A Physical-Guided Deep Learning Method for Cloud Detection in Remote Sensing Data

Abstract

The present invention discloses a physical-guided deep learning remote sensing data cloud detection method, belonging to the technical field of data processing, including: Step 1, the preprocessing step of satellite remote sensing observation data and background field data; Step 2, the cloud probability calculation step based on the physical process of radiative transfer; Step 3, the cloud detection step of the physical-guided image segmentation deep learning large model. The physical-guided satellite remote sensing data cloud detection method of the present invention adopts radiative transfer physical simulation and uses Bayesian maximum a posteriori optimization to estimate the probability of cloud occurrence. On this basis, the constraint information is introduced into the prompt encoder of the SAM large model in the form of point prompts and box prompts, enabling the deep learning image segmentation large model SAM to possess physical prior knowledge, thereby achieving more accurate cloud detection and being able to flexibly adapt to different atmospheric and ocean conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and specifically relates to a physical-guided deep learning remote sensing data cloud detection method. Background Technique

[0002] Clouds will absorb and scatter the infrared radiation of the sun, interfering with the remote sensing inversion of surface parameters. Therefore, cloud detection of satellite remote sensing data is crucial for the analysis and application of satellite remote sensing data. Since clouds usually exhibit spectral characteristics of high reflectivity and low brightness temperature. Therefore, the observed pixel values are generally classified as clouds or non-clouds by setting fixed thresholds. However, due to the influence of environmental variables such as surface types and regional ranges, the threshold method cannot adapt to different environments and dynamic changes, resulting in a high false detection rate.

[0003] Therefore, recent mainstream research mostly uses deep neural networks to extract features in different scenarios, and uses functions to map the feature vectors into dynamic thresholds for cloud determination. The latest computer vision research shows that the feature extraction ability of neural networks mainly depends on the size of the learnable parameters in the network. When the network parameters reach hundreds of millions, its ability to extract and express features will be greatly improved. Such networks are called "large models" because the number of parameters is much larger than that of conventional deep networks.

[0004] The cloud detection task belongs to image segmentation in computer vision, and the corresponding large model is usually called "Segment Anything Model (SAM)". However, although the SAM large model has strong image information extraction ability, it cannot be directly used for satellite remote sensing cloud detection tasks. This is because when the model is reasoning, it lacks the guidance of physical prior knowledge and can only distinguish clouds through the color and texture of images, resulting in the model being difficult to accurately detect clouds under complex weather conditions. However, the physical process of remote sensing, that is, atmospheric radiation transmission, can provide additional prior knowledge for the reasoning of the SAM model. Summary of the Invention

[0005] Aiming at the problems of limited model representation ability and lack of physical prior knowledge in the current deep network cloud detection method, the present invention proposes a physical-guided deep learning remote sensing data cloud detection method, which can solve the above problems.

[0006] To solve the above technical problems, the present invention adopts the following technical solutions to achieve:

[0007] A physical-guided deep learning remote sensing data cloud detection method, including:

[0008] Step 1, satellite observation data and background field data Preprocessing steps, including:

[0009] Obtain the reflectance of the visible light channel in the satellite observation data;

[0010] Obtain the brightness temperature of the thermal infrared channel in the satellite observation data, which is the observed brightness temperature ;

[0011] Background field data Preprocessing is used to process the spatial resolution of the background field data to be consistent with the resolution of the satellite observation data.

[0012] Step 2, the cloud probability calculation step based on the physical process of radiative transfer, includes:

[0013] Given the satellite observation data and the background field data calculate the clear sky probability according to Bayesian theory :

[0014] where, is the probability that the satellite observation value appears under the given background field data and assuming it is in a clear sky state, is the probability that the satellite observation value appears under the given background field data and assuming it is in a cloud state, is the prior clear sky probability,

[0015] The calculation method of

[0016] ;

[0017] where, is the probability density function of the spectral part in the clear sky state, is the probability density function of the texture part in the clear sky state;

[0018] The calculation method of

[0019] ;

[0020] is the probability density function of the spectral part in the cloud state, is the probability density function of the texture part in the cloud state, and are both obtained by using a look-up table based on empirical statistics or numerical simulation;

[0021] The calculation method of

[0022] First, determine the prior cloud probability through the cloud cover ratio TCC in the numerical weather prediction data. , if TCC is between 0.5 and 0.95, then assign TCC to ; if TCC is less than 0.5, then ; if TCC is greater than 0.95, , calculate to obtain ;

[0023] Calculate the cloud probability under the given satellite observation data and background field data as = 1 - .

