A joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data

Abstract

The present invention discloses a method for joint inversion of atmospheric and ocean color parameters under sparse remote sensing satellite data. The method first obtains a variety of satellite remote sensing and ocean color data in a unified format. Secondly, the acquired satellite remote sensing and ocean color data are used to construct a sparse atmospheric and ocean color data set through pixel masking and combined with the observation geometry information of the satellite remote sensing data. Then, based on the remote sensing band data processing, a sparse remote sensing depth inversion network is constructed, which includes a sparse processing network, an external depth estimation network and a scale fusion network. Finally, the sparse remote sensing depth inversion network is trained using the sparse atmospheric and ocean color data set, the inversion result image is output, and the model is evaluated. The present invention can realize the joint estimation of multiple parameters of atmosphere and ocean color, improve the inversion accuracy in complex ocean and atmosphere interaction scenarios, and provide reliable technical support for high-precision atmospheric correction and ocean carbon flux estimation.

Description

Technical Field

[0001] This method belongs to the field of deep learning and atmospheric remote sensing, and specifically involves a joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data. Background Art

[0002] As the core data source for global marine ecological parameter monitoring, the multi-source ocean color satellite collaborative observation system provides key technical support for improving the level of understanding of marine ecosystems, optimizing the sustainable utilization of marine resources, formulating climate response strategies and implementing marine environmental protection. It has important scientific value and practical significance.

[0003] Traditional ocean color remote sensing atmospheric correction algorithms are based on the assumption that reflectance off water in the near-infrared and shortwave infrared bands tends to zero. Using a radiative transfer model, they construct a lookup table for aerosol scattered radiation, thereby inverting aerosol optical properties. This method demonstrates good applicability in waters with non-absorbing and weakly absorbing aerosols, but exhibits significant limitations in complex atmospheric environments dominated by strongly absorbing aerosols, restricting its universal application.

[0004] Satellite remote sensing technology has become an effective means of obtaining information on the vertical distribution characteristics of aerosols and ocean water composition. However, due to the limited spatial coverage and temporal resolution of the observation system, existing high-vertical-resolution aerosol products cannot meet the requirements for refined atmospheric correction of ocean color remote sensing. It is worth noting that although advanced sensors such as tropospheric monitors and multi-angle polarimetric radiometers incorporate the Gaussian vertical distribution assumption in aerosol retrieval, this prior information has not yet been systematically integrated into the atmospheric correction algorithm framework.

[0005] Satellite remote sensing data, subject to cloud interference and revisit cycles, generally suffer from low effective observation rates and insufficient temporal continuity, severely impacting the accuracy of quantitative inversion of ocean water components. Atmospheric correction, a core preprocessing step in ocean color remote sensing, aims to eliminate the influence of the atmosphere on apparent reflectance. By decoupling the atmosphere-ocean coupling process through physical modeling, key parameters such as water-leaving emissivity can be accurately derived. Traditional methods face the dual challenges of difficulty in decoupling complex physical processes and insufficient computational stability when satellite data is severely lacking. Summary of the Invention

[0006] To address the difficulties in decoupling complex physical processes and insufficient computational stability faced when satellite data is severely lacking, this method provides a joint inversion method for atmospheric and ocean color parameters using sparse remote sensing satellite data. By constructing a deep learning model that couples the vertical distribution characteristics of aerosols, this method breaks through the parameterization limitations of traditional algorithms and aims to improve the inversion accuracy and algorithm robustness of ocean color elements under complex atmospheric conditions. The method includes the following steps:

[0007] 1. Obtain a variety of satellite remote sensing and ocean color data in a unified format;

[0008] 2. Use the satellite remote sensing and ocean color data obtained in step 1 to construct a sparse atmospheric and ocean color dataset through pixel masking and combining the observation geometry information of the satellite remote sensing data.

[0009] 3. Based on remote sensing band data processing, a sparse remote sensing depth inversion network is constructed, including a sparse processing network, an external depth estimation network, and a scale fusion network;

[0010] 4. The sparse remote sensing depth inversion network is trained using a sparse atmospheric and ocean color dataset, and the inversion result image is output. After training, a single satellite image at a specific time is obtained for model evaluation.

