Ocean remote sensing image atmospheric correction method based on multi-scale transformer

Abstract

This invention relates to the field of atmospheric correction technology, and in particular provides a method for atmospheric correction of marine remote sensing images based on multi-scale Transformer. The method includes acquiring multi-source remote sensing observation data and auxiliary parameters, performing joint preprocessing, and constructing a standardized training dataset; constructing a multi-scale Transformer atmospheric correction model containing parallel convolutional branches; constructing a composite loss function containing physical soft constraints, and training the atmospheric correction model using the training dataset to obtain the trained model; and using the trained model to perform high-throughput end-to-end atmospheric correction inference. In complex large-scale marine environments, this method achieves high-precision, robust, and physically consistent atmospheric correction.

Description

Technical Field

[0001] This invention relates to the field of atmospheric correction technology, and in particular to an atmospheric correction method for marine remote sensing images based on multi-scale Transformer. Background Technology

[0002] Water color remote sensing plays a crucial role in global marine ecosystem monitoring, carbon cycle research, and climate change assessment. However, due to the strong scattering and absorption of electromagnetic waves by atmospheric components (such as gases, aerosols, and water vapor), 80% to 90% of the signals detected by satellite sensors originate from the atmosphere rather than the water body itself, severely interfering with the extraction of water body signals. This atmospheric interference constitutes a major technical bottleneck limiting the accuracy and reliability of water quality parameter inversion. Therefore, to obtain the true spectral reflectance of water bodies, atmospheric correction is a core preprocessing step in water color remote sensing, and its accuracy directly determines the reliability of subsequent water quality inversion and environmental monitoring results.

[0003] Currently, existing atmospheric correction methods are mainly divided into three categories, but all of them have significant technical shortcomings in large-scale, automated operational applications. The first category is methods based on traditional physical models, such as 6S, MODTRAN, SeaDAS, and Polymer. These methods describe the propagation mechanism of light in the atmosphere and water bodies in detail through radiative transfer models, and theoretically can achieve high correction accuracy. However, these methods face serious computational bottlenecks in practical applications: they are highly dependent on atmospheric parameters such as aerosol optical depth (AOD) and columnar total water vapor content (CWV), which are difficult to obtain in real time. At the same time, the complex radiative transfer calculations and the cumbersome lookup table (LUT) generation process are extremely time-consuming, resulting in extremely low data processing efficiency, which is completely unable to meet the current needs for automated processing of large-scale, global-level high-throughput satellite data.

[0004] The second category is image-based assumption-based methods, such as the dark pixel assumption method and the empirical line method. These methods utilize only the information in the image itself for rapid correction, and the calculation process is relatively simple. However, their fatal flaw lies in their extremely limited applicability. When faced with turbid nearshore waters or complex scenarios with significant environmental and seasonal changes, these assumptions often fail, easily leading to correction failure and making it impossible to provide stable and reliable data support.

[0005] The third category comprises machine learning and conventional deep learning methods that have emerged in recent years. Traditional machine learning methods primarily rely on fitting statistical relationships within local regions, lacking the ability to generalize across regions and seasons. While conventional deep learning networks (such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) have, to some extent, avoided complex physical parameter dependencies and demonstrated the potential to accelerate atmospheric correction, most current research remains limited to small sample scenarios on land or nearshore areas, lacking systematic validation across large scales and multiple time phases. Specifically, CNN-based networks are limited by their inherent local receptive field, making it difficult to effectively capture complex ocean gradients and extract global features; while GAN-based networks, when dealing with complex spectral environments with strong aerosol scattering, often produce severe numerical distortion and physically inconsistent reflectance artifacts in shorter wavelength bands.

[0006] In summary, existing technical solutions are not only limited by complex physical parameter inversion and huge computational costs, but also face serious numerical distortion and insufficient generalization ability when dealing with large-scale complex marine environments. Summary of the Invention

[0007] In view of this, the present invention provides an atmospheric correction method for marine remote sensing images based on multi-scale Transformer, which can achieve high-precision, robust and physically consistent atmospheric correction in complex large-scale marine environments.

[0008] In a first aspect, the present invention provides an atmospheric correction method for ocean remote sensing images based on a multi-scale Transformer, the method comprising:

[0009] Step 1: Acquire multi-source remote sensing observation data and auxiliary parameters, and perform joint preprocessing to construct a standardized training dataset; Step 2: Construct a multi-scale Transformer atmospheric correction model that includes parallel convolutional branches; Step 3: Construct a composite loss function that includes physical soft constraints, and train the atmospheric correction model using the training dataset to obtain the trained model; Step 4: Utilize the trained model to perform high-throughput end-to-end atmospheric correction inference.

