A thin cloud morphology field generation method and system for landsat 8 / 9 satellite sample statistical prior

Abstract

This invention provides a method and system for generating thin cloud morphological fields based on statistical priors for Landsat 8 / 9 satellite samples, belonging to the fields of remote sensing image processing and computer vision. It addresses the lack of statistically significant (but not visually-based) methods for generating thin cloud morphological fields, and the absence of explicit statistical constraints and consistency checks in existing methods, which leads to inconsistencies between generated results and real Landsat thin cloud samples in terms of marginal distribution, multi-scale spatial structure, or topological properties. This invention models the statistical characteristics of real thin cloud samples and combines a diffusion generation model with a statistical consistency check mechanism to achieve continuous thin cloud morphological field generation with a controllable statistical structure. This invention significantly improves statistical consistency, enhances the ability to maintain multi-scale spatial structure, and improves verifiability and controllability. It can be applied to fields such as thin cloud impact simulation, degradation modeling, algorithm robustness testing, and scene construction for remote sensing data processing systems.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, and more specifically, to a method and system for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples. Background Technology

[0002] Landsat 8 / 9, as a long-term, continuous, medium-to-high resolution remote sensing satellite system, is widely used for surface monitoring, resource surveys, and environmental assessments. However, thin cloud contamination is a persistent and difficult-to-avoid problem in optical remote sensing imagery. Thin clouds are typically semi-transparent, exhibiting multi-scale, striped, or patchy spatial distributions, and introducing complex nonlinear degradation effects in brightness temperature or reflectivity space. Unlike thick clouds, thin clouds often do not completely obscure the surface, but rather cause localized radiation attenuation and spatial blurring, and their morphology exhibits significant statistical structural characteristics.

[0003] In existing technologies, research on thin clouds mainly focuses on two directions:

[0004] (1) Thin cloud detection and cloud mask generation, such as thin cloud recognition algorithms based on convolutional neural networks or cirrus cloud bands;

[0005] (2) Thin cloud removal or image restoration, such as cloud removal methods based on deep learning or physical modeling.

[0006] However, within the aforementioned technical framework, thin clouds are typically viewed as "disturbing factors to be eliminated" rather than as modelable objects with statistical structure. Existing research rarely focuses on the generation of the "thin cloud morphological field itself," and in particular, lacks a controllable generation mechanism that can statistically maintain consistency with real thin cloud samples from Landsat 8 / 9.

[0007] In terms of generation methods, diffusion models have made significant progress in image generation in recent years and have been applied to tasks such as cloud removal and super-resolution reconstruction of remote sensing images. However, these methods mainly focus on restoring or generating natural images, rather than addressing the statistical generation problem of thin cloud morphology itself. Directly using diffusion models to generate thin cloud morphological fields, without explicit statistical constraints and consistency checks, often only guarantees visual similarity, but cannot guarantee that the generated samples are consistent with real Landsat thin cloud samples in terms of marginal distribution, multi-scale spatial structure, or topological properties.

[0008] Furthermore, existing generation methods typically lack closed-loop control mechanisms. Even if the model learns the data distribution during training, its sampling results may still deviate from the statistical range of the true samples, and existing technologies lack quantitative verification and feedback adjustment mechanisms for "statistical consistency." This makes it difficult to use the generated results for rigorous algorithm verification or simulation studies.

[0009] In summary, existing technologies have significant shortcomings in the following aspects:

[0010] 1. No systematic modeling was performed for the statistical priors of the Landsat 8 / 9 thin cloud samples;

[0011] 2. There is a lack of statistically constrained generation methods for thin cloud morphological fields;

[0012] 3. Lack of a verifiable closed-loop mechanism between the generated results and the statistical structure of the actual samples.

[0013] Therefore, it is necessary to propose a thin cloud morphological field generation method based on the statistical prior of Landsat 8 / 9 satellite samples, which can ensure the diversity of generation while ensuring that the generated results are consistent with the real thin cloud samples in a statistical sense. Summary of the Invention

[0014] The technical problem to be solved by this invention is:

[0015] To address the lack of statistically significant rather than visual methods for generating thin cloud morphological fields, and the fact that existing generation methods lack explicit statistical constraints and consistency checks, resulting in inconsistent generated results with real Landsat thin cloud samples in terms of marginal distribution, multi-scale spatial structure, or topological characteristics.

[0016] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0017] This invention provides a method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples, comprising the following steps:

[0018] S100. Obtain thin cloud sample regions from Landsat 8 / 9 public data; the sample source is based on reproducible thin cloud detection or official quality labeling products, thus obtaining a thin cloud binary mask; on this basis, construct a continuous thin cloud morphological field through distance transformation or continuous mapping, so that the interior and boundary of the thin cloud present a smooth transition structure; subsequently, perform statistical analysis on the samples to extract marginal distribution, multi-scale spatial structure and morphological topological features, forming a thin cloud statistical prior set;

[0019] S200. After the statistical prior extraction is completed, the diffusion model is trained to learn the distribution of the thin cloud morphological field. The diffusion model uses continuous morphological field patches as training samples and learns the data distribution by gradually adding and removing noise. After training, the model generates new thin cloud morphological field samples from random noise.

[0020] S300. A statistical consistency verification mechanism is introduced during the generation stage. For each generated sample, its marginal distribution, power spectrum structure, and connected component statistical indicators are calculated and compared with the statistical prior of the real thin cloud sample. If the statistical characteristics of the generated sample exceed the allowable range of the real sample, it is judged as an unqualified sample and resampling or condition adjustment is triggered. This closed-loop mechanism is used to ensure that the final output sample is consistent with the real thin cloud morphology of Landsat 8 / 9 in a statistical sense.

