Deep sea color inversion method fusing satellite radiation simulation under multi-level sea color background

Abstract

The invention discloses a deep sea color inversion method fusing satellite radiation simulation under a multi-level sea color background, and the method comprises the steps: carrying out the simulation through employing a simulation algorithm based on an OSOAA radiation transmission model, obtaining simulation data including simulation wave band apparent reflectivity, and constructing a multi-level sea color data set; secondly, constructing a deep sea color inversion network with the orientation and wave band feature joint characterization capability, including a residual network and a binary attention mechanism, and performing sea color inversion based on multi-level sea color data; and finally, evaluating by using a test part of the data set, collecting real satellite remote sensing and sea color data, and testing. According to the method, the problems of low robustness and reliability of deep learning in a complex scene and the like caused by scarcity of current sea color marking data are solved, the problem of end-to-end inversion of sea color parameters such as chlorophyll concentration is solved, and the inversion efficiency is improved.

Description

Technical Field

[0001] This method belongs to the field of deep learning and atmospheric remote sensing, and specifically involves a deep ocean color inversion method that integrates satellite radiation simulation under a multi-level ocean color background. Background Art

[0002] Satellite ocean color observations are an important vehicle for analyzing ocean-atmosphere coupled signals, and their application effectiveness is highly dependent on the accuracy of atmospheric correction techniques. A key step in atmospheric decoupling is the inversion of atmospheric aerosol optical depth (AOT), which requires assumptions about aerosol types. However, deviations in aerosol type assumptions directly affect the accuracy of scattered radiation calculations, and the difficulty in decoupling their dynamic optical properties from satellite observations exacerbates inversion uncertainty. To address these issues, a fast aerosol inversion algorithm based on a lookup table (LUT) achieves efficient inversion of aerosol optical depth (AOT) through multispectral channel data and pre-calculated LUT interpolation. However, it relies on a limited number of band combinations and has difficulty coping with complex aerosol-ocean coupled scenarios. Another aerosol inversion method based on optimal estimation uses a multi-parameter collaborative inversion method. By minimizing the difference between observed and simulated spectra and parameter prior errors, it jointly inverts parameters such as AOT and fine particle fraction. However, its multi-parameter inversion is prone to solution instability due to kernel matrix ill-conditioning, and the frequent Jacobian matrix calculations and iterative optimization algorithms greatly increase the computational cost.

[0003] More critically, the traditional step-by-step processing model of "atmospheric correction first, parameter inversion later" leads to the gradual accumulation of errors, seriously affecting the inversion accuracy of key ecological parameters such as chlorophyll concentration. Due to problems such as misjudgment of aerosol types and interference from sea surface flares, the cascaded step-by-step processing of "atmospheric correction first, parameter inversion later" used in traditional ocean color inversion leads to step-by-step error amplification. In complex nearshore waters, the separation error of water-leave radiance can reach as high as 35%-40%. When the corrected data is nonlinearly transferred to the parameter inversion stage through the bio-optical model, the cumulative error further increases by 20%-25%. Secondly, the ocean color inversion process separates the atmosphere-water system and ignores the coupling effect of aerosol scattering on the underwater light field, which leads to systematic biases in dynamic scenarios such as red tides and dust deposition. In addition, the time-consuming combination of step-by-step iterative calculations and parameter optimization makes it difficult to meet the timeliness requirements of high-frequency monitoring such as red tide warnings, and errors continue to accumulate with increasing frequency. These problems directly lead to insufficient credibility of monitoring data in nearshore ecologically sensitive areas and the reliability of conclusions from long-term studies such as the global carbon cycle.

[0004] Deep learning technology, with its powerful nonlinear modeling capabilities, has provided new insights for complex aerosol inversion in ocean color remote sensing. However, it is limited by the dual challenges of scarcity of annotated data and lack of physical interpretability. Although machine learning-based methods can circumvent complex radiation transfer calculations and achieve efficient inversion by directly mapping satellite observation spectra to ground-based data, their training data relies heavily on ground-based in-situ measurements. This type of data is limited by uneven geographical distribution and satellite observation geometry, resulting in insufficient representativeness at the global scale. Authoritative databases such as SeaBASS have a sample size of less than 50,000 sets. In addition, the physical foundation of existing datasets is weak, with more than 30% of satellite images lacking synchronous field verification and single parameter coverage, which seriously restricts the generalization ability of models. Data quality is significantly affected by environmental interference. Approximately 70% of pixels in daily satellite images worldwide fail due to noise pollution such as cloud obstruction, solar flares, or large sensor viewing angles. These problems collectively lead to the problem that the robustness and credibility of existing machine learning algorithms in complex scenarios are difficult to meet the needs of operational applications. Summary of the Invention