[0010] Optionally, step 1 includes: Step 11, Multi-source data acquisition and feature set construction: Acquire visible light band TOA radiation data observed by satellite sensors, and simultaneously acquire the corresponding geometric parameters Geo and meteorological parameters Meteo as feature inputs; acquire corrected surface reflectance data as the true BOA label of the lower atmosphere; Step 12, Pixel-level quality filtering based on quality control flags: To ensure the consistency and effectiveness of model input and output, the acquired remote sensing images are filtered at the pixel level using quality control flags L2_flags. Step 13, Resampling and Spatial Alignment of Auxiliary Variables: To ensure the consistency of geometric parameters Geo, meteorological parameters Meteo and spectral band data in spatial dimension, bilinear interpolation is used to resample them to the same resolution as the radiation data. Step 14, Patch Construction and Invalid Value Filling: In the data that has been quality filtered and spatially aligned, the track-level imagery is divided into fixed-size image patches; an effective pixel ratio threshold is set, and only image patches with an effective pixel ratio greater than 90% are retained; for the qualified image patches after filtering, the mean of each channel is calculated, and the mean is used to fill the NaN invalid values ​​in the patch. Step 15: Z-score standardization of data: After completing the tile construction, perform Z-score standardization on all L1B input features and L2 output labels, that is, subtract the mean of each channel's data and then divide it by its standard deviation.

[0011] Optionally, step 2 includes: Step 21: Construct an overall network architecture based on an encoder-decoder: Construct a multi-scale Transformer framework that integrates windowed multi-head self-attention (W-MHSA) and multilayer perceptron (MLP); the overall architecture is based on an encoder-decoder structure: In the encoder part, the network progressively extracts global features from the input image through multiple atmospheric correction blocks (AtmoCorr Blocks) and two down-sampling modules (Down-sample); In the decoder part, the network restores the spatial resolution of the image through an up-sampling module (Up-sample), and combines SK Fusion and subsequent atmospheric correction blocks for detail reconstruction to ensure the preservation of local textures in the image; Step 22: Constructing a windowed multi-head self-attention mechanism W-MHSA-PC in the atmospheric correction block: In order to enhance the model's ability to model features at different scales, a parallel convolutional branch is introduced into the standard multi-head self-attention mechanism MHSA to construct the W-MHSA module. Step 23: Adaptive feature fusion based on gated weighting mechanism: In the fusion stage of the W-MHSA module, the features output by the self-attention branch and the parallel convolution branch are adaptively fused through the gated weighting fusion mechanism. Step 24: Reconstruct channel features using a multilayer perceptron (MLP): Introduce a multilayer perceptron (MLP) after each attention block to enhance the nonlinear interaction between channels.

[0012] Optionally, step 3 includes: Step 31: Construct a composite objective function containing multiple penalties: Construct a function incorporating Charbonnier loss. Mean Square Error (MSE) Loss and range of penalty losses The composite loss function; Step 32: Set the weights for the composite loss function: The final training objective function is the weighted sum of the three losses mentioned above, and its expression is: ; Weight the MSE loss Weights of range constraint loss All are set to the optimal balance value of 0.05; Step 33: Configure the adaptive optimizer and dynamic learning rate scheduling strategy: When optimizing network parameters through backpropagation, the Adam optimizer is used, with an initial learning rate of 0.001 and momentum parameters configured. =0.9 and =0.999; To stabilize the training process, a dynamic learning rate adjustment strategy is introduced. When the root mean square error (RMSE) on the validation set does not improve within a set number of consecutive epochs, the learning rate is automatically reduced by a predetermined decay factor; at the same time, the lower bound of the learning rate is set to 1×10. -6 ; Step 34: Introduce a stable training mechanism with mixed precision and gradient clipping: During the batch training of the model, introduce automatic mixed precision (AMP) technology; apply the gradient clipping mechanism to limit the maximum gradient norm to 1.

[0013] Optionally, step 4 includes: Step 41: Obtain multi-source real-time remote sensing data of the area to be corrected: Obtain satellite remote sensing images of the target area to be corrected, extract the raw TOA radiation data of the visible light band, and simultaneously extract the real-time auxiliary parameters corresponding to the images, including the corresponding geometric parameters and meteorological parameters. Step 42: Perform consistent data preprocessing: Perform preprocessing on the data to be corrected, using quality control flags to remove invalid pixels affected by clouds and land; perform bilinear interpolation on sparse grid auxiliary variables to achieve pixel-by-pixel alignment of spatial resolution; and perform Z-score standardization on all input features to ensure that the data distribution of the input features to be corrected is consistent with that during model training. Step 43: End-to-end fast inference and output of bottom atmospheric reflectance (BOA): The preprocessed TOA radiation data to be corrected and the corresponding geometric and meteorological parameters are used as feature tensors and directly input into the trained multi-scale Transformer atmospheric correction model; the model maps and outputs the predicted bottom atmospheric reflectance (BOA) surface spectral reflectance through a forward propagation network.

[0014] Optionally, step 31 includes: Charbonnier loss is used as the primary loss term, and a smoothing constant is introduced in its mathematical expression. MSE loss is used as an auxiliary loss term to emphasize pixel-level numerical accuracy, aiming to narrow the gap between predicted and true values. Range constraint loss is introduced to apply a soft penalty to the model's prediction results to avoid negative or outlier values ​​greater than 1, and to force the output reflectivity to be within the physically reasonable range of [0,1] to ensure the physical consistency of the output.

[0015] Thirdly, embodiments of the present invention provide a computer-readable storage medium comprising a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to execute the atmospheric correction method for ocean remote sensing images based on multi-scale Transformer, as described in the first aspect or any possible implementation thereof.

[0016] Fourthly, embodiments of the present invention provide an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to perform the multi-scale Transformer-based atmospheric correction method for marine remote sensing images in the first aspect or any possible implementation thereof.