[0021] Further, in step S100, the following is included:

[0022] S110, Thin Cloud Sample Construction and Statistical Prior Extraction

[0023] Let the nth Landsat 8 / 9 scene multispectral image be... ,in Represents pixel coordinates, The image domain is represented by k, and the spectral index is represented by k. The thin cloud binary mask is:

[0024] (1)

[0025] in, This indicates that pixel x is marked as a thin cloud region. Indicates non-thin cloud regions;

[0026] Construction of S120 and the continuous thin cloud morphological field W(x)

[0027] Introducing a continuous thin cloud morphology field , used to indicate the intensity or extent of the influence of thin clouds in the observation space;

[0028] A continuous processing method for thin cloud binary masks is applied based on distance transformation; assuming... Let x be the Euclidean distance to the nearest non-thin cloud pixel, and define the continuous thin cloud morphology field as:

[0029] (2)

[0030] in, exp represents the scale parameter used to control the boundary transition width;

[0031] Obtained through this method The value approaches 1 within the thin cloud region and continuously decreases to 0 at the boundary, thus forming a continuous thin cloud morphological expression.

[0032] S130. Definition and extraction of statistical priors

[0033] After obtaining a set of continuous thin cloud morphological fields Then, statistical priors are extracted, including marginal distribution priors, multi-scale spatial structure priors, and morphological and topological priors.

[0034] Furthermore, in step S120, the method of constructing the continuous field must satisfy:

[0035] (1) It maintains a high value inside thin clouds and a low value in non-thin cloud areas;

[0036] (2) A smooth transition is presented near the boundary to depict the gradual characteristics of thin clouds.

[0037] Furthermore, in step S130, the statistical prior design satisfies:

[0038] (1) It can be used as a computable constraint during the generation phase;

[0039] (2) It can objectively reflect the statistical characteristics of the thin cloud morphology of Landsat 8 / 9.

[0040] Furthermore, the statistical prior includes,

[0041] S131, Marginal distribution prior, for all samples Statistically analyze its marginal distribution and estimate its empirical probability density function. ;

[0042] S132, Multi-scale spatial structure priors, calculation The power spectral density; let its two-dimensional Fourier transform be... Then the power spectrum for:

[0043] (3)

[0044] right Averaging in the radial frequency domain yields a one-dimensional spectrum. The slope of the spectrum in logarithmic coordinates is calculated to characterize the multi-scale attenuation characteristics; the spectral slope and related length parameters constitute the spatial structure prior.

[0045] S133, Morphological and Topological Priors, for or Perform connected component analysis on the high-value regions, statistically analyze the area distribution A of the connected components, and determine the perimeter-area relationship; let the area of ​​the i-th connected component be A. i The perimeter is L i Calculate its power-law fitting relationship:

[0046] (4)

[0047] The exponent β or its distribution range is used to characterize boundary roughness and morphological complexity. These statistics together constitute the topological prior.

[0048] S134. Formal expression of the statistical prior set: Combining the above statistics, the statistical prior of the thin cloud sample is expressed as follows:

[0049] 5)

[0050] in, This represents the power spectrum slope or related length parameter. Indicates the pattern index, This represents the range of area distribution of the connected domain.

[0051] Further, in step S200, the following are included:

[0052] After obtaining thin cloud samples of Landsat 8 / 9 and their continuous thin cloud morphological field W(x), a diffusion generation model is used to learn the data distribution of W and generate new thin cloud morphological field samples during the sampling stage.

[0053] The training data consists of continuous morphological field samples obtained in step S100, and a continuous morphological field is constructed based on these samples. and from A fixed-size patch is cropped from the middle and used as a training sample, denoted as . To adapt to the commonly used numerical range of diffusion models, the following will be implemented: A linear mapping to [-1, 1] is denoted as During training, a standard forward diffusion process is used. Gradual noise addition And learn a denoising network to recover noise or recover clean samples;

[0054] The forward diffusion process is defined as a Markov chain. Its form is a Gaussian perturbation:

[0055] (6) Among them, For the number of diffusion steps, For noise schedule;

[0056] Equivalently, sampled in closed form:

[0057] (7)

[0058] in, ,in, Represents the symbol for a product. Represents a normal distribution. The noise scheduling parameters represent the noise distribution process; the denoising network is denoted as... The input is a noise sample. The time step t and optional condition variable c are used to output the response to noise. The estimate;

[0059] When using a noisy prediction paradigm, the training objective is to minimize the mean square error. :

[0060] (8)

[0061] in, Expressing expectations;

[0062] Sampling phase from pure noise Starting from this point, we obtain the result through inverse denoising iteration. The sample approximation; based on the DDPM denoising diffusion probability model sampling, its reverse update is:

[0063] (9)

[0064] in, The sampling noise figure;

[0065] After sampling is completed Mapping back to [0,1] yields And crop it to the legal range:

[0066] (10)

[0067] Final output This is the generated thin cloud morphology field sample.

[0068] Further, in step S300, the following are included:

[0069] S310. Establish a statistical prior reference representation on the real sample set; suppose a thin cloud binary mask is obtained from the Landsat 8 / 9 thin cloud samples, and construct a continuous thin cloud morphology field based on this mask. And extract the statistical prior set S from all training samples;

[0070] S320. For each thin cloud morphological field sample generated by the diffusion model Statistical consistency verification is achieved by calculating three types of statistical indicators; including,

[0071] S321, Marginal Distribution Index, used to measure Does the intensity distribution match that of the real sample?

[0072] Will Divide the [0,1] region into B bins to obtain a normalized histogram. The reference histogram for the real sample is denoted as... Then the marginal distribution deviation Defined as:

[0073] (11)

[0074] S322, Spatial Structure Index, is used to measure whether multi-scale texture structures match;

[0075] right Perform mean removal and calculate the two-dimensional power spectrum radial averaging yields Fit the slope of the logarithmic spectrum over a specified frequency band. and the reference slope interval Comparison, defining deviation for:

[0076] (12)

[0077] S323, Morphology / Topology Indicators, used to constrain the connectivity and scale distribution of cloud morphology;

[0078] exist A binary morphology set is obtained by selecting a fixed threshold τ. , Perform connected component analysis to obtain the set of connected component areas. and its empirical distribution The real sample yields the reference distribution. ; morphological deviation is measured by distribution distance definition:

[0079] (13)

[0080] S330. After obtaining the above deviation measures, combine them into an overall consistency scoring function. To decide whether to accept the generated sample:

[0081] (14) Among them, Used to balance the dimensions and importance of different indicators;

[0082] If a total threshold δ is set, then the consistency decision is: when Accepting samples Otherwise, reject the request and trigger resampling;

[0083] S340. Closed-loop resampling strategies are divided into two categories based on whether a condition variable c is introduced during diffusion generation. If unconditional diffusion is used, samples are rejected and resampling is performed directly until the condition is met. Or reach the maximum number of attempts N maxIf conditional diffusion is used, after rejecting samples, the condition variable c should be adjusted first, or a more suitable prior cluster reference distribution should be selected. .

