Reversible neural network based remote sensing calibration correction bidirectional coupling deep learning method

Abstract

The present application relates to the technical field of atmospheric correction, and more particularly to a remote sensing calibration correction bidirectional coupling deep learning method based on reversible neural network. The method comprises the following steps: extracting the apparent radiance pixel sequence of the target water area; constructing an environmental prior data set; using an FT-Transformer network to perform deep heterogeneous feature coding on multi-source auxiliary physical data to generate a physical condition embedding vector; constructing a reversible neural network architecture to establish an equal-dimension mapping channel; performing adaptive modulation on the radiation transmission process of the real physical environment within the reversible network; performing end-to-end bidirectional physical constraint training based on a joint loss function, forcing the network to learn the radiation transmission physical law in the forward direction and compressing the atmospheric disturbance to the latent variable space; and outputting high-precision remote sensing reflectivity in the actual inference stage. The method overcomes the problems of missing physical mechanism, weak generalization ability and low calculation efficiency in the existing atmospheric correction technology for two types of turbid water bodies.

Description

Technical Field

[0001] This invention relates to the field of atmospheric correction technology, and in particular to a two-way coupled deep learning method for remote sensing calibration correction based on a reversible neural network. Background Technology

[0002] With the increasing global demand for water environment monitoring, high-precision remote sensing inversion of water color for Case II waters, including nearshore seas, estuaries, and inland lakes, has become a crucial aspect of marine science, environmental monitoring, and carbon cycle research. The core task of water color remote sensing is to extract atmospheric path radiation and water surface reflectance components from the entrance pupil radiance (TOA) received from satellite sensors, thereby obtaining the water-leaving radiance or remote sensing reflectance (Ro) carrying biological optical information of the water body. rs This process is called atmospheric correction. For Class I water bodies such as open oceans, the existing atmospheric correction algorithms are relatively mature because the optical properties of the water are relatively simple and absorption by phytoplankton is the main factor. However, for Class II turbid water bodies, the concentrations of suspended particulate matter (SPM) and colored soluble organic matter (CDOM) are extremely high and vary drastically, and their optical properties are extremely complex. This makes atmospheric correction a typical ill-posed inversion problem: more than 90% of the signal received by the sensor comes from atmospheric scattering and water surface reflection, with less than 10% of the weak signal originating from the water body. Furthermore, it is highly susceptible to interference from the variable types of aerosols and the nonlinearity of solar geometry.

[0003] Existing atmospheric correction methods for Class II turbid water bodies are mainly divided into two categories: traditional physical model methods and conventional deep learning methods, but both have significant limitations. Although major space agencies such as NASA and ESA have released standard Level-2 product algorithms based on physical models for satellites such as MODIS and Sentinel-3 (such as NASA's NIR-SWIR iterative algorithm or ESA's C2RCC neural network algorithm), these general algorithms still face significant challenges when dealing with nearshore Class II turbid water bodies. Traditional physics algorithms typically rely on pre-built massive lookup tables (LUTs) or complex iterative optimization processes. Their drawbacks are twofold: firstly, to reduce computational complexity, lookup tables often discretize aerosol patterns and geometric angles, leading to unavoidable interpolation errors; secondly, these standard algorithms are usually based on specific ideal assumptions (such as the deep-sea dark pixel assumption), which often fail in complex nearshore environments with extreme turbidity or high aerosol optical thickness. This results in negative or abnormally high values ​​of remote sensing reflectance in official standard products, and the computation process is extremely time-consuming, making it difficult to meet the real-time processing needs of massive satellite data.

[0004] In recent years, while data-driven methods, represented by Convolutional Neural Networks (CNNs), have improved computational efficiency, they still suffer from serious theoretical flaws. First, most existing deep learning models employ an end-to-end unidirectional regression strategy, directly mapping satellite observation data to water reflectance. This black-box approach completely ignores the physical mechanisms of photon propagation in the atmosphere and water, resulting in a lack of physical interpretability and a tendency to produce outputs that violate physical laws (such as negative values). Second, when processing multi-source heterogeneous information, existing methods typically simply concatenate auxiliary data such as solar zenith angle and wind speed with spectral data. This coarse fusion method fails to effectively capture the nonlinear modulation effects of environmental variables on spectral features. Furthermore, because training data is often limited to specific regions, conventional models exhibit extremely poor generalization ability when faced with unseen sea areas or extreme weather conditions. Summary of the Invention

[0005] In view of this, the present invention provides a two-way coupled deep learning method for remote sensing calibration correction based on reversible neural networks, in order to overcome the problems of lack of physical mechanism, weak generalization ability and low computational efficiency in existing atmospheric correction techniques for Class II turbid water bodies.

[0006] In a first aspect, the present invention provides a bidirectional coupled deep learning method for remote sensing calibration and correction based on a reversible neural network, the method comprising:

[0007] Step 1: Acquire satellite multispectral observation image data, perform sub-pixel-level spatiotemporal matching and bitmask quality control, remove cloud and abnormal interference, and extract the apparent radiance pixel sequence of the target water area; Step 2: Acquire multi-source auxiliary physical data that is spatiotemporally synchronized with satellite observations, perform spatial interpolation and feature alignment, and construct an environmental prior dataset; Step 3: Use the feature-based meta-transformer network FT-Transformer to perform deep heterogeneous feature encoding on multi-source auxiliary physical data to generate physical condition embedding vectors containing global environmental semantics. Step 4: Construct a reversible neural network architecture and establish an equal-dimensional mapping channel between the water remote sensing reflectance state space and the satellite observation space by alternating affine coupling blocks and orthogonal Householder channel permutation layers. Step 5: Within the reversible network, the spectral cross-attention mechanism is used to embed the physical condition vector as the key and value, and the spectral features of the main data stream are dynamically queried and fused to adaptively modulate the radiative transfer process in the real physical environment. Step 6: Perform end-to-end bidirectional physical constraint training based on the joint loss function. In the positive direction, force the network to learn the physical laws of radiative transfer and compress atmospheric disturbances into the latent variable space. In the actual inference stage, perform posterior mode optimization based on the observed signal and random sampling of the latent space to output high-precision remote sensing reflectance with atmospheric interference removed.

[0008] Optionally, step 3 includes: The design incorporates a heterogeneous feature extraction mechanism based on the FT-Transformer. First, each scalar environmental parameter... The feature tokenizer layer maps the data to high-dimensional semantic vectors. : ; in, and For learnable weights and biases; The high-dimensional token is then processed by the self-attention interaction of the Transformer encoder to generate a conditional embedding that includes the global context. This provides a physical background constraint for subsequent spectral decoupling.

[0009] Optionally, step 4 includes: A reversible neural network architecture between the physical parameter space and the observation data space is constructed. Based on the theory of reversible neural networks (INN), a bidirectional closed-loop system is constructed. The system utilizes the mathematical properties of reversible transformation to define the atmospheric correction process as the inverse mapping of the radiative transfer process. Define physical state vector That is, remote sensing reflectance and observation state vector That is, the radiance of satellite L1B, introducing potential variables. Explicit modeling of atmospheric path radiation and sensor noise; under environmental condition vectors Under the control of [the model], the inverse inference process of the model, namely atmospheric correction, is described by the following mathematical bijective function: ; in, For parameterized invertible inverse transformation, For the sampled random noise, These are the embedded environmental features.

[0010] Optionally, step 4 further includes: The Householder reflection transform is introduced to replace random matrix permutation or 1×1 convolution; the Householder reflection transform uses the product of orthogonal reflection matrices to obtain eigenvectors. The global rotation of is expressed as: ; in, To output the feature vector, For the input feature vector, It is the identity matrix. For continuous multiplication from k=1 to K, Let K be the learnable reflection plane normal vector, and K be the number of reflections. for transpose, for The square of the second norm.