[0017] The technical solution provided by this invention includes acquiring multi-source remote sensing observation data and auxiliary parameters, performing joint preprocessing, and constructing a standardized training dataset; constructing a multi-scale Transformer atmospheric correction model containing parallel convolutional branches; constructing a composite loss function containing physical soft constraints, and training the atmospheric correction model using the training dataset to obtain the trained model; and using the trained model to perform high-throughput end-to-end atmospheric correction inference. In complex large-scale ocean environments, this method achieves high-precision, robust, and physically consistent atmospheric correction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of an atmospheric correction method for ocean remote sensing images based on multi-scale Transformer provided in an embodiment of the present invention; Figure 2 A schematic diagram of the multi-scale Transformer atmospheric correction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0023] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0024] Figure 1This is a schematic diagram of an atmospheric correction method for ocean remote sensing images based on multi-scale Transformer provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Step 1: Obtain multi-source remote sensing observation data and auxiliary parameters, and perform joint preprocessing to construct a standardized training dataset.

[0025] In this embodiment of the invention, step 1 includes: Step 11, Multi-source data acquisition and feature set construction: Acquire visible light band TOA radiation data observed by satellite sensors, and simultaneously acquire the corresponding geometric parameters Geo and meteorological parameters Meteo as feature inputs; acquire corrected surface reflectance data as the true BOA label of the lower atmosphere; Observational data from satellite sensors (such as the Ocean and Land Color Instrument (OLCI) on the Sentinel-3B satellite) are acquired. Specifically, raw Top of Atmosphere (TOA) radiance data in the visible light band is extracted from Level 1B (L1B) products as key input; in this embodiment, six visible light bands are specifically selected: 400 nm, 412 nm, 442 nm, 490 nm, 510 nm, and 560 nm. Corresponding auxiliary variables are acquired simultaneously, including geometric parameters (Geo) and meteorological parameters (Meteo). Geometric parameters include solar zenith angle (SZA), solar azimuth angle (SAA), observed zenith angle (OZA), and observed azimuth angle (OAA), used to reflect the observational geometry between the light source and the sensor; meteorological parameters include total columnar water vapor content (CWV), total ozone (TOZ), and sea level pressure (SLP), used to mitigate the uncertainties caused by spatiotemporal variations in atmospheric conditions. Simultaneously, the water reflectance corresponding to the aforementioned six visible light bands was obtained from the Level-2 (L2) product after rigorous atmospheric correction, serving as a label for the true bottom atmosphere (BOA) surface spectral reflectance. To meet the computational efficiency requirements of large-scale studies, the above data was uniformly downsampled to a specific spatial resolution (e.g., 1200m).

[0026] The multi-source data is preprocessed, including quality filtering, patch construction, invalid value imputation, and Z-score standardization, to eliminate the dimensional differences between different variables and remove low-quality pixels affected by clouds, land, etc., and generate aligned input-output tensor pairs.

[0027] In this embodiment of the invention, atmospheric correction (AC) refers to the process of eliminating reflectance distortion in remote sensing images caused by the scattering and absorption of electromagnetic waves by atmospheric components (such as gases, aerosols, and water vapor) in order to obtain the true spectral reflectance of the target (such as a water body). Top-of-Atmosphere Radiance (TOA): Raw radiation data directly observed by satellite sensors before any atmospheric interference is eliminated. This data is greatly affected by atmospheric scattering and absorption. In one specific embodiment of the present invention, raw visible light radiation from a satellite may be used, but is not limited to this.

[0028] Bottom-of-Atmosphere Reflectance (BOA): refers to the spectral reflectance reaching the water or land surface after atmospheric correction to eliminate atmospheric molecular scattering, aerosol scattering, and absorption. In the embodiments of this invention concerning marine remote sensing imagery, it is specifically represented as Remote Sensing Reflectance (Rrs). In the broader context of surface remote sensing, the BOA described in this invention has an equivalent physical meaning to Surface Reflectance (SR).

[0029] Window-based Multi-Head Self-Attention (W-MHSA): This invention constructs a feature extraction module that introduces parallel convolutional branches on top of standard multi-head self-attention. This module can simultaneously capture global information through local windows and enhance the extraction capability of local high-frequency textures (such as edges, water details, etc.) through parallel convolutional branches.

[0030] Soft Range Constraints: This invention introduces a penalty mechanism when constructing the loss function. By penalizing the reflectance results predicted by the model, it clips and restricts the output value to a physically reasonable range (such as [0,1]), thereby effectively preventing the occurrence of outliers and ensuring the physical consistency and stability of the model output.

[0031] Meteorological parameters (Meteo): data reflecting the physical state of the atmosphere. In specific embodiments of this invention, these include, but are not limited to, total columnar water vapor content (CWV), total ozone (TOZ), and sea level pressure (SLP). These parameters are used to mitigate the uncertainties caused by spatiotemporal variations in atmospheric conditions and enhance the model's cross-temporal and spatiotemporal generalization ability.

[0032] Geometric parameters (Geo): These parameters reflect the observational geometry between the light source and the sensor, directly affecting the propagation path of light in the atmosphere-water system. Specifically, they include, but are not limited to, solar zenith angle (SZA), solar azimuth angle (SAA), observation zenith angle (OZA), and observation azimuth angle (OAA).