[0011] Preferably, the features of the sparse atmospheric and ocean color dataset in step 2 include observation geometry, land surface information, and remote sensing band information after 95% pixel masking, and the label is the ocean color data in step 1. The observation geometry includes solar azimuth angle (SAA), solar zenith angle (SZA), observation azimuth angle (VAA), and observation zenith angle (VZA). The land surface information includes normalized difference vegetation index (NDVI). The remote sensing band information consists of 7 bands, and the central wavelengths are respectively in the following nanometer (nm) ranges: [0.45, 0.52], [0.52, 0.60], [0.63, 0.69], [0.77, 0.90], [1.55, 1.75], [2.08, 2.35], and [0.52, 0.90].

[0012] Preferably, the sparse remote sensing depth inversion network in step 3 includes a sparse processing network (I), an external depth estimation network (II) and a scale fusion network (III). The sparse processing network is used to extract sparse remote sensing band information, and the external depth estimation network jointly encodes the land surface information and the observation geometry. The two networks perform parallel calculations to obtain sparse remote sensing band information and joint encoding guidance information respectively. The scale fusion network guides the reconstruction of the sparse remote sensing band information according to the joint encoding guidance information and finally obtains the inversion result image. The three parts are introduced below:

[0013] (I) The sparse processing network extracts and completes features of sparse satellite remote sensing data, shortens the mapping distance from sparse features to labels through a pre-reconstruction method, and has a structure of a discrete encoder-decoder network, wherein the encoder is a vision transfer Vision Transformer (ViT) network with a multi-layer perceptron network (MLP) as an output head, and the decoder is a convolutional neural network (CNN) network with a U-Net architecture; the discrete encoder-decoder network uses a discrete training method, that is, the encoder network is independently trained using the remote sensing band information and dataset labels of the sparse atmospheric and ocean color dataset proposed by this method as training data, and the decoder network is trained using the output of the trained encoder network and the labels of the sparse atmospheric and ocean color dataset as training data, and the trained encoder and decoder are connected in series to obtain a pre-trained discrete encoder-decoder network; a network consisting of n layers of convolutional layers and normalized residual networks and one layer of convolutional layers in series is used to follow the discrete encoder-decoder network to obtain pending reconstruction features; the discrete encoder-decoder network is pre-trained and its parameters are frozen after training.

[0014] (II) The external depth estimation network jointly encodes the land surface information and the observed geometry to obtain jointly encoded information and assists in pixel reconstruction and mapping of atmospheric correction and ocean color prediction tasks in network III; the same discrete encoder-decoder network as (I) is used for the land surface information, and a non-discrete encoder-decoder with the same architecture as (I) that does not use a discrete training method is used for the observed geometry, and the query and key of ViT in the non-discrete encoder-decoder come from network I; the output of the land surface information through the discrete encoder-decoder network and the output of the observed geometry through the non-discrete encoder-decoder network are spliced on the channel through n-layer joint external information encoding modules and then through a convolutional layer for channel dimensionality reduction to obtain the jointly encoded guidance information.

[0015] The joint external information encoding module includes a self-attention module, specifically: the input passes through a convolution layer and normalization, and passes through a convolution layer with itself as the query, key and value respectively. The query and key values are multiplied and then Softmax weighted to obtain the attention weight, which is multiplied with the value passed through the convolution layer to obtain the output. This output is residually connected with the input passed through a convolution layer to obtain the output of the layer module.

[0016] (III) The scale fusion network scale-fuses the pending reconstruction features obtained by network I and the joint coding guidance information obtained by network II to obtain the final atmospheric or ocean color component data, i.e., the inversion result image. The scale fusion network includes several layers of semantic prior fusion (SPF) and guidance reconstruction layers. The first guidance reconstruction layer receives the pending reconstruction features, the nth guidance reconstruction layer receives the output of the n-1th SPF layer (6≥n≥2), the first SPF layer receives the joint coding guidance information output in (II) and the output of the first guidance reconstruction layer, the nth SPF layer receives the output of the n-1th SPF layer and the output of the nth guidance reconstruction layer (5≥n≥2), and the last guidance reconstruction layer receives the output of the last SPF layer to obtain the final atmospheric correction or ocean color component data; the SPF uses the same self-attention module as in network II, and the guidance reconstruction layer is a U-Net network;

