Deep learning-based thin cloud atmosphere top reflectivity estimation method and system

Abstract

The invention discloses a thin cloud atmosphere top reflectivity estimation method and system based on deep learning. The method comprises the following steps: firstly, performing radiometric calibration on obtained remote sensing data to obtain a reflectivity image; secondly, 1 / 2 of a wave band is input, end-to-end feature extraction and cloud TOA reflectivity inversion are achieved through a convolutional neural network, and a preliminary inversion result is output; and finally, sampling and generating a denoised cloud TOA reflectivity image by using the trained denoising diffusion model according to the preliminary inversion result. According to the method, thin cloud reflectivity estimation is modeled as an image recovery problem, and preliminary inversion of thin cloud radiation characteristics is realized through an empirical relationship between wavebands. The advantages of the diffusion model in the aspects of image feature expression and reconstruction are combined, the radiation contribution of the cloud is accurately separated from complex earth surface and atmosphere mixed signals, effective estimation of the TOA reflectivity of thin cloud areas of different types and concentrations is achieved, and therefore a more robust physical basis is provided for downstream cloud detection and cloudless image recovery.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image analysis and atmospheric physics inversion technology, and in particular to a method and system for estimating the reflectance of the top of thin clouds based on deep learning. Background Technology

[0002] Clouds are diverse and complex, with significant differences in spectral, spatial, and temporal characteristics among different types. Cirrus clouds, in particular, are a challenging type of thin, high-altitude cloud, typically composed of ice crystals and exhibiting a translucent, filamentous structure. Research in cloud detection can be categorized into four types: those based on physical rules, those based on temporal variations, those based on machine learning, and those based on deep learning.

[0003] These models, ranging from traditional handcrafted features to end-to-end deep networks and joint learning of physical priors, all demonstrate the trend of continuous iteration and integration of machine learning methods. However, it is important to note that current mainstream cloud detection methods mostly focus on generating cloud masks to eliminate areas affected by cloud interference; their essence remains "determining the existence of clouds" rather than "characterizing the physical properties of clouds." Although methods such as Fmask and TCTR have improved thin cloud detection capabilities by utilizing the cirrus band, the cirrus band suffers from low signal-to-noise ratios and susceptibility to interference from bright ground features (such as deserts and snowfields) in arid regions, resulting in a large number of false positives and false negatives in practical applications. Especially in thin cloud regions, existing methods struggle to accurately reconstruct their radiation characteristics, affecting the extraction and utilization of high-precision cloudless imagery downstream.

[0004] Therefore, there is a need for a thin cloud TOA reflectivity estimation algorithm and system that can combine physical models with deep learning. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method and system for estimating the atmospheric top reflectance of thin clouds based on deep learning. This method combines physical modeling and deep learning to estimate the atmospheric top reflectance (TOAreflectance) of thin clouds, and is applicable to high-precision cloud reflectance inversion under various land surface, cloud types and Band 9 failure scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The present invention provides a deep learning-based method for estimating the atmospheric reflectance of thin clouds at their top, comprising the following steps:

[0008] S1: Radiometric calibration is performed on the acquired remote sensing image data to obtain a dual-channel cloud pollution image containing reflectance information of band 1 and band 2;

[0009] S2: Input the dual-channel cloud pollution image into a pre-trained convolutional neural network (CNN) module, and learn the nonlinear mapping relationship between band 1 and band 2 through the CNN module, and output the preliminary inversion result of the TOA reflectance of the thin cloud atmosphere top;

[0010] S3: Input the preliminary inversion results into the pre-trained residual denoising diffusion model, optimize the preliminary inversion results through the residual denoising diffusion model, separate the surface interference signal and retain the cloud radiation characteristics, and generate the final thin cloud TOA reflectance estimation image.

[0011] The residual denoising diffusion model employs a dual-path diffusion framework, including:

[0012] Deterministic residual diffusion path: Modeling the systematic deviation between the preliminary inversion results and the ideal cloud reflectivity;

[0013] Random noise propagation path: Modeling random interference introduced by atmospheric scattering and sensor noise.