[0084] A thin cloud morphology field generation system based on Landsat 8 / 9 satellite sample statistical priors. The system has program modules corresponding to the above steps, and executes the steps in the above-described method for generating thin cloud morphology fields based on Landsat 8 / 9 satellite sample statistical priors during runtime.

[0085] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement steps of a method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples.

[0086] Compared with the prior art, the beneficial effects of the present invention are:

[0087] 1. Significantly improved statistical consistency: Traditional generation methods only focus on visual similarity and lack statistical constraints. This invention, through a consistency verification mechanism, significantly reduces the deviation of generated samples from real Landsat 8 / 9 thin cloud samples in terms of indicators such as marginal distribution, power spectrum slope, and connected region area distribution. Experiments show that within the set statistical distance threshold range, the pass rate of generated samples can reach over 90%, while the pass rate of unconstrained generation methods is significantly reduced.

[0088] 2. Enhanced ability to preserve multi-scale spatial structure: By explicitly modeling the spatial spectrum slope and correlation length, the thin cloud morphology field generated by this invention is closer to the real sample in terms of multi-scale texture features, avoiding random noise structure or overly smoothed structure.

[0089] 3. Improved verifiability and controllability: This invention establishes a clear statistical consistency evaluation system, which provides the generated results with a basis for repeatable experimental verification, avoiding the problem of traditional generation models relying solely on subjective visual judgment.

[0090] 4. It is suitable for algorithm simulation and degradation testing. The generated continuous thin cloud morphological field can be directly used as degradation simulation input to verify the robustness of the algorithm under the cloud and provide stable and controllable experimental conditions for subsequent research.

[0091] In summary, this invention elevates the generation process from "visual generation" to "statistically controllable generation," significantly improving the reliability and scientific rigor in research and simulation scenarios. Attached Figure Description

[0092] Figure 1 This is a flowchart of a thin cloud morphological field generation method based on statistical priors of Landsat 8 / 9 satellite samples in an embodiment of the present invention. Detailed Implementation

[0093] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0094] This invention proposes a thin cloud morphological field generation framework based on statistical priors from Landsat 8 / 9 satellite samples. The goal of this method is not to reconstruct realistic cloud microphysical parameters, but rather to generate a continuous thin cloud morphological field with realistic statistical properties within the observation space, for subsequent research tasks such as degradation simulation, algorithm verification, or scene construction.

[0095] The overall process consists of four stages: sample construction and statistical modeling, diffusion distribution learning, statistical consistency verification, and closed-loop control generation.

[0096] Specific Implementation Plan 1: Combining Figure 1 As shown, this invention provides a method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples, comprising the following steps:

[0097] S100. Obtain thin cloud sample regions from Landsat 8 / 9 public data; the sample source is based on reproducible thin cloud detection or official quality labeling products, thus obtaining a thin cloud binary mask; on this basis, construct a continuous thin cloud morphological field through distance transformation or continuous mapping, so that the interior and boundary of the thin cloud present a smooth transition structure; subsequently, perform statistical analysis on a large number of samples to extract marginal distribution, multi-scale spatial structure and morphological topological features, forming a thin cloud statistical prior set;

[0098] Specifically, including,

[0099] S110, Thin Cloud Sample Construction and Statistical Prior Extraction

[0100] Before proposing a thin cloud morphology field generation model, it is necessary to first construct a reliable and reproducible set of thin cloud samples and extract statistical priors from them. Since this invention is aimed at Landsat 8 / 9 data, the sample source must be based on publicly available and standardized data products to ensure the objectivity and verifiability of the statistical priors.

[0101] Let the nth Landsat 8 / 9 scene multispectral image be... ,in Represents pixel coordinates, The image domain is represented as x, meaning the pixel coordinate x belongs to the image region Ω, and k represents the spectral band index. To obtain the thin cloud sample region, this invention uses a publicly available and reproducible cloud mask or thin cloud category product (e.g., a thin cloud detection algorithm based on the cirrus band or official quality layer annotation) as the initial thin cloud region indicator; this thin cloud binary mask is denoted as:

[0102] (1)

[0103] in, This indicates that pixel x is marked as a thin cloud region. Indicates non-thin cloud regions;

[0104] It should be emphasized that this thin cloud binary mask is used for statistical sample extraction and does not assume that it is a "real cloud field" in a physical sense, but rather as an approximate representation of thin clouds in the observation space.

[0105] Construction of S120 and the continuous thin cloud morphological field W(x)

[0106] Directly using thin cloud binary mask Statistical analysis can easily lose the transitional nature of thin cloud boundaries; therefore, this invention introduces a continuous thin cloud morphology field. This is used to represent the intensity or coverage of thin clouds in the observation space; the construction of the continuous field must meet two principles:

[0107] (1) It maintains a high value inside thin clouds and a low value in non-thin cloud areas;

[0108] (2) A smooth transition is presented near the boundary to depict the gradual characteristics of thin clouds;

[0109] A reproducible construction method is based on the continuity processing of the thin cloud binary mask using distance transformation; let... The Euclidean distance from pixel x to the nearest non-thin cloud pixel (in (calculated within the region), then the continuous thin cloud morphological field can be defined as:

[0110] (2)

[0111] in, exp represents the scale parameter used to control the boundary transition width;

[0112] Obtained through this method The value approaches 1 within the thin cloud region and continuously decreases to 0 at the boundary, thus forming a continuous thin cloud morphological expression.