[0005] This method addresses the inherent defects of traditional ocean parameter inversion methods, such as limited computational efficiency, gradual error accumulation, and insufficient training samples in the step-by-step process. A deep ocean color inversion method that integrates satellite radiation simulation under a multi-level ocean color background is proposed: a multi-dimensional, large-sample ocean optical simulation dataset is constructed through a radiation transfer model, and a deep ocean color inversion network is proposed to replace the traditional complex physical modeling process, thereby eliminating the error transmission in the physical decoupling step-by-step calculation process and improving the inversion efficiency; this method realizes end-to-end integrated inversion from raw observation data to five ocean environmental parameters, balances the accuracy of the physical model with computational efficiency, solves the practical problem of data-driven needs and scarce annotations, meets the needs of multi-parameter inversion and eliminates the limitations of single-task models, and significantly enhances the interpretability of the model based on physically constrained simulation data, providing technical support for the optimization of atmospheric correction algorithms and multi-parameter ocean environment prediction.

[0006] To achieve the above objectives, this project adopts the following technical solutions:

[0007] Step 1. In order to obtain a large amount of satellite remote sensing data and ocean color data that are close to reality, the present invention uses a simulation algorithm based on the OSOAA radiation transfer model to simulate and obtain a total of 23 groups of simulation data including the apparent reflectivity of the simulated band. The simulation algorithm includes a parameter selection method and a simulation method.

[0008] Step 2. To address the problem of lack of integrated and extensive data sets in the field of ocean color inversion, the present invention uses the simulation data described in step 1 to construct a multi-level ocean color data set, whose features include azimuthal features and band features, and whose labels include remote sensing reflectance, water-leaving radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration.

[0009] Step 3. To solve the error accumulation problem from remote sensing images to ocean color prediction and realize end-to-end ocean color inversion task, a deep ocean color inversion network with the ability to jointly represent azimuth-band features is proposed, including a residual network and a binary attention mechanism.

[0010] Step 4. To test the deep ocean color inversion network's ability to invert ocean color components, the model is evaluated using the test portion of the dataset. To verify the dataset's relevance to the real world, real satellite remote sensing and ocean color data are collected, tested using the model, and compared with the real ocean color data to indirectly verify the dataset's authenticity.

[0011] Preferably, the simulation algorithm based on the OSOAA radiation transfer model in step 1 includes a parameter selection method and a simulation method;

[0012] For the parameter selection method, variable parameters and fixed parameters are configured respectively. The variable parameters are: aerosol optical depth, solar zenith angle, relative azimuth, chlorophyll a concentration, colored soluble organic matter concentration, debris absorption spectrum variation coefficient, suspended particulate matter concentration; the fixed parameters are: SPM particle refractive index, wind speed, seabed depth, seabed composition, phytoplankton-like particle refractive index, aerosol distribution characteristic model, aerosol model type, and relative humidity.

[0013] For the simulation method, this method uses a Python wrapper based on the OSOAA code, namely the pyOSOAA library, to configure sea color parameters and simulate data. This method simulates atmospheric apparent reflectance, water-off radiance, and remote sensing reflectance respectively. A simulation includes the following steps: given variable parameters and fixed parameters 1) For atmospheric apparent reflectance, the pyOSOAA library records the observed zenith angle and its corresponding apparent reflectance values ​​in the range of [-90°, 90°], a total of 102 segments, and the central bands are selected as 0.412, 0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.940, 1.375, 1.610, and 2.190 (nm) for simultaneous simulation. The observed zenith angle and its corresponding apparent reflectance values ​​are interpolated every 0.1° and a random value of the observed zenith angle and its corresponding apparent reflectance data is taken. The observation zenith angle and the apparent reflectance of all central bands are recorded; 2) For the water-off radiance and remote sensing reflectance, the pyOSOAA library records the height from sea level to the ocean bottom and the corresponding water surface global radiation and specular reflection radiance values. The central bands selected by this method are 0.412, 0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.940, 1.37 5, 1.610, 2.190 (nm) were used for simulation, and the values ​​of total radiation and specular reflection radiance of water surface at sea level were selected. The off-water radiance was expressed as the difference between the total radiation and specular reflection radiance of water surface, and the remote sensing reflectance was expressed as the ratio of the off-water radiance to the downward flux of the total radiation of water surface at sea level. A central band and its corresponding off-water radiance and remote sensing reflectance were recorded; 3) the variable parameters of chlorophyll a concentration, aerosol optical depth and suspended particulate matter concentration were recorded.