[0011] Optionally, step 5 includes: To ensure reversibility, spectral features and environmental information are deeply fused, and the backbone network employs stacked affine coupling blocks. Within each affine coupling block, the input features are split into two parts. and The model utilizes conditional embedding. The scaling and translation transformations of spectral features are performed; the forward propagation of this process, i.e., the radiative transfer simulation, is defined as follows: ; in, and These represent the two output features obtained after affine coupled block transformation. It is an exponential scaling function. and This is the nonlinear transformation function fitted by the internal spectral Transformer subnet. This represents the Hadamard product, which is the product of elements.

[0012] Optionally, step 6 includes: Constructing a hybrid loss function system corresponding to the physical process The hybrid loss function system utilizes the bidirectional characteristics of the reversible neural network to apply physical constraints to both the forward radiative transfer process and the reverse atmospheric correction process. Total loss function Defined as: ; in, To simulate the constraint loss for forward radiative transfer, for The corresponding weighting coefficients, For asymmetric reconstruction loss, for The corresponding weighting coefficients, For spectral angle mapping loss, for The corresponding weighting coefficients, For the maximum mean difference loss in the potential space, for The corresponding weighting coefficients.

[0013] Optionally, it includes: I. In the forward path, the network will display the actual remote sensing reflectance. With zero-filled vector As input, the output is predicted satellite observations. and latent variables Define the following two losses: First, we introduce the simulated loss of forward radiative transfer. Simulated observation data generated by the forward radiative transfer simulation loss forced model Approximates real satellite observations This allows the neural network to learn the radiative transfer equation (RTE) physical mechanism of photons traveling from water to the sensor, and its expression is: ; Secondly, the maximum mean difference loss of the potential space is introduced. Latent variables output by the latent space maximum mean difference loss constrained model Distribution and standard Gaussian distribution The expression for minimizing the statistical distance between them is: ; in, This represents the summation of the MMD results from k kernel functions, where MMD... k This represents the maximum mean difference calculated under the k-th kernel function. Indicates the standard Gaussian distribution Prior latent variables obtained from sampling; II. In the reverse path, the network will transmit satellite observation data. With random sampling noise As input, the output is the predicted remote sensing reflectance. Define the following two losses: First, an asymmetric reconstruction loss is introduced into the logarithmic space. Define prediction error When the predicted value Less than the true value At that time, apply a doubling penalty. This is to encourage the model to output non-negative results during training, thus suppressing negative values. Its expression is: ; in, This represents the number of samples involved in the loss calculation, where i represents the sample index. This indicates that the error terms are summed over all samples. This represents the prediction error on the i-th sample. This represents the asymmetric weight corresponding to the i-th error. This represents the multiplication penalty coefficient; Secondly, spectral angle mapping loss is introduced. Spectral angle mapping loss constrained prediction of spectral vector With the true spectral vector The angle between the two points is minimized, and its expression is: ; Among them, arccos Represents the inverse cosine function. Represents extremely small positive numbers. The L2 norm of the predicted spectral vector is represented. The L2 norm represents the true spectral vector.

[0014] Optionally, step 6 further includes: To design a robust inference paradigm based on test-time enhanced TTA and mean-shift clustering, and to obtain the maximum a posteriori probability estimate (MAP), i.e., to find the mode with the highest probability density in the solution space under the joint constraints of current observations and environment, the steps are as follows: First, for each fixed observation input, the model independently samples multiple times from the standard normal distribution in the latent space, and generates a set of candidate solutions through an inverse network. ; Subsequently, the mean-shift algorithm is used to perform a pattern search in the solution space. To improve robustness to outliers and accelerate convergence, the algorithm first calculates the median of the dimension of all candidate solutions as an initial estimate. Next, the local density gradient is estimated using a Gaussian kernel function, and the estimated value is iteratively shifted towards the density maxima; The update in the next iteration follows the formula: ; in, The center position at the t-th iteration. Let h be a Gaussian kernel function with bandwidth h; The iteration stops when the center displacement is less than a preset threshold, and the iteration eventually converges. This refers to the high-precision remote sensing reflectance output.

[0015] In a second aspect, 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 where the computer-readable storage medium is located to execute the bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network, as described in the first aspect or any possible implementation thereof.

[0016] Thirdly, embodiments of the present invention provide an electronic device, including: 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 bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network in the first aspect or any possible implementation of the first aspect.

[0017] The technical solution provided by this invention includes the following steps: acquiring satellite multispectral observation image data; performing sub-pixel-level spatiotemporal matching and bitmask quality control to remove cloud and anomalous interference; extracting the apparent radiance pixel sequence of the target water area; acquiring multi-source auxiliary physical data synchronized with satellite observations; performing spatial interpolation and feature alignment to construct an environmental prior dataset; using the feature lexical transformation Transformer network FT-Transformer to perform deep heterogeneous feature encoding on the multi-source auxiliary physical data to generate physical condition embedding vectors containing global environmental semantics; and constructing a reversible neural network architecture to establish a water remote sensing reflectance state space and a dynamical property space through alternating affine coupling blocks and orthogonal Householder channel permutation layers. This method employs an equal-dimensional mapping channel between satellite observation spaces. Within the reversible network, a spectral cross-attention mechanism is used to embed physical conditions into vectors as keys and values, dynamically querying and fusing the spectral features of the main data stream to adaptively modulate the radiative transfer process in the real physical environment. End-to-end bidirectional physical constraint training is performed based on a joint loss function, forcing the network to learn the physical laws of radiative transfer in the positive direction and compressing atmospheric disturbances into the latent variable space. In the actual inference stage, posterior mode optimization is performed based on random sampling of the observed signal and the latent space, outputting a high-precision remote sensing reflectance with atmospheric interference removed. This method overcomes the problems of missing physical mechanisms, weak generalization ability, and low computational efficiency in existing atmospheric correction techniques for Class II turbid water bodies. 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 a bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network provided in an embodiment of the present invention; Figure 2 A data processing flowchart provided for embodiments of the present invention; Figure 3 The convergence curves of each loss function of the model provided in the embodiments of the present invention during training are shown, wherein (a) is the total loss convergence curve, (b) is the forward physical simulation loss convergence curve, (c) is the inverse spectral reconstruction loss convergence curve, and (d) is the forward process latent space noise constraint convergence curve. Figure 4 The scatter plots of the positive predicted atmospheric top radiance of the model provided in the embodiments of the present invention on the test set are as follows: (a) is the scatter plot corresponding to the 400nm band; (b) is the scatter plot corresponding to the 412nm band; (c) is the scatter plot corresponding to the 443nm band; (d) is the scatter plot corresponding to the 490nm band; (e) is the scatter plot corresponding to the 510nm band; and (f) is the scatter plot corresponding to the 560nm band. Figure 5 The following are comparison diagrams of the distributions of four randomly selected dimensions of the 8-dimensional latent variable z obtained by the forward model in the embodiments of the present invention with the standard normal distribution. Among them, (a) is a comparison diagram of the distribution characteristics of the first latent variable z1 with the standard normal distribution; (b) is a comparison diagram of the distribution characteristics of the second latent variable z2 with the standard normal distribution; (c) is a comparison diagram of the distribution characteristics of the third latent variable z3 with the standard normal distribution; and (d) is a comparison diagram of the distribution characteristics of the fourth latent variable z4 with the standard normal distribution. Figure 6 The following are scatter plots comparing the actual and predicted remote sensing reflectance of the model provided in this embodiment of the invention across eight representative bands on the test set: (a) is the scatter plot for the 400nm band; (b) is the scatter plot for the 412nm band; (c) is the scatter plot for the 443nm band; (d) is the scatter plot for the 490nm band; (e) is the scatter plot for the 510nm band; (f) is the scatter plot for the 560nm band; (g) is the scatter plot for the 620nm band; and (h) is the scatter plot for the 665nm band. Figure 7The following are the spectral correction uncertainty quantification and reconstruction results of the model provided in this embodiment of the invention based on Monte Carlo sampling and mean shift strategy. (a) is a comparison of the correction results for turbid Class II water sample 13427; (b) is a comparison of the correction results for turbid Class II water sample 449112; (c) is a comparison of the correction results for turbid Class II water sample 276126; (d) is a comparison of the correction results for turbid Class II water sample 720289; (e) is a comparison of the correction results for turbid Class II water sample 776151; (f) is a comparison of the correction results for turbid Class II water sample 503606; (g) is a comparison of the correction results for turbid Class II water sample 670695; and (h) is a comparison of the correction results for turbid Class II water sample 596759. Figure 8 The images show a comparison between the predicted remote sensing reflectance and the in-situ measured values ​​of the model provided in this embodiment of the invention on representative bands of the AERONET-OC dataset; wherein, (a) is a comparison between the predicted and measured values ​​for the 490nm band; (b) is a comparison between the predicted and measured values ​​for the 510nm band; (c) is a comparison between the predicted and measured values ​​for the 560nm band; and (d) is a comparison between the predicted and measured values ​​for the 620nm band. Figure 9 The cumulative distribution function of the absolute prediction error of the inverse atmospheric correction model provided in this embodiment of the invention; Figure 10 A statistical histogram showing the median absolute percentage difference of the model provided in this embodiment of the invention across different bands on an independent test set; Figure 11 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 1 This is a schematic diagram of a bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Step 1: Acquire satellite multispectral observation image data, perform subpixel-level spatiotemporal matching and bitmask quality control, remove cloud and abnormal interference, and extract the apparent radiance pixel sequence of the target water area.

