to generate an countermeasure network and the pre-trained fused data to generate an countermeasure network, re-inputting the pre-trained estimated data to generate the countermeasure network and the pre-trained remote sensing seawater transparency data to generate the countermeasure network, and performing fine tuning training on the pre-trained estimated data to generate the countermeasure network and the pre-trained fused data to generate the countermeasure network.

Further, the obtaining of the historical seawater transparency fusion data m days before the target area includes:

acquiring third history seawater transparency fusion data generated by the initial fusion data generator when the fine tuning training of the trained fusion data generation countermeasure network is completed;

and screening out the historical seawater transparency fusion data of the previous m days from the third historical seawater transparency fusion data.

Further, when the second historical seawater transparency fusion data is re-input into the pre-trained estimated data to generate an confrontation network, the pre-trained estimated data is generated into the confrontation network again and fine tuning training is performed to obtain trained estimated data to generate the confrontation network, the first loss function of the estimated data discriminator in the pre-estimated data generation confrontation network comprises:

/>

in the formula, D represents an estimated data discriminator;G ₁ a representation prediction data generator; l is _D A first loss function representing a predicted data discriminator; x represents historical seawater transparency fusion data;

representing historical pre-estimated seawater transparency data; p _data Represents the distribution of historical seawater transparency fusion data>

Representing the distribution of historical pre-estimated seawater transparency data;

the generating of the second loss function against the predictive data generator in the network comprises:

in the formula (I), the compound is shown in the specification,

a second loss function representing a prediction data generator; m is _land Representing a land mask for characterizing a land distribution of the target area; m is _ocean Representing a sea mask for characterizing a sea distribution of the target area; an indication of multiplication of the corresponding element; II (8230); II ₁ Is represented by ₁ A distance; λ and μ are constants.

Further, when the historical satellite remote sensing seawater transparency data is input into the fusion data generated by the pre-training again to generate the countermeasure network, and the fusion data generated by the pre-training is subjected to fine tuning training again to obtain the trained fusion data generated countermeasure network, a third loss function of a global data discriminator in the fusion data generated countermeasure network is expressed as:

in the formula, D _a Representing global data predicatesA discriminator; g ₂ A generator of the fused data is represented,

a third loss function representing a global data arbiter; x represents historical seawater transparency fusion data; x' represents historical seawater transparency Poisson fusion data; p is _data1 Representing the distribution of the Poisson fusion data of historical seawater transparency; />

Representing the distribution of historical seawater transparency fusion data;

the fourth loss function of the satellite data discriminator in the fusion data generation countermeasure network is expressed as:

in the formula D _s A representation satellite data discriminator;

a fourth loss function representing a satellite data arbiter; />

Representing historical satellite remote sensing seawater transparency data; p is _data2 Representing the distribution of historical satellite remote sensing seawater transparency data;

the fifth loss function of the fused data generator in the converged data generation countermeasure network is expressed as:

in the formula (I), the compound is shown in the specification,

a fifth loss function representing a fused data generator; II (8230); II ₁ Represents l ₁ A distance; II (8230); II ₂ Is represented by ₂ A distance; m is _sat Representing a satellite observation region in the target region; />

Representing a gradient operator; />

Denotes structural similarity, μ _x Denotes the mean value of x>

Represents->

Is based on the mean value of (4)>

Represents the variance of x +>

Represents->

Is greater than or equal to>

Denotes x and->

Covariance of c ₁ And c ₂ Is a constant; δ, α, β, and γ are constants.

Furthermore, the prediction data generator comprises a prediction encoder of an n-layer network structure and a prediction decoder of the n-layer network structure; the pre-estimated encoder comprises m encoding paths; the first coding path comprises a spatial dimension reduction unit and n sequentially connected first residual error structures; each of the other coding paths comprises a space dimension reduction unit, a first residual error structure, n convolution gating circulation units and n-1 second residual error structures; the space dimension reduction unit is connected with the first residual error structure, and the convolution gating circulating unit of each layer is connected with the first residual error structure or connected with the convolution gating circulating unit of the upper layer through a second residual error structure; the convolution gating circulation unit of each layer in the former coding path is connected with the convolution gating circulation unit of the corresponding layer in the latter coding path; the pre-estimation decoder comprises n third residual error structures and an output block; the convolution gating circulation unit of each layer in the m-th coding path is connected to the third residual error structure of the corresponding layer in the pre-estimation decoder; the m coding paths correspond to historical seawater transparency fusion data of previous m days one by one;

the inputting the historical seawater transparency fusion data into a pre-trained estimation data generation estimation data generator in an countermeasure network to obtain the estimation seawater transparency fusion data of the target area at the m +1 th day comprises the following steps:

correspondingly inputting the historical seawater transparency fusion data of the previous m days into m coding paths in the pre-estimation encoder to obtain seawater transparency fusion data output by each layer of convolution gating circulation unit in the last coding path;

inputting the seawater transparency fusion data output by each layer of convolution gating circulation unit into the second residual error structure of the corresponding layer in the pre-estimation decoder to obtain the pre-estimation seawater transparency fusion data of the m +1 th day output by the output block of the pre-estimation decoder.

Further, the fusion data generator comprises a fusion encoder of a p-layer network structure and a fusion decoder of the p-layer network structure; the fusion encoder comprises a spatial dimension reduction unit and p second residual error structures; the fused decoder comprises p third residual structures and an output block; the second residual structure of each layer in the fused encoder is connected to the third residual structure of the corresponding layer in the fused decoder;

the method for inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into the fusion data which is trained in advance to generate the fusion data generator in the countermeasure network to obtain the target seawater transparency fusion data of the m +1 th day of the target area comprises the following steps:

inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into the fusion encoder to obtain seawater transparency fusion data output by the second residual error structure of each layer in the fusion encoder;

and inputting the seawater transparency fusion data output by the second residual structure of each layer in the fusion encoder into the third residual structure of the corresponding layer in the fusion decoder to obtain the target seawater transparency fusion data of the m +1 th day output by the output block of the fusion decoder.

The embodiment of the present application further provides a generation apparatus for seawater transparency fusion data, the generation apparatus includes:

the acquisition module is used for acquiring historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days after the target area; m is a positive integer;

the first determination module is used for inputting the historical seawater transparency fusion data into pre-trained pre-estimated data to generate pre-estimated data generators in an countermeasure network to obtain pre-estimated seawater transparency fusion data of the target area at day m + 1; the predicted data generation countermeasure network comprises a predicted data generator and a predicted data discriminator, and is obtained by mutual countermeasure training of the predicted data generator and the predicted data discriminator;

the second determination module is used for inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into fusion data which are trained in advance to generate fusion data generators in the countermeasure network, and obtaining target seawater transparency fusion data of the m +1 th day of the target area; the fusion data generation countermeasure network comprises the fusion data generator, a global data discriminator and a satellite data discriminator; the fusion data generation countermeasure network is obtained by the mutual countermeasure of the fusion data generator and the global data arbiter and the mutual countermeasure training of the fusion data generator and the satellite data arbiter.