[0014] Preferably, each central band in 2) is simulated for 80,000 steps, and the variable parameters are resampled every 600 simulations to simulate real changes, ultimately obtaining 14 simulated bands of apparent reflectance, solar zenith angle, observation zenith angle, relative azimuth, 1 target central band wavelength and its simulated band remote sensing reflectance and water-leaving radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration;

[0015] Preferably, the multi-level ocean color dataset described in step 2 includes azimuth features and band features. The azimuth features include 4 features, namely, the wavelength of the target central band, the solar zenith angle, the observation zenith angle, and the relative azimuth angle. The band features include 14 features, namely, the simulated band apparent reflectance. The labels include remote sensing reflectance, water-leaving radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration.

[0016] Preferably, the deep ocean color inversion network described in step 3 includes a residual network and a binary attention module; the residual network uses an n-layer residual module connected in series, and the residual module is composed of several layers of sub-modules composed of linear layers and activation functions connected in series, and the input of the latter sub-module is composed of the sum of the outputs of all the previous sub-modules (for example: the input of the third-layer sub-module is the numerical sum of the output of the second layer and the output of the first layer); the binary attention module targets the azimuth-band coupling effect unique to ocean remote sensing, and explicitly establishes a directional path from azimuth features to band features in the band attention layer. The binary attention module requires two query inputs and value inputs of the same feature length. The query input and value input are respectively passed through an independent linear layer as query and key for scaled dot product attention calculation of the attention score. The attention score is obtained after Softmax to obtain the attention weight matrix, and the value input is passed through a linear layer and then matrix multiplied with the attention weight matrix to obtain the output. The output is passed through a linear layer and an activation function to obtain the output of the binary attention module.

[0017] Preferably, the deep ocean color inversion network designs a parallel azimuth residual network and a pre-trained band feature network, and the structures of the two are consistent with the residual network, and are used to process the azimuth features and band features respectively. The azimuth features and band features are respectively expanded to the hidden layer dimension through a linear layer, and then respectively input into the parallel azimuth residual network and band feature network. The outputs of the two networks are respectively subjected to a linear layer feature transformation to obtain a feature azimuth output and a feature band output consistent with the label feature; the method designs a parallel n-layer azimuth binary attention module and a band binary attention module, and the structures of the two are consistent with the binary attention module. The query input and value input of the azimuth binary attention module are both feature azimuth outputs, the query input of the band binary attention module is feature azimuth output, and the value input is feature band output. The output of the last layer of band binary attention module is passed through a linear layer to obtain the output of the final deep ocean color inversion network; the pre-trained band feature network will be trained using only band features and labels on the data set proposed by this method.

[0018] Preferably, step 4 performs performance evaluation and quality inspection on the deep ocean color inversion network and multi-level ocean color dataset proposed by the present method respectively; the deep ocean color inversion network is connected to the trained residual network and parameters are frozen; the apparent reflectance data of the Sentinel series satellites are selected as features, and the central band to be inverted is selected and arranged into inputs in the same format as the training data, and the ocean color data of measured sites such as Aeronet are selected as the ocean color true value. The trained deep ocean color inversion network is used for testing under the true value to obtain the inverted ocean color data and compare it with the ocean color true value to check the domain gap between the data.

[0019] Beneficial effects of the present invention:

[0020] The present invention uses the OSOAA radiation transfer model to generate a multi-level ocean color dataset through multi-parameter collaborative simulation, which makes up for the current scarcity of ocean color annotation data, resulting in low robustness and reliability of deep learning in complex scenarios. In particular, it solves the end-to-end inversion problem of ocean color parameters such as chlorophyll concentration.