[0025] In this embodiment of the invention, Step 2: Acquire multi-source auxiliary physical data that is spatiotemporally synchronized with satellite observations, perform spatial interpolation and feature alignment, and construct an environmental prior dataset.

[0026] In this embodiment of the invention, Step 3: Use the feature-based metamorphic Transformer network FT-Transformer to perform deep heterogeneous feature encoding on multi-source auxiliary physical data to generate physical condition embedding vectors containing global environmental semantics.

[0027] In this embodiment of the invention, step 3 includes: Given the complex nonlinear modulation effect of meteorological and geometrically-aided data on radiative transfer, and the significant differences in their distributions compared to spectral data, this invention designs a heterogeneous feature extraction mechanism based on the FT-Transformer. Unlike traditional simple dimensional concatenation, this mechanism first extracts each scalar environmental parameter... (e.g., the j-th meteorological variable) is mapped to a high-dimensional semantic vector through the Feature Tokenizer layer. : ; in, and For learnable weights and biases; The high-dimensional token is then processed by the self-attention interaction of the Transformer encoder to generate a conditional embedding that includes the global context. This provides a physical background constraint for subsequent spectral decoupling.

[0028] Step 4: Construct a reversible neural network architecture. By alternating affine coupling blocks and orthogonal Householder channel permutation layers, establish an equal-dimensional mapping channel between the water remote sensing reflectance state space and the satellite observation space.

[0029] In this embodiment of the invention, step 4 includes: This invention constructs a reversible neural network architecture between the physical parameter space and the observation data space. For the typical ill-conditioned inversion problem of atmospheric correction, this invention completely abandons the black-box strategy of unidirectional approximation of traditional deep learning models. Instead, it constructs a bidirectional closed-loop system based on the theory of reversible neural networks (INN). The system utilizes the mathematical properties of reversible transformation to define the atmospheric correction process as the inverse mapping of the radiative transfer process. Define physical state vector That is, remote sensing reflectance and observation state vector That is, the radiance of satellite L1B, introducing potential variables. Explicit modeling of atmospheric path radiation and sensor noise; under environmental condition vectors Under the control of factors such as solar zenith angle and air pressure, the inverse inference process of the model, namely atmospheric correction, is described by the following mathematical bijective function: ; in, For parameterized invertible inverse transformation, For the sampled random noise, These represent the embedded environmental features. Through the above formula, the model mathematically guarantees that the mapping from the observation space to the physical space is unique and stable, thus fundamentally avoiding the multi-solution problem of traditional algorithms. For example... Figure 1 The diagram illustrates the overall architecture of the model and introduces representative model modules.

[0030] In this embodiment of the invention, step 4 further includes: To address the numerical stability issue of deep networks when feature channels are mixed and to enhance the model's ability to capture correlations between spectral bands, a Householder reflection transform is introduced to replace random matrix permutations or 1×1 convolutions. The Householder reflection transform uses the product of orthogonal reflection matrices to generate feature vectors. The global rotation is expressed as: ; in, To output the feature vector, For the input feature vector, It is the identity matrix. For continuous multiplication from k=1 to K, Let K be the learnable reflection plane normal vector, and K be the number of reflections. for transpose, for The square of the second norm.

[0031] Step 5: Within the reversible network, the spectral cross-attention mechanism is used to embed the physical conditions into vectors as keys and values, and the spectral features of the main data stream are dynamically queried and fused to adaptively modulate the radiative transfer process in the real physical environment.

[0032] In this embodiment of the invention, step 5 includes: To ensure reversibility, spectral features and environmental information are deeply fused, and the backbone network employs stacked affine coupling blocks. Within each affine coupling block, the input features are split into two parts. and The model utilizes conditional embedding. The forward propagation of spectral features, which involves scaling and translation transformations, is known as radiative transfer simulation and is defined as follows: ; in, and These represent the two output features obtained after affine coupled block transformation. It is an exponential scaling function. and This is the nonlinear transformation function fitted by the internal spectral Transformer subnet. This represents the Hadamard product, which is element-wise multiplication. The above design cleverly embeds complex nonlinear feature extraction (done by Transformer) into a strictly reversible architectural framework, ensuring both mathematical invertibility and endowing the model with powerful nonlinear expressive capabilities.

[0033] Step 6: Perform end-to-end bidirectional physical constraint training based on the joint loss function. In the positive direction, force the network to learn the physical laws of radiative transfer and compress atmospheric disturbances into the latent variable space. In the actual inference stage, perform posterior mode optimization based on the observed signal and random sampling of the latent space to output high-precision remote sensing reflectance with atmospheric interference removed.

[0034] In this embodiment of the invention, step 6 includes: To overcome the ill-conditioned ambiguity in atmospheric correction for Class II water bodies, a hybrid loss function system corresponding to the physical process is constructed. The hybrid loss function system utilizes the bidirectional characteristics of the reversible neural network to apply physical constraints to the forward radiative transfer process and the reverse atmospheric correction process, thereby ensuring that the model conforms to physical laws in both directions. Total loss function Defined as: ; in, To simulate the constraint loss for forward radiative transfer, for The corresponding weighting coefficients, For asymmetric reconstruction loss, for The corresponding weighting coefficients, For spectral angle mapping loss, for The corresponding weighting coefficients, For the maximum mean difference loss in the potential space, for The corresponding weighting coefficients.