[0024] Step 3, the cloud detection step of the physics-guided image segmentation model, includes the following steps.

[0025] (31) Load the pre-trained SAM image segmentation model;

[0026] (32) Input the preprocessed reflectivity TR into the SAM image segmentation model to generate the feature representation of the image;

[0027] (33) Take as the point prompt and input it into the prompt encoder of the SAM image segmentation model. The SAM image segmentation model outputs the predicted cloud probability of each pixel, which is the first predicted cloud probability;

[0028] (34) Find out the pixel points in whose cloud probability is not less than the set threshold to form a point set, and generate all possible squares composed of four pixel points from the point set;

[0029] (35) Establish an integer linear programming model with the goal of maximizing the sum of the vertex cloud probabilities of the selected squares;

[0030] (36) Use an integer linear programming solver or a heuristic algorithm to obtain the optimal or approximately optimal square combination;

[0031] (37) Take the square combination as the box prompt and input it into the prompt encoder of the SAM image segmentation model. The SAM image segmentation model outputs the predicted cloud probability of each pixel, which is the second predicted cloud probability;

[0032] (38) Calculate the average value of the first predicted cloud probability and the second predicted cloud probability, compare the average value with the set threshold, and determine whether the corresponding pixel points are cloud pixels or non-cloud pixels to complete the cloud detection.

[0033] In some embodiments, the conversion method of the reflectivity TR in step 1 is:

[0034] ;

[0035] Among them, ER is the Earth-view reflectance, and SZA is the solar zenith angle.

[0036] In some embodiments, in step one, the data in the numerical weather prediction is used as the background field data, and the background field data The preprocessing steps include:

[0037] (11) Identify the satellite observation data pixels closest in distance according to the latitude and longitude information of the numerical weather prediction data;

[0038] (12) Assign the satellite zenith angle of the satellite observation data pixels to the corresponding numerical weather prediction data. The numerical weather prediction data with the satellite zenith angle assigned is the preprocessed background field data .

[0039] In some embodiments, the calculation method of the spectral part probability density function under clear sky conditions is as follows:

[0040] ;

[0041] Among them, is the observed brightness temperature and the simulated brightness temperature vector difference, that is is the Jacobian matrix of the simulated brightness temperature of the atmospheric radiative transfer; is the covariance matrix of the background field data; is the sum of the covariance of the simulated brightness temperature and the covariance of the observed brightness temperature under the assumption of ignoring the background field data error; The acquisition method of is: input the atmospheric profile parameters and sea surface information into the atmospheric radiative transfer model MODTRAN, and the brightness temperature at the top of the atmosphere output is the simulated brightness temperature

[0042] In some embodiments, the texture part probability density function under clear sky conditions is obtained through the historical experience look-up table.

[0043] In some embodiments, the method for generating a square in step (34) includes:

[0044] For each pair of pixel points in the point set , calculate the positions of the other two vertices that can form a square;

[0045] Determine whether there are pixel points of the other two vertices in the point set. If there are pixel points, a square is generated by these four pixel points, otherwise a square cannot be generated.

[0046] In some embodiments, the objective function of the integer linear programming model in step (35) is:

[0047] ;

[0048] where is the set of squares, is the set of vertices of the square , is the existence probability of the vertex ;

[0049] The constraint conditions of the integer linear programming model include: each pixel can be used to construct at most one square.

[0050] In some embodiments, the visible light channel includes visible light in the bands of 0.436 μm - 0.454 μm, 0.545 μm - 0.565 μm, and 0.662 μm - 0.682 μm.

[0051] In some embodiments, the thermal infrared channel includes the bands of 10.763 μm and 12.013 μm.

[0052] In some embodiments, after step three, it further includes performing morphological operations or connected component analysis on the data with cloud detection completed to remove small noise regions and isolated pixels. For cloud pixel regions larger than a preset region threshold, the connected component analysis method is used to merge adjacent cloud pixel regions or separate misdetected non-cloud pixel regions; the processed cloud detection data is converted into a standard format, and a final cloud detection result map is generated in combination with the original remote sensing data image.

[0053] Compared with the prior art, the advantages and positive effects of the present invention are:

[0054] 1. Improve the accuracy of cloud detection: The physically guided process provides additional prior knowledge, and through the methods of point prompts and box prompts, the mask decoding results of the SAM large model can be effectively constrained in the physically feasible space, thereby improving the accuracy of cloud detection.