[0014] Furthermore, the CNN module in step S2 is constructed based on the physical radiative transfer equation, which is expressed as:

[0015]

[0016] Where, ρ modeled_ci ρ represents the cloud TOA reflectivity obtained from modeling. i * ρ represents the TOA reflectivity of band i; j * a represents the TOA reflectivity of band j; ij and b ij It is a constant.

[0017] Furthermore, the training process of the residual denoising diffusion model includes:

[0018] a. Define the degenerate state equation at time t:

[0019]

[0020] The key parameters are defined as follows:

[0021] Residual term: ρ res =ρ SC ―ρ9, representing the preliminary inversion result ρ SC Systematic deviation from ideal cloud reflectivity;

[0022] Noise item: Characterize random disturbances;

[0023] Cumulative coefficient: The incremental injection intensity of residuals and noise is controlled separately.

[0024] b. Train the residual prediction networks separately.

[0025] By training the residual prediction network To predict residuals

[0026] Training the noise prediction network ε θ (ρ t ,t,ρ SC To estimate noise

[0027] During the sampling phase, preliminary cloud TOA reflectance images ρ, which contain some ground object interference, are used. SC Perform gradual noise reduction.

[0028] Furthermore, the specific process of the optimization treatment in step S3 is as follows:

[0029] Using the preliminary inversion results as the initial input state, multi-time-step iterative sampling is performed in the residual denoising diffusion model;

[0030] At each time step, the predicted values ​​output by the residual prediction network and the noise prediction network are used simultaneously for residual removal and noise removal;

[0031] The final output is an optimized TOA reflectance image of thin clouds.

[0032] Furthermore, the dual-channel cloud pollution image only contains band 1 and band 2 data in the visible to near-infrared bands.

[0033] Furthermore, the preliminary inversion results of the reflectance at the top of the thin cloud atmosphere (TOA) are calculated according to the following formula:

[0034] ρ SC =LeakeyRelu(Conv 1×1 (ρ1 * ,ρ2 * ));

[0035] Where, ρ SC This represents the cloud TOA reflectance obtained from the initial inversion.

[0036] LeakeyRelu() represents the activation function;

[0037] ρ1 * This indicates the TOA reflectivity of band 1;

[0038] ρ2 * This indicates the TOA reflectivity of band 2.

[0039] Furthermore, the CNN module in step S2 adopts a lightweight architecture. The input data of the CNN module is dual-channel 2D data with the data shape being [batch_size, 2, height, width].

[0040] The number of channels is set to compress the channel dimension layer by layer;

[0041] Activation function: LeakyReLU activation function with a preset negative slope is used after each convolution layer.

[0042] Furthermore, the target loss function in the residual denoising diffusion model is calculated according to the following formula:

[0043]

[0044] Where, λ i (i = 1, 2, 3, 4) represents the weighting coefficients, which are used to adjust the importance of different losses.

[0045] The present invention provides a deep learning-based system for estimating the atmospheric reflectance of thin clouds at the top, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the program.

[0046] The beneficial effects of this invention are as follows:

[0047] This invention provides a deep learning-based method and system for estimating the reflectance of the top of the atmosphere in thin clouds. The method first performs radiometric calibration on the acquired remote sensing data to obtain a reflectance image. Then, it inputs band 1 and band 2, and uses a convolutional neural network to perform end-to-end feature extraction and cloud TOA reflectance inversion, outputting a preliminary inversion result. Finally, based on the preliminary inversion result, a trained denoising diffusion model is used to sample and generate a denoised cloud TOA reflectance image. This method combines physical rules with deep learning to estimate the reflectance of the top of the atmosphere in thin clouds, analyzing typical differences in clouds across spectral bands based on physical rules, and classifying them using fixed or dynamic thresholds.