[0035] In this embodiment of the invention, it includes: I. In the forward path, the network will display the actual remote sensing reflectance. With zero-filled vector As input, the output is predicted satellite observations. and latent variables The following two losses are defined to ensure the accuracy of the physical process and the statistical normality of the latent space: First, to address the issue that traditional deep learning models are often treated as black boxes and lack physical interpretability, a forward radiative transfer constraint is introduced. Simulated observation data generated by the forward radiative transfer simulation loss forced model. Approximates real satellite observations This allows the neural network to learn the radiative transfer equation (RTE) physical mechanism of photons traveling from water to the sensor, and its expression is: ; Secondly, to address the problem of chaotic distribution and difficulty in sampling of latent variables due to information loss during atmospheric correction, a latent spatial maximum mean difference loss is introduced. Latent variables output by the latent space maximum mean difference loss constrained model Distribution and standard Gaussian distribution Minimizing the statistical distance between them ensures that the sampled random noise is statistically effective during reverse reasoning, thus guaranteeing the stability of the solution. Its expression is: ; in, This represents the summation of the MMD results from k kernel functions, where MMD... k This represents the maximum mean difference calculated under the k-th kernel function. Indicates the standard Gaussian distribution Prior latent variables obtained from sampling; II. In the reverse path, the network will transmit satellite observation data. With random sampling noise As input, the output is the predicted remote sensing reflectance. The following two losses are defined to ensure the physical nonnegativity of the correction result is consistent with the spectral consistency: First, addressing the core challenge of frequent negative remote sensing reflectance values ​​due to aerosol over-correction in Class II turbid water bodies, an inverse asymmetric correction constraint is introduced in logarithmic space. ), define prediction error When the predicted value (i.e., overcorrected) less than the true value At that time, apply a doubling penalty. (like =2.0), to make the model tend to output non-negative results during training, in order to suppress negative values, its expression is: ; in, This represents the number of samples involved in the loss calculation, where i represents the sample index. This indicates that the error terms are summed over all samples. This represents the prediction error on the i-th sample. This represents the asymmetric weight corresponding to the i-th error. This represents the multiplication penalty coefficient; Secondly, to address the problem that traditional mean square error (MSE) focuses only on numerical approximation while ignoring spectral waveforms, leading to distortion of the relative proportions between bands (such as the blue-green band ratio), a spectral angle mapping loss is introduced. Spectral angle mapping loss constrained prediction of spectral vector With the true spectral vector Minimizing the included angle ensures the accuracy of the corrected spectral shape, thereby guaranteeing the accuracy of subsequent inversion of water color parameters such as chlorophyll and suspended matter. Its expression is: ; Among them, arccos Represents the inverse cosine function. Represents extremely small positive numbers. The L2 norm of the predicted spectral vector is represented. The L2 norm represents the true spectral vector.

[0036] In this embodiment of the invention, step 6 further includes: To address the inherent ill-posed ambiguity of atmospheric correction and the potential multimodal posterior distributions arising from inverse mapping in deep networks, this invention abandons traditional single-step deterministic inference or simple arithmetic averaging during the inference phase. Instead, it designs a robust inference paradigm based on Test-Time Augmentation (TTA) and Mean-Shift clustering. Since reversible neural networks map simple latent variable distributions to complex water body parameter spaces, their posterior distributions often exhibit high non-convexity. In this case, a simple arithmetic mean may fall into low-density regions between different probability peaks, leading to physically infeasible smooth solutions. To obtain the maximum posterior probability estimate (MAP), i.e., to find the mode (Mode) with the highest probability density in the solution space under the combined constraints of current observations and the environment, the steps are as follows: First, for each fixed observation input, the model performs multiple independent samplings from the standard normal distribution in the latent space (e.g., sampling N=50 times), and generates a set of candidate solutions through an inverse network. ; Subsequently, the mean-shift algorithm is used to perform a pattern search in the solution space. To improve robustness to outliers and accelerate convergence, the algorithm first calculates the element-wise median of all candidate solutions as an initial estimate. Next, the local density gradient is estimated using a Gaussian kernel function, and the estimated value is iteratively shifted towards the density maxima; The update in the next iteration follows the formula: ; in, The center position at the t-th iteration. Let h be a Gaussian kernel function with bandwidth h; The iteration stops when the center displacement is less than a preset threshold, and the iteration eventually converges. This refers to the high-precision remote sensing reflectance output.

[0037] In this embodiment of the invention, a multi-source remote sensing and meteorological fusion dataset is used for the construction and validation of a universal atmospheric correction model for Class II water bodies worldwide: To ensure that the model can learn comprehensive and representative physical mechanisms of radiative transfer in type II turbid water bodies, and to objectively and fairly evaluate its performance, the following two types of datasets are constructed: ①A large-scale training dataset of typical turbid water bodies worldwide; The core observational data for model training in this invention primarily comes from the OLCI sensor carried by the Sentinel-3B satellite. This sensor not only provides 21 high-quality spectral bands covering 400 to 1020 nanometers, but its 300-meter full-resolution capability is also crucial for achieving pixel-by-pixel precise correction. It can effectively resolve strong suspended matter gradients and microscale dynamic characteristics in turbid water bodies, significantly reducing the mixed-pixel effect physically. To constrain the solution space of radiative transfer, this invention simultaneously incorporates ERA 5-hour single-layer reanalysis data released by ECMWF as atmospheric prior information. Specifically, total column water vapor and ozone are extracted to correct gas absorption, surface air pressure is used to accurately calculate Rayleigh scattering optical thickness, and 10-meter wind field components are combined to characterize sea surface roughness and assist in solar flare estimation, thus providing complete physical boundary conditions for model training.

[0038] In embodiments of the present invention, such as Figure 2 As shown, in order to ensure the physical authenticity of the training target and the spatial consistency of multivariate data, this invention constructs a strict hierarchical bitmask screening system, the specific definitions of which are shown in Table 1. For L1B radiance data, saturated and nonlinear response pixels are removed by quality identification to ensure that the input radiance (LTOA) is within the linear range of the sensor. For L2 remote sensing reflectance data, in addition to the conventional removal of cloud, shadow and land interference, environmental noise such as stray light from cloud edges and proximity effects is further removed.

[0039] Table 1. Bitmasks used when processing Sentinel-3B L1B and L2 data. .

[0040] To address the significant scale differences between ERA5 reanalysis data (spatial resolution of approximately 30 km) and OLCI observation data (spatial resolution of 300 m), this invention designs a high-precision spatiotemporal alignment scheme. In the temporal dimension, since satellite overpasses acquire instantaneous observation signals, while ERA5 provides hourly meteorological fields, linear temporal interpolation is performed by retrieving the nearest adjacent ERA5 time intervals to the satellite imaging time to accurately reconstruct the instantaneous atmospheric state. In the spatial dimension, bilinear interpolation is employed to precisely map the coarse-resolution meteorological grid (including ozone and synthetic wind speed converted to Dobson units) to the center of each 300-meter OLCI pixel, thus fully preserving the environmental heterogeneity of turbid water bodies at the microscale while incorporating physical priors.

[0041] Subsequently, the k-D tree (KDTree) algorithm was used to perform sub-pixel-level spatial matching between L1B observations, L2 labels and interpolated ERA5 parameters (the search radius was set to be less than the pixel spacing), ensuring strict geospatial correspondence between multi-source input features and reference labels.

[0042] To further improve the spectral purity of the dataset, a second round of rigorous quality control (Strict QC) was implemented after matching. Specifically, this included: (1) calculating the solar flare angle (θg) based on geometric angles and removing strong flare regions where θg < 15° to eliminate specular reflection contamination; (2) forcibly removing any bands with negative values ​​or non-physically high values ​​(Rg). rs R rs (3) Examine the spectral shape of the near-infrared band to eliminate residual noise. The final generated sample library contains millions of high-quality pixel pairs covering a wide dynamic range, laying the physical foundation for robust training of deep learning models.

[0043] To ensure the samples have global physical generalization, this invention selected seven representative extremely turbid water bodies worldwide and systematically constructed a dataset by coupling their optical properties with the model's constraint target depth.

[0044] In strong backscattering environments, the high inorganic suspended matter content of the Bohai Sea effectively maintains the model's stability; while the cloud-like reflectance characteristics of the La Plata estuary are specifically used to enhance the model's robustness in water-cloud confusion scenarios. For complex optical gradients, the extensive sediment gradient of the East China Sea is used to improve the model's ability to capture smooth optical transitions; in the southern North China Sea, tidal resuspension and a high CDOM environment constrain the model to achieve accurate separation of scattering and absorption signals. Regarding the challenges of high absorption and low signal-to-noise ratio, the CDOM-rich blackwater of the northern Amazon estuary forcibly improves the model's reflectance estimation accuracy under low SNR conditions such as the blue light band; the organic-rich plume of the Mississippi River estuary strengthens the model's ability to distinguish between organic and inorganic components. Finally, considering the complexity of air-sea coupling, the significant influence of terrestrial aerosols on the Venice Lagoon provides stringent robustness constraints for aerosol-water coupling modeling in atmospheric correction. This sample distribution spanning complete optical gradients and typical atmospheric environments provides solid physical support for achieving pixel-level atmospheric correction.