[0113] S130. Definition and extraction of statistical priors

[0114] After obtaining a set of continuous thin cloud morphological fields Subsequently, this invention extracts statistical priors; the design of statistical priors must meet two requirements:

[0115] (1) It can be used as a computable constraint during the generation phase;

[0116] (2) It can objectively reflect the statistical characteristics of the thin cloud morphology of Landsat 8 / 9;

[0117] This invention constructs statistical priors from three levels:

[0118] S131, Marginal Distribution Prior

[0119] For all samples Statistically analyze its marginal distribution and estimate its empirical probability density function. ;

[0120] To avoid the influence of scene differences, the data is first normalized within each sample before aggregation; this distribution reflects the sparsity and central tendency of thin cloud intensity and can be used as a marginal constraint for the generative model.

[0121] S132, Multi-scale Spatial Structure Priorities

[0122] To characterize the spatial relational structure of thin cloud morphology, calculations were performed. The power spectral density; let its two-dimensional Fourier transform be... Then the power spectrum for:

[0123] (3)

[0124] right Averaging in the radial frequency domain yields a one-dimensional spectrum. The slope of the spectrum in logarithmic coordinates is calculated to characterize the multi-scale attenuation characteristics; the spectral slope and related length parameters constitute the spatial structure prior.

[0125] S133, Morphological and Topological Priors

[0126] right or Perform connected component analysis on the high-value regions, statistically analyze the area distribution A of the connected components, and determine the perimeter-area relationship; let the area of ​​the i-th connected component be A. i The perimeter is L i Then its power-law fitting relationship can be calculated:

[0127] (4)

[0128] The exponent β or its distribution range is used to characterize boundary roughness and morphological complexity. These statistics together constitute the topological prior.

[0129] S134. Formal Representation of Statistical Prior Sets

[0130] Based on the above statistics, the statistical prior of the thin cloud sample can be represented as a set:

[0131] (5)

[0132] in, This represents the power spectrum slope or related length parameter. Indicates the pattern index, Indicates the range of area distribution of connected components;

[0133] This statistical prior set will serve as a constraint basis for subsequent diffusion model generation and consistency verification stages;

[0134] The above parts constitute the statistical basis of the thin cloud morphological field generation method. The core goal of this stage is not the generation model itself, but to ensure that the generation model has clear, computable and reliable statistical constraints, thereby avoiding the uncontrollability caused by unconstrained random generation.

[0135] S200. After the statistical prior is extracted, the diffusion model is trained to learn the distribution of the thin cloud morphological field. The diffusion model uses continuous morphological field patches as training samples and learns the data distribution by gradually adding and removing noise. The model training phase only relies on real thin cloud samples and does not require additional manual annotation. After training, the model can generate new thin cloud morphological field samples from random noise.

[0136] Specifically, including,

[0137] After obtaining thin cloud samples from Landsat 8 / 9 and their continuous thin cloud morphological field W(x), this invention employs a diffusion generation model to learn the data distribution of W and generates new thin cloud morphological field samples during the sampling phase. Unlike directly generating remote sensing images, this invention limits the generated objects to continuous morphological fields. Its semantics are an equivalent expression of the influence intensity of thin clouds, rather than the true value of cloud microphysical parameters; this setting makes the generation task more focused on the statistical consistency of "morphological structure and intensity distribution", and can directly serve subsequent degradation simulation and algorithm verification.

[0138] The training data consists of continuous morphological field samples obtained in step S100; specifically, the thin cloud mask for each scene n is defined as:

[0139] (6)

[0140] Based on this, a continuous morphological field is constructed. (For example, monotonic mappings based on distance transformation, such as...) Where D represents the distance value obtained from the distance transformation, that is, the distance from pixel x to the thin cloud boundary in the nth scene; and from A fixed-size patch is cropped from the middle and used as a training sample, denoted as . To adapt to the commonly used numerical range of diffusion models, the following will be implemented: A linear mapping to [-1, 1] is denoted as ,in, An image patch representing a continuous thin cloud morphological field; training uses a standard forward diffusion process. Gradual noise addition And learn a denoising network to recover noise or recover clean samples;

[0141] The forward diffusion process is defined as a Markov chain. Its form is a Gaussian perturbation:

[0142] (7) Among them, For the number of diffusion steps, For noise schedules (linear or cosine are both acceptable);

[0143] Equivalently, it can be sampled directly from the closed form:

[0144] (8)

[0145] in, , Represents the symbol for a product. Represents a normal distribution. The noise scheduling parameters represent the noise distribution process; the denoising network is denoted as... The input is a noise sample. With time step t (and optional condition variable c), the output is a response to noise. The estimate;

[0146] When using a noisy prediction paradigm, the training objective is to minimize the mean square error. :

[0147] (9)

[0148] in, Expressing expectations;

[0149] The network structure can adopt the lightweight U-Net: it has multi-scale feature extraction and skip connections, which is suitable for the model to learn the multi-scale cloud morphology and texture of W. At the same time, the number of parameters is controllable, which makes it easy to train stably on the public Landsat 8 / 9 samples.

[0150] To demonstrate the controllable generation capability "oriented towards statistical priors," the condition variable c is used as an optional design item to distinguish the distribution of thin cloud morphology in different statistical clusters; this condition variable should come from the reproducible statistics or metadata classification in step S100, such as thin cloud coverage classification (by...). or The data can be categorized by statistical data, morphological type (strips / patches, which can be coarsely classified by anisotropic indices), or scene category (surface type). In implementation, the condition variable c can be one-hot or continuous scalar input, injected into the network through conditional embedding and temporal embedding, thereby generating morphological field samples that satisfy the specified statistical clusters during the sampling phase. If the condition variable is not used, the condition variable c is omitted, and the model learns the overall mixture distribution.

[0151] Sampling phase from pure noise Starting from this point, we obtain the result through inverse denoising iteration. The sample approximation; taking the DDPM denoising diffusion probability model sampling as an example, its reverse update can be written as:

[0152] (10)

[0153] in, The sampling noise figure;

[0154] DDIM deterministic sampling can also be used to accelerate generation and improve morphological stability; after sampling, Mapping back to [0,1] yields And crop it to the legal range:

[0155] (11)

[0156] Final output The generated thin cloud morphological field samples can be directly used for subsequent statistical consistency verification and closed-loop resampling, or they can be used as inputs for thin cloud influence conditions into degradation simulation or algorithm verification processes.