[0048] This method models thin cloud reflectance estimation as an image restoration problem. Through empirical relationships between bands, it achieves preliminary inversion of thin cloud radiative characteristics without relying on the highly sensitive but easily interfered-with cirrus band. Combining the advantages of diffusion models in image feature representation and reconstruction, it accurately separates the radiative contribution of clouds from complex surface and atmospheric mixed signals, enabling effective estimation of TOA reflectance generalized to different types and concentrations of thin clouds. This provides a more robust physical basis for downstream cloud detection and cloudless image restoration.

[0049] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0050] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0051] Figure 1 This is a structural diagram of a thin cloud TOA reflectivity estimation model.

[0052] Figure 2 The physical cloud TOA inversion module (PCIM).

[0053] Figure 3 Flowchart for residual denoising and diffusion recovery.

[0054] Figure 4 This is a comparison chart of the experiments. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, Figure 1 The diagram shows the structure of the thin cloud TOA reflectance estimation model. This embodiment provides a deep learning-based method for estimating the reflectance of the thin cloud atmospheric top, which includes the following steps:

[0058] S1: Radiometric calibration is performed on the acquired remote sensing image data to obtain a dual-channel cloud pollution image containing reflectance information of band 1 and band 2;

[0059] S2: Input the dual-channel cloud pollution image into a pre-trained convolutional neural network (CNN) module, learn the nonlinear mapping relationship between band 1 and band 2 through the CNN module, and output the preliminary inversion result of the reflectance of the top of the thin cloud atmosphere (TOA);

[0060] S3: Input the preliminary inversion results into the pre-trained residual denoising diffusion model, optimize the preliminary inversion results through the residual denoising diffusion model, separate the surface interference signal and retain the cloud radiation characteristics, and generate the final thin cloud TOA reflectance estimation image.

[0061] The residual denoising diffusion model employs a dual-path diffusion framework, including:

[0062] Deterministic residual diffusion path: Modeling the systematic deviation between the preliminary inversion results and the ideal cloud reflectivity;

[0063] Random noise propagation path: Modeling random interference introduced by atmospheric scattering and sensor noise.

[0064] The deterministic residual diffusion path forms a residual prediction network; the random noise diffusion path forms a noise prediction network.

[0065] The specific implementation process of the estimation method provided in this embodiment is as follows:

[0066] Assuming the thin clouds lie above a "virtual surface," and considering cloud transmission and multiple reflections between the clouds and the "virtual surface," the reflectivity of the top atmosphere (TOA) observed by the satellite in the visible light band can be derived as follows:

[0067]

[0068] in, Indicates the reflectivity of the top atmosphere (TOA);

[0069] ρ ci This indicates the TOA reflectance of thin clouds;

[0070] T i This indicates the bidirectional transmittance through the cloud layer (typically greater than 0.9 for thin clouds);

[0071] ρ i The TOA reflectivity represents the "virtual surface";

[0072] S ci Indicates the reflectivity of upward radiation from the cloud base;

[0073] This indicates the TOA reflectivity of the satellite sensor, with the subscript i indicating the spectral band;

[0074] Due to S ci For thin clouds, the value is typically less than 0.1, therefore S ci ρ i Since it is much smaller than 1, equation (1) can be approximated as:

[0075]

[0076] For band j, the simplified radiative transfer equation is:

[0077] ρ j * =ρ cj +Tj ·ρ j (3)

[0078] Therefore, the top-of-atmosphere (TOA) reflectance of thin clouds in band i was modeled, and ρ modeled_ci The model is as follows:

[0079]

[0080] in, ρ represents the TOA reflectivity of band i; j * T represents the TOA reflectivity of band j; i ρ represents the bidirectional (direct plus diffuse) transmittance through clouds; cj ρ represents the TOA reflectivity of thin clouds in band j; j ρ represents the reflectivity of the "virtual surface" TOA in band j. modeled_ci a represents the cloud TOA reflectivity obtained from modeling. ij and b ij It is a constant;

[0081] Generally speaking, the TOA reflectance should be zero when there are no clouds; when there are clouds, the TOA reflectance should be greater than zero.