[0033] Z-score normalization is a data normalization method that uses a linear transformation based on the mean and standard deviation of the original data. In this invention, this method is used to transform radiance, geometric parameters (Geo), and meteorological parameters (Meteo) with different physical meanings and dimensions to a unified numerical distribution range, thereby eliminating dimensional differences between features and accelerating model convergence.

[0034] Step 12: Pixel-level quality filtering based on quality control flags: To ensure the consistency and effectiveness of model input and output, pixel-level filtering is performed on the acquired remote sensing images using quality control flags L2_flags. Specifically, pixels affected by the following factors are strictly excluded: land, coastline, fresh inland water, tidal region, bright region, stray light risk, invalid, cosmetic, duplicated, sun-glint risk, dubious, saturated, and partially saturated. All pixels judged as invalid are uniformly marked as NaN, thereby reducing the interference of noisy data on model learning from the source.

[0035] Step 13, Resampling and Spatial Alignment of Auxiliary Variables: Since the acquired geometric parameters (Geo) and meteorological parameters (Meteo) are usually stored in the form of sparse tie-point grids, in order to ensure the consistency of the geometric parameters Geo, meteorological parameters Meteo and spectral band data in spatial dimension, a bilinear interpolation method is used to resample them to the same resolution as the radiation data, ensuring pixel-by-pixel alignment of the input features and avoiding systematic errors introduced by spatial mismatch; Step 14, Patch Construction and Invalid Value Filling: In the quality-filtered and spatially aligned data, the track-level imagery is divided into fixed-size image patches (e.g., 128×128 pixel patches); an effective pixel ratio threshold (e.g., 0.9) is set, and only image patches with an effective pixel ratio greater than 90% are retained to ensure sufficient supervision information during model training; for the qualified image patches after filtering, the mean of each channel is calculated, and the mean is used to fill the NaN invalid values ​​in the patch, thereby ensuring the integrity and numerical stability of the model input tensor; Step 15: Z-score standardization of data: After completing the tile construction, Z-score standardization is performed on all L1B input features (including radiance, geometric parameter Geo, and meteorological parameter Meteo) and L2 output labels (true reflectance Rrs). This involves subtracting the mean of each channel's data and then dividing by its standard deviation. This effectively eliminates the dimensional differences between different physical variables and significantly enhances the numerical stability and convergence speed in the subsequent model optimization process.

[0036] Step 2: Construct a multi-scale Transformer atmospheric correction model that includes parallel convolutional branches.

[0037] A deep learning network based on an encoder-decoder architecture is constructed. In the feature extraction stage of the encoder and decoder, a window multi-head self-attention mechanism (W-MHSA-PC) with parallel convolutional branches is adopted. This mechanism aims to capture global dependency information through self-attention within a local window and simultaneously extract local high-frequency textures (such as water details and edges) using parallel convolutional branches. Subsequently, a multilayer perceptron (MLP) is combined to reconstruct and reduce the dimensionality of nonlinear features between channels, thereby constructing a multi-scale, two-layer feature representation, such as... Figure 2 As shown, the atmospheric correction model network architecture is described in detail, which consists of a macroscopic encoder-decoder overall process and a microscopic internal structure of the atmospheric correction block: In terms of the overall macroscopic process, the initial input to the model is a concatenated multi-source tensor (with dimensions of H×W×13, where H and W represent the height and width of the image, respectively, and 13 represents the total number of channels including visible light, geometric, and meteorological parameters). This input tensor is then processed by a 3×3 convolutional layer for shallow feature extraction, mapping it to a feature map of size H×W×C (where C is the number of basic channels). During the encoder stage, the feature map alternately passes through an AtmoCorr Block and two downsampling modules, progressively reducing its spatial size. and This process progressively extracts multi-scale global deep features. In the decoder stage, deep features are alternately passed through two upsampling modules and an atmospheric correction block (AtmoCorr Block). After each upsampling, a selective convolutional kernel fusion (SK Fusion) module is introduced to perform skip connections and adaptive fusion of features at corresponding scales from the encoder. The feature map size is sequentially restored to [size missing]. And H×W×C. Finally, the feature map restored to the original resolution is processed by a 3×3 convolution and a soft reconstruction module to output a final atmospheric bottom visible light reflectance data tensor with a size of H×W×6.

[0038] In terms of the micro-module structure, the internal feature flow of the atmospheric correction block is demonstrated. After being scaled and normalized (RescaleNorm), the input features are split into a windowed multi-head self-attention module (W-MHSA-PC) containing parallel convolutions. Its self-attention branch generates queries and keys (Q&K) and values ​​(V) through linear layers, calculates attention weights through matrix multiplication (MatMul) and the Softmax activation function, and then performs matrix multiplication and merging operations again to capture the global context. Simultaneously, the parallel convolution branch extracts local high-frequency textures through linear layers and convolution (Conv) operations. The outputs of both branches are element-wise added and dimensionality reduced by linear layers, then added to the original input with the first residual. Subsequently, the output passes through an affine transformation layer and a multilayer perceptron (MLP, consisting of two linear layers and a ReLU activation function in the middle). Finally, the output of the MLP is added to the features before the affine transformation with the second residual, yielding the output features of the atmospheric correction block.