[0157] It is important to emphasize that the diffusion model plays the role of "learning and generating thin cloud morphological field distribution" in this invention, but does not guarantee that the generated samples necessarily satisfy all statistical prior constraints. Therefore, the key innovation of this method lies in the subsequent steps: performing a statistical consistency test on the generated samples, and using a feedback resampling mechanism to make the generated results statistically aligned with the prior of the real thin cloud samples of Landsat 8 / 9, thereby ensuring the controllability and verifiability of the generation process.

[0158] S300. A statistical consistency verification mechanism is introduced during the generation stage. For each generated sample, its marginal distribution, power spectrum structure, and connected component statistical indicators are calculated and compared with the statistical prior of the real thin cloud sample. If the statistical characteristics of the generated sample exceed the allowable range of the real sample, it is judged as an unqualified sample and resampling or condition adjustment is triggered. This closed-loop mechanism ensures that the final output sample is statistically consistent with the real thin cloud morphology of Landsat 8 / 9.

[0159] Specifically, including,

[0160] The thin cloud morphology field generated by the diffusion model can visually present a multi-scale structure, but the generation model alone cannot guarantee that every sample is consistent with the statistical characteristics of the real thin cloud samples in Landsat 8 / 9. To make the generation results controllable and verifiable, this invention introduces a statistical consistency check and closed-loop resampling mechanism in the sampling stage: for each generated sample, a set of statistical indicators isomorphic to the "sample statistical prior" in step S100 is calculated, and the deviation between these indicators and the allowable interval or reference distribution of the real sample prior is measured; when the deviation exceeds the threshold, the sample is rejected and resampling or condition adjustment is triggered, thereby upgrading the generation process from "unconstrained sampling" to "generation constrained by statistical prior".

[0161] S310. Establish a statistical prior reference representation on the real sample set; let the thin cloud binary mask obtained from the Landsat 8 / 9 thin cloud sample be:

[0162] (12)

[0163] Based on this, a continuous thin cloud morphological field is constructed. The statistical prior set S is extracted from all training samples. To facilitate consistency verification, the prior is stored in two forms: "distribution / interval". On the one hand, empirical distributions (e.g., histograms or cumulative distribution functions) are stored for statistics on marginal distribution and connected component scale. On the other hand, confidence intervals (e.g., for power spectrum slope, correlation length, and anisotropy statistics) are stored on the training set. (quantile intervals), thereby transforming "statistical prior" into an executable discrimination criterion;

[0164] S320. For each thin cloud morphological field sample generated by the diffusion model (or patch form) ), statistical consistency verification first calculates three types of statistical indicators:

[0165] One is the marginal distribution index, used to measure... Whether the intensity distribution is consistent with the real sample; specifically, whether the intensity distribution is consistent with the real sample. Divide the [0,1] region into B bins (intervals) to obtain a normalized histogram. Correspondingly, the reference histogram of the real sample is denoted as... (This can be either the average histogram of the training set or a clustered histogram); then the marginal distribution bias... It can be defined as:

[0166] (13)

[0167] Alternatively, Earth Mover's Distance (EMD) can be used when more robustness is required;

[0168] The second is the spatial structure index, used to measure whether multi-scale texture structures match; for Perform mean removal and calculate the two-dimensional power spectrum radial averaging yields Fit the slope of the logarithmic spectrum over a specified frequency band. and the reference slope interval The deviation is defined as follows:

[0169] (14)

[0170] Thirdly, there are morphological / topological indicators, used to constrain the connectivity and scale distribution of cloud morphology; specifically, in... A binary morphological set is obtained by selecting a fixed threshold τ (or an adaptive threshold based on quantiles). and to Perform connected component analysis to obtain the set of connected component areas. and its empirical distribution Correspondingly, the real sample obtains a reference distribution. ; morphological deviations can be represented by distribution distance definition:

[0171] (15)

[0172] Furthermore, porosity and elongation indices can be introduced as supplementary constraints to suppress non-realistic forms such as "strip noise pseudo-clouds" or "fragmented pseudo-clouds".

[0173] S330. After obtaining the above deviation measures, combine them into an overall consistency scoring function. To decide whether to accept the generated sample; a reproducible combination is a weighted sum:

[0174] (16) Among them, Weights are used to balance the dimensions and importance of different indicators;

[0175] If an overall threshold δ is set (which can be determined through cross-validation on the training samples), then the consistency decision is: when Accepting samples Otherwise, the sample set is rejected and resampling is triggered. This rejection sampling mechanism ensures that the output sample set is not only visually reasonable, but also statistically consistent with the prior thin cloud samples of Landsat 8 / 9.

[0176] S340. Closed-loop resampling strategies are divided into two categories based on whether a conditional variable c is introduced during diffusion generation. If unconditional diffusion is used, samples are rejected and then directly resampled (by changing the random noise seed or using different sampling step sizes / randomness parameters) until the condition is met. Or reach the maximum number of attempts N max This strategy is simple to implement but may increase sampling costs. If conditional diffusion is used, after rejecting samples, priority should be given to adjusting the condition variable c (e.g., cloud cover grading, anisotropy grading) or selecting a more suitable prior cluster reference distribution. This makes it easier for samples to fall into the target statistical cluster, thereby reducing the number of resampling times. Regardless of the strategy used, the closed-loop mechanism ensures that the final output thin cloud morphological field sample set meets the requirements of "generable, controllable, and verifiable", and provides statistically consistent thin cloud influence conditions as input for subsequent degradation modeling and algorithm verification.

[0177] It should be noted that this statistical consistency check does not claim to restore the true value of cloud microphysics, but rather ensures that the generated thin cloud morphology field is statistically consistent with the real thin cloud samples of Landsat 8 / 9 at the observation space level. This avoids the unrigorous problem of "judging the generation quality solely based on subjective visual perception" in academic research and provides the generation process with objective criteria for repeatable experiments.

[0178] The S400 method generates a continuous thin cloud morphological field that can be directly input into the downstream model as a continuous influence field, or it can be converted into a binary thin cloud mask through a unified threshold strategy for comparison and verification with traditional cloud detection methods. The whole method combines "learning generation" with "statistical control" to maintain verifiability while ensuring diversity.