[0055] 2. Flexibly adapt to different weather scenarios: While the image encoder of the SAM large model captures texture features, physical prior features of radiative transfer calculation are also provided, and the two types of features can complement each other, thus flexibly adapting to different weather scenarios.

[0056] After reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become clearer. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a block diagram of the principle of an embodiment of the physically-guided deep learning satellite remote sensing cloud detection method proposed by the present invention;

[0058] Figure 2 It is a diagram of the SAM image segmentation model in an embodiment of the physically-guided deep learning satellite remote sensing cloud detection method proposed by the present invention. Detailed implementation manners

[0059] The following further describes in detail the specific implementation manners of the present invention with reference to the accompanying drawings.

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] It should be noted that in the description of the present invention, the terms indicating directions or positional relationships such as "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are based on the directions or positional relationships shown in the accompanying drawings. This is only for convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0062] In the present invention, unless otherwise clearly defined and limited, the terms "installed", "connected", "connected", "fixed", etc. shall be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0063] Embodiment 1. Refer to Figure 1 As shown, the present invention proposes a physically-guided deep learning satellite remote sensing cloud detection method, including:

[0064] Step 1, satellite observation data and background field data Preprocessing steps, the preprocessing steps include:

[0065] Since clouds usually exhibit spectral characteristics of high reflectivity and low brightness temperature in remote sensing data, it is necessary to obtain the reflectivity and brightness temperature data observed by satellites in this embodiment.

[0066] For the visible light channels (0.445μm, 0.555μm, and 0.673μm bands), the Earth-view reflectivity ER provided in the original data product is converted to the true reflectivity TR:

[0067] ;

[0068] where ER is the Earth-view reflectivity and SZA is the solar zenith angle.

[0069] For the thermal infrared channels (10.763μm and 12.013μm), considering that the brightness temperature has a larger dynamic range and can show more detailed features on the image, the radiance data provided in the original data product is converted to the brightness temperature according to the "radiance - brightness temperature" look-up table, denoted as the observed brightness temperature. .

[0070] Background field data Preprocessing is used to process the spatial resolution of the background field data to be consistent with the resolution of the remote sensing data.

[0071] Step 2, the cloud probability calculation step based on the physical process of radiative transfer, includes:

[0072] Given the remote sensing data and the background field data , according to Bayes' theory, calculate the clear-sky probability :

[0073] .

[0074] where is the conditional likelihood probability in the Bayesian framework, defined as the probability of the satellite observation value occurring given the background field data and assuming it is in a clear-sky state. is the probability of the satellite observation value occurring given the background field data and assuming it is in a cloud state. The prior clear-sky probability in Bayes' formula can be obtained by calculating the proportion of cloud pixels in the background field data. Through the Bayesian framework, combined with the physical radiative transfer process and supplemented by an empirical look-up table, the clear-sky probability can be obtained, and the probability of clouds is The calculation result will be used to guide the next large model cloud detection. Through the universality of physical laws, the large model is constrained to ensure that its results meet the requirements of cloud detection accuracy while guaranteeing generalization, and no additional model training is required.

[0075] The following will introduce respectively 、 and 's specific calculation methods.

[0076] (1) 's calculation method is:

[0077] .

[0078] Among them, is the probability density function of the spectral part in the clear sky state, is the probability density function of the texture part in the clear sky state.

[0079] (2) 's calculation method is:

[0080] .

[0081] is the probability density function of the spectral part in the cloud state, is the probability density function of the texture part in the cloud state, and are both obtained using a lookup table based on empirical statistics or numerical simulation.

[0082] (3) 's calculation method is:

[0083] First, determine the prior cloud probability through the cloud cover ratio TCC in the numerical weather prediction data. If the value of TCC is between 0.5 - 0.95, then assign the TCC value to ; if TCC is less than 0.5, then ; if TCC is greater than 0.95, further calculate to obtain .

[0084] Step three, the cloud detection steps of the physically guided image segmentation model, including:

[0085] (31), Download the SAM code and pre-trained model from and load the pre-trained SAM image segmentation model through the pytorch library.

[0086] (32) Input the preprocessed reflectance TR into the image encoding module of the SAM model to generate a feature representation of the image. All the image encoder codes used are provided by the SAM model.