[0021] The present invention proposes a deep ocean color inversion network, which includes a residual network and a binary attention mechanism. It optimizes the band feature extraction capability and combines the observation geometry to achieve end-to-end ocean color inversion, eliminating the error transmission in the physically decoupled step-by-step calculation process and improving the inversion efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of deep ocean color inversion method integrating satellite radiation simulation with multi-level ocean color background;

[0023] Figure 2 Flowchart for ocean color component simulation using the OSOAA radiative transfer model;

[0024] Figure 3 Schematic diagram of the deep ocean color inversion network proposed for this method;

[0025] Figure 4 This is the scatter density plot of the band residual network of this method in the evaluation of aerosol optical depth;

[0026] Figure 5 This is the scatter density plot of the band residual network of this method in the evaluation of chlorophyll a concentration;

[0027] Figure 6 This is the scatter density map of the band residual network of this method in the evaluation of water-offset radiance;

[0028] Figure 7 This is the scatter density map of the band residual network of this method in remote sensing reflectance assessment;

[0029] Figure 8 This is the scatter density plot of the band residual network used in this method in the evaluation of suspended particulate matter concentration;

[0030] Figure 9 Small-scale scatter density map of the band residual network of this method on water-offset radiance and remote sensing reflectance. DETAILED DESCRIPTION

[0031] In order to better illustrate the invention and advantages of this project, the invention content is further described below with reference to the accompanying drawings and examples.

[0032] Example 1:

[0033] This method provides a deep ocean color inversion method that integrates satellite radiation simulation under multi-level ocean color background, such as Figure 1 As shown, the following steps are included:

[0034] Step 1. To obtain a wide range of satellite remote sensing data and ocean color data that are close to reality, this method uses a simulation algorithm based on the OSOAA radiation transfer model to obtain a total of 23 sets of simulated data, including the apparent reflectance of the simulated band. The simulation algorithm includes a parameter selection method and a simulation method.

[0035] Step 2. To address the lack of integrated and extensive datasets in the field of ocean color retrieval, this method uses the simulated data described in step 1 to construct a multi-level ocean color dataset, whose features include azimuthal features and band features, and whose labels include remote sensing reflectance, water-offset radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration.

[0036] Step 3. To solve the error accumulation problem from remote sensing images to ocean color prediction and realize end-to-end ocean color inversion task, this method proposes a deep ocean color inversion network with the ability to jointly represent azimuth-band features, including a residual network and a binary attention mechanism.

[0037] Step 4. To test the deep ocean color inversion network's ability to invert ocean color components, the model is evaluated using the test portion of the dataset. To verify the dataset's relevance to the real world, real satellite remote sensing and ocean color data are collected, tested using the model, and compared with the real ocean color data to indirectly verify the dataset's authenticity.

[0038] The simulation algorithm based on the OSOAA radiation transfer model described in step 1 includes a parameter selection method and a simulation method.

[0039] For the parameter selection method, the following parameters were configured respectively: i) Variable parameters were: aerosol optical depth (N(U(0.01,0.5),0.3)), solar zenith angle (U(0,70)[°]), relative azimuth angle (U(0,180)[°]), chlorophyll a concentration (N(U(0.01,2),0.5)[mg m -3 ]), colored soluble organic matter concentration (N(U(0.001,1),0.1)[m -1 ]), coefficient of variation of debris absorption spectrum (N(U(0.02,0.06),0.1)[m -1 ]), suspended particulate matter concentration (N(U(0.4,10),2)[gm -3 ]), ii) fixed parameters are: SPM particle refractive index (1.15), wind speed (5 [m·s -1]), seafloor depth (50 [m]), seafloor composition (sand), phytoplankton-like particle refractive index (0.0176), aerosol distribution characteristic model (Maritime S&F model), aerosol model type (Shettle & Fenn dual-mode model), relative humidity (98 [%]), where U (a, b) represents a uniform distribution on (a, b), N (μ, σ 2 ) means the mean is μ and the variance is σ 2 Gaussian distribution.