[0179] Specifically, including,

[0180] The object of generation in this invention is a continuous thin cloud morphology field. The semantics of this expression refer to the equivalent influence intensity of thin clouds in the observation space, rather than the true value of cloud microphysics. To support different downstream tasks (such as degradation simulation, cloud impact stratification assessment, or comparison with traditional cloud masking methods), this invention provides a unified output expression and establishes a verification protocol for reproducible comparison with Landsat 8 / 9 real samples. Real thin cloud samples are still represented based on binary masks obtained from publicly available and reproducible thin cloud annotations or products, i.e.:

[0181] (17)

[0182] This mask is used for sample construction and statistical reference, and is not interpreted as cloud microphysical truth.

[0183] In terms of output representation, continuous morphological fields As the main output, it can be directly used to describe the spatial continuity and boundary transition of thin cloud effects; in order to align and compare with traditional cloud detection / segmentation methods (which typically output binary masks), and to facilitate topological and area statistics, this invention will... Further conversion to a binary thin cloud mask The binarization strategy must be consistent with the statistical test in step S330 and be used uniformly in both the real and generated samples; specifically, a fixed threshold τ can be used.

[0184] (18)

[0185] Where τ can be 0.5; or a quantile threshold can be used to match different cloud coverage levels, for example, let τ be... The qth quantile, thus The coverage rate is consistent with the target statistical cluster; regardless of the strategy used, the threshold selection rules must be clearly given in the experiment to ensure reproducibility and comparability;

[0186] Verifiability analysis mainly includes two layers:

[0187] Statistical consistency assessment and comparative verification protocol; statistical consistency assessment is used to quantify whether the generated sample falls within the statistically permissible range of the real sample, and must simultaneously cover three types of indicators: marginal distribution, spatial structure, and topological morphology; marginal distribution consistency is verified through comparison. The intensity histogram or cumulative distribution function is used to generate the distribution distance of the generated samples (such as L1 or EMD), which should fall within the fluctuation range of the real samples; spatial structure consistency is achieved through... The statistical implementation of the power spectrum requires calculating the radial spectrum P(k) after removing the mean and applying a window / block average, then fitting the logarithmic spectrum slope α or estimating the correlation length, and aligning it with the quantile intervals of the real sample; topological consistency is based on a binary mask. The connected component statistics are implemented, such as the connected component area distribution p(A), porosity, boundary roughness, etc., and reported in the form of distribution distance or interval deviation; to avoid uncertainty introduced by threshold selection, the calculation of topological indicators should always use τ or quantile rules consistent with the binarization strategy;

[0188] To make the claim of "statistical prior for Landsat 8 / 9 samples" testable, this invention employs a rigorous comparative verification scheme: the real sample set is divided into a statistical prior extraction set and an independent validation set. The prior extraction set is used to estimate the reference distribution and threshold intervals (e.g., quantile intervals). The validation set is only used to evaluate the consistency between generated samples and real samples in terms of statistical metrics, and does not participate in model training or threshold selection. For each statistical metric, the distance statistics (mean, standard deviation, and high quantile) between the generated sample distribution and the validation set distribution are reported, and the "pass rate" is given as a measure of the effectiveness of closed-loop resampling, i.e., given a maximum number of sampling times N. max Below, the proportion of generated samples that meet the statistical consistency criterion; furthermore, to ensure that the results are not only numerically verifiable but also morphologically interpretable, it is recommended to simultaneously display several typical generated samples. Visualization, corresponding binary mask The power spectrum and connected component statistics curves are compared with real verification samples to form a complete and verifiable closed loop of "numerical indicators + morphological examples + protocol verification".

[0189] Through the above output expression and verifiability analysis, this invention ensures that the generated results can serve subsequent degradation simulation and algorithm verification in the form of a continuous morphological field, and can also align with the traditional cloud processing process at the binary mask level. At the same time, it rigorously proves with multi-level statistical indicators and independent verification protocols that the generated samples statistically conform to the Landsat 8 / 9 real thin cloud sample prior, thereby avoiding the unrigorous conclusion of judging the generation quality solely by subjective visual judgment.

[0190] It is important to emphasize that this method only requires learning parameters during the diffusion model training phase; the rest of the process involves deterministic calculations and statistical tests. During the testing or application phase, no annotation information is required; only random noise and conditional variables need to be input to generate thin cloud morphological field samples that conform to statistical priors.

[0191] Specific Implementation Scheme 2: The present invention provides a thin cloud morphology field generation system based on statistical priors of Landsat 8 / 9 satellite samples. This system has program modules corresponding to the above steps, and executes the steps in the above method for generating thin cloud morphology fields based on statistical priors of Landsat 8 / 9 satellite samples during runtime.

[0192] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0193] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to, when called by a processor, implement the steps of a method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples.

[0194] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0195] The core improvement of this invention lies not in simply using a diffusion model to generate images, but in explicitly introducing statistical priors into the thin cloud morphology field generation process and constructing a closed-loop control mechanism through consistency verification. This includes the following technical points:

[0196] The continuous thin cloud morphology field construction mechanism of this invention represents thin clouds as a continuous morphology field. Rather than a simple binary mask, this continuous field is constructed by monotonically mapping or transforming a publicly available thin cloud sample mask, giving it spatially gradual characteristics and statistical modeling capabilities. This continuous representation provides a unified foundation for subsequent distribution learning and statistical consistency verification.

[0197] Based on the statistical prior extraction method of Landsat 8 / 9 samples, this invention proposes to extract multidimensional statistical priors from real thin cloud samples, including marginal distribution features, power spectrum slope and spatial correlation length, connected domain area distribution, and topological parameters, and formalizes these statistical features into a computable set of constraints. This statistical modeling mechanism is the core of this invention, distinguishing it from general generative models.

[0198] The generative framework combining diffusion model and statistical prior is not simply about using diffusion model to generate thin cloud morphological fields. Instead, it uses diffusion model as a probability distribution learner and introduces statistical consistency constraints in the generation stage, so that the generated samples are not only visually similar, but also statistically consistent with the Landsat thin cloud sample distribution.