[0045] In terms of handling the temporal and spatial dimensions, this invention adopts a discontinuous observation strategy that samples one scene every five days across a complete calendar year. This aims to overcome cloudy and rainy weather windows and disrupt the short-term autocorrelation between aerosol and water mass evolution, comprehensively capturing the seasonal fluctuations of solar observation geometry and meteorological fields. Furthermore, this invention introduces strict location-independent constraints into the input feature space, explicitly removing latitude, longitude, and absolute time information from the data. This forces the model to purely learn the physical nature of radiative transfer rather than memorizing geographical priors for specific regions, providing a solid data foundation for training a universal atmospheric correction model for global Class II water bodies.

[0046] ② Independent benchmark validation and generalization test dataset; To rigorously evaluate the generalization ability of the model in real and unknown complex optical environments, this invention uses data from the AERONET-OC (Aerosol Robotic Network - Ocean Color) global automatic ocean color observation network, completely independent of the model training set, as absolute ground truth for non-homogeneous verification. Addressing the core challenge of atmospheric correction for Class II water bodies, 24 high-turbidity or complex aerosol stations located in typical nearshore, estuarine, and inland lakes worldwide were selected. These stations not only cover the European coast, which is affected by both strong tides and industrial aerosols (such as AAOT in Italy, Thornton_C-power in Belgium, and Irbe_Lighthouse in the Baltic Sea), but also the nearshore areas of the Americas and Asia, dominated by alluvial deposits from major rivers (such as RdP-EsNM in Argentina, Sacramento_River in the United States, and Socheongcho in South Korea), ensuring the global representativeness and extreme condition coverage of the verification dataset across optical water body types.

[0047] To extract high-quality satellite-to-ground matching pairs, this invention strictly adheres to internationally recognized satellite verification protocols. In the time dimension, the time difference between satellite transit and ground-based measurements is limited to within 3 hours. In the spatial dimension, effective pixels within the macroscopic window at the station center are extracted and their mean values ​​are calculated to overcome point-to-area scale effects caused by sensor geolocation errors and spatial heterogeneity of water bodies. After rigorous spatiotemporal filtering and quality control, a three-way comparison dataset is constructed, comprising the absolute true values ​​from field measurements, the inferred values ​​from this invention's model, and Sentinel-3 official standard L2 products (such as those produced by the C2RCC algorithm). This dataset is independent of the training phase and is used solely for the final model performance evaluation. It aims to objectively quantify the technical advantages of this invention in suppressing negative reflectivity and improving correction accuracy, ensuring the fairness of the evaluation results.

[0048] In this embodiment of the invention, the experimental setup is as follows: The parameters that need to be set for this model include: number of reversible blocks, physical and observation dimensions, zero-padding and noise dimensions, number of feature tokenization layers, conditional embedding dimensions, number of spectral Transformer subnet layers and hidden layer dimensions, number of Householder reflections, weight coefficients of each directional loss, learning rate, learning rate scheduling parameters, number of warm-up steps, weight decay, batch size, total number of training steps, number of augmented samplings during testing, Mean-Shift bandwidth, and computational accuracy.

[0049] In this embodiment of the invention, model training, evaluation, and testing are as follows: To verify the effectiveness of the technical solution of this invention, a standard experimental procedure was designed, including bidirectional closed-loop training and global independent site validation. Model training was performed on a high-performance computing platform, using the AdamW optimizer in conjunction with linear preheating and cosine annealing strategies to ensure stable convergence of the model within the bidirectional mapping space.

[0050] Model training is driven by a multi-objective hybrid loss function. In the forward path, the network will input the true remote sensing reflectance. With zero-filled vector As input, the output is predicted satellite observations. and latent variables Forward simulation loss and latent space MMD loss are defined to ensure the accuracy of the physical processes and the statistical normality of the latent space. The forward simulation loss forces the model to grasp the physical mapping from water color parameters to satellite signals. The effects of forward radiative transfer and the distribution of dependent variables are as follows: Figure 4 (a) to (f) and Figure 5 As shown in (a) to (d) in the diagram. In the reverse path, the network will use satellite observation data. With random sampling noise As input, the output is the predicted remote sensing reflectance. Asymmetric reconstruction loss and spectral angle loss are defined to ensure the physical nonnegativity and spectral fidelity of the correction results. Addressing the challenge of negative values ​​in atmospheric correction for Class II water bodies, the asymmetric reconstruction loss suppresses atmospheric overcorrection through differentiated penalty weights; for example... Figure 3 (a), (b), (c), and (d) in the figure show the convergence curves of the various loss functions during model training.

[0051] In the inference phase, this invention employs posterior mode estimation based on Mean-Shift clustering. For the input observations, candidate solution sets are generated by sampling from the latent space through time-of-test (TTA) enhancement, and the Mean-Shift algorithm is used to search for the mode center with the highest probability density in the solution space as the final remote sensing reflectance correction value, thereby effectively solving the problem of multiple solutions in the atmospheric correction process. The R obtained from the posterior mode estimation based on Mean-Shift clustering... rs Spectral curves as follows Figure 7 As shown in (a) to (h) in the figure, the horizontal axis of each sub-figure is wavelength, the vertical axis is remote sensing reflectance, the red semi-transparent scatter or line clusters represent the distribution of candidate solutions generated by Monte Carlo sampling, which is used to characterize the uncertainty of the correction, the red dashed line represents the final predicted spectrum after applying the mean shift strategy, and the black solid line represents the true spectrum.

[0052] The following four indicators are used to quantify remote sensing reflectance. Prediction accuracy: Mean Absolute Error (MAE): ; Root Mean Square Error (RMSE): ; Median absolute percentage difference (MdAPD): ; Systematic bias (Bias): ; A multi-source, multi-scale validation strategy is employed to comprehensively evaluate the technical advantages of the model in this invention. First, in-situ measured data from the AERONET-OC global automatic ocean color observation network are used as absolute ground truth for non-homogeneous source validation to assess the model's cross-scene generalization ability under different geographical regions and extreme optical conditions. Second, a large-scale pixel-level correction accuracy test is conducted using an independent test set constructed from Sentinel-3 imagery. This test set is strictly isolated from the training data in terms of spatiotemporal distribution to verify the model's ability to capture satellite sensor features and its reliability in operational production.

[0053] In addition to the four numerical indicators mentioned above, this invention visually demonstrates the atmospheric correction effect by plotting 1:1 density scatter plots. For example... Figure 6 As shown in (a) to (h), the x-axis of each subplot represents the true remote sensing reflectance, and the y-axis represents the predicted remote sensing reflectance. The black dashed line in each subplot is the 1:1 reference line, and the red solid line is the linear fitting line. The text box contains R... 2 The statistical indicators are RMSE and MdAPD, with the bottom color bar representing the sample density distribution. In the scatter plot, the horizontal axis represents the field measurement or the true benchmark value. The vertical axis represents the model correction value. In addition, such as Figure 8 As shown in (a) to (d), the horizontal axis of each subplot represents in-situ remote sensing reflectance, and the vertical axis represents predicted remote sensing reflectance. The green scatter plots and corresponding statistical indicators correspond to the PCSA-INN model proposed in this invention, while the orange scatter plots and corresponding statistical indicators correspond to the existing C2RCC algorithm. On the AERONET-OC independent validation set and the Sentinel-3 internal test set, the calibration data points for each spectral band of this invention are highly clustered near the 1:1 standard line and exhibit extremely high correlation coefficients. This demonstrates the excellent consistency between the correction results and the true values.