[0040] As for the simulation method, the present invention uses a python wrapper based on OSOAA code, namely pyOSOAA library, to perform sea color parameter configuration and data simulation; the present invention simulates atmospheric apparent reflectance, water-leaving radiance and remote sensing reflectance respectively, and a simulation includes the following steps: given variable parameters and fixed parameters (1) for atmospheric apparent reflectance, pyOSOAA library records 102 segments of observation zenith angles and their corresponding apparent reflectance values ​​in the range of [-90°, 90°], and selects central bands of 0.412, 0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.940, 1.375, 1.610, 2.190 (nm) for simultaneous simulation, interpolation processing is performed on the observation zenith angle and its corresponding apparent reflectance values ​​every 0.1°, and a random value of the observation zenith angle and its corresponding apparent reflectance data is taken, and recorded. The observation zenith angle and the apparent reflectance of all central bands are recorded; (2) For the off-water radiance and remote sensing reflectance, the pyOSOAA database records the height from sea level to the bottom of the ocean and the corresponding water surface total radiation and specular reflection radiance values. The central bands selected by this method are 0.412, 0.443, 0.490, 0.560, 0.665, 0.705, 0.740, 0.783, 0.842, 0.865, 0.940, 1.375 , 1.610, 2.190 (nm) for simulation, select the water surface total radiation and specular reflection radiance values ​​at sea level, the water-leaving radiance is expressed as the difference between the water surface total radiation and the specular reflection radiance, the remote sensing reflectance is expressed as the ratio of the water-leaving radiance to the downward flux of the water surface total radiation at sea level, record a central band and its corresponding water-leaving radiance and remote sensing reflectance; (3) record the variable parameters of chlorophyll a concentration, aerosol optical depth and suspended particulate matter concentration.

[0041] Each central band in (2) is simulated for 80,000 steps, and the above variable parameters are resampled every 600 simulations to simulate real changes. Finally, the apparent reflectance of 14 simulated bands, solar zenith angle, observation zenith angle, relative azimuth, 1 target central band wavelength and its simulated band remote sensing reflectance and water-leaving radiance, chlorophyll a concentration, aerosol optical depth and suspended particulate matter concentration are obtained. The flow chart of ocean color component simulation using OSOAA radiation transfer model is as follows: Figure 2 As shown;

[0042] The multi-level ocean color dataset described in step 2 includes azimuth features and band features. The azimuth features include the wavelength of the target center band, the solar zenith angle, the observed zenith angle, and the relative azimuth angle, a total of four features. The band features include 14 features of the simulated band apparent reflectance. The labels include remote sensing reflectance, water-offset radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration. 90% of the total data is used as the training set and 10% as the test set. The dataset of this method allows for the real-world tasks of atmospheric correction and end-to-end ocean color prediction, and has obvious practical significance.

[0043] In order to combine the observation geometry auxiliary information and band information, such as Figure 3 As shown, the present invention proposes the deep ocean color inversion network described in step 3, including a residual network and a binary attention module; the residual network uses an n-layer residual module in series, and the residual module is composed of 3 layers of sub-modules composed of linear layers and activation functions in series, and the input of the third layer sub-module is the sum of the output of the second layer and the output of the first layer; the binary attention module targets the azimuth-band coupling effect unique to ocean remote sensing, and explicitly establishes a directional path from azimuth features to band features in the band attention layer. The binary attention module requires two query inputs and value inputs of the same feature length. The query input and value input are respectively passed through an independent linear layer as query and key to perform scaled dot product attention to calculate the attention score. The attention score is obtained after Softmax to obtain the attention weight matrix, and the value input is passed through a linear layer and then matrix multiplied with the attention weight matrix to obtain the output. The output is passed through a linear layer and an activation function to obtain the output of the binary attention module; the scaled click attention formula is as follows:

[0044]

[0045] Where q represents the query, k represents the key, d represents the feature length, and a(q,k) represents the attention score of q and k.

[0046] The forward step of the deep ocean color inversion network is as follows: the method designs a parallel azimuth residual network and a pre-trained band feature network, the structures of which are consistent with the residual network, and are used to process azimuth features and band features respectively. The azimuth features and band features are respectively expanded to the hidden layer dimension through a linear layer, and then respectively input into the parallel azimuth residual network and band feature network. The outputs of the two networks are respectively subjected to a linear layer feature transformation to obtain a feature azimuth output and a feature band output consistent with the label feature; the method designs a parallel n-layer azimuth binary attention module and a band binary attention module, the query input and value input of the azimuth binary attention module are both feature azimuth outputs, the query input of the band binary attention module is feature azimuth output, and the value input is feature band output, the output of the last layer of band binary attention module is passed through a linear layer to obtain the output of the final deep ocean color inversion network; the pre-trained band feature network will be trained using only band features and labels on the data set proposed by the method.