An embodiment of the present application further provides an electronic device, including: a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, the processor and the memory communicating via the bus when the electronic device is running, the machine readable instructions when executed by the processor performing the steps of a method of generating seawater transparency fusion data as described above.

Embodiments of the present application also provide a computer-readable storage medium, on which a computer program is stored, where the computer program is executed by a processor to perform the steps of the method for generating seawater transparency fusion data.

According to the method and the device for generating the seawater transparency fusion data, historical seawater transparency fusion data of m days before a target area and satellite remote sensing seawater transparency data of m +1 days are input into pre-trained pre-estimation data to generate an anti-network, and pre-estimation seawater transparency fusion data of m +1 days generated by a pre-estimation data generator are obtained; inputting the satellite remote sensing seawater transparency data of the m +1 day and the estimated seawater transparency fusion data of the m +1 day into pre-trained fusion data to generate an antagonistic network, and obtaining target seawater transparency fusion data of the m +1 day of the target area generated by the fusion data generator. Therefore, a time-space change rule for fighting against the network excavation seawater transparency can be generated through the estimated data, estimated seawater transparency fusion data are generated, and missing data reconstruction is achieved; and then, fusion of seawater transparency fusion data and high-precision satellite remote sensing seawater transparency is realized through a countermeasure network generated by fusing data, and finally, full-coverage and high-precision single-day seawater transparency fusion data is generated.

In order to make the aforementioned objects, features and advantages of the present application more comprehensible, preferred embodiments accompanied with figures are described in detail below.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are required to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can be obtained from the drawings without inventive effort.

Fig. 1 is a flowchart illustrating a method for generating seawater transparency fusion data according to an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram illustrating a forecast data generation countermeasure network and a converged data generation countermeasure network provided by an embodiment of the present application;

FIG. 3 is a schematic diagram illustrating a network structure of a prediction data generator according to an embodiment of the present disclosure;

fig. 4 (a) - (c) show a second schematic network structure of a prediction data generator according to the embodiment of the present application;

FIG. 5 is a schematic diagram illustrating a network structure of a prediction data discriminator according to an embodiment of the present application;

fig. 6 is a schematic diagram illustrating a network structure of a converged data generator provided in an embodiment of the present application;

fig. 7 is a schematic diagram illustrating a network structure of a satellite data discriminator according to an embodiment of the present application;

fig. 8 is a schematic structural diagram illustrating an apparatus for generating seawater transparency fusion data according to an embodiment of the present application;

fig. 9 shows a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The components of the embodiments of the present application, as generally described and illustrated in the figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, presented in the accompanying drawings, is not intended to limit the scope of the claimed application, but is merely representative of selected embodiments of the application. Every other embodiment that can be obtained by a person skilled in the art without making creative efforts based on the embodiments of the present application falls within the protection scope of the present application.

Through research, the seawater transparency refers to the maximum visible depth of the transparent scale which is vertically sunk into seawater, and is closely related to the physical properties, chemical compositions, suspended matters in the seawater and the dynamic state of the ocean. The method for obtaining the seawater transparency with full coverage, high space-time resolution and high precision has important significance for researching the physical and chemical properties of the seawater, monitoring the marine ecology, producing the marine fishery and guaranteeing the marine military activities.

At present, in the prior art, seawater transparency data is often obtained through satellite remote sensing and data interpolation processing technologies. However, due to factors such as cloud occlusion, solar flare pollution and satellite orbit clearance, the seawater transparency data acquired in this way is insufficient in terms of spatial coverage, detail characteristics and the like, resulting in poor data quality.

Based on this, the embodiment of the application provides a method and a device for generating seawater transparency fusion data, which generate a temporal-spatial change rule for resisting the network excavation of seawater transparency through pre-estimated data, generate pre-estimated seawater transparency fusion data, and realize missing data reconstruction; and then an antagonistic network is generated through the fusion data to realize the fusion of the seawater transparency fusion data and the high-precision satellite remote sensing seawater transparency, and finally the full-coverage high-precision single-day seawater transparency fusion data is generated.

Referring to fig. 1, fig. 1 is a flowchart illustrating a method for generating seawater transparency fusion data according to an embodiment of the present disclosure. As shown in fig. 1, a generation method provided in an embodiment of the present application includes:

s101, obtaining historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days after the target area. m is a positive integer.

In specific implementation, the ranges of 103.0 to 124.5 ° E and 2.5 to 24.0 ° N may be selected as the target region. The region has complex coastlines, numerous islands, large water depth change, numerous rivers afflux into coastal sea regions, complex spatial distribution characteristics of seawater transparency and obvious seasonal change, and is suitable for research and development and experimental verification of a seawater transparency fusion model (consisting of a predicted data generation countermeasure network and a fusion data generation countermeasure network).

S102, inputting the historical seawater transparency fusion data into a pre-trained estimation data generation generator in an countermeasure network to obtain estimation seawater transparency fusion data of the target area in the (m + 1) th day.

The pre-estimated data generation countermeasure network comprises the pre-estimated data generator and a pre-estimated data discriminator, and the pre-estimated data generation countermeasure network is obtained by mutual countermeasure training of the pre-estimated data generator and the pre-estimated data discriminator. The estimated data discriminator is used for discriminating the authenticity of the estimated seawater transparency fusion data of the m +1 day.

S103, inputting the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into fusion data which is trained in advance to generate fusion data generators in the countermeasure network, and obtaining target seawater transparency fusion data of the m +1 th day of the target area.

The fusion data generation countermeasure network comprises the fusion data generator, a global data discriminator and a satellite data discriminator; the fusion data generation countermeasure network is obtained by the mutual countermeasure of the fusion data generator and the global data arbiter and the mutual countermeasure training of the fusion data generator and the satellite data arbiter. The global data discriminator is used for discriminating the data consistency and continuity between an observed region and an unobserved region of satellite remote sensing in the target seawater transparency fusion data of the m +1 day; and the satellite data discriminator is used for discriminating the authenticity of data of an observation area remotely sensed by a satellite in the target seawater transparency fusion data of the m +1 day.

Referring to fig. 2, fig. 2 is a schematic structural diagram of an advance data generation countermeasure network and a converged data generation countermeasure network according to an embodiment of the present application. The training process of generating the confrontation network by the predicted data and generating the confrontation network by the fusion data will be described in detail with reference to fig. 2.

In one possible embodiment, the pre-estimated data generating countermeasure network and the fused data generating countermeasure network are trained by:

in specific implementation, because the historical seawater transparency fusion data of a long time sequence are not generated, the countermeasure network generated by the initial estimation data and the countermeasure network generated by the initial fusion data can be pre-trained on the basis of the historical seawater transparency data; and then according to the seawater transparency fusion data generated by the pre-training, generating an confrontation network for the pre-trained initial estimation data and generating the confrontation network for the initial fusion data to perform fine tuning training so as to complete the training of generating the confrontation network for the initial estimation data and generating the confrontation network for the initial fusion data. Specifically, the training step comprises:

s201, inputting the historical seawater transparency data of the target area into an initial prediction data generator in a countermeasure network to obtain first historical prediction seawater transparency data predicted by the initial prediction data generator.