[0039] In this embodiment of the invention, step 2 includes: Step 21: Construct an overall network architecture based on an encoder-decoder: Construct a multi-scale Transformer framework that integrates windowed multi-head self-attention (W-MHSA) and multilayer perceptron (MLP); the overall architecture is based on an encoder-decoder structure: In the encoder part, the network progressively extracts global features from the input image through multiple atmospheric correction blocks (AtmoCorr Blocks) and two down-sampling modules (Down-sample); In the decoder part, the network restores the spatial resolution of the image through an up-sampling module (Up-sample), and combines SK Fusion and subsequent atmospheric correction blocks for detail reconstruction to ensure the preservation of local textures in the image; Step 22: Construct a windowed multi-head self-attention mechanism (W-MHSA-PC) in the atmospheric correction block: To enhance the model's ability to model features at different scales, a parallel convolutional branch is introduced into the standard multi-head self-attention mechanism (MHSA) to construct the W-MHSA module. During the operation of this module, the self-attention mechanism is responsible for capturing global information within a local window, while the parallel convolutional branch is responsible for enhancing the extraction of local high-frequency textures (such as edges, textures, and water details). Step 23: Adaptive Feature Fusion Based on Gated Weighted Fusion Mechanism: In the fusion stage of the W-MHSA module, the features output by the self-attention branch and the parallel convolution branch are adaptively fused through a gated weighted fusion mechanism. This fusion mechanism effectively balances the relationship between global dependency modeling and local detail preservation, enabling the model to effectively learn the two-layer features (i.e., atmospheric scattering features and water reflection features) under complex atmospheric conditions, and significantly enhances the stability and generalization ability of the model. Step 24: Channel Feature Reconstruction Using a Multilayer Perceptron (MLP): An MLP is introduced after each attention block to enhance nonlinear interactions between channels. The MLP enhances cross-band correlation modeling capabilities under multispectral input by performing channel reconstruction processes involving dimensionality upscaling, nonlinear activation, and dimensionality reduction. In the detail recovery stage, the MLP further optimizes the information flow between channels and enhances training stability through residual connections. Particularly in the decoding stage, the MLP acts as a detail shaper, compressing the output features of the attention and convolution branches into a physically reasonable channel space, effectively reducing over-smoothing and ringing artifacts, and further improving the accuracy of surface reflectance reconstruction.

[0040] Step 3: Construct a composite loss function that includes physical soft constraints, and train the atmospheric correction model using the training dataset to obtain the trained model.

[0041] In this embodiment of the invention, a composite training objective function combining Charbonnier loss, mean squared error (MSE) loss, and range constraint loss is designed. The Charbonnier loss is used to suppress interference from anomalous atmospheric noise, while the range constraint loss applies a soft penalty to the model's prediction results, forcing the output BOA reflectance to be limited to a physically reasonable range (e.g., [0,1]). The constructed training dataset is input into the model, and the backpropagation algorithm is used to optimize the network parameters until the model converges, resulting in the trained multi-scale Transformer atmospheric correction model.

[0042] In this embodiment of the invention, step 3 includes: Step 31: Construct a composite objective function containing multiple penalties: Construct a function incorporating Charbonnier loss. Mean Square Error (MSE) Loss and range of penalty losses The composite loss function; In this embodiment of the invention, step 31 includes: Using Charbonnier loss as the primary loss term, compared to traditional L1 loss, Charbonnier loss is insensitive to outliers and can effectively reduce the impact of significant atmospheric noise, thereby enhancing the model's robustness in environments with significant noise or distortion. Its mathematical expression incorporates a smoothing constant (e.g., set as...). MSE loss is used as an auxiliary loss term to emphasize pixel-level numerical accuracy, aiming to narrow the gap between predicted and true values ​​and thus accelerate model convergence. Range constraint loss is introduced to apply a soft penalty to the model's prediction results to avoid negative or outlier values ​​greater than 1, and to force the output reflectivity to be within the physically reasonable range of [0,1] to ensure the physical consistency of the output.

[0043] Step 32: Set the weights for the composite loss function: The final training objective function is the weighted sum of the three losses mentioned above, and its expression is: ; In practical model optimization configurations, to avoid the model overfitting minute pixel-level differences due to excessively high MSE weights, thereby weakening the Charbonnier loss's ability to suppress outliers and maintain physical consistency through range constraints, the weights of the MSE loss are adjusted accordingly. Weights of range constraint loss All are set to the optimal balance value of 0.05; Step 33: Configure the adaptive optimizer and dynamic learning rate scheduling strategy: When optimizing network parameters through backpropagation, the Adam optimizer is used, with an initial learning rate of 0.001 and momentum parameters configured. =0.9 and =0.999; To stabilize the training process, a dynamic learning rate adjustment strategy (ReduceLROnPlateau) is introduced. When the root mean square error (RMSE) on the validation set does not improve within a set number of consecutive epochs (e.g., 7 epochs), the learning rate is automatically reduced by a predetermined decay factor (e.g., 0.1); at the same time, the lower bound of the learning rate is set to 1×10. -6This effectively ensures rapid convergence of the model in the early stages of training and enables fine-grained optimization of parameters in the later stages of training. Step 34: Introduce a stable training mechanism with mixed precision and gradient clipping: During the batch training of the model, automatic mixed precision (AMP) technology is introduced to significantly reduce memory consumption and accelerate the computation process without compromising the model's prediction accuracy. In addition, in order to effectively reduce the risk of gradient explosion during the backpropagation process of complex networks, a gradient clipping mechanism is applied to limit the maximum gradient norm to 1.