[0087] (33) Input as a point prompt into the prompt encoder of the SAM image segmentation model. The SAM image segmentation model outputs the predicted cloud probability for each pixel, which is the first predicted cloud probability.

[0088] (34) Find the pixel points in whose cloud probability is not less than the set threshold to form a point set, and generate all possible squares composed of four pixel points from the point set.

[0089] (35) Establish an integer linear programming model with the goal of maximizing the sum of the vertex cloud probabilities of the selected squares.

[0090] (36) Use an integer linear programming solver or a heuristic algorithm to obtain an optimal or approximately optimal square combination.

[0091] (37) Input the square combination as a box prompt into the prompt encoder of the SAM image segmentation model. The SAM image segmentation model outputs the predicted cloud probability for each pixel, which is the second predicted cloud probability.

[0092] (38) Calculate the average of the first predicted cloud probability and the second predicted cloud probability, compare the average with the set threshold, and determine whether the corresponding pixel points are cloud pixels or non-cloud pixels to complete cloud detection.

[0093] In this embodiment, the SAM image segmentation model is used for cloud detection. Since the SAM image segmentation model has hundreds of millions of parameters, far exceeding conventional cloud detection neural networks, it has strong feature extraction capabilities, providing a prerequisite for high-precision cloud detection.

[0094] When implementing this solution, it is necessary to ensure that the working environment has an NVIDIA GPU and CUDA and cuDNN are correctly installed to support deep learning acceleration. Install Python 3.6 or a higher version, and use pip to install the necessary libraries. Visit the official PyTorch website, obtain the appropriate installation command according to the CUDA version, and install PyTorch and torchvision. Subsequently, install the Transformers library of Hugging Face. When verifying the installation, check the versions of PyTorch and Transformers, and the availability of CUDA is True. According to the requirements of the solution, other dependent libraries such as numpy and pandas may need to be installed. Finally, set the CUDA-related environment variables to ensure that the GPU can be used normally. These steps will ensure that the environment is ready to accelerate the model training and inference process.

[0095] Start with the pre-trained SAM (SegmentAnything Model) image segmentation model and load the base model parameters. The code and pre-trained model of the SAM model can be downloaded from and the pre-trained model is loaded through pytorch. Input the pre-processed image into the image encoder of the SAM model to generate the feature representation of the image. The structure of the SAM image segmentation model is as Figure 2 shown.

[0096] The cloud detection method of this embodiment uses a physically-guided image segmentation model for cloud detection. It adopts radiative transfer physical simulation, combines Bayes' theorem, physically constrains the probability of cloud appearance, and introduces the constraint information into the prompt encoder of the SAM large model in the form of point prompts and box prompts to generate a high-dimensional feature representation rich in physical information for subsequent pixel-by-pixel cloud detection. Due to the additional physical prior knowledge, the accuracy of cloud detection is improved by this method.

[0097] In some embodiments, the conversion of radiance to reflectance in step one is based on the principle of radiative transfer. Specifically, the conversion method of reflectance TR is as follows.

[0098] ;

[0099] where ER is the Earth-view reflectance and SZA is the solar zenith angle.

[0100] In some embodiments, in step one, the conversion of brightness temperature uses the radiance-brightness temperature conversion lookup table provided in the VIIRS data product to convert the radiance of the thermal infrared channel into brightness temperature.

[0101] In some embodiments, the data in Numerical Weather Prediction (NWP) in step one is used as the background field data. Since the spatial resolutions of satellite remote sensing data and NWP data are different, background field data preprocessing is required.

[0102] Background field data The preprocessing steps include:

[0103] (11) Identify the remotely sensed data pixel closest in distance according to the latitude and longitude information of the NWP data.

[0104] (12) Assign the satellite zenith angle of the remotely sensed data pixel to the corresponding NWP data. The NWP data with the satellite zenith angle assigned is the preprocessed background field data. .

[0105] In some embodiments, the probability density function of the spectral part under clear sky conditions can be derived from the Gaussian distribution assumption, and the calculation method is as follows:

[0106] .

[0107] Where is the vector difference between the observed brightness temperature and the simulated brightness temperature , that is , is the Jacobian matrix of the simulated brightness temperature of atmospheric radiative transfer, is the covariance matrix of the background field data, is the sum of the covariance of the simulated brightness temperature and the covariance of the observed brightness temperature under the assumption of ignoring the background field data error, The method for obtaining is: input the atmospheric profile parameters and sea surface information into the atmospheric radiative transfer model MODTRAN, and the brightness temperature at the top of the atmosphere output is the simulated brightness temperature.