In this step, the OCEANCOLOUR _ GLO _ OPTICS _ L4_ REP _ OBSERVATIONS _009 081 daily seawater transparency data set of the CMEMS global marine water color product may be used as historical seawater transparency data for pre-training of the model. Specifically, as shown in fig. 2, historical seawater transparency data of previous p days (T-p-T days, where T and p are positive integers) are input into an initial estimation data generator in the countermeasure network to obtain first historical estimated seawater transparency data estimated by the initial estimation data generator, that is, estimated seawater transparency data of T +1 day.

S202, inputting the first historical estimated seawater transparency data and the historical seawater transparency data into initial estimated data to generate an initial estimated data discriminator in an countermeasure network, and obtaining a first discrimination result of the initial estimated data discriminator on the first historical estimated seawater transparency data.

Inputting estimated seawater transparency data of the T +1 day and seawater transparency data of the T +1 calendar history into initial estimated data to generate an initial estimated data discriminator in the countermeasure network, extracting spatial features by the estimated data discriminator, and discriminating the truth of the estimated seawater transparency data of the T +1 day according to the feature difference between the estimated seawater transparency data of the T +1 day and the seawater transparency data of the T +1 calendar history.

S203, pre-training the confrontation network generated by the initial estimation data based on the historical seawater transparency data, the first historical estimation seawater transparency data and the first judgment result to obtain the confrontation network generated by the estimation data after the pre-training is finished and the second historical estimation seawater transparency data estimated by the initial estimation data generator when the pre-training is finished.

Here, a first discrimination result of the first historical estimated seawater transparency data by the estimated data discriminator is fed back to the estimated data generator, so that the estimated seawater transparency data is further improved. The estimated data of T +1 day output by the estimated seawater transparency data after improvement is more in line with the actual situation than the estimated seawater transparency data of the previous T +1 day, so that the estimated data discriminator is enabled to further improve the discrimination capability. Repeating the process, the estimated data generator and the estimated data discriminator mutually confront each other, and finally, the two are balanced to finish the pre-training of the generation of the confrontation network for the estimated data.

S204, inputting the second historical estimated seawater transparency data and the historical satellite remote sensing seawater transparency data of the target area into initial fusion data to generate an initial fusion data generator in a countermeasure network, and obtaining first historical seawater transparency fusion data generated by the initial fusion data generator.

In one possible implementation, satellite remote sensing seawater transparency data such as SeaWiFS/OrbView-2, MERIS/ENVISAT, MODIS/Aqua, VIIRS/NPP, VIIRS/JPSS1, OLCI/Sentinel-3A and OLCI/Sentinel-3B can be selected for model development and experimental verification. The data are from the global color project of the European space agency (http:// www. GlobColour. Info /), the spatial resolution is 4km, and the time resolution is 24 hours, it needs to be stated that the spatial coverage of the seawater transparency data of the satellite remote sensing is low due to the influence of factors such as cloud shielding, solar flare pollution, satellite orbit clearance and the like, for example, 5/1/2022, and the daily spatial coverage of seawater transparency of MODIS/Aqua, VIIRS/NPP, VIIRS/JPSS1, OLCI/Sentein-3A, OLCI/Sentein-3B is 10.1%, 10.4%, 6.97%, 25.1% and 28.3%, even if the seawater transparency data of the satellite remote sensing are combined to obtain the seawater transparency data of the multi-source satellite remote sensing, the combined daily spatial coverage is only 56.3%. In order to construct a high-precision seawater transparency fusion model, the multi-source satellite remote sensing seawater transparency data obtained in a combined mode can be selected as historical satellite remote sensing seawater transparency data.

In this step, as shown in fig. 2, the second historical estimated seawater transparency data (estimated seawater transparency data for T +1 day estimated by the initial estimated data generator when pre-training is completed) and the remote sensing seawater transparency data of the T +1 calendar history satellite may be input into the initial fusion data generator to generate initial fusion data in the countermeasure network, so as to obtain the first historical seawater transparency fusion data, i.e., the seawater transparency fusion data for T +1 calendar history, generated by the initial fusion data generator.

S205, poisson fusion is carried out on the historical satellite remote sensing seawater transparency data and the second historical estimated seawater transparency data, and historical seawater transparency Poisson fusion data are obtained.

In the step, poisson fusion is carried out on the historical satellite remote sensing seawater transparency data of the day T +1 and the second historical estimated seawater transparency data through a Poisson equation, and historical seawater transparency Poisson fusion data, namely the historical seawater transparency Poisson fusion data of the calendar T +1, is obtained.

S206, inputting the historical seawater transparency Poisson fusion data and the first historical seawater transparency fusion data into initial fusion data to generate an initial global data discriminator in a countermeasure network, and obtaining a second discrimination result of the initial global data discriminator on the first historical seawater transparency fusion data.

S207, inputting the historical satellite remote sensing seawater transparency data and the first historical seawater transparency fusion data into initial fusion data to generate an initial satellite data discriminator in a countermeasure network, and obtaining a third discrimination result of the initial satellite data discriminator on the first historical seawater transparency fusion data.

S208, pre-training the initial fusion data generation countermeasure network based on the historical seawater transparency Poisson fusion data, the historical satellite remote sensing seawater transparency data, the first historical seawater transparency fusion data, the second judgment result and the third judgment result to obtain a fusion data generation countermeasure network after pre-training and second historical seawater transparency fusion data generated by the initial fusion data generator after pre-training.

Here, the global data discriminator discriminates authenticity through spatial feature difference between the T +1 calendar history sea water transparency fusion data and the T +1 calendar history sea water transparency Poisson fusion data, the satellite data discriminator discriminates authenticity through spatial feature difference between the T +1 calendar history sea water transparency fusion data and the T +1 calendar history sea water transparency satellite remote sensing sea water transparency data, and the authenticity is fed back to the fusion data generator after the two are connected, so that the fusion data generator is further improved. The T +1 day seawater transparency fusion data output by the improved fusion data generator is more in line with the actual situation than the previous seawater transparency fusion data, so that the global data discriminator and the satellite data discriminator are enabled to further improve the discrimination capability. Repeating the process, the fused data generator, the global data arbiter and the satellite data arbiter mutually confront each other, and finally balance is achieved, at this time, the T +1 calendar history sea water transparency fused data has better global consistency, continuity and satellite observation authenticity. Because the space coverage of the satellite remote sensing seawater transparency data every day has randomness, the authenticity judgment of the coverage of the whole target area can be realized through the training of a large number of samples.

S209, respectively re-inputting the second historical seawater transparency fusion data and the historical satellite remote sensing seawater transparency data into the pre-trained estimated data generation countermeasure network and the pre-trained fusion data generation countermeasure network, and performing fine tuning training on the pre-trained estimated data generation countermeasure network and the pre-trained fusion data generation countermeasure network again to obtain the trained estimated data generation countermeasure network and the trained fusion data generation countermeasure network.