[0082] Therefore, in the modeling, the negative TOA value is set to zero.

[0083] Equation (4) shows that the TOA reflectance of clouds can be inverted using the linear relationship between the two bands. However, this method involves multiple approximations. To improve accuracy, this embodiment uses a CNN to learn the nonlinear relationship between the two bands based on a simple linear model. The input to the CNN includes two-channel cloud contamination images (b1 and b2), and the output is a preliminary estimate of the cloud TOA reflectance.

[0084] In this embodiment, the input is 2-channel 2D data (shape [batch_size, 2, height, width]). The output is 1-channel 2D data (shape [batch_size, 1, height, width]). This lightweight network uses only 1×1 convolution kernels, does not change the spatial resolution (stride = 1), performs nonlinear fitting, and preserves spatial information. The network has 3 layers, with the number of channels varying from 2 to 16 to 8 to 1. The activation function is LeakyReLU (negative slope = 0.2) to alleviate the gradient vanishing problem. The parameters are initialized using Kaiming uniform initialization. The core of this module is defined in (5) as using stacked 1×1 convolutions and activation functions to learn the nonlinear mapping between two bands, thereby improving the cloud TOA reflectance (ρ). SC A preliminary estimate was made.

[0085] In this embodiment, the activation function is LeakyReLU (negative slope = 0.2) to alleviate the gradient vanishing problem. Parameter initialization uses Kaiming uniform initialization.

[0086] The flowchart of the physical cloud TOA inversion module PCIM is as follows: Figure 2 As shown, Figure 2 The physical cloud TOA inversion module (PCIM).

[0087] ρ SC =LeakeyRelu(Conv 1×1 (ρ1 * ,ρ2 * (5)

[0088] Where, ρ SC This represents the cloud TOA reflectance obtained from the initial inversion.

[0089] LeakeyRelu() represents the activation function, and in this embodiment, the negative slope is preset to 0.2;

[0090] ρ1 * This indicates the TOA reflectivity of band 1;

[0091] ρ2 * This indicates the TOA reflectivity of band 2;

[0092] Although the CNN-based estimation results are satisfactory, ground object information is still observed to be retained in the initial cloud TOA reflectance ρ. SC middle.

[0093] To further optimize the results, a diffusion model is introduced to improve the estimation. Traditional diffusion models generate results by sampling pure noise ε. However, in the initial cloud TOA reflectivity result ρ... SC In this context, cloud information dominates, requiring only the removal of a small amount of ground information. Therefore, there's no need to start the recovery process from pure noise, because ρ SC Useful cloud information has already been provided.

[0094] Therefore, the decoupling requirement of the "surface cloud" mixed signal in cloud TOA reflectivity inversion is addressed.

[0095] like Figure 3 As shown, this embodiment employs a dual-path diffusion framework. Figure 3 This is a flowchart for residual denoising and diffusion recovery. The core idea is to decompose the forward degradation process into two parts: a deterministic residual path and a random noise path.

[0096] Deterministic residual diffusion explicit modeling of surface reflectance ρ iSystemic biases generated by cloud transmission provide clear guidance for the gradual recovery process;

[0097] Random noise diffusion is used to characterize random interferences such as atmospheric scattering and sensor noise.

[0098] In this embodiment, both the residual prediction network and the noise prediction network adopt the UnetRes architecture, the core of which is an encoder-decoder structure. The network first processes the input image through a 7x7 initial convolutional layer (init_conv), which contains the original image and temporal information. The temporal information is transformed into a high-dimensional vector through a configurable sinusoidal positional encoding (learnable RandomOrLearnedSinusoidalPosEmb), and then processed through a two-layer MLP (time_mlp) containing the GELU activation function, and finally injected into the residual block of each network layer.