[0199] This invention proposes a statistical consistency check and closed-loop resampling mechanism. It calculates statistical indicators (distribution distance, spectral structure deviation, topological differences, etc.) for the generated samples, and triggers resampling or condition adjustment when the deviation exceeds a preset range. This closed-loop control structure ensures the verifiability and controllability of the generated results, and is one of the key innovations of this invention.

[0200] Controllable condition variable design (optional extension): By introducing condition variables (such as thin cloud coverage grading, morphological type grading, or scene type), controllable generation of distribution is achieved, thereby enhancing the generalization ability of the generated model under different statistical clusters.

[0201] Example

[0202] We selected publicly available data containing thin cloud scenes from the Landsat 8 Collection 2 Level-1 product as our sample source. First, thin cloud regions were extracted using the official quality layer or publicly available thin cloud detection algorithms to obtain a binary thin cloud mask. To ensure sample quality, only scenes with high cloud confidence and excluding areas severely obscured by thick clouds were selected. For each scene, a continuous thin cloud morphological field representation was constructed. Specifically, a distance transformation was applied to the binary thin cloud mask, calculating the distance from pixels inside the thin cloud to the nearest non-thin cloud boundary. This distance was then converted into a continuous intensity value within the [0,1] interval using a monotonic mapping function, resulting in high values ​​inside the thin cloud and a smooth transition at the boundary. The continuous morphological field constructed in this way preserves the spatial structure of the thin cloud while possessing statistically modelable continuity.

[0203] Subsequently, statistical analysis was performed on all sample scenarios to extract the statistical prior of the thin cloud morphological field. The statistical prior includes three levels: First, marginal distribution statistics, which statistically analyzes the probability distribution and quantile intervals of pixel values ​​in the continuous morphological field; second, multi-scale spatial structure statistics, which calculates the radial average power spectrum by performing a two-dimensional Fourier transform on the morphological field and fitting its slope in logarithmic coordinates to characterize the spatial scale attenuation features of the cloud field; third, morphological topology statistics, which performs connected component analysis on the continuous morphological field after binarization at a fixed threshold, statistically analyzing indicators such as connected component area distribution, hole ratio, and boundary complexity. These statistical results constitute the statistical reference interval for real samples, used for compliance judgment of subsequently generated samples.

[0204] In the generation phase, a diffusion model is used to learn the distribution of continuous thin cloud morphological fields. During the training phase, real samples are cropped into fixed-size image patches and standardized. The diffusion model fits the real sample distribution by progressively adding Gaussian noise to the real morphological field and training the neural network to learn the denoising process. After training, the model can progressively denoise from random noise to generate new thin cloud morphological field samples.

[0205] After generating samples, they are not directly used as the final output, but instead enter the statistical consistency verification stage. For each generated sample, its marginal distribution, power spectrum slope, and connected region area distribution are calculated and compared with the statistical intervals of the real samples. If the generated sample deviates from a preset threshold in any core statistical indicator, the sample is determined to be a statistically inconsistent sample and is not accepted, and resampling is performed; if all statistical indicators fall within the statistical intervals of the real samples, the sample is accepted as a valid generation result. Through this closed-loop resampling mechanism, it is ensured that the final output sample is statistically consistent with the real thin cloud samples of Landsat 8 / 9.

[0206] In experimental verification, the method of this invention was compared with diffusion generation methods without statistical constraints. Experimental results show that, under the same number of generation attempts, the sample generated by the method of this invention significantly reduces the bias in marginal distribution, power spectrum structure, and connected domain scale distribution. The generated results not only visually present the multi-scale structure of realistic thin clouds, but also maintain a high degree of consistency with the real samples in a statistical sense.

[0207] This embodiment illustrates that the present invention achieves controllable generation of thin cloud morphological fields through a complete process of "statistical modeling - diffusion generation - consistency verification - closed-loop control", which can provide a reliable data foundation for research tasks such as degradation simulation, algorithm robustness testing and scene construction.

[0208] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples, characterized in that, Includes the following steps: S100. Obtain thin cloud sample areas from Landsat 8 / 9 public data; The samples are sourced from reproducible thin cloud detection or official quality-labeled products, thus obtaining a thin cloud binary mask. On this basis, a continuous thin cloud morphological field is constructed through distance transformation or continuous mapping, so that the interior and boundary of the thin cloud present a smooth transition structure. Subsequently, statistical analysis is performed on the samples to extract marginal distribution, multi-scale spatial structure and morphological topological features, forming a statistical prior set of thin clouds. S200. After the statistical prior extraction is completed, the diffusion model is trained to learn the distribution of the thin cloud morphological field. The diffusion model uses continuous morphological field patches as training samples and learns the data distribution by gradually adding and removing noise. After training, the model generates new thin cloud morphological field samples from random noise. S300. A statistical consistency verification mechanism is introduced during the generation stage. For each generated sample, its marginal distribution, power spectrum structure, and connected component statistical indicators are calculated and compared with the statistical prior of the real thin cloud sample. If the statistical characteristics of the generated sample exceed the allowable range of the real sample, it is judged as an unqualified sample and resampling or condition adjustment is triggered. This closed-loop mechanism is used to ensure that the final output sample is consistent with the real thin cloud morphology of Landsat 8 / 9 in a statistical sense.

2. The method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples according to claim 1, characterized in that: In step S100, the following are included: S110, Thin Cloud Sample Construction and Statistical Prior Extraction Let the nth Landsat 8 / 9 scene multispectral image be... ,in Represents pixel coordinates, The image domain is represented by k, and the spectral index is represented by k. The thin cloud binary mask is: （1） in, This indicates that pixel x is marked as a thin cloud region. Indicates non-thin cloud regions; Construction of S120 and the continuous thin cloud morphological field W(x) Introducing a continuous thin cloud morphology field , used to indicate the intensity or extent of the influence of thin clouds in the observation space; A continuous processing method for thin cloud binary masks is applied based on distance transformation; assuming... Let x be the Euclidean distance to the nearest non-thin cloud pixel, and define the continuous thin cloud morphology field as: （2） in, exp represents the scale parameter used to control the boundary transition width; Obtained through this method The value approaches 1 within the thin cloud region and continuously decreases to 0 at the boundary, thus forming a continuous thin cloud morphological expression. S130. Definition and extraction of statistical priors After obtaining a set of continuous thin cloud morphological fields Then, statistical priors are extracted, including marginal distribution priors, multi-scale spatial structure priors, and morphological and topological priors.