[0017] Preferably, step 4 selects 90% of the data in the dataset as a training dataset and 10% of the data as a test dataset; first, the discrete encoder-decoder network is trained: for the ViT-based encoder network, the satellite image patch_size is 8, the number of layers is 5, the number of self-attention heads is 8, the dropout is 0.1, the batch size is 4, the learning rate is 1e-4, the number of iterations is 50, and training is performed on 2 RTX4070supers; for the CNN-based decoder network, the number of input channels and the number of output channels are the same as the number of label channels, the batch size is 12, and the remaining parameters are consistent with the encoder training parameters. The output and label of the encoder network are used as training data to train the decoder network; the sparse remote sensing depth inversion network freezes the parameters of the trained discrete network and trains it on the sparse atmospheric and ocean color dataset proposed in this method. The training parameters and loss function are the same as the ViT network training parameters; the trained network is used to perform feasibility testing under real single satellite images.

[0018] Beneficial results of the present invention:

[0019] (1) This paper proposes a joint inversion method for atmosphere-ocean color parameters based on sparse remote sensing satellite data. By constructing a cross-modal coupling architecture that integrates physical constraints with deep learning, it effectively solves the problem of radiation transmission decoupling caused by insufficient spatial coverage of satellite data. This method can realize the joint estimation of multiple parameters of atmosphere-ocean color, improve the inversion accuracy in complex ocean-atmosphere interaction scenarios, and provide reliable technical support for high-precision atmospheric correction and ocean carbon flux estimation.

[0020] (2) The present invention proposes a sparse remote sensing depth inversion network with multimodal feature interaction. By constructing a hybrid structure of sparse feature extraction, external prior encoding and scale fusion, it can effectively solve the problem of pathological inversion of radiation transfer under sparse observation conditions and achieve high-precision joint reconstruction of atmosphere-ocean color parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flowchart of the deep atmospheric correction and ocean color inversion method based on sparse remote sensing satellite data;

[0022] Figure 2 This is a diagram of the network architecture proposed by the present invention;

[0023] Figure 3 This is a display of remote sensing band information with a central wavelength of [0.45, 0.52] nm that passes 95% of the pixel masks;

[0024] Figure 4 The 550nm aerosol inversion results of the discrete encoder-decoder network of the present invention in atmospheric correction;

[0025] Figure 5 The scatter density plot of the 550nm aerosol inversion for the atmospheric correction of the discrete encoder-decoder network of the present invention;

[0026] Figure 6 The inversion results of the water-offset emissivity at 405-420 nm for the discrete encoder-decoder network of the present invention;

[0027] Figure 7 Scatter density plot of water-exit radiance inversion at 405-420 nm for the discrete encoder-decoder network of the present invention. DETAILED DESCRIPTION

[0028] 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.

[0029] Example 1:

[0030] This method provides a joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data. Figure 1 As shown, the following steps are included:

[0031] 1. Use a predefined JavaScript algorithm to obtain a variety of satellite remote sensing and ocean color data in a unified format on the Google Earth Engine platform;

[0032] 2. Use the satellite remote sensing and ocean color data obtained in step 1 to simulate the severe lack of satellite remote sensing data through 95% pixel masking. Combined with the observation geometry information of the satellite remote sensing data, a sparse atmospheric and ocean color dataset is constructed.

[0033] 3. Based on remote sensing band data processing, a sparse remote sensing depth inversion network is constructed, including a sparse processing network, an external depth estimation network, and a scale fusion network;

[0034] 4. The sparse remote sensing depth inversion network is trained using a sparse atmospheric and ocean color dataset to obtain an inversion result image. After training, a single satellite image at a specific time is obtained for model evaluation.

[0035] The step 1 uses the planetary-level geospatial analysis platform Google Earth Engine platform (GEE platform) to obtain ocean color data, wherein the ocean color data includes apparent reflectance data, water-offset radiance data, and atmospheric aerosol data. The median of the effective values in the study area is taken within one acquisition cycle. The study area is the Bohai Sea area of China. The GEE platform is used to collect ocean color data for 18 years from 2002 to 2020, with an acquisition cycle of 8 days and a downsampling resolution of 1000m. The images collected using GEE contain valid parts, missing parts, and invalid parts. The valid part is the pixel with a value greater than 0 in the study area, the missing part is the part with a value less than or equal to 0 in the study area, and the invalid part is the part outside the study area in the image.