[0099] The encoder path consists of multiple stages, each containing two residual blocks (ResNetBlock, using grouped convolutions, with a default group size of 8), a linear attention module, and a downsampling layer. The number of channels increases progressively according to the dim_mults parameter (e.g., 1, 2, 4, 8). The decoder path gradually restores resolution through upsampling layers and skip connections, with an additional global attention mechanism added to the bottleneck layer to capture long-range dependencies. The network ultimately outputs a residual block and a 1x1 convolution (final_conv), supporting learned variance to predict noise distribution.

[0100] All experiments were conducted using the same PyTorch architecture. The UNet architecture was used for residual and noise prediction in the network. UNet settings remained consistent across all tasks, including channel size (64) and channel multipliers (1, 2, 4, 8). Training and testing for all experiments were performed on a single Nvidia GTX 3090. Specific experimental details are shown in Table 1 below. The training iterations were set to 80,000, with 5 sampling steps, and DDIM was used to accelerate sampling.

[0101] Table 1 Experimental Setup

[0102]

[0103] Given the ideal cloud reflectivity ρ 0 =ρ9, define the degenerate state at time t as:

[0104]

[0105] The key parameters are defined as follows:

[0106] Residual term: ρ res =ρ SC ―ρ9, representing the preliminary inversion result ρ SC Systematic deviation from ideal cloud reflectivity;

[0107] Noise item: Characterize random disturbances;

[0108] Cumulative coefficient: The incremental injection intensity of residuals and noise is controlled separately.

[0109] Wherein, ρ9 represents the reflectivity of the virtual surface TOA in band 9; ρ res The TOA reflectance represents the residual; ρ SC ρ represents the cloud TOA reflectance obtained from the preliminary inversion. t The TOA reflectance image represents time step t; α t Table showing the time coefficients for controlling residual addition; β t This table represents the time coefficient for controlling the addition of noise; ρ 0 Represents the reflectance image of an ideal thin cloud at TOA;

[0110] α i The residual coefficient represents the time step t. This represents the sum of the time coefficients for the control residuals; β t The noise figure represents the time step t; The sum of the time coefficients for controlling noise; ε t―1 ε represents the noise at time step t1; ε represents the noise that follows a standard normal distribution; t represents the time step.

[0111] When t = T, the cumulative residual The total attenuation of surface signals through clouds; cumulative noise. Represents the superposition effect of atmospheric scattering and sensors, etc.; degraded state ρ T =ρ SC +ε represents the initial inversion result with noise. By adjusting... The intensity of noise disturbance can be controlled to adapt to different recovery needs.

[0112] Where T represents the final time step;

[0113] During the training process, the residual prediction network and the noise prediction network are trained separately.

[0114] By training the residual prediction network To predict residuals

[0115] Training the noise prediction network ε θ (ρ t ,t,ρ SC To estimate noise

[0116] in, ε represents the predicted residual; θ This represents the noise obtained from the prediction.

[0117] During the sampling (generation) phase, the target image ρ is no longer used. 0 Instead, it only utilizes preliminary cloud TOA reflectance images ρ with some ground feature interference. SC Perform gradual noise reduction.

[0118] The specific method for constructing the dataset in this embodiment includes the process of generating synthetic data (such as the physical model or simulation tool used), the source and preprocessing steps of real data, and the data annotation standards.

[0119] A diverse dataset was constructed using Landsat 8 and Landsat 9 data to improve generalization ability. Since cirrus performance is inconsistent in arid, cold seasons, more accurate training data with a more uniform numerical distribution was obtained, reducing the workload of dataset construction. Referring to the Köppen global climate classification, and focusing on the temperate and tropical regions, the wet season (March and May) and geographical location (the Great Lakes region of the United States) were selected first. Eighteen images were chosen from these, covering various environments including snow-capped mountains, urban areas, coastal areas, forests, and arid landscapes. The parameters of the selected scenes are listed in Table 2. These selected scenes were then carefully cropped into 2146 images, each with a resolution of 512×512 pixels. The final dataset includes 1716 images designated for training and 430 images reserved for testing. To ensure fairness in subsequent experiments, 256 randomly selected scene images from other regions were prepared to construct a validation dataset, which underwent the same preprocessing steps and were used for subsequent evaluation analysis.