3. The method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples according to claim 2, characterized in that: In step S120, the construction method of the continuous field must satisfy: (1) It maintains a high value inside thin clouds and a low value in non-thin cloud areas; (2) A smooth transition is presented near the boundary to depict the gradual characteristics of thin clouds.

4. The method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples according to claim 3, characterized in that: In step S130, the statistical prior design satisfies: (1) It can be used as a computable constraint during the generation phase; (2) It can objectively reflect the statistical characteristics of the thin cloud morphology of Landsat 8 / 9.

5. The method for generating thin cloud morphological fields based on statistical priors for Landsat 8 / 9 satellite samples according to claim 3, characterized in that: The statistical priors include, S131, Marginal distribution prior, for all samples Statistically analyze its marginal distribution and estimate its empirical probability density function. ; S132, Multi-scale spatial structure priors, calculation The power spectral density; Let its two-dimensional Fourier transform be Then the power spectrum for: （3） right Averaging in the radial frequency domain yields a one-dimensional spectrum. The slope of the spectrum in logarithmic coordinates is calculated to characterize the multi-scale attenuation characteristics; the spectral slope and related length parameters constitute the spatial structure prior. S133, Morphological and Topological Priors, for or Perform connected component analysis on the high-value regions, statistically analyze the area distribution A of the connected components, and determine the perimeter-area relationship; let the area of ​​the i-th connected component be A. i The perimeter is L i Calculate its power-law fitting relationship: （4） The exponent β or its distribution range is used to characterize boundary roughness and morphological complexity. These statistics together constitute the topological prior. S134. Formal expression of the statistical prior set: Combining the above statistics, the statistical prior of the thin cloud sample is expressed as follows: （5） in, This represents the power spectrum slope or related length parameter. Indicates the pattern index, This represents the range of area distribution of the connected domain.

6. The method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples according to claim 5, characterized in that: Step S200 includes, After obtaining thin cloud samples of Landsat 8 / 9 and their continuous thin cloud morphological field W(x), a diffusion generation model is used to learn the data distribution of W and generate new thin cloud morphological field samples during the sampling stage. The training data consists of continuous morphological field samples obtained in step S100, and a continuous morphological field is constructed based on these samples. and from A fixed-size patch is cropped from the middle and used as a training sample, denoted as . To adapt to the commonly used numerical range of diffusion models, the following will be implemented: A linear mapping to [-1, 1] is denoted as ; The training process uses the standard forward diffusion process, Gradual noise addition And learn a denoising network to recover noise or recover clean samples; The forward diffusion process is defined as a Markov chain. Its form is a Gaussian perturbation: (6) Among them, For the number of diffusion steps, For noise schedule; Equivalently, sampled in closed form: （7） in, ,in, Represents the symbol for a product. Represents a normal distribution. The noise scheduling parameters represent the noise distribution process; the denoising network is denoted as... The input is a noise sample. The time step t and optional condition variable c are used to output the response to noise. The estimate; When using a noisy prediction paradigm, the training objective is to minimize the mean square error. : （8） in, Expressing expectations; Sampling phase from pure noise Starting from this point, we obtain the result through inverse denoising iteration. The sample approximation; based on the DDPM denoising diffusion probability model sampling, its reverse update is: （9） in, The sampling noise figure; After sampling is completed Mapping back to [0,1] yields And crop it to the legal range: （10） Final output This is the generated thin cloud morphology field sample.

7. A method for generating thin cloud morphological fields based on statistical priors for Landsat 8 / 9 satellite samples, as described in claim 6, is characterized in that: Step S300 includes, S310. Establish a statistical prior reference representation on the real sample set; suppose a thin cloud binary mask is obtained from the Landsat 8 / 9 thin cloud samples, and construct a continuous thin cloud morphology field based on this mask. And extract the statistical prior set S from all training samples; S320. For each thin cloud morphological field sample generated by the diffusion model Statistical consistency verification is achieved by calculating three types of statistical indicators; include ， S321, Marginal Distribution Index, used to measure Does the intensity distribution match that of the real sample? Will Divide the [0,1] region into B bins to obtain a normalized histogram. The reference histogram for the real sample is denoted as... Then the marginal distribution deviation Defined as: （11） S322, Spatial Structure Index, is used to measure whether multi-scale texture structures match; right Perform mean removal and calculate the two-dimensional power spectrum radial averaging yields Fit the slope of the logarithmic spectrum over a specified frequency band. and the reference slope interval Comparison, defining deviation for: （12） S323, Morphology / Topology Indicators, used to constrain the connectivity and scale distribution of cloud morphology; exist A binary morphology set is obtained by selecting a fixed threshold τ. and to Perform connected component analysis to obtain the set of connected component areas. and its empirical distribution The real sample yields the reference distribution. ; morphological deviation is measured by distribution distance definition: （13） S330. After obtaining the above deviation measures, combine them into an overall consistency scoring function. To decide whether to accept the generated sample: (14) Among them, Used to balance the dimensions and importance of different indicators; If a total threshold δ is set, then the consistency decision is: when Accepting samples Otherwise, reject the request and trigger resampling; S340. Closed-loop resampling strategies are divided into two categories based on whether a condition variable c is introduced during diffusion generation. If unconditional diffusion is used, samples are rejected and resampling is performed directly until the condition is met. Or reach the maximum number of attempts N max If conditional diffusion is used, after rejecting samples, the condition variable c should be adjusted first, or a more suitable prior cluster reference distribution should be selected. .

8. A thin cloud morphology field generation system based on statistical priors for Landsat 8 / 9 satellite samples, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-7 above, and executes the steps in the above-described method for generating thin cloud morphological fields based on statistical priors of Landsat 8 / 9 satellite samples when it is run.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the thin cloud morphological field generation method according to any one of claims 1-7, which is based on statistical priors for Landsat 8 / 9 satellite samples.

Images