In specific implementation, after the pre-training is completed, the second historical seawater transparency fusion data and the historical satellite remote sensing seawater transparency data generated during the pre-training are input into the seawater transparency fusion model again as training data for fine tuning training, so that the seawater transparency fusion model is adaptive to the seawater transparency fusion data. The fine tuning training process is similar to the pre-training process, the network structure is the same, but the input data for training is changed; specifically, step S209 may include:

s2091, re-inputting the second historical seawater transparency fusion data into pre-trained estimation data to generate an confrontation network, and obtaining third historical estimation seawater transparency data estimated by the initial estimation data generator and a fourth judgment result of the initial estimation data discriminator on the third historical estimation seawater transparency data.

S2092, determining a first loss function of the initial prediction data discriminator and a second loss function of the initial prediction data generator based on the second historical seawater transparency fusion data, the third historical prediction seawater transparency data and the fourth discrimination result.

Wherein, the prediction data generates a loss function for the network, including the resistance loss and the content loss. The combined training direction between the estimated data generator and the estimated data discriminator is restrained by the loss resistance, the T-day seawater transparency fusion data and the T-day estimated data are respectively used as true and false samples to be input into the discriminator, the game between the generator and the discriminator is carried out, the generator is forced to generate the T-day estimated data deceiving the discriminator as far as possible, and the discriminator needs to distinguish the T-day estimated data from the T-day seawater transparency fusion data as far as possible. To minimize the loss of the generator, maximize the loss of the arbiter, the penalty loss is expressed as:

in the formula, G ₁ Representing the estimated data generator, D representing the estimated data discriminator, x representing the seawater transparency fusion data,

representing estimated seawater transparency data, P _data Represents the distribution of the sea water transparency fusion data>

Indicating the distribution of the estimated data and E indicating the expectation.

And the content loss is used for measuring the difference between the T-day seawater transparency fusion data and the T-day estimation data, wherein l is adopted ₁ And distance, minimizing the sum of absolute differences of the seawater transparency fusion data and the estimated data of the T day. The research area of the embodiment of the application covers ocean and land, the transparency gradient of the seawater near the ocean-land boundary is large, and if the ocean and the land are not distinguished when the content loss is calculated, the error of the transparency of the seawater near the ocean-land boundary is large. Thus, the content loss is calculated separately as ocean and land:

L _content ＝L _land +μL _ocean

/>

wherein μ is a constant, and in the case of concrete implementation, it is preferably 6; m is _ocean Denotes a sea mask, m _land Representing a land mask; II (8230); II ₁ Is represented by ₁ A distance; an indication of multiplication of the corresponding element; wherein the ocean mask is used for describing ocean distribution of the target area, and specifically, the ocean mask can be representedShown as a matrix; the matrix corresponds to the target area, and each element in the matrix corresponds to one grid area in the target area; if a certain grid region in a certain target region belongs to the ocean, the element value corresponding to the grid region in the ocean mask is 1; if a certain grid region in a certain target region belongs to land, the element value corresponding to the grid region in the ocean mask is 0; accordingly, the land mask is used to describe the land distribution of the target area, and specifically, the land mask can also be represented as a matrix; if a certain grid area in a certain target area belongs to the ocean, the element value corresponding to the grid area in the land mask is 0; if a certain grid region in a certain target region belongs to the land, the element value corresponding to the grid region in the land mask is 1.

In summary, the first loss function of the design prediction data discriminator in the present application can be expressed as:

in the formula, D represents an estimated data discriminator; g ₁ A representation prediction data generator; l is _D A first loss function representing a predictor; x represents historical seawater transparency fusion data;

representing historical pre-estimated seawater transparency data; p _data Represents the distribution of historical seawater transparency fusion data, and>

representing the distribution of historical estimated seawater transparency data;

the second loss function of the prediction data generator includes an immunity loss and a content loss, and the formula can be expressed as:

in the formula (I), the compound is shown in the specification,

a second loss function representing a prediction data generator; m is _land Representing a land mask for characterizing a land distribution of the target area; m is _ocean Representing a ocean mask for characterizing an ocean distribution of the target area; an indication of multiplication of the corresponding element; II (8230); II ₁ Represents l ₁ A distance; λ and μ are constants.

S2093, adjusting network parameters of the initial estimated data discriminator based on the first loss function, and adjusting network parameters of the initial estimated data generator based on the second loss function to perform fine tuning training on the pre-trained estimated data generation confrontation network until confrontation training between the initial estimated data discriminator and the initial estimated data generator is balanced, and obtaining trained estimated data to generate the confrontation network.

S2094, obtaining the trained estimated data to generate fourth historical estimated seawater transparency data estimated by the estimated data generator when the fine tuning training of the countermeasure network is completed.

S2095, inputting the fourth historical estimated seawater transparency data and the historical satellite remote sensing seawater transparency data into pre-trained fusion data to generate an confrontation network, and obtaining third historical seawater transparency fusion data generated by the initial fusion data generator, a fifth judgment result of the initial global data discriminator on the third historical seawater transparency fusion data and a sixth judgment result of the initial satellite data discriminator on the first historical seawater transparency fusion data.

S2096, based on the fourth history estimated seawater transparency data, the third history seawater transparency fusion data and the fifth discrimination result, determining a third loss function of the initial global data discriminator; determining a fourth loss function of the initial satellite data discriminator based on the third history estimated seawater transparency data and the sixth discrimination result; and determining a fifth loss function of the initial fusion data generator based on the third history estimated seawater transparency data, the fifth discrimination result and the sixth discrimination result.

Likewise, fusing data to generate a loss function against the network includes both opposed losses and content losses. The challenge loss consists of two parts: the countermeasure loss of the fusion data generator and the global data discriminator and the countermeasure loss of the fusion data generator and the satellite data discriminator; the formula can be expressed as:

in the formula, L _adv (G ₂ ) Representing the penalty of the fusion data generator, G ₂ Representing fused data generators, D _a Representing global data discriminators, D _s A satellite data discriminator is shown in which the satellite data,

the distribution of the seawater transparency fusion data is shown, and x represents the seawater transparency fusion data.

The content loss is used for measuring the difference between the T +1 calendar history seawater transparency fusion data and the T +1 calendar history seawater transparency Poisson fusion data and is also used for measuring the difference between the T +1 calendar history seawater transparency fusion data and the T +1 calendar history satellite remote sensing seawater transparency data in a satellite observation area. When the difference between the two is measured, in order to ensure global consistency and continuity, l is adopted ₂ The distance and gradient calculate the data difference. When the difference between the two is measured, l is adopted ₁ Distance and structural similarity calculate data differences. The content loss is expressed as:

in the formula, x' represents estimated data Poisson fusion data;

representing multi-source satellite remote sensing seawater transparency data; alpha, beta and gamma are constants, and can be respectively 4, 2 and 80 in specific implementation; />

Representing a gradient operator; II (8230); II ₁ Is represented by ₁ A distance; II (8230); II ₂ Is represented by ₂ A distance; m is _sat Representing the observation area of the satellite, likewise, m _sat Can also be represented as a matrix; the matrix corresponds to a target area, each element in the matrix corresponds to one grid area in the target area, and if one grid area in one target area belongs to an area observed by satellite remote sensing, m is _sat The element value corresponding to the grid area is 1; if a certain grid region in a certain target region does not belong to a region observed by satellite remote sensing, m is _sat The element value corresponding to the grid area is 1; />

Denotes structural similarity, μ _x Denotes the mean value of x>

Represents->

Is based on the mean value of (4)>

Represents the variance of x +>

Represents->

In (b) based on the variance of (c), in>

Denotes x and->

Covariance of c ₁ And c ₂ As a constant, 0.0004 and 0.0036, respectively, can be adopted in specific implementation.