[0120] Table 2 Sensors and Acquired Parameters

[0121]

[0122] First, using the preliminary inverted cloud TOA reflectance image ρ SC As input;

[0123] Simultaneously, the predicted residuals and noise are used for residual and noise removal after multiple time steps;

[0124] Finally, the denoised cloud TOA reflectance image is obtained.

[0125] Where, ρ t―1 It can be obtained from ρ in the following ways t Medium sampling:

[0126]

[0127] in, In this task, η = 0 is used for the directional recovery process;

[0128] ρ t―1 Represents the TOA reflectance image at time step t1; ρ t This represents the TOA reflectance image at time step t.

[0129] σ t ε represents the parameter that controls whether the generation is random or directed; θ This indicates that the predicted noise is present.

[0130] It is important to emphasize that the goal of the network during training is to learn the estimation capabilities of residuals and noise separately. However, during the sampling (inference) process, the real target image ρ9 is no longer used. Instead, the model is relied upon to iteratively recover the final cloud TOA reflectance image.

[0131] The input used during the training phase is ρ T =ρ SC +ε+ρ res , where ρ res =ρ SC ―ρ9;

[0132] The input during the sampling phase is ρ T =ρ SC +ε.

[0133] Where, ρ T A noisy image representing the final time step;

[0134] This design ensures that the model can learn accurate residual and noise information during training, while recovering high-quality cloud TOA reflectivity from noisy preliminary results during the inference phase.

[0135] This embodiment enhances the utilization of prior information by employing a dual-diffusion structure in the residual denoising diffusion model, building upon traditional diffusion models and making the recovery process more efficient and reasonable. Deterministic residual diffusion provides a clear recovery direction, while noise diffusion ensures the diversity and detail fidelity of the generated data. Ultimately, this method effectively improves the inversion quality of cloud TOA reflectivity and reduces noise interference.

[0136] The estimation method provided in this embodiment reverses the cloud detection task into a "ground object signal separation" problem, using a generative denoising network to separate surface interference while preserving cloud radiation characteristics. A physics-based cloud TOA inversion module (PCIM) is constructed. Based on the radiative transfer equation, and considering the strong correlation between bands 1 and 2, as well as sensitivity to clouds, the cloud TOA reflectivity is physically modeled, achieving a preliminary end-to-end inversion of cloud TOA reflectivity. A residual denoising diffusion model is employed, treating the directional degradation process from the ideal cloud reflectivity image to the ground object signal residue as residual diffusion, explicitly guiding the reverse generation process for cloud TOA reflectivity inversion. Noise diffusion is preserved as a random perturbation. Taking advantage of the small radiative difference between the two bands (Band 1 and 2), a dataset is constructed to obtain general patterns, generalizing to the inversion of other types of clouds and fog, such as stratus and cumulus.

[0137] Compared with existing cloud detection algorithms, the method in this embodiment can achieve quantitative estimation of cloud physical properties and has stronger robustness and generalization performance, adapting to complex scenarios and cloud and fog types. Compared with deep learning methods, this patent integrates a physical model to enhance generalization ability, reduces dependence on large amounts of data annotation, and introduces residual-guided diffusion, which can remove ground object interference while maintaining cloud characteristics, and realize an end-to-end estimation network structure.

[0138] This embodiment utilizes this method, setting the training parameters according to Table 3, to obtain the following... Figure 4 The experimental comparison results shown are as follows:

[0139] Table 3 shows the training parameters.