[0044] Step 4: Utilize the trained model to perform high-throughput end-to-end atmospheric correction inference.

[0045] In practical large-scale remote sensing applications, the original visible light TOA radiation data of the area to be corrected, along with the corresponding real-time geometric and meteorological parameters, are input into the trained model. The model can then directly and quickly output the corresponding bottom atmospheric area (BOA) reflectance, thereby bypassing the traditional physical inversion calculation process and achieving efficient and high-precision marine atmospheric correction.

[0046] In this embodiment of the invention, step 4 includes: Step 41: Obtain multi-source real-time remote sensing data of the area to be corrected: Obtain satellite remote sensing images of the target area to be corrected, extract the raw TOA radiation data of the visible light band, and simultaneously extract the corresponding real-time auxiliary parameters of the images, including the corresponding geometric parameters (such as the solar zenith angle and azimuth angle of the observation) and meteorological parameters (such as total columnar water vapor content, total ozone and sea level pressure). Step 42: Perform consistent data preprocessing: Perform preprocessing on the data to be corrected, using quality control flags to remove invalid pixels affected by clouds and land; perform bilinear interpolation on sparse grid auxiliary variables to achieve pixel-by-pixel alignment of spatial resolution; and perform Z-score standardization on all input features to ensure that the data distribution of the input features to be corrected is consistent with that during model training. Step 43: End-to-end rapid inference and output of bottom atmospheric reflectance (BOA): The preprocessed TOA radiation data to be corrected, along with the corresponding geometric and meteorological parameters, are used as feature tensors and directly input into the trained multi-scale Transformer atmospheric correction model. The model directly and rapidly maps and outputs the predicted bottom atmospheric reflectance (BOA) surface spectral reflectance through a forward propagation network. This inference process completely bypasses the extremely time-consuming and complex atmospheric parameter inversion and lookup table (LUT) generation stages in traditional physical models, thus providing a highly computationally efficient solution for automated, high-throughput global large-scale satellite remote sensing data processing.

[0047] This invention aims to significantly improve computational efficiency by directly establishing an end-to-end mapping of reflectance from the top of the atmosphere (TOA) to the bottom of the atmosphere (BOA), completely bypassing the time-consuming physical inversion process. Simultaneously, this invention aims to enhance the extraction of local details while effectively capturing global environmental information by introducing specific feature extraction structures and auxiliary meteorological parameters, ensuring that the model output is strictly limited to a physically reasonable range. This achieves high-precision, robust, and physically consistent atmospheric correction in complex, large-scale ocean environments.

[0048] Compared with the prior art, the present invention has the following significant advantages: (1) Significantly improves computational efficiency and breaks through the bottleneck of large-scale remote sensing data processing. Existing methods based on traditional physical models rely heavily on complex atmospheric parameter inversion and cumbersome lookup table (LUT) generation processes, resulting in extremely high computational costs. This invention directly establishes a mapping from radiation from the top of the visible atmosphere (TOA) to the bottom of the atmosphere (BOA) by constructing an end-to-end multi-scale Transformer architecture. This method completely bypasses the time-consuming physical inversion and lookup table search process, providing a highly efficient solution for automated, high-throughput global large-scale satellite remote sensing data processing.

[0049] (2) Effectively balances global features and local high-frequency details, significantly improving the accuracy of spatial structure reconstruction. Addressing the limitations of existing convolutional neural networks (CNNs) in capturing complex ocean gradients due to their inherent local receptive fields, and generative adversarial networks (GANs) prone to artifacts, this invention innovatively introduces a windowed multi-head self-attention (W-MHSA) module in the feature extraction stage. This module utilizes a self-attention mechanism to capture global dependency information while compensating for the lack of local inductive bias through parallel convolutional branches. Experiments show that, compared to the pure Transformer architecture, the structural similarity index (SSIM) of the model increases from 0.82 to 0.86 after introducing parallel convolutions, effectively preserving complex coastal boundaries and fine-grained ocean textures, and overcoming the over-smoothing problem.

[0050] (3) Possesses strong cross-temporal and spatial generalization ability and physical interpretability. Traditional deep learning models are mostly limited to local nearshore small sample scenarios, resulting in poor generalization ability. This invention not only introduces geometric parameters, but also innovatively uses real-time meteorological parameters (total columnar water vapor content, total ozone, and sea level pressure) as physical atmospheric state constraints input to the model. Ablation experiments demonstrate that the introduction of meteorological parameters effectively eliminates the global systematic radiation bias caused by gas absorption and aerosol scattering, reducing RMSE by 50%. Furthermore, on completely independent test sets for the four seasons of 2024, this model still maintains excellent performance (RMSE can still reach 0.006 in winter, and SSIM is 0.92), and shows high consistency in the AERONET-OC measured data validation, proving that the model has successfully learned the underlying radiative transfer optical principles, rather than simply the systematic bias of fitting official data.

[0051] (4) Ensure the physical consistency and extremely high numerical fidelity of the output results. In remote sensing data processing, negative reflectance values ​​and other abnormal values ​​are easily affected by environmental noise. When constructing the composite loss function, this invention introduces soft range constraints, which strictly limit the output to a physically reasonable range of [0,1] by penalizing the prediction results. At the same time, combined with the Charbonnier loss which is robust to outliers, this invention achieves extremely high numerical fidelity on challenging multi-source test sets, with an overall performance of excellent levels of root mean square error (RMSE) = 0.005, mean relative deviation (MRD) = 0.12, and peak signal-to-noise ratio (PSNR) = 40.9.