In summary, the third loss function for designing the global data arbiter in the present application can be expressed as:

in the formula, D _a A representation global data arbiter; g ₂ A generator of the fused data is represented,

a third penalty function representing a global data arbiter; x represents historical seawater transparency fusion data; x' represents historical seawater transparency poisson fusion data; p _data1 Representing the distribution of the Poisson fusion data of historical seawater transparency; />

Representing the distribution of historical seawater transparency fusion data.

The fourth loss function of the fusion data generation countermeasure network satellite data arbiter is expressed as:

in the formula, D _s A representation satellite data discriminator;

a fourth loss function representing a satellite data arbiter; />

Representing historical satellite remote sensing seawater transparency data; p is _data2 And the distribution of the historical satellite remote sensing seawater transparency data is shown.

The converged data generation countermeasure networkThe fifth loss function of the medium fused data generator is expressed as:

in the formula (I), the compound is shown in the specification,

a fifth loss function representing a fused data generator; II | \8230 | ₁ Is represented by ₁ A distance; II | \8230 | ₂ Is represented by ₂ A distance; m is a unit of _sat Representing a satellite observation region in the target region; />

Representing a gradient operator; />

Denotes structural similarity, μ _x Denotes the mean value of x>

Represents->

Is based on the mean value of (4)>

Variance representing x, <' > based on>

Represents->

The variance of (a) is determined,

denotes x and->

Covariance of c ₁ And c ₂ Is a constant; δ, α, β, and γ are constants.

S2097, adjusting network parameters of the initial global data discriminator based on the third loss function, adjusting network parameters of the initial satellite data discriminator based on the fourth loss function, adjusting network parameters of the initial fusion data generator based on the fifth loss function, and generating an confrontation network and performing fine tuning training on fusion data which are trained in advance until confrontation training between the initial global data discriminator and the initial fusion data generator and confrontation training between the initial satellite data discriminator and the initial fusion data generator are balanced, and obtaining trained fusion data to generate the confrontation network.

It should be noted that, because the fine tuning training and the pre-training are similar in process and the network structure is the same, the expressions of the loss functions used in the fine tuning training and the pre-training are also the same, except that the change of the input data of the training causes the change of the data input to the loss functions. Therefore, the loss function in the pre-training process can be designed with reference to the loss function in the fine-tuning training process, and the same technical effect can be achieved, which is not described in detail herein.

Further, after obtaining the trained pre-estimated data to generate the countermeasure network, step S101, obtaining historical seawater transparency fusion data m days before the target area, includes:

s1011, acquiring the trained fused data to generate third history seawater transparency fused data generated by the initial fused data generator when the fine tuning training of the confrontation network is completed;

s1012, screening out historical seawater transparency fusion data of the previous m days from the third historical seawater transparency fusion data.

Referring to fig. 3 and fig. 4 (a) - (c), fig. 3 is a schematic diagram of a network structure of an estimated data generator according to an embodiment of the present disclosure; fig. 4 (a) - (c) are schematic diagrams of network structures of an estimation data generator according to the second embodiment of the present application. As shown in fig. 3, an estimated data generator provided in an embodiment of the present application includes: a prediction encoder of an n-layer network structure and a prediction decoder of the n-layer network structure; the pre-estimated encoder comprises m encoding paths; the first coding path comprises a spatial dimension reduction unit (SpaceToDepth) and n first residual error structures (ResBlock I ↓) which are connected in sequence; each of the rest encoding paths comprises a spatial dimension reduction unit (SpaceToDepth), a first residual structure (ResBlock I ↓), n convolution gating cycle units (convGRU) and n-1 second residual structures (ResBlock II + ResBlock II ↓); the spatial dimension reduction unit is connected with the first residual error structure, and the convolution gating circulating unit of each layer is connected with the first residual error structure or connected with the convolution gating circulating unit of the upper layer through a second residual error structure; the convolution gating circulation unit of each layer in the former coding path is connected with the convolution gating circulation unit of the corresponding layer in the latter coding path; the pre-prediction decoder comprises n third residual error structures (ResBlock II ≠ and) and an output block (OutputLock); the convolution gating circulation unit of each layer in the mth coding path is connected to the third residual error structure of the corresponding layer in the pre-estimated decoder; the m coding paths correspond to historical seawater transparency fusion data of previous m days one by one.

As shown in fig. 4 (a) - (c), fig. 4 (a) is a schematic structural diagram of a first residual structure (ResBlock I ↓); the first residual structure consists of a rectifying linear unit (Relu) and a convolution and down-sampling unit and is used for realizing the down-sampling of seawater transparency input data and the initialization of the memory state of the convGRU unit; fig. 4 (b) is a schematic structural diagram of a configurable residual structure (ResBlock II), which is composed of group normalization, relu activation function, convolution unit, and is configurable to be upsampled, downsampled, and null; specifically, when the Sample unit in the ResBlock II is configured to be empty, the configurable residual structure does not up-down Sample data; when the Sample cell in ResBlock II is configured as an average pooling cell, the configurable residual structure (i.e., resBlock II ↓) is used to implement downsampling of data; when the Sample unit in ResBlock II is configured as a bilinear interpolation unit, a configurable residual structure (i.e., a third residual structure ResBlock II ×) is used to implement upsampling of data. Fig. 4 (c) is a schematic structural diagram of an output block.

Illustratively, as shown in fig. 3, the prediction data generator is composed of an encoder having an encoding function and a decoder having a decoding function. The encoder and the decoder are connected to form a network structure with the depth of 5 layers, the reduction path of the encoder comprises a space dimension reduction unit, a convolution gating cycle unit (convGRU), a first residual error structure and a second residual error structure, and the training speed of the network and the processing speed of the model are increased due to the fact that the size of data processed by each layer is reduced. The spatial dimension reduction unit rearranges the spatial data dimensions to depth dimensions, implementing downsampling. The convGRU unit changes full-connection calculation during information transmission in the gating circulation unit into convolution calculation, and feature information extraction of seawater transparency data in time and space dimensions is achieved. In order to relieve the gradient disappearance problem caused by increasing the network depth, a first residual structure and a second residual structure are designed in a pre-estimated data generator, wherein the first residual structure consists of a rectifying linear unit (Relu), a convolution unit and a down-sampling unit and is used for realizing the down-sampling of seawater transparency input data and the initialization of a memory state of a convGRU unit; the second residual structure is composed of group normalization, relu activation function and convolution unit, and can be configured with up-sampling and down-sampling units. In the reduction path, the second residual structure is used for realizing connection between different levels of convGRU units, and the configurable average pooling unit realizes down-sampling of data.