[0108] The atmospheric profile parameters and sea surface information are read from the NWP data. The atmospheric profile parameters include air pressure, temperature, humidity, ozone content, and carbon dioxide content. The sea surface information includes sea skin temperature, wind speed at 10 m height, and humidity at 2 m height.

[0109] In some embodiments, the probability density function of the texture part under clear sky conditions is obtained through an empirical look-up table. In this look-up table, under different satellite zenith angles and different texture features can be obtained.

[0110] In some embodiments, the method for generating the square in step (34) includes:

[0111] For each pair of pixel points in the point set calculate the positions of the other two vertices that can form a square.

[0112] In this embodiment, data structures such as KD - tree or quadtree can be used to accelerate the query of neighboring points.

[0113] Determine whether there are pixel points in the point set for the other two vertices. If there are pixel points, the four pixel points form a square; otherwise, a square cannot be generated.

[0114] In some embodiments, the objective function of the integer linear programming model in step (35) is:

[0115] ;

[0116] where is the set of squares, is a square is the set of vertices of is a vertex is the probability of existence;

[0117] The constraint conditions of the integer linear programming model include: each pixel point is used to construct at most one square.

[0118] In some embodiments, the visible light channel includes visible light in the wavelength bands of 0.436μm - 0.454μm, 0.545μm - 0.565μm, and 0.662μm - 0.682μm.

[0119] In some embodiments, the thermal infrared channel includes the wavelength bands of 10.763μm and 12.013μm.

[0120] In some embodiments, after step three, it further includes performing morphological operations or connected - component analysis on the data after cloud detection to remove small noise regions and isolated pixels. For cloud pixel regions larger than a preset area threshold, the connected - component analysis method is used to merge adjacent cloud pixel regions or separate mis - detected non - cloud pixel regions.

[0121] Convert the processed cloud - detection data into a standard format and generate a final cloud - detection result map in combination with the original remote - sensing data image.

[0122] In some embodiments, it further includes evaluating the cloud - detection result.

[0123] Verify the model performance on an independent test set and evaluate its actual effect in the cloud - detection task. The calculated metrics include:

[0124] IoU (Intersection over Union): Measures the overlap between the predicted segmentation region and the ground truth segmentation region. The calculation formula is:

[0125] .

[0126] Where TP (True Positive) is the number of pixels predicted as cloud and actually being cloud. FP (False Positive) is the number of pixels predicted as cloud but actually not being cloud. FN (False Negative) is the number of pixels predicted not as cloud but actually being cloud.

[0127] Dice Coefficient: Measures the similarity between the predicted segmentation and the ground truth segmentation region.

[0128] .

[0129] Where the definitions of TP, FP, and FN are the same as above.

[0130] Pixel Accuracy: The proportion of correctly predicted pixels to the total number of pixels.

[0131] .

[0132] Where TN is the number of pixels predicted not as cloud and actually not being cloud.

[0133] Precision: The proportion of correctly predicted positive samples to the predicted positive samples.

[0134] .

[0135] Recall: The proportion of correctly predicted actual positive samples:

[0136] .