[0052] In this embodiment of the invention, although the OLCI sensor data from the Sentinel-3 satellite is used as an example for selecting the data input source and frequency band, this end-to-end atmospheric correction framework is not limited to a specific satellite. Those skilled in the art can directly extend it to other remote sensing platforms with multispectral detection capabilities, such as MODIS and VIIRS. Furthermore, although six visible light bands from 400 nm to 560 nm are selected in this specific embodiment, the input and output bands can be expanded according to the actual sensor specifications, for example, by adding near-infrared or short-wave infrared bands to meet the detection requirements of different water quality parameters.

[0053] In this embodiment of the invention, meteorological parameters such as total columnar water vapor content (CWV), total ozone (TOZ), and sea level pressure (SLP), as well as related geometric parameters, are used for auxiliary feature parameters and their acquisition methods. As an alternative, in addition to using satellite-provided auxiliary data, third-party meteorological reanalysis datasets can be introduced for spatiotemporal matching and then used as input. Furthermore, the input features can be further expanded, for example, by adding environmental variables such as wind speed, estimated aerosol optical thickness, or digital elevation models, to enhance the physical constraints of the model under extreme climates or complex terrain.

[0054] In this embodiment of the invention, regarding the local extraction module of the core network architecture, to prevent direct circumvention of the network structure, the parallel convolutional branch introduced by the W-MHSA module can be equivalently replaced. For example, this local feature extraction branch can be replaced with depthwise separable convolution, deformable convolution, or local perceptual blocks to achieve the same high-frequency texture extraction effect. Simultaneously, the gated weighted fusion mechanism used between self-attention features and convolutional features can also be replaced with conventional feature channel concatenation, element-wise addition, or the introduction of channel attention mechanisms, etc., as feature fusion methods.

[0055] In this embodiment of the invention, regarding the loss function and numerical constraint mechanism, a composite objective function combining Charbonnier loss, mean squared error loss, and range penalty loss is designed. As an equivalent replacement, the main loss term, Charbonnier loss, can be replaced with Huber loss or mean absolute error. Furthermore, to prevent anomalies such as negative reflectance values, the embodiment introduces a soft range constraint penalty to limit the output to the [0,1] interval. As an alternative to hard constraints, a specific activation function (such as Sigmoid) or truncation function can be directly added to the final output layer of the model decoder to enforce physical consistency at the network structure level.

[0056] Finally, regarding the transfer of application scenarios, although this invention is designed for large-scale complex marine environments in water color remote sensing, the multi-scale Transformer architecture and its physical meteorological constraint mechanism can be directly transferred and applied to atmospheric correction tasks of land surface reflectance to solve the problem of restoring the true reflectance of land vegetation, cities and other land features.

[0057] The technical solution provided by this invention includes acquiring multi-source remote sensing observation data and auxiliary parameters, performing joint preprocessing, and constructing a standardized training dataset; constructing a multi-scale Transformer atmospheric correction model containing parallel convolutional branches; constructing a composite loss function containing physical soft constraints, and training the atmospheric correction model using the training dataset to obtain the trained model; and using the trained model to perform high-throughput end-to-end atmospheric correction inference. In complex large-scale ocean environments, this method achieves high-precision, robust, and physically consistent atmospheric correction.

[0058] The various steps in the embodiments of the present invention can be performed by an electronic device. This electronic device includes, but is not limited to, tablet computers, portable PCs, and desktop computers.

[0059] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the electronic device containing the computer-readable storage medium to execute the above-described embodiment of the atmospheric correction method for marine remote sensing images based on multi-scale Transformer.

[0060] Figure 3 A schematic diagram of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the electronic device 21 includes a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, it implements the atmospheric correction method for marine remote sensing images based on multi-scale Transformer in the embodiment. To avoid repetition, it will not be described in detail here.

[0061] Electronic device 21 includes, but is not limited to, processor 211 and memory 212. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 21 and does not constitute a limitation on electronic device 21. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0062] The processor 211 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0063] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or RAM of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the electronic device 21. Furthermore, the memory 212 can include both internal and external storage units of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0064] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for atmospheric correction of ocean remote sensing images based on multi-scale Transformer, characterized in that, The method includes: Step 1: Acquire multi-source remote sensing observation data and auxiliary parameters, and perform joint preprocessing to construct a standardized training dataset; Step 2: Construct a multi-scale Transformer atmospheric correction model that includes parallel convolutional branches; Step 3: Construct a composite loss function that includes physical soft constraints, and train the atmospheric correction model using the training dataset to obtain the trained model; Step 4: Utilize the trained model to perform high-throughput end-to-end atmospheric correction inference.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 11, Multi-source data acquisition and feature set construction: Acquire visible light band TOA radiation data observed by satellite sensors, and simultaneously acquire the corresponding geometric parameters Geo and meteorological parameters Meteo as feature inputs; acquire corrected surface reflectance data as the true BOA label of the lower atmosphere; Step 12, Pixel-level quality filtering based on quality control flags: To ensure the consistency and effectiveness of model input and output, the acquired remote sensing images are filtered at the pixel level using quality control flags L2_flags. Step 13, Resampling and Spatial Alignment of Auxiliary Variables: To ensure the consistency of geometric parameters Geo, meteorological parameters Meteo and spectral band data in spatial dimension, bilinear interpolation is used to resample them to the same resolution as the radiation data. Step 14, Patch Construction and Invalid Value Filling: In the data that has been quality filtered and spatially aligned, the track-level imagery is divided into fixed-size image patches; an effective pixel ratio threshold is set, and only image patches with an effective pixel ratio greater than 90% are retained; for the qualified image patches after filtering, the mean of each channel is calculated, and the mean is used to fill the NaN invalid values ​​in the patch. Step 15: Z-score standardization of data: After completing the tile construction, perform Z-score standardization on all L1B input features and L2 output labels, that is, subtract the mean of each channel's data and then divide it by its standard deviation.