[0140]

[0141] The objective loss function of the residual denoising diffusion model in this embodiment consists of four parts: The complete expression of the objective function is defined as follows:

[0142]

[0143] Where, λ i (i = 1, 2, 3, 4) represent weighting coefficients, used to adjust the importance of different losses;

[0144] L1 represents the standard L1 loss function; L 1perceptual This represents the perceived L1 loss; L 1edge This represents the edge-aware L1 loss;

[0145] ρ res TOA reflectance represents the true residual; This represents the TOA reflectance of the predicted residual;

[0146] 1. Perpixel Loss: To measure the consistency between the measured cloud TOA reflectance image and the generated result, L1 loss is used across the entire image to calculate the difference between the target image and the reconstructed image. L11 evaluates the consistency between predicted noise and residuals and actual noise and residuals to ensure the removal of ground interference and coherent recovery of cloud TOA reflectance. L12 evaluates the distance between the initial inversion result and the actual cloud TOA reflectance data to correct the initial inversion process of cloud TOA reflectance. The expressions for these two losses are given below:

[0147]

[0148] Where N is the number of pixels in the image;

[0149] L l―1 L represents the loss between predicted noise and residuals and actual noise and residuals. l―2 This indicates the loss between the preliminary inversion results and the actual cloud TOA reflectivity;

[0150] 2. Perceptual Loss: Perceptual loss is introduced to evaluate the perceptual similarity between the predicted residual and the true residual. It extracts feature representations from a pre-trained VGG16 network and calculates the L1 distance between the predicted image and the target image in the deep feature space. Perceptual loss aims to restore all texture and perceptual information. The perceptual loss is defined as follows:

[0151]

[0152] Where, ψ p This represents the feature extractor of the p-th layer in the VGG16 network, where p is the selected number of depth layers. Perceptual loss can effectively restore texture and perceptual information, improving the visual quality of the generated image.

[0153] L perceptual Represents the perceptual L1 loss; P represents the selected number of network layers; ψ p This represents the feature extractor of the p-th layer in the VGG16 network;

[0154] 3. Edge Loss: Edge loss is used to preserve edge details of the image, ensuring that the predicted residual is structurally consistent with the target residual. It is defined as follows: By calculating the L1 distance between the edge maps of the predicted and target residuals, edge loss is defined as:

[0155]

[0156] Here, Edge(·) represents the edge extraction operation, implemented using the Sobel operator;

[0157] L edge The L1 loss represents the margin plot between the predicted residual and the target residual;

[0158] Figure 4 This is a comparison chart of the experiments. The first row shows the results when b9 is working properly. Figure 4 In the first row, a, b, and c represent the true-color image, band 9, and the final inversion result for the same scene, respectively; in the second row, d, e, and f represent the true-color image, band 9, and the final inversion result when there are missed detections in band 9, respectively.

[0159] In the first scene, we can see a complex mix of towns, farmland, land features, and water bodies. When band 9 is functioning correctly, the algorithm, relying only on the b1 and b2 channels, can accurately derive the TOA reflectivity of the clouds, comparable to b9. Figure 4 (c) in the middle; however, in the sea scene in the second line, one can see Figure 4 In the red box (e), band 9 did not detect thick low-altitude clouds; however, the algorithm... Figure 4 The (f) in the formula compensates for this problem and can accurately reflect the TOA reflectivity of the cloud in the entire scene.

[0160] Table 4 shows the quantitative accuracy evaluation indicators.

[0161]

[0162] As shown in Table 4, the data presents a comparative result of quantization accuracy evaluation. The method shown in Table 4 performs excellently in three key metrics: MAE (Mean Absolute Error) is 0.01 (↓ indicates lower is better), indicating minimal error between the predicted and actual values, demonstrating high accuracy; PSNR (Peak Signal-to-Noise Ratio) reaches 42.5dB (↑ indicates higher is better), indicating excellent image quality preservation with minimal noise; and SSIM (Structural Similarity) is 0.95 (↑ indicates higher is better, with a maximum of 1), indicating that the reconstructed image is structurally almost identical to the original image. These data demonstrate that the method performs exceptionally well in quantization accuracy, particularly achieving near-perfect levels in preserving image quality and structural similarity (SSIM 0.95 is close to the theoretical maximum of 1).