[0137] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those of ordinary skill in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A deep learning remote sensing data cloud detection method based on physical guidance, characterized in that: include: Step 1: Satellite observation data and background field data x b Preprocessing steps include: Obtain the reflectivity of the visible light channel in satellite observation data; Get the brightness temperature of the thermal infrared channel in the satellite observation data, which is the observed brightness temperature y o ; Background field data x b The preprocessing is used to convert the background field data x b The spatial resolution of the satellite observation data is processed to be consistent with the resolution of the satellite observation data; Step 2, the cloud probability calculation step based on the radiation transfer physical process, includes: Given the satellite observation data y o and background field data x b In the case of , according to Bayesian theory, calculate the clear sky probability P(c|y o , x b ): Among them, P(y o |x b , c) is the given background field data x b Assuming that the sky is clear, the probability of the satellite observation value appearing is: Given the background field data x b Assuming that it is in the cloud state, the probability of the satellite observation value appearing, P(c) is the prior clear sky probability, is the prior probability of cloudiness; P(y o |x b , c) is calculated as: in, is the probability density function of the spectrum under clear sky conditions, is the probability density function of the texture part under clear sky conditions; The calculation method is: is the probability density function of the spectrum under cloud conditions, is the probability density function of the texture part under the cloud state, and They are all obtained using lookup tables based on empirical statistics or numerical simulation; The calculation method of P(c) is: Determine the prior probability of cloud cover by using the cloud cover ratio TCC in numerical weather forecast data If TCC is between 0.5 and 0.95, assign TCC to If TCC is less than 0.5, then If TCC is greater than 0.95, Calculated Calculate the given satellite observation data and background field data x b The cloud probability under Step 3, the physical guided image segmentation model cloud detection step, includes: (31) Load the pre-trained SAM image segmentation model; (32), inputting the preprocessed reflectivity TR into the SAM image segmentation model to generate a feature representation of the image; (33) As a point cue input into the cue encoder of the SAM image segmentation model, the SAM image segmentation model outputs the predicted cloud probability of each pixel as the first predicted cloud probability; (34) Find out The pixels whose cloud probability is not less than the set threshold constitute a point set, and all possible squares consisting of four pixels are generated from the point set; (35) Establish an integer linear programming model with the goal of maximizing the sum of the vertex cloud probabilities of the selected square; (36) Using an integer linear programming solver or a heuristic algorithm to obtain an optimal or approximately optimal square combination; (37), inputting the square combination as a box prompt into the prompt encoder of the SAM image segmentation model, and the SAM image segmentation model outputs the predicted cloud probability of each pixel as the second predicted cloud probability; (38) calculating an average of the first predicted cloud probability and the second predicted cloud probability, comparing the average with a set threshold, determining whether the corresponding pixel is a cloud pixel or a non-cloud pixel, and completing cloud detection; The method for generating the square in step (34) includes: For each pair of pixels in the point set (P i ,P j ), calculate the positions of the other two vertices that can form a square; It is determined whether there are pixels in the point set at the other two vertices. If there are pixels, the four pixels generate a square. Otherwise, a square cannot be generated.

2. The method according to claim 1, characterized in that The conversion method of reflectivity TR in step 1 is: Among them, ER is the reflectivity of the earth's visual angle, and SZA is the solar zenith angle.

3. The method according to claim 1, characterized in that In step 1, the data in numerical weather forecast is used as background field data. The background field data x b The preprocessing steps include: (11) Identify the nearest satellite observation data pixel based on the latitude and longitude information of the numerical weather forecast data; (12) Assign the satellite zenith angle of the satellite observation data pixel to the corresponding numerical weather forecast data. The numerical weather forecast data with the satellite zenith angle assignment is the preprocessed background field data x b .

4. The method according to claim 1, characterized in that: Probability density function of the spectrum under clear sky conditions The calculation method is: Where Δy is the observed brightness temperature y o and simulated brightness temperature y b The vector difference, that is, Δy = y o -y b ; H is the Jacobian matrix of the brightness temperature simulated by atmospheric radiation transfer; B is the covariance matrix of the background field data; R is the sum of the covariance of the simulated brightness temperature and the covariance of the observed brightness temperature under the assumption that the background field data error is ignored; y b The method for obtaining is: input the atmospheric profile parameters and sea surface information into the atmospheric radiation transfer model MODTRAN, and the output of the top of the atmosphere brightness temperature is the simulated brightness temperature y b .

5. The method according to claim 1, characterized in that Probability density function of texture part under clear sky condition Obtained through historical experience lookup table.

6. The method according to any one of claims 1 to 5, characterized in that: The objective function of the integer linear programming model in step (35) is: Where S is a set of squares, V s is the vertex set of square s, P p is the existence probability of vertex p; The constraint conditions of the integer linear programming model include: each pixel point is used to construct at most one square.

7. The method according to any one of claims 1 to 5, characterized in that: The visible light channel includes wavelength bands of 0.436 μm-0.454 μm, 0.545 μm-0.565 μm and 0.662 μm-0.682 μm.

8. The method according to any one of claims 1 to 5, characterized in that: The thermal infrared channel includes wavelength bands of 11 μm and 12 μm.

9. The method according to any one of claims 1 to 5, characterized in that: After step three, it also includes using morphological operations or connected domain analysis on the data after cloud detection to remove small noise areas and isolated pixels. For cloud pixel areas larger than a preset area threshold, the connected domain analysis method is used to merge adjacent cloud pixel areas, or to separate the misdetected non-cloud pixel areas; the processed cloud detection data is converted into a standard format, and combined with the original remote sensing data image to generate the final cloud detection result map.

Images