[0036] Remote sensing band information contains the most important and rich data for remote sensing processing. Land surface information provides water conditions, which helps guide atmospheric correction and ocean color prediction research when the geographical and water environment changes. Observation geometry is an important component of remote sensing image basic data and one of the essential parameters for atmospheric and ocean color inversion. The sparse atmospheric and ocean color dataset features described in step 2 include observation geometry, land surface information, and remote sensing band information after 95% pixel masking. The label is the ocean color data described in step 1. The observation geometry includes solar azimuth (SAA), solar zenith angle (SZA), observation azimuth (VAA), and observation zenith angle (VZA). The land surface information includes normalized difference vegetation index (NDVI). The remote sensing band information consists of 7 bands, whose central wavelengths are in the following nanometer (nm) ranges: [0.45, 0.52], [0.52, 0.60], [0.63, 0.69], [0.77, 0.90], [1.55, 1.75], [2.08, 2.35], [0.52, 0.90]; 95% pixel mask constitutes interpolation invalidity and non-obvious meaningful remote sensing inversion task, which can greatly eliminate the domain gap between the data collected by the collection algorithm and the real world. The remote sensing band information under the central wavelength of 95% pixel mask is [0.45, 0.52]nm, such as Figure 3 shown.

[0037] To perform atmospheric correction and ocean color inversion tasks from remote sensing band data, land surface information and observation geometry, such as Figure 2 As shown, this method proposes the sparse remote sensing depth inversion network described in step 3, including a sparse processing network (I), an external depth estimation network (II) and a scale fusion network (III). The sparse processing network is used to extract sparse remote sensing band information, and the external depth estimation network jointly encodes the land surface information and the observation geometry. The two networks calculate in parallel to obtain sparse remote sensing band information and joint encoding guidance information respectively. The scale fusion network guides the reconstruction of the sparse remote sensing band information according to the joint encoding guidance information and finally obtains the inversion component. The three parts are introduced below:

[0038] (I) Atmospheric correction and ocean color inversion require as much relevant and effective information as possible. The sparse processing network performs feature extraction and completion tasks on sparse remote sensing satellite data, and shortens the mapping distance from sparse features to labels through pre-reconstruction methods. Transformer has a natural advantage in sparse processing tasks due to its powerful global modeling capabilities and broader inductive bias. Convolutional neural networks (CNNs) have better pixel-level output capabilities and can further improve pixel prediction accuracy. However, simply training the two in series will cause the CNN network to affect the modeling ability of the Transformer network. Therefore, this method proposes a sparse processing network; specifically, the sparse processing network is a discrete encoder-decoder network, and the encoder is a VisionT with a multi-layer perceptron network (MLP) as the output head. Transformer (ViT) network, the decoder is a CNN network with U-Net architecture; the discrete encoder-decoder network represents a discrete training method, that is, the encoder network is independently trained using the remote sensing band information and dataset labels of the sparse atmospheric and ocean color dataset proposed by this method as training data, and the decoder network is trained using the output of the trained encoder network and the labels of the sparse atmospheric and ocean color dataset as training data. The trained encoder and decoder are connected in series to obtain a pre-trained discrete encoder-decoder network; the discrete network is pre-trained and its parameters are frozen after training; this network can effectively combine the global modeling capability of the Transformer and the pixel-level output capability of the CNN network, providing effective prior information for the final atmospheric correction and ocean color prediction.

[0039] A network consisting of n convolutional layers and normalized residual networks and one convolutional layer in series is used following the discrete encoder-decoder network to obtain the proposed reconstructed features.

[0040] The external depth estimation network (II) is designed to jointly encode land surface information and observation geometry to obtain high-quality guidance information and assist in pixel reconstruction and mapping for atmospheric correction and ocean color prediction tasks in network III. Sparse land surface information requires feature extraction and completion, so the same discrete encoder-decoder network as (I) is used for land surface information. The observation geometry usually does not include high-precision requirements, and only a non-discrete encoder-decoder with the same architecture is used to maintain semantic space consistency, and its query and key in ViT come from network I. The outputs of the two are spliced on the channel and passed through an n-layer joint external information encoding module. After channel dimensionality reduction through a convolutional layer, the jointly encoded guidance information is obtained.

[0041] The joint external information encoding module includes a self-attention module, specifically: the input passes through a convolution layer and normalization, and passes through a convolution layer with itself as the query, key and value respectively. The query and key values are multiplied and then Softmax weighted to obtain the attention weight, which is multiplied with the value passed through the convolution layer to obtain the output. This output is residually connected with the input passed through a convolution layer to obtain the output of the layer module.