[0054] Most importantly, the distribution characteristics of the scatter plot in the low-value region (approaching the origin of the coordinate system) demonstrate the significant effectiveness of this invention in suppressing negative reflectance values: compared with the phenomenon that traditional atmospheric correction algorithms often fall into the negative value region (below the vertical axis) in the blue light band, this invention, through the constraint of asymmetric reconstruction loss, ensures that the corrected scatter points are strictly distributed in the first quadrant, thus completely avoiding the generation of invalid solutions from both physical and statistical dimensions.

[0055] Beyond pixel-level accuracy verification, this invention further evaluates the model's predictive robustness in complex hydrological environments through statistical distribution characteristics and large-scale inversion tests. First, quantitative evaluation on independent test sets shows that the model of this invention possesses extremely high inversion accuracy across the entire spectrum. For example... Figure 10 As shown, the horizontal axis represents spectral wavelength (in nm), and the vertical axis represents error level (in %). The bar chart displays the MdAPD distribution of the model across eight spectral bands: 400 nm, 412 nm, 443 nm, 490 nm, 510 nm, 560 nm, 620 nm, and 665 nm. The values ​​above the bars correspond to the specific error values ​​for each spectral band. The model's median absolute percentage difference (MdAPD) is extremely low across all observed spectral bands, demonstrating its high-precision inversion capability across the entire spectrum. Figure 9As shown, the cumulative distribution function (CDF) curve of the absolute prediction error of remote sensing reflectance with inverse atmospheric correction of the proposed model shows an extremely steep upward trend. The prediction error of most samples is highly concentrated in a very small numerical range, reflecting that the model prediction results have extremely strong stability and reliability.

[0056] In robustness tests of large-scale spatial inversion, this invention made extensive predictions for typical Class II turbid water bodies worldwide (such as the Yangtze River Estuary and the La Plata Estuary). Experimental results show that, when dealing with challenging areas with extremely high suspended particulate matter concentrations, the remote sensing reflectance products generated by this invention significantly outperform the European Space Agency's (ESA) official C2RCC (Case 2 Regional Coast Colour) model in terms of logical consistency of numerical distribution.

[0057] Crucially, traditional C2RCC models often suffer from large-area negative values ​​in inversion results due to spectral saturation in short-wavelength bands such as blue light, leading to severe physical failures and loss of observational data. In contrast, the model of this invention significantly improves the valid pixel rate under the same extreme conditions, effectively solving the problem of being unable to obtain data under complex optical conditions. This superior performance is not only reflected in the lower average error index, but also in its ability to ensure the output of physically meaningful valid inversion solutions even in extremely turbid environments, thus supporting highly reliable large-scale operational applications.

[0058] By combining numerical index evaluation, multi-source dataset cross-validation, and 1:1 scatter plot analysis, this invention objectively and comprehensively demonstrates the outstanding advantages of the model in terms of atmospheric correction accuracy, spectral shape fidelity, and physical consistency when dealing with Class II turbid water bodies.

[0059] The present invention aims to overcome the shortcomings of existing atmospheric correction techniques for Class II turbid water bodies, such as lack of physical mechanisms, weak generalization ability, and low computational efficiency. The present invention abandons the traditional deep learning model's one-way fitting approach that can only establish end-to-end black box mapping, and instead adopts a two-way coupling architecture that strictly follows the physical laws of radiative transfer. Furthermore, the decoupling task of complex heterogeneous environmental light fields is entrusted to the heterogeneous feature interaction and reversible inference module within the model.

[0060] To achieve the above objectives, the present invention proposes the following technical solution: (1) To construct a bidirectional reversible mapping architecture constrained by the radiative transfer mechanism to replace the unidirectional black-box fitting framework that lacks physical interpretability. The primary problem this invention aims to solve is the fundamental limitation of current deep learning models in handling ill-conditioned inversion problems—that is, the model only learns the statistical correlation between data and ignores the physical process of photon transfer, resulting in frequent negative values ​​in the output and a lack of interpretability. This invention constructs a mathematically rigorous bidirectional closed-loop system by introducing a reversible neural network (INN). Its core contribution lies in transforming the inverse solution task of atmospheric correction into an accurate simulation and reversible inference of the forward radiative transfer process (from water bodies to satellite signals), thereby fundamentally eliminating the physical inconsistencies caused by purely data-driven approaches within a unified physical constraint framework, ensuring the non-negativity and physical rationality of the correction results.

[0061] (2) A heterogeneous meteorological data feature tokenizer mechanism based on FT-Transformer is established to address the technical bottleneck of efficient deep fusion between multi-source auxiliary data and spectral data. Traditional algorithms typically use simple multilayer perceptrons (MLPs) for dimensional concatenation when processing meteorological and geometric auxiliary data such as solar zenith angle, relative humidity, and air pressure. This approach struggles to effectively capture the nonlinear modulation relationship between continuous numerical environmental data and other high-dimensional spectral features. This invention innovatively introduces the FT-Transformer (Feature Tokenizer Transformer) module to transform heterogeneous scalar meteorological data into high-dimensional token vectors with rich semantic information. This mechanism enables the model to elevate physical environmental constraints from simple numerical inputs to semantic-level contextual information, providing accurate environmental background representation for subsequent spectral decoupling.

[0062] (3) A spectral-environment deep interaction paradigm based on cross-attention and Householder reflection is proposed to achieve adaptive removal of complex atmospheric path radiation interference. Addressing the challenge of high and varied aerosol optical thickness in Class II water bodies, this invention abandons traditional convolution operations and constructs an interaction layer based on cross-attention, using meteorological tokens as query keys to dynamically adjust the feature weights of each spectral band. Simultaneously, Householder reflection transformation is introduced for global channel mixing, replacing traditional random matrices or 1×1 convolutions. This orthogonal transformation not only ensures the numerical stability of features during deep network transmission but also achieves lossless coupling between spectral features and environmental constraints, thereby fundamentally improving the model's robustness in strong interference environments.

[0063] (4) This paper proposes an efficient global computation paradigm based on asymmetric noise injection and a single forward inference, aiming to decouple the high-precision requirements of physical models from the huge computational overhead of traditional iterative algorithms. Traditional physical methods usually rely on massive lookup table (LUT) retrieval or time-consuming pixel-by-pixel iterative optimization, and are difficult to handle the multiple solutions problem in ill-conditioned inversion. This invention designs an asymmetric dimension alignment strategy in the INN architecture and explicitly introduces Gaussian noise channels to absorb information loss, enabling the model to handle one-to-many mappings. At the same time, the model is trained using a dataset covering seven typical turbid water areas around the world, and only one millisecond-level backward inference is required to complete the correction in the application stage. This paradigm successfully decouples the high-precision standards of physical models from the efficient inference capabilities of deep learning, and ultimately contributes a practical water color remote sensing processing model that combines physical high fidelity, global generalization, and computational efficiency.