3. The method according to claim 2, characterized in that, Step 2 includes: Step 21: Construct an overall network architecture based on an encoder-decoder: Construct a multi-scale Transformer framework that integrates windowed multi-head self-attention (W-MHSA) and multilayer perceptron (MLP); the overall architecture is based on an encoder-decoder structure: In the encoder part, the network progressively extracts global features from the input image through multiple atmospheric correction blocks (AtmoCorr Blocks) and two down-sampling modules (Down-sample); In the decoder part, the network restores the spatial resolution of the image through an up-sampling module (Up-sample), and combines SK Fusion and subsequent atmospheric correction blocks for detail reconstruction to ensure the preservation of local textures in the image; Step 22: Constructing a windowed multi-head self-attention mechanism W-MHSA-PC in the atmospheric correction block: In order to enhance the model's ability to model features at different scales, a parallel convolutional branch is introduced into the standard multi-head self-attention mechanism MHSA to construct the W-MHSA module. Step 23: Adaptive feature fusion based on gated weighting mechanism: In the fusion stage of the W-MHSA module, the features output by the self-attention branch and the parallel convolution branch are adaptively fused through the gated weighting fusion mechanism. Step 24: Reconstruct channel features using a multilayer perceptron (MLP): Introduce a multilayer perceptron (MLP) after each attention block to enhance the nonlinear interaction between channels.

4. The method according to claim 3, characterized in that, Step 3 includes: Step 31: Construct a composite objective function containing multiple penalties: Construct a function incorporating Charbonnier loss. Mean Square Error (MSE) Loss and range of penalty losses The composite loss function; Step 32: Set the weights for the composite loss function: The final training objective function is the weighted sum of the three losses mentioned above, and its expression is: ； Weight the MSE loss Weights of range constraint loss All are set to the optimal balance value of 0.05; Step 33: Configure the adaptive optimizer and dynamic learning rate scheduling strategy: When optimizing network parameters through backpropagation, the Adam optimizer is used, with an initial learning rate of 0.001 and momentum parameters configured. =0.9 and =0.999; To stabilize the training process, a dynamic learning rate adjustment strategy is introduced. When the root mean square error (RMSE) on the validation set does not improve within a set number of consecutive epochs, the learning rate is automatically reduced by a predetermined decay factor; at the same time, the lower bound of the learning rate is set to 1×10. -6 ; Step 34: Introduce a stable training mechanism with mixed precision and gradient clipping: During the batch training of the model, introduce automatic mixed precision (AMP) technology; apply the gradient clipping mechanism to limit the maximum gradient norm to 1.

5. The method according to claim 4, characterized in that, Step 4 includes: Step 41: Obtain multi-source real-time remote sensing data of the area to be corrected: Obtain satellite remote sensing images of the target area to be corrected, extract the raw TOA radiation data of the visible light band, and simultaneously extract the real-time auxiliary parameters corresponding to the images, including the corresponding geometric parameters and meteorological parameters. Step 42: Perform consistent data preprocessing: Perform preprocessing on the data to be corrected, using quality control flags to remove invalid pixels affected by clouds and land; perform bilinear interpolation on sparse grid auxiliary variables to achieve pixel-by-pixel alignment of spatial resolution; and perform Z-score standardization on all input features to ensure that the data distribution of the input features to be corrected is consistent with that during model training. Step 43: End-to-end fast inference and output of bottom atmospheric reflectance (BOA): The preprocessed TOA radiation data to be corrected and the corresponding geometric and meteorological parameters are used as feature tensors and directly input into the trained multi-scale Transformer atmospheric correction model; the model maps and outputs the predicted bottom atmospheric reflectance (BOA) surface spectral reflectance through a forward propagation network.

6. The method according to claim 5, characterized in that, Step 31 includes: Charbonnier loss is used as the primary loss term, and a smoothing constant is introduced in its mathematical expression. MSE loss is used as an auxiliary loss term to emphasize pixel-level numerical accuracy, aiming to narrow the gap between predicted and true values. Range constraint loss is introduced to apply a soft penalty to the model's prediction results to avoid negative or outlier values ​​greater than 1, and to force the output reflectivity to be within the physically reasonable range of [0,1] to ensure the physical consistency of the output.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the atmospheric correction method for marine remote sensing images based on multi-scale Transformer as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs including instructions that, when executed by the device, cause the device to perform the atmospheric correction method for ocean remote sensing images based on multi-scale Transformer as described in any one of claims 1 to 6.

Images