[0042] (III) The scale fusion network scale-fuses the pending reconstruction features obtained by network I and the joint coding guidance information obtained by network II to obtain the final atmospheric correction or ocean color component data; specifically, it includes 5 layers of semantic prior fusion (SPF) and 6 layers of guidance reconstruction layers, the 1st guidance reconstruction layer receives the pending reconstruction features, the nth guidance reconstruction layer receives the output of the n-1th SPF layer (6≥n≥2), the 1st SPF layer receives the joint coding guidance information output in (II) and the output of the 1st guidance reconstruction layer, the nth SPF layer receives the output of the n-1th SPF layer and the output of the nth guidance reconstruction layer (5≥n≥2), and finally the 6th guidance reconstruction layer receives the output of the 5th SPF layer to obtain the final atmospheric correction or ocean color component data; the SPF uses the same self-attention module as in network II, and the guidance reconstruction layer is a U-Net network.

[0043] According to the above embodiment method, the performance analysis of the deep atmospheric correction and ocean color inversion method under sparse remote sensing satellite data is performed according to step 4 below:

[0044] Example 2:

[0045] 1. Discrete Encoder-Decoder Network Training and Validation

[0046] Select 90% of the data in the dataset as the training dataset and 10% of the data as the test dataset;

[0047] For the ViT-based encoder network, the satellite image patch_size is 8, the number of layers is 5, the number of self-attention heads is 8, the dropout is 0.1, the batch size is 4, the learning rate is 1e-4, the number of iterations is 50, and the training is performed on two RTX4070supers. The loss function As shown below:

[0048]

[0049] in represents the model parameters, represents the network input, Represents various dataset labels, Represents the image mask, 1 represents the valid part, 0 represents the invalid part and the missing part, Take 1.0.

[0050] For the CNN-based decoder network, the number of input channels and output channels is the same as the number of label channels, such as 9 for the atmospheric correction task and 1 for the aerosol optical depth inversion task. The batch size is 12, and the remaining parameters are consistent with the encoder training parameters. The output and labels of the encoder network are used as training data to train the decoder network, and the training is performed on the loss function of formula (1).

[0051] Figure 4 The discrete encoder-decoder network of this method is used to invert the atmospherically corrected 550nm aerosol. Figure 6 The paper presents the inversion results of the water-offset emissivity at 405-420nm using the discrete encoder-decoder network of this method, which respectively includes sparse feature input, ViT encoder output, CNN decoder output, label truth, and pixel scatter density map. Experiments show that the ViT encoder can extract relatively complete and accurate target information from sparse information, but ViT is affected by the cutting of image patches, and it exhibits a strong "splitting block feeling" at the macro level. The CNN decoder network can refine and reconstruct the encoder output at the pixel level, and ultimately obtain a predicted image that is highly similar to the label. Figure 5 and Figure 7 Shows the corresponding Figure 4 and Figure 6 The scatter density plot records the correlation coefficient (R 2 ), root mean square error (RMSE), mean absolute percentage error (MAPE) and number of test points (Number). Experiments show that the network composed of discrete encoders and decoders can invert close to the true atmospheric correction information or ocean color component information from sparse satellite data, which can significantly shorten the mapping distance from sparse features to labels;

[0052] 2. Sparse Remote Sensing Depth Inversion Network Training Method

[0053] The sparse remote sensing depth inversion network freezes the parameters of the discrete network in Example 2.1, reducing the computing power required by the model while maintaining the feature extraction capability of the discrete network. The training is performed on the sparse atmospheric and ocean color dataset proposed in this method. The training parameters are the same as those of the ViT network in Example 2.1, and the formula (1) is used as the loss function for training.

[0054] The deep atmospheric correction and ocean color inversion method based on sparse remote sensing satellite data proposed in this method introduces external auxiliary data of land surface information and observation geometry, which can reflect the changing state of the real-world environment. With the assistance of external data of land surface information and observation geometry, the inversion capability from sparse satellite data can be further improved, achieving efficient atmospheric correction and ocean color inversion capabilities.