The amplifying path of the decoder comprises a third residual error structure (ResBlock II ℃) and an output block, the third residual error structure is provided with a bilinear interpolation unit and used for realizing layer-by-layer restoration of the data size, the output block consists of group normalization, a Relu activation function, convolution and a hyperbolic tangent unit activation function, the data size and the channel are further restored, and the estimated data is output. The second residual structure on each layer of amplification path is in jump connection with the convGRU on the same layer of reduction path, so that the shallow feature on the reduction path is effectively fused with the deep feature on the amplification path, the estimated data generator obtains better feature extraction capability, and the loss generated by the amplification path during decoding is reduced.

In a specific implementation, step S102 may include:

and S1021, correspondingly inputting the historical seawater transparency fusion data of the previous m days into m coding paths in the pre-estimation encoder to obtain the seawater transparency fusion data output by each layer of convolution gating circulation unit in the last coding path.

And S1022, inputting the seawater transparency fusion data output by each layer of convolution gating circulation unit into the second residual error structure of the corresponding layer in the pre-estimation decoder to obtain the pre-estimation seawater transparency fusion data of the m +1 th day output by the output block of the pre-estimation decoder.

Referring to fig. 5, fig. 5 is a schematic diagram of a network structure of an estimated data discriminator according to an embodiment of the present application. As shown in fig. 5, the prediction data discriminator illustratively includes 6 layers of convolutional neural networks, except for the last layer of convolutional layer, a leakage relu activation function is used after each layer of convolutional, the discriminator separately discriminates all the Patch by dividing the input data into N × N sized matrices, obtains the discrimination result of each Patch, and finally outputs true or false according to the average value of the discrimination results of all the patches.

Referring to fig. 6, fig. 6 is a schematic diagram of a network structure of a converged data generator according to an embodiment of the present application. As shown in fig. 6, the fused data generator includes a fused encoder of a p-layer network structure and a fused decoder of the p-layer network structure; the fusion encoder comprises a spatial dimension reduction unit and p second residual error structures; the fused decoder comprises p third residual structures and an output block; the second residual structure of each layer in the fused encoder is connected to the third residual structure of the corresponding layer in the fused decoder;

as shown in fig. 6, illustratively, the fused data generator is composed of a fusion encoder and a fusion decoder of a 5-layer network structure, as shown in fig. 6, the fusion encoder includes a spatial dimension reduction unit (SpaceToDepth), a second residual structure (ResBlock II + ResBlock II ↓); the fused decoder comprises a third residual structure (ResBlock II ×) and an output block (outputbulk). The second residual structure is mainly used for connecting a shallow layer and a deep layer network in the model, so that the problem that the gradient is difficult to flow to the shallow layer network in the back propagation process to cause gradient disappearance is solved. The fusion encoder adopts ResBlock II ↓ configured with an average pooling unit to realize data size reduction, and the training speed of the network and the processing speed of the model are increased. And the fusion decoder adopts a third residual error structure configured with average bilinear interpolation to realize the restoration of the data size. The same level of fusion encoder and the same level of fusion decoder adopt jump connection, and the shallow feature on the reduction path is fused with the deep feature on the amplification path, so that the fusion data generator obtains better feature extraction capability, and the loss generated by the amplification path during decoding is reduced.

In one possible implementation, the global data discriminator is used for distinguishing the T +1 day seawater transparency fusion data from the T +1 day seawater transparency Poisson fusion data; the network structure of the global data discriminator is similar to that of the estimated data discriminator in fig. 5, and the same technical effect can be achieved, which is not described in detail herein.

Referring to fig. 7, fig. 7 is a schematic diagram of a network structure of a satellite data discriminator according to an embodiment of the present disclosure. As shown in fig. 7, the satellite data arbiter employs a 6-layer convolutional neural network, and adds a spatial attention module SpatialAttention at layers 2 and 4. A maximum pooling unit Max Pooling and a random pooling unit StochasticPooling are adopted in the spatial attention module spatialAttention to enhance the generalization capability of the model.

In experimental verification, an experimental hardware platform adopted in the embodiment of the application includes an Intel Xeon Silver 4212R as a CPU and NVIDIA Tesla V100s as a GPU, and a CUDA parallel framework and a cuDNN acceleration library are configured. The deep learning framework used for training was PyTorch, the batch size was 64, and the epoch was 1000 times. The prediction data generation countermeasure network and the fusion data generation countermeasure network both use an ADAM optimizer, the learning rate of the generated model is 0.0002, the learning rate of the discrimination model is 0.0002, and the prediction data generation countermeasure network and the fusion data generation countermeasure network are alternately optimized.

The time to train the dataset ranged from 1/2002 to 31/2017/12/2019, the time to validate the dataset ranged from 1/2018 to 31/2019/12/2020 to 31/2020/1/2020. The pre-estimation data generation countermeasure network is pre-trained by adopting the CMEMS seawater transparency, and fine adjustment is carried out after the seawater transparency fusion data is generated, so that the pre-estimation data model is adapted to the seawater transparency fusion data. And the fusion data generation confrontation network model is trained by adopting multi-source satellite remote sensing seawater transparency data and pre-estimated seawater transparency fusion data to generate seawater transparency fusion data.

The experiment uses indexes such as Root Mean Square Error (RMSE), mean Absolute Error (MAE) and mean relative error (ARE) to evaluate the effect of seawater transparency fusion. The fusion data source is satellite remote sensing seawater transparency data such as SeaWiFS/OrbView-2, MERIS/ENVISAT, MODIS/Aqua, VIIRS/NPP, OLCI/Sentinel-3A and OLCI/Sentinel-3B, and the inspection data source is VIIRS/JPSS1 satellite remote sensing seawater transparency.

Experiments prove that the space coverage of the seawater transparency fusion data generated in the embodiment of the application is 100%, and the reconstruction of missing data can be realized; meanwhile, the error of the seawater transparency fusion data generated by the embodiment of the application is smaller than that of the seawater transparency fusion data generated by means of averaging, kriging interpolation and the like, and the seawater transparency fusion data generated by the embodiment of the application has higher precision.

According to the method for generating the seawater transparency fusion data, historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days are input into pre-trained estimation data to generate an anti-network, and estimation seawater transparency fusion data m +1 days generated by an estimation data generator are obtained; and inputting the satellite remote sensing seawater transparency data of the m +1 day and the estimated seawater transparency fusion data of the m +1 day into pre-trained fusion data to generate an antagonistic network, and obtaining target seawater transparency fusion data of the m +1 day of a target area generated by the fusion data generator. Therefore, a time-space change rule for fighting against the network excavation seawater transparency can be generated through the estimated data, estimated seawater transparency fusion data are generated, and missing data reconstruction is achieved; and then, fusion of seawater transparency fusion data and high-precision satellite remote sensing seawater transparency is realized through a countermeasure network generated by fusing data, and finally, full-coverage and high-precision single-day seawater transparency fusion data is generated.