[0064] This invention is a deep learning method specifically designed to address the extremely complex optical properties of Class II turbid water bodies globally, the susceptibility of traditional algorithms to failure in strong aerosol environments, and the frequent occurrence of negative values. Its core technological breakthrough lies in completely abandoning the limitation of traditional deep learning models that can only establish a one-way black-box mapping from satellite-observed radiance to remotely sensed reflectance. It innovatively utilizes Invertible Neural Networks (INNs) to construct a bidirectional mapping paradigm that strictly adheres to the physical laws of radiative transfer. In this system, the forward process is defined as radiative transfer simulation, which is no longer a simple feature extraction but a precise simulation of the physical process of photon radiative transfer in the atmosphere and water bodies, i.e., generating satellite-observed radiance from water remotely sensed reflectance. Based on the mathematically rigorous reversibility of INNs, once the model grasps the forward radiative transfer law, the inverse process is defined as atmospheric correction inference. This transforms atmospheric correction, a typical ill-conditioned, multi-solution inverse problem, into a benign solution process with clear physical constraints, fundamentally eliminating the physical inconsistencies caused by purely data-driven approaches. To address the challenges of diverse aerosol types and severe atmospheric path radiation interference in Class II turbid water bodies, this invention achieves architectural innovation in spectral feature processing and multi-source information fusion. The model completely abandons the inefficient approach of traditional algorithms that simply concatenate spectral features with auxiliary data or perform only feature-wise linear modulation. Instead, it innovatively constructs a deep interactive architecture based on cross-attention. In this architecture, each spectral band is treated as an independent sequence without positional encoding. Heterogeneous environmental features extracted by the FT-Transformer (such as solar zenith angle and wind speed at 10 meters) are used as keys and values, dynamically interacting with the spectral feature sequence as a query. This mechanism can capture deep nonlinear spectral dependencies between bands and adaptively adjust the feature weights of each band according to environmental conditions. Simultaneously, a Householder reflection transform is introduced for global channel mixing, replacing the traditional random matrix or 1×1 convolution, significantly improving the numerical stability and orthogonality of feature mixing. Furthermore, to explicitly address the information loss problem in atmospheric correction (i.e., the extremely low proportion of water signals in the sensor-received signal), the model designs a unique asymmetric dimension alignment strategy: employing 8-dimensional remote sensing reflectance (R0) in the latent space. rsThe method employs a 21-dimensional zero-filling vector, corresponding to a 21-dimensional L1B radiance in the observation space, coupled with 8-dimensional Gaussian noise. This design provides the model with a dedicated noise channel to absorb and separate atmospheric path radiation noise, enabling the model to accurately reconstruct weak water body signals from complex total radiation signals during inverse inference. In terms of data-driven strategy, the method constructs a complete dataset covering seven typical turbid water bodies globally for training, giving the model strong cross-regional generalization capabilities. Ultimately, while maintaining accuracy comparable to or even better than traditional physical algorithms, this invention completes correction with a single inverse inference, significantly improving computational efficiency. It also mathematically guarantees the physical rationality of the output results, substantially reducing the frequency of overcorrection negative values ​​common in Class II turbid water bodies, and significantly enhancing the model's generalization stability in unobserved sea areas. This achieves highly robust and physically consistent accurate atmospheric correction for complex Class II water bodies globally.

[0065] Compared with the prior art, the present invention has the following advantages: (1) Successfully achieved a breakthrough in atmospheric correction technology for complex turbid water bodies based on physical mechanisms.

[0066] Addressing the challenge of highly coupled water-air signals in Class II turbid water bodies, where traditional algorithms struggle to accurately decouple them, this invention, for the first time, explicitly utilizes the physical symmetry in atmospheric correction tasks by constructing a reversible neural network (INN) architecture. This invention transforms the complex radiative transfer process into a bidirectional reversible mapping: the forward path uses deep learning to simulate the evolution of the radiative transfer equation (RTE) from remotely sensed reflectance x to satellite-observed radiance y, allowing the model to fully internalize the physical mechanisms; the reverse path directly utilizes the learned physical mechanisms to perform high-precision real-time atmospheric correction tasks. This unique feature of forward simulating physics and reverse execution of correction fundamentally guarantees the physical success and numerical accuracy of this invention in atmospheric correction tasks.

[0067] (2) The accuracy of signal decoupling is significantly improved by using a two-way physical closed loop.

[0068] Existing models often employ unidirectional black-box mapping, making it difficult to guarantee that the output results conform to physical realities. This invention constructs a rigorous physical closed loop by jointly constraining the simulation loss of the forward path and the reconstruction loss of the reverse path. This mechanism forces the model to simultaneously satisfy the energy conservation of radiative transfer and the spectral consistency law when performing atmospheric correction, thereby enabling precise decoupling of atmospheric molecular scattering, aerosol contributions, and water component signals superimposed on the same signal. This bidirectional closed-loop design significantly improves the model's correction accuracy under extreme turbidity and multi-source noise interference, achieving high-fidelity restoration of the physical truth.

[0069] (3) It fundamentally solves the problem of negative reflectance values ​​that frequently occur in atmospheric correction of highly turbid water bodies.

[0070] To address the core drawback of traditional algorithms, which are prone to generating invalid negative values ​​due to aerosol over-correction in complex atmospheric environments, this invention introduces an asymmetric reconstruction loss mechanism in the inverse correction path. By imposing a doubling penalty on the error region that generates negative values, a natural physical nonnegativity barrier is constructed at the algorithm architecture level. Combined with a Mean-Shift-based probabilistic inference paradigm, this invention can effectively identify and filter physically infeasible outliers, ensuring that the corrected scatter points are strictly distributed in the first quadrant, thus completely avoiding the generation of invalid solutions from both statistical and physical dimensions.

[0071] (4) It possesses excellent global generalization capabilities and large-scale business processing efficiency.

[0072] This invention, by fusing ERA5 global meteorological data and implementing strict location-agnostic feature space constraints, forces the model to learn universal physical laws independent of geographical priors. Trained on a large-scale dataset covering seven typical turbid water bodies globally, this invention demonstrates exceptional zero-sample transfer capability and can be directly applied to any unknown turbid water body scenario worldwide. Furthermore, based on a pixel-by-pixel parallel processing architecture, this invention significantly reduces computational overhead while maintaining 300-meter spatial resolution detail, providing an efficient technical means for the operational processing of real-time atmospheric correction of large-scale satellite data.

[0073] (5) It provides a high-fidelity spectral reference, which significantly enhances the application value of downstream water color business.

[0074] This invention introduces spectral angle mapping (SAM) and band scaling constraints during the training process to ensure that the atmospherically corrected remote sensing reflectance x maintains extremely high spectral shape fidelity across the entire band. 1:1 density scatter plot verification shows that the spectrum corrected by this invention maintains excellent consistency with the ground-measured true value in both numerical value and waveform, solving the historical problem of signal distortion in the blue light band encountered by traditional algorithms. This high-quality correction output provides high-performance, plug-and-play data input for subsequent accurate monitoring of water color elements such as chlorophyll and suspended matter, significantly enhancing the practical value of atmospheric correction products in environmental early warning and resource management.

[0075] The technical solution provided by this invention includes the following steps: acquiring satellite multispectral observation image data; performing sub-pixel-level spatiotemporal matching and bitmask quality control to remove cloud and anomalous interference; extracting the apparent radiance pixel sequence of the target water area; acquiring multi-source auxiliary physical data synchronized with satellite observations; performing spatial interpolation and feature alignment to construct an environmental prior dataset; using the feature lexical transformation Transformer network FT-Transformer to perform deep heterogeneous feature encoding on the multi-source auxiliary physical data to generate physical condition embedding vectors containing global environmental semantics; and constructing a reversible neural network architecture to establish a water remote sensing reflectance state space and a dynamical property space through alternating affine coupling blocks and orthogonal Householder channel permutation layers. This method employs an equal-dimensional mapping channel between satellite observation spaces. Within the reversible network, a spectral cross-attention mechanism is used to embed physical conditions into vectors as keys and values, dynamically querying and fusing the spectral features of the main data stream to adaptively modulate the radiative transfer process in the real physical environment. End-to-end bidirectional physical constraint training is performed based on a joint loss function, forcing the network to learn the physical laws of radiative transfer in the positive direction and compressing atmospheric disturbances into the latent variable space. In the actual inference stage, posterior mode optimization is performed based on random sampling of the observed signal and the latent space, outputting a high-precision remote sensing reflectance with atmospheric interference removed. This method overcomes the problems of missing physical mechanisms, weak generalization ability, and low computational efficiency in existing atmospheric correction techniques for Class II turbid water bodies.

[0076] 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.

[0077] 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 bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network.