Claims

1. A joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data, characterized in that: The following steps are involved: Step 1. Obtain a variety of satellite remote sensing and ocean color data; Step 2. Use satellite remote sensing and ocean color data to construct a sparse atmospheric ocean color dataset through pixel masking and combining the observation geometry information of satellite remote sensing data. Step 3. Based on remote sensing band data processing, a sparse remote sensing depth inversion network is constructed, which includes a sparse processing network, an external depth estimation network, and a scale fusion network; The sparse processing network extracts and completes features of sparse satellite remote sensing data in the sparse atmospheric and ocean color dataset, shortens the mapping distance from sparse features to labels through a pre-reconstruction method, and is used to extract sparse remote sensing band information. The external depth estimation network uses the same discrete encoder-decoder network as the sparse processing network for land surface information in the sparse atmospheric and ocean color dataset. The encoder is a ViT network with a multi-layer perceptron network as the output head, and the decoder is a convolutional neural network with a U-Net architecture. The external depth estimation network jointly encodes the land surface information and the observed geometry. The two networks perform parallel calculations to obtain sparse remote sensing band information and joint encoding guidance information, respectively. The scale fusion network performs scale fusion on the undetermined reconstruction features obtained by the sparse processing network and the joint encoding guidance information obtained by the external depth estimation network, and guides the reconstruction of the sparse remote sensing band information according to the joint encoding guidance information to ultimately obtain an inversion result image. Step 4. Use the sparse atmospheric and ocean color dataset to train the sparse remote sensing depth inversion network, output the inversion result image, and evaluate it.

2. The joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data according to claim 1, characterized in that: The features of the sparse atmospheric and ocean color dataset include pixel-masked observation geometry, land surface information, and remote sensing band information, and the label is the ocean color data described in step 1; the observation geometry includes solar azimuth, solar zenith angle, observation azimuth, and observation zenith angle, and the land surface information includes the Normalized Difference Vegetation Index (NDVI).

3. The joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data according to claim 2, characterized in that: The specific implementation process of the sparse processing network is as follows: The sparse processing network structure is a pre-trained discrete encoder-decoder network. The discrete encoder-decoder network uses a discrete training method, that is, the encoder network uses the remote sensing band information and dataset labels of the sparse atmospheric and ocean color dataset as training data for independent training, and uses the output of the trained encoder network and the labels of the sparse atmospheric and ocean color dataset as training data to train the decoder network. The trained encoder and decoder are connected in series to obtain a pre-trained discrete encoder-decoder network; a network including n layers of convolutional layers and normalized residual networks and one layer of convolutional layers in series is used to follow the discrete encoder-decoder network to obtain pending reconstruction features.

4. The joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data according to claim 3, characterized in that: The specific implementation process of the external depth estimation network is as follows: The external depth estimation network uses a non-discrete encoder-decoder with the same architecture as the sparse processing network for the observation geometry without using a discrete training method, and the query and key of ViT in the non-discrete encoder-decoder come from the sparse processing network; The output of the land surface information through the discrete encoder-decoder network and the output of the observation geometry through the non-discrete encoder-decoder network are spliced on the channel. After passing through an n-layer joint external information encoding module and a channel dimensionality reduction through a convolutional layer, the joint coding guidance information is obtained.

5. The joint inversion method for atmospheric and ocean color parameters based on sparse remote sensing satellite data according to claim 4, characterized in that: The joint external information encoding module is a self-attention module, which is specifically implemented as follows: the input passes through a convolution layer and normalization, and the query, key and value itself pass through a convolution layer respectively. The query and key values are multiplied and then weighted by Softmax to obtain the attention weight, which is multiplied by the value passed through the convolution layer to obtain the output. This output is residually connected with the input passed through a convolution layer to obtain the output of the self-attention module.

6. The method for joint inversion of atmospheric and ocean color parameters using sparse remote sensing satellite data according to claim 5, characterized in that: The specific implementation process of the scale fusion network is as follows: The scale fusion network includes several layers of semantic prior fusion SPF and guidance reconstruction layers, the first guidance reconstruction layer receives the pending reconstruction features, the nth guidance reconstruction layer receives the output of the n-1th SPF layer, the first SPF layer receives the joint coding guidance information and the output of the first guidance reconstruction layer, the nth SPF layer receives the output of the n-1th SPF layer and the output of the nth guidance reconstruction layer, and the last guidance reconstruction layer receives the output of the last SPF layer to obtain the final atmospheric correction or ocean color component data; the SPF uses the same self-attention module as in the external depth estimation network, and the guidance reconstruction layer is a U-Net network.

Images