[0163] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A deep learning-based method for estimating the reflectivity of the atmospheric top of thin clouds, characterized by: Includes the following steps: S1: Radiometric calibration is performed on the acquired remote sensing image data to obtain a dual-channel cloud pollution image containing reflectance information of band 1 and band 2; S2: Input the dual-channel cloud pollution image into a pre-trained convolutional neural network (CNN) module, and learn the nonlinear mapping relationship between band 1 and band 2 through the CNN module, and output the preliminary inversion result of the TOA reflectance of the thin cloud atmosphere top; S3: Input the preliminary inversion results into the pre-trained residual denoising diffusion model, optimize the preliminary inversion results through the residual denoising diffusion model, separate the surface interference signal and retain the cloud radiation characteristics, and generate the final thin cloud TOA reflectance estimation image. The residual denoising diffusion model employs a dual-path diffusion framework, including: Deterministic residual diffusion path: Modeling the systematic deviation between the preliminary inversion results and the ideal cloud reflectivity; Random noise propagation path: Modeling random interference introduced by atmospheric scattering and sensor noise.

2. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 1, characterized in that, In step S2, the CNN module is constructed based on the physical radiative transfer equation, which is expressed as: Where, ρ modeled_ci ρ represents the cloud TOA reflectivity obtained from modeling. i * ρ represents the TOA reflectivity of band i; j * a represents the TOA reflectivity of band j; ij and b ij It is a constant.

3. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 1, characterized in that, The training process of the residual denoising diffusion model includes: a. Define the degenerate state equation at time t: The key parameters are defined as follows: Residual term: ρ res =ρ SC ―ρ9, representing the preliminary inversion result ρ SC Systematic deviation from ideal cloud reflectivity; Noise item: Characterize random disturbances; Cumulative coefficient: The incremental injection intensity of residuals and noise is controlled separately; b. Train the residual prediction networks separately. By training the residual prediction network To predict residuals Training the noise prediction network ε θ (ρ t ,t,ρ SC To estimate noise During the sampling phase, preliminary cloud TOA reflectance images ρ, which contain some ground object interference, are used. SC Perform gradual noise reduction.

4. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 1, characterized in that, The specific process of the optimization process in step S3 is as follows: Using the preliminary inversion results as the initial input state, multi-time-step iterative sampling is performed in the residual denoising diffusion model; At each time step, the predicted values ​​output by the residual prediction network and the noise prediction network are used simultaneously for residual removal and noise removal; The final output is an optimized TOA reflectance image of thin clouds.

5. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 1, characterized in that, The dual-channel cloud pollution image contains only band 1 and band 2 data in the visible to near-infrared bands.

6. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 2, characterized in that, The preliminary inversion results of the reflectance at the top of the thin cloud atmosphere (TOA) are calculated according to the following formula: r SC =LeakeyRelu(Conv 1×1 (p1 * ,p2 * )); Where, ρ SC This represents the cloud TOA reflectance obtained from the initial inversion. LeakeyRelu() represents the activation function; ρ1 * This indicates the TOA reflectivity of band 1; ρ2 * This indicates the TOA reflectivity of band 2.

7. The method for estimating the reflectance of the atmospheric top of thin clouds based on deep learning according to claim 1, characterized in that, The CNN module in step S2 adopts a lightweight architecture. The input data of the CNN module is dual-channel 2D data with the data shape of [batch_size,2,height,width]. The number of channels is set to compress the channel dimension layer by layer; Activation function: LeakyReLU activation function with a preset negative slope is used after each convolution layer.

8. The method for estimating the atmospheric reflectance of thin clouds based on deep learning according to claim 1, characterized in that, The target loss function in the residual denoising diffusion model is calculated according to the following formula: Where, λ i (i = 1, 2, 3, 4) represents the weighting coefficients, which are used to adjust the importance of different losses.

9. A deep learning-based system for estimating the reflectivity of the atmospheric top of thin clouds, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 8.

Images