[0047] According to the above embodiment method, the performance evaluation and quality detection of the deep ocean color inversion network and the multi-level ocean color dataset proposed by this method are performed according to step 4 below.

[0048] Example 2:

[0049] 1. Band residual network training and verification

[0050] The activation function of the network is Softplus activation function, batch_size is 8192, the model hidden layer size is 256, dropout is 0.3, the learning rate is 2e-4, the weight decay is 1e-5, the optimization algorithm is Adam algorithm, and the loss function uses the L2 distance between the model output and the label. Training is performed on a single GeForce RTX 4070super graphics card with 200 iterations.

[0051] The model is tested and evaluated based on the band features and labels on the test set. This method randomly selects 20,000 sample points on the test set for model evaluation. The evaluation indicators include R 2 , RMSE, Figures 4 to 8 The scatter density plots show the evaluation of the proposed method on aerosol optical depth, chlorophyll a concentration, water-leaving radiance, remote sensing reflectance, and suspended particle concentration. The non-continuous aerosol optical depth, chlorophyll a concentration, and suspended particle concentration are updated every 600 steps in R 2 The indicators all reached 1.0, and the RMSE reached 0.01. This may be the result of the 600-step fixed sampling of non-continuous data, which resulted in an inflated accuracy rate. 2The indicators are 0.94 and 0.93 respectively, and the RMSE is 0.02, which proves the effectiveness of the model in data extraction and ocean color prediction. From the scatter density diagram of off-water radiance and remote sensing reflectance, it can be seen that the simulation results are divided into a bunch near the 1:1 line and a bunch that is significantly lower than the true value. This may be because when simulating the ocean color components, the parameters are randomly taken to form a parameter combination that does not exist in reality, which affects the effect of model fitting.

[0052] Figure 9 The small-scale scatter density map of the model of this method on the water-leaving radiance and remote sensing reflectance is shown. For the simulated continuous remote sensing reflectance and water-leaving radiance, most pixels are concentrated in the range of 0-0.2 [W·m -2 ·sr -1 μm -1 ] range, and is clearly divided into a bundle close to the 1:1 line and a bundle significantly lower than the true value, with corresponding RMSE values ​​of 0.44 and 0.29, respectively. The overall image presents a "spindle shape". Observing the bundle close to the 1:1 line, the model presents a "horizontal line" shape in the part where the value is greater than 0.2, indicating that the model has some difficulties in continuous prediction of small-scale data, which may be caused by the insufficient fitting ability of the band residual network. Figure 4 、 Figure 5 、 Figure 8 It can be seen that for the discontinuous aerosol optical depth, chlorophyll a concentration, and suspended particulate matter concentration with random values ​​for a fixed number of steps, they are always near the 1:1 line. They show a "vertical line-like" feature in various ranges, which may be caused by the random combination of the configured parameters not conforming to the real-world distribution. However, all points are distributed near the 1:1 line, with RMSEs of 0.06, 0.06, and 0.07, respectively, and the overall effect is good.

[0053] 2. Deep Ocean Color Inversion Network Training

[0054] The deep ocean color inversion network is connected to the trained residual network and its parameters are frozen, which can reduce the computing power required by the model while maintaining the feature extraction capability of the band residual network. The training is performed on the dataset proposed by this method, with a learning rate of 1 / 5 of that in Example 2.1. The rest of the training parameters and loss function are the same as those of the band residual network in Example 2.1.

[0055] 3. Multi-level ocean color dataset performance test

[0056] Since the central band of the apparent reflectance simulation of this method is consistent with that of the Sentinel series satellites, the apparent reflectance data of the Sentinel series satellites are selected as features. The central band to be inverted is selected and arranged into an input format with the same format as the training data. Ocean color data from actual measurement sites such as Aeronet are selected as the true ocean color value. The trained deep ocean color inversion network is used for testing under the true value to obtain the inverted ocean color data and compare it with the true ocean color value. This indirect comparison method can detect the domain gap between the data.