Referring to fig. 8, fig. 8 is a schematic structural diagram of a device for generating seawater transparency fusion data according to an embodiment of the present application. As shown in fig. 8, the generating means 800 comprises:

the acquisition module 810 is used for acquiring historical seawater transparency fusion data m days before a target area and satellite remote sensing seawater transparency data m +1 days of the target area; m is a positive integer;

a first determining module 820, configured to input the historical seawater transparency fusion data into a pre-trained pre-estimated data generator in an countermeasure network to obtain pre-estimated seawater transparency fusion data of the target area at day m + 1; the predicted data generation countermeasure network comprises a predicted data generator and a predicted data discriminator, and is obtained by mutual countermeasure training of the predicted data generator and the predicted data discriminator;

the second determining module 830 is configured to input the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into fusion data which is trained in advance to generate fusion data generators in the countermeasure network, so as to obtain target seawater transparency fusion data of the m +1 th day of the target area; the fusion data generation countermeasure network comprises the fusion data generator, a global data discriminator and a satellite data discriminator; the fusion data generation countermeasure network is obtained by the mutual countermeasure of the fusion data generator and the global data arbiter and the mutual countermeasure training of the fusion data generator and the satellite data arbiter.

Further, the generating device 800 further includes a training module; the training module is used for training to obtain the pre-estimated data generation countermeasure network and the fusion data generation countermeasure network in the following modes:

inputting the historical seawater transparency data of the target area into initial prediction data to generate an initial prediction data generator in an antagonistic network, and obtaining first historical prediction seawater transparency data predicted by the initial prediction data generator;

inputting the second historical pre-estimated seawater transparency data and historical satellite remote sensing seawater transparency data of the target area into initial fusion data to generate an initial fusion data generator in a countermeasure network, and obtaining first historical seawater transparency fusion data generated by the initial fusion data generator;

and respectively inputting the second historical seawater transparency fusion data and the historical satellite remote sensing seawater transparency data again into the pre-trained pre-estimated data generation countermeasure network and the pre-trained fusion data generation countermeasure network, and performing fine tuning training on the pre-trained pre-estimated data generation countermeasure network and the pre-trained fusion data generation countermeasure network again to obtain a trained pre-estimated data generation countermeasure network and a trained fusion data generation countermeasure network.

Further, when the obtaining module 810 is configured to obtain the historical seawater transparency fusion data m days before the target area, the obtaining module 810 is configured to:

acquiring third history seawater transparency fusion data generated by the initial fusion data generator when the fine-tuning training of the trained fusion data generation countermeasure network is completed;

Further, when the training module is configured to re-input the second historical seawater transparency fusion data into pre-trained estimated data to generate an confrontation network, and re-generate the confrontation network and perform fine tuning training on the pre-trained estimated data to obtain trained estimated data to generate the confrontation network, the first loss function of the estimated data discriminator in the pre-estimated data generation confrontation network includes:

representing historical predicted seawater transparency data; p _data Represents the distribution of historical seawater transparency fusion data, and>

in the formula (I), the compound is shown in the specification,

a second penalty function representing a generator of the prediction data; m is _land Representing a land mask for characterizing a land distribution of the target area; m is _ocean Representing a ocean mask for characterizing an ocean distribution of the target area; an indication of multiplication of the corresponding element; II (8230); II ₁ Is represented by ₁ A distance; λ and μ are constants.

Further, when the training module is configured to re-input the historical satellite remote sensing seawater transparency data into the fusion data generated as the pre-training is completed to generate the countermeasure network, and perform fine tuning training on the fusion data generated as the pre-training is completed again to obtain the trained fusion data generated countermeasure network, a third loss function of the global data discriminator in the fusion data generated countermeasure network is represented as:

Representing the distribution of historical seawater transparency fusion data;

in the formula D _s A representation satellite data discriminator;

a fourth loss function representing a satellite data arbiter; />

in the formula (I), the compound is shown in the specification,

a fifth loss function representing a fused data generator; II (8230); II ₁ Is represented by ₁ A distance; II | \8230 | ₂ Is represented by ₂ A distance; m is a unit of _sat Representing a satellite observation region in the target region; />

Representing a gradient operator; />

Denotes structural similarity, μ _x Denotes the mean value of x>

Represents->

In the mean value of (a), based on>

Variance representing x, <' > based on>

Represents->

The variance of (a) is determined,

denotes x and->

Covariance of c ₁ And c ₂ Is a constant; δ, α, β, and γ are constants.

Furthermore, the prediction data generator comprises a prediction encoder of an n-layer network structure and a prediction decoder of the n-layer network structure; the pre-estimated encoder comprises m encoding paths; the first coding path comprises a space dimension reduction unit and n sequentially connected first residual error structures; each of the other coding paths comprises a space dimension reduction unit, a first residual error structure, n convolution gating circulation units and n-1 second residual error structures; the space dimension reduction unit is connected with the first residual error structure, and the convolution gating circulating unit of each layer is connected with the first residual error structure or connected with the convolution gating circulating unit of the upper layer through a second residual error structure; the convolution gating circulation unit of each layer in the former coding path is connected with the convolution gating circulation unit of the corresponding layer in the latter coding path; the pre-estimation decoder comprises n third residual error structures and an output block; the convolution gating circulation unit of each layer in the m-th coding path is connected to the third residual error structure of the corresponding layer in the pre-estimation decoder; the m coding paths correspond to historical seawater transparency fusion data of previous m days one by one;

further, when the first determining module 820 is configured to input the historical seawater transparency fusion data into a pre-trained predictive data generator in an confrontation network to obtain the predictive seawater transparency fusion data of the target area at day m +1, the first determining module 820 is configured to:

inputting the seawater transparency fusion data output by each layer of convolution gating circulation unit into the second residual error structure of the corresponding layer in the pre-estimated decoder to obtain the pre-estimated seawater transparency fusion data of the m +1 th day output by the output block of the pre-estimated decoder.

when the second determining module 830 is configured to input the satellite remote sensing seawater transparency data of the m +1 th day and the estimated seawater transparency fusion data of the m +1 th day into the fusion data generator in the pre-trained fusion data generation countermeasure network to obtain the target seawater transparency fusion data of the m +1 th day of the target area, the second determining module 830 is configured to:

inputting the seawater transparency fusion data output by the second residual structure of each layer in the fusion encoder into the third residual structure of the corresponding layer in the fusion decoder to obtain the target seawater transparency fusion data of the (m + 1) th day output by the output block of the fusion decoder.

Referring to fig. 9, fig. 9 is a schematic structural diagram of an electronic device according to an embodiment of the present disclosure. As shown in fig. 9, the electronic device 900 includes a processor 910, a memory 920, and a bus 930.