[0078] Figure 11 A schematic diagram of an electronic device provided in an embodiment of the present invention, such as... Figure 11 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 bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network in the embodiment. To avoid repetition, it will not be described in detail here.

[0079] Electronic device 21 includes, but is not limited to, processor 211 and memory 212. Those skilled in the art will understand that... Figure 11This 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.

[0080] 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.

[0081] 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.

[0082] 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.

[0083] 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 bidirectional coupled deep learning method for remote sensing calibration and correction based on a reversible neural network, characterized in that, The method includes: Step 1: Acquire satellite multispectral image data, perform sub-pixel-level spatiotemporal matching and bitmask quality control, and remove cloud and abnormal interference; Step 2: Acquire multi-source auxiliary physical data that is spatiotemporally synchronized with satellite observations, perform spatial interpolation and feature alignment, and construct an environmental prior dataset; Step 3: Use the feature-based meta-transformer network FT-Transformer to perform deep heterogeneous feature encoding on multi-source auxiliary physical data to generate physical condition embedding vectors containing global environmental semantics. Step 4: Construct a reversible neural network architecture and establish an equal-dimensional mapping channel between the water remote sensing reflectance state space and the satellite observation space by alternating affine coupling blocks and orthogonal Householder channel permutation layers. Step 5: Inside the reversible neural network, the spectral cross-attention mechanism is used to embed the physical condition vector as the key and value to dynamically query and fuse the spectral features of the multispectral observation image data, so as to adaptively modulate the radiation transfer process of the real physical environment. Step 6: Perform end-to-end bidirectional physical constraint training based on the joint loss function. In the positive direction, force the network to learn the physical laws of radiative transfer and compress atmospheric disturbances into the latent variable space. In the actual inference stage, perform posterior mode optimization based on the observed signal and random sampling of the latent space to output high-precision remote sensing reflectance with atmospheric interference removed. Step 3 includes: The design incorporates a heterogeneous feature extraction mechanism based on the FT-Transformer. First, each scalar environmental parameter... The feature tokenizer layer maps the data to high-dimensional semantic vectors. : ； in, and For learnable weights and biases; The high-dimensional token is then processed by the self-attention interaction of the Transformer encoder to generate a conditional embedding that includes the global context. This provides a physical background constraint for subsequent spectral decoupling; Step 4 includes: A reversible neural network architecture between the physical parameter space and the observation data space is constructed. Based on the theory of reversible neural networks (INN), a bidirectional closed-loop system is constructed. The system utilizes the mathematical properties of reversible transformation to define the atmospheric correction process as the inverse mapping of the radiative transfer process. Define physical state vector That is, remote sensing reflectance, and define the observation state vector. That is, the radiance of satellite L1B, introducing potential variables. Explicit modeling of atmospheric path radiation and sensor noise; under environmental condition vectors Under the control of [the model], the inverse inference process of the model, namely atmospheric correction, is described by the following mathematical bijective function: ； in, For parameterized inverse transform, For the sampled random noise, These are the embedded environmental features.

2. The method according to claim 1, characterized in that, Step 4 also includes: The Householder reflection transform is introduced to replace random matrix permutation or 1×1 convolution; the Householder reflection transform uses the product of orthogonal reflection matrices to obtain eigenvectors. The global rotation of is expressed as: ； in, To output the feature vector, For the input feature vector, It is the identity matrix. For continuous multiplication from k=1 to K, Let K be the learnable reflection plane normal vector, and K be the number of reflections. for transpose, for The square of the second norm.

3. The method according to claim 1, characterized in that, Step 5 includes: To ensure reversibility, spectral features and environmental information are deeply fused, and the backbone network employs stacked affine coupling blocks. Within each affine coupling block, the input features are split into two parts. and The model utilizes conditional embedding. The scaling and translation transformations of spectral features are performed; the forward propagation of this process, i.e., radiative transfer simulation, is defined as follows: ； in, and These represent the two output features obtained after affine coupled block transformation. It is an exponential scaling function. and This is the nonlinear transformation function fitted by the internal spectral Transformer subnet. This represents the Hadamard product, which is the product of elements.

4. The method according to claim 1, characterized in that, Step 6 includes: Constructing a hybrid loss function system corresponding to the physical process The hybrid loss function system utilizes the bidirectional characteristics of the invertible neural network to apply physical constraints to both the forward radiative transfer process and the reverse atmospheric correction process. Total loss function Defined as: ； in, To simulate the constraint loss for forward radiative transfer, for The corresponding weighting coefficients, For asymmetric reconstruction loss, for The corresponding weighting coefficients, For spectral angle mapping loss, for The corresponding weighting coefficients, For the maximum mean difference loss in the potential space, for The corresponding weighting coefficients.

5. The method according to claim 4, characterized in that, include: I. In the forward path, the network will display the actual remote sensing reflectance. With zero-filled vector As input, the output is predicted satellite observations. and latent variables Define the following two losses: First, we introduce forward radiative transfer to simulate constraint loss. Simulated observation data generated by the forward radiative transfer simulation constraint loss forced model. Approximates real satellite observations This allows the neural network to learn the radiative transfer equation (RTE) physical mechanism of photons traveling from water to the sensor, and its expression is: ； Secondly, the maximum mean difference loss of the potential space is introduced. Latent variables output by the latent space maximum mean difference loss constrained model Distribution and standard Gaussian distribution The expression for minimizing the statistical distance between them is: ； in, This represents the summation of the MMD results from k kernel functions, where MMD... k This represents the maximum mean difference calculated under the k-th kernel function. Indicates the standard Gaussian distribution Prior latent variables obtained from sampling; II. In the reverse path, the network will transmit satellite observation data. With random sampling noise As input, the output is the predicted remote sensing reflectance. Define the following two losses: First, an asymmetric reconstruction loss is introduced into the logarithmic space. Define prediction error When the predicted value Less than the true value At that time, apply a doubling penalty. This is to encourage the model to output non-negative results during training, thus suppressing negative values. Its expression is: ； in, This represents the number of samples involved in the loss calculation, where i represents the sample index. This indicates that the error terms are summed over all samples. This represents the prediction error on the i-th sample. This represents the asymmetric weight corresponding to the i-th error. This represents the multiplication penalty coefficient; Secondly, spectral angle mapping loss is introduced. Spectral angle mapping loss constrained prediction of spectral vector With the true spectral vector The angle between the two points is minimized, and its expression is: ； Among them, arccos Represents the inverse cosine function. Represents extremely small positive numbers. The L2 norm of the predicted spectral vector is represented. The L2 norm represents the true spectral vector.

6. The method according to claim 1, characterized in that, Step 6 also includes: To design a robust inference paradigm based on test-time enhanced TTA and mean-shift clustering, and to obtain the maximum a posteriori probability estimate (MAP), i.e., to find the mode with the highest probability density in the solution space under the joint constraints of current observations and environment, the steps are as follows: First, for each fixed observation input, the model independently samples multiple times from the standard normal distribution in the latent space, and generates a set of candidate solutions through an inverse network. ; Subsequently, the mean-shift algorithm is used to perform a pattern search in the solution space. To improve robustness to outliers and accelerate convergence, the algorithm first calculates the median of the dimension of all candidate solutions as an initial estimate. Next, the local density gradient is estimated using a Gaussian kernel function, and the estimated value is iteratively shifted towards the density maxima; The update in the next iteration follows the formula: ； in, The center position at the t-th iteration. Let h be a Gaussian kernel function with bandwidth h; The iteration stops when the center displacement is less than a preset threshold, and the iteration eventually converges. This refers to the high-precision remote sensing reflectance 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 containing the computer-readable storage medium to perform the bidirectional coupled deep learning method for remote sensing calibration correction based on a reversible neural network 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 remote sensing calibration correction bidirectional coupled deep learning method based on a reversible neural network as described in any one of claims 1 to 6.

Images