[0057] Data simulated using the OSOAA radiative transfer model can serve as a rich and high-quality training source. Simple neural networks can simultaneously perform end-to-end prediction tasks for atmospheric correction and different ocean color components, achieving good results and making it possible to establish a unified and efficient task processing system.

Claims

1. A deep ocean color inversion method integrating satellite radiation simulation under multi-level ocean color background, characterized by: The steps include: Step 1. Use the simulation algorithm based on the OSOAA radiative transfer model to obtain simulation data including the apparent reflectivity of the simulation band; Step 2. Use simulated data to construct a multi-level ocean color dataset; Step 3. Build a deep ocean color inversion network capable of jointly representing azimuth and band features, including a residual network and a binary attention mechanism, to process the data in the multi-level ocean color dataset and output the deep ocean color inversion results. Step 4. Use the test part of the multi-level ocean color dataset and real data to evaluate the output results of the deep ocean color inversion network.

2. The deep ocean color inversion method integrating satellite radiation simulation under multi-level ocean color background according to claim 1 is characterized in that: The simulation algorithm includes a parameter selection method and a simulation method; For the parameter selection method, variable parameters and fixed parameters are configured respectively. The variable parameters are: aerosol optical depth, solar zenith angle, relative azimuth, chlorophyll a concentration, colored soluble organic matter concentration, debris absorption spectrum variation coefficient, suspended particulate matter concentration; the fixed parameters are: SPM particle refractive index, wind speed, seabed depth, seabed composition, phytoplankton-like particle refractive index, aerosol distribution characteristic model, aerosol model type, and relative humidity; For the simulation method, a wrapper based on the OSOAA code was used for sea color parameter configuration and data simulation.

3. The deep ocean color inversion method integrating satellite radiation simulation under multi-level ocean color background according to claim 2 is characterized in that: The multi-level ocean color dataset described in step 2 includes azimuth features and band features. The azimuth features include the wavelength of the target center band, the solar zenith angle, the observation zenith angle, and the relative azimuth angle, a total of four features. The band features include the simulated band apparent reflectance, and the labels include remote sensing reflectance, water-leaving radiance, chlorophyll a concentration, aerosol optical depth, and suspended particulate matter concentration.

4. The deep ocean color inversion method integrating satellite radiation simulation under multi-level ocean color background according to claim 3 is characterized in that: The deep ocean color inversion network includes a residual network and a binary attention module; The residual network uses n-layer serial residual modules, which are composed of several layers of linear layers and activation function submodules in series, and the input of the latter submodule is the sum of the outputs of all the previous submodules; The binary attention module targets the azimuth-band coupling effect unique to ocean remote sensing, and explicitly establishes a directional path from azimuth features to band features in the band attention layer. The binary attention module requires two query inputs and value inputs of the same feature length. The query input and value input are respectively passed through independent linear layers as queries and keys, and scaled dot product attention is performed to calculate the attention score. The attention score is passed through Softmax to obtain the attention weight matrix. The value input is then passed through the linear layer and matrix multiplied with the attention weight matrix to obtain the output. The output is passed through the linear layer and the activation function to obtain the output of the binary attention module.

5. The deep ocean color inversion method based on satellite radiation simulation integrated with multi-level ocean color background according to claim 4 is characterized in that: The specific implementation process of the deep ocean color inversion network is as follows: Step 3.

1. Construct a parallel orientation residual network and a pre-trained band feature network. The structures of the two networks are consistent with the residual network, and are used to process orientation features and band features respectively. After the orientation features and band features are expanded to the hidden layer dimension through the linear layer, they are respectively input into the parallel orientation residual network and band feature network. The outputs of the two networks are respectively subjected to the feature transformation of the linear layer to obtain the feature orientation output and feature band output consistent with the label feature. Step 3.

2. Construct a parallel n-layer orientation binary attention module and a band binary attention module. The structures of the two are consistent with the binary attention module. The query input and value input of the orientation binary attention module are both feature orientation outputs, the query input of the band binary attention module is the feature orientation output, and the value input is the feature band output. The output of the last layer of band binary attention module is passed through a linear layer to obtain the output of the final deep ocean color inversion network; the pre-trained band feature network will be trained on the multi-level ocean color dataset using only band features and labels.

Images