The memory 920 stores machine-readable instructions executable by the processor 910, when the electronic device 900 runs, the processor 910 communicates with the memory 920 through the bus 930, and when the machine-readable instructions are executed by the processor 910, the steps of the method for generating seawater transparency fusion data in the method embodiments shown in fig. 1 to fig. 7 may be executed.

An embodiment of the present application further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the step of the method for generating seawater transparency fusion data in the method embodiments shown in fig. 1 to fig. 7 may be executed.

It can be clearly understood by those skilled in the art that, for convenience and simplicity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus and method may be implemented in other ways. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one logical division, and there may be other divisions when actually implemented, and for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed coupling or direct coupling or communication connection between each other may be through some communication interfaces, indirect coupling or communication connection between devices or units, and may be in an electrical, mechanical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such understanding, the technical solution of the present application or portions thereof that substantially contribute to the prior art may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, an optical disk, or other various media capable of storing program codes.

Finally, it should be noted that: the above-mentioned embodiments are only specific embodiments of the present application, and are used to illustrate the technical solutions of the present application, but not to limit the technical solutions, and the scope of the present application is not limited to the above-mentioned embodiments, although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: those skilled in the art can still make modifications or changes to the embodiments described in the foregoing embodiments, or make equivalent substitutions for some features, within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the exemplary embodiments of the present application, and are intended to be covered by the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A generation method of seawater transparency fusion data is characterized by comprising the following steps:

inputting the historical seawater transparency fusion data into pre-trained pre-estimated data to generate pre-estimated data generators in an antagonistic network, and obtaining pre-estimated seawater transparency fusion data of the m +1 th day of the target area; the predicted data generation countermeasure network comprises a predicted data generator and a predicted data discriminator, and is obtained by mutual countermeasure training of the predicted data generator and the predicted data discriminator;

2. The generation method of claim 1, wherein the pre-estimation data generation countermeasure network and the fusion data generation countermeasure network are trained by:

3. The generation method according to claim 2, wherein the obtaining historical seawater transparency fusion data m days before the target area comprises:

4. The generation method of claim 2, wherein when the second historical seawater transparency fusion data is re-input into the pre-trained predicted data to generate the countermeasure network, and the pre-trained predicted data is re-generated into the countermeasure network and is subjected to the fine tuning training again to obtain the trained predicted data to generate the countermeasure network, the generating of the first loss function of the predicted data discriminator in the countermeasure network by the predicted data comprises:

in the formula, D represents an estimated data discriminator; g ₁ A representation prediction data generator; l is a radical of an alcohol _D A first loss function representing a predictor; x represents historical seawater transparency fusion data;

representing historical pre-estimated seawater transparency data; p is _data Represents the distribution of historical seawater transparency fusion data>

the second loss function of the prediction data generation countermeasure network comprises:

in the formula (I), the compound is shown in the specification,

a second penalty function representing a generator of the prediction data; m is _land Representing a land mask for characterizing a land distribution of the target area; m is a unit of _ocean Representing ocean masks ofCharacterizing a marine distribution of the target area; an indication of multiplication of the corresponding element; II (8230); II ₁ Represents l ₁ A distance; λ and μ are constants.

5. The generation method of claim 4, wherein when the historical satellite remote sensing seawater transparency data is re-input into the pre-trained fusion data generation countermeasure network, and the pre-trained fusion data generation countermeasure network is re-subjected to fine tuning training again to obtain the trained fusion data generation countermeasure network, a third loss function of a global data discriminator in the fusion data generation countermeasure network is represented as:

a third loss function representing a global data arbiter; x represents historical seawater transparency fusion data; x' represents historical seawater transparency poisson fusion data; p _data1 Representing the distribution of the Poisson fusion data of historical seawater transparency; />

Representing the distribution of historical seawater transparency fusion data;

in the formula, D _s A representation satellite data discriminator;

a fourth loss function representing a satellite data arbiter; />

Representing historical satellite remote sensing seawater transparency data; p is a radical of _data2 Representing the distribution of historical satellite remote sensing seawater transparency data;

in the formula (I), the compound is shown in the specification,

a fifth loss function representing a fused data generator; II | \8230 | ₁ Is represented by ₁ A distance; II (8230); II ₂ Represents l ₂ A distance; m is _sat Representing a satellite observation region in the target region; />

Representing a gradient operator; />

Denotes structural similarity, μ _x Means representing x>

Represents->

Is based on the mean value of (4)>

Represents the variance of x +>

Represents->

Is greater than or equal to>

Represents x and +>

Covariance of c ₁ And c ₂ Is a constant; δ, α, β, and γ are constants.

6. A generation method according to claim 1, wherein the prediction data generator comprises a prediction encoder of an n-layer network structure and a prediction decoder of an n-layer network structure; the pre-estimated encoder comprises m encoding paths; the first coding path comprises a space dimension reduction unit and n sequentially connected first residual error structures; each of the other coding paths comprises a space dimension reduction unit, a first residual error structure, n convolution gating circulation units and n-1 second residual error structures; the space dimension reduction unit is connected with the first residual error structure, and the convolution gating circulating unit of each layer is connected with the first residual error structure or connected with the convolution gating circulating unit of the upper layer through a second residual error structure; the convolution gating circulation unit of each layer in the former coding path is connected with the convolution gating circulation unit of the corresponding layer in the latter coding path; the pre-estimation decoder comprises n third residual error structures and an output block; the convolution gating circulation unit of each layer in the m-th coding path is connected to the third residual error structure of the corresponding layer in the pre-estimation decoder; the m coding paths correspond to historical seawater transparency fusion data of previous m days one by one;

7. The generation method according to claim 1, wherein the fused data generator includes a fused encoder of a p-layer network structure and a fused decoder of the p-layer network structure; the fusion encoder comprises a spatial dimension reduction unit and p second residual error structures; the fused decoder comprises p third residual structures and an output block; the second residual structure of each layer in the fused encoder is connected to the third residual structure of the corresponding layer in the fused decoder;

inputting the satellite remote sensing seawater transparency data of the m +1 day and the estimated seawater transparency fusion data of the m +1 day into the fusion encoder to obtain seawater transparency fusion data output by a second residual error structure of each layer in the fusion encoder;

8. A generation apparatus for seawater transparency fusion data, the generation apparatus comprising:

the first determining module is used for inputting the historical seawater transparency fusion data into pre-trained pre-estimated data to generate pre-estimated data generators in an confrontation network to obtain pre-estimated seawater transparency fusion data of the m +1 th day of the target area; the predicted data generation countermeasure network comprises a predicted data generator and a predicted data discriminator, and is obtained by mutual countermeasure training of the predicted data generator and the predicted data discriminator;

9. An electronic device, comprising: a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, the processor and the memory communicating via the bus when the electronic device is running, the machine readable instructions when executed by the processor performing the steps of a method of generating seawater transparency fusion data according to any one of claims 1 to 7.

10. A computer-readable storage medium, having stored thereon a computer program which, when being executed by a processor, carries out the steps of a method for generating seawater transparency fusion data according to any one of claims 1 to 7.