Thick cloud-thin cloud detection method and system based on multi-scale depth model

Abstract

The invention discloses a thick cloud-thin cloud integrated detection method and system based on spectral features and a multi-scale depth model, and belongs to the technical field of remote sensing image processing, and the method comprises the steps: 1, training a plurality of scene-level models of different scales based on a cloud detection data set, generating a plurality of scene-level cloud probability graphs and scene-level binary cloud masks of different scales; step 2, based on a plurality of scene-level cloud probability graphs of different scales and dark pixel features of the image, pixel-level cloud probability graphs facing thick cloud and thin cloud are generated respectively; and step 3, combining the scene-level binary cloud mask, fusing the pixel-level cloud probability graphs facing the thick cloud and the thin cloud by adopting dark pixel gradient, and then combining an adaptive threshold and distance weighting to generate a pixel-level binary cloud mask so as to realize thick cloud-thin cloud integrated detection. According to the method, manual segmentation threshold setting is avoided, and fusion of the thick cloud probability graph and the thin cloud probability graph and full-automatic accurate detection of the thick cloud, the thin cloud and the cloud edge are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and particularly relates to a thick cloud-thin cloud detection method and system based on a multi-scale deep model. Background Art

[0002] When an optical remote sensing satellite acquires images, it is often interfered by clouds, resulting in information loss problems of varying degrees in the images. The area covered by thick clouds will cause complete loss of information, while the area covered by thin clouds will cause radiation distortion of the images, affecting the usability of the images. Therefore, in the processing of optical remote sensing images, generating an accurate cloud mask through cloud detection can provide a reliable basis for subsequent image interpretation and analysis and product production, and enhance the application value of remote sensing images.

[0003] Currently, there are mainly two types of cloud detection methods: multi-temporal methods and single-temporal methods. The multi-temporal methods achieve cloud detection by analyzing pixel mutations in time series images. However, due to their dependence on multiple images of the same area at different time periods, their practical applications are greatly limited. The single-temporal methods mainly use the characteristics of clouds in a single image to achieve cloud detection, which can be further divided into physical rule-based methods and deep learning-based methods. Among them, the physical rule-based methods design rules and segmentation thresholds through the physical characteristics of clouds (such as high reflectivity and low temperature) to achieve cloud detection. However, in the actual application process, this method usually requires manual selection of rules and thresholds, and it is difficult to obtain optimal parameters, resulting in limited detection accuracy. In contrast, the deep learning-based methods use large-scale datasets to train image classification models, which can automatically learn the characteristics of clouds and achieve high-precision cloud detection, with stronger adaptability and accuracy.

[0004] The accuracy of existing deep learning-based cloud detection methods depends on large-scale and high-quality training samples. However, the existing training samples often lack thin cloud samples, resulting in most current cloud detection methods can only achieve thick cloud detection and are difficult to achieve accurate detection of thin clouds, greatly limiting the practical applications of such methods.

[0005] Therefore, it is necessary to design a thick cloud-thin cloud detection method and system based on a multi-scale deep model for the above problems. Summary of the Invention

[0006] The object of the present invention is to provide a thick cloud - thin cloud detection method based on a multi - scale deep model to address the problem that existing cloud detection methods are often limited by factors such as the complexity of thin cloud features and the quality of training samples, making it difficult to accurately detect thin clouds. The thick cloud - thin cloud integrated detection technology based on spectral features and a multi - scale deep model is based on a multi - scale scene - level deep model. By using the differential optical features of thick clouds and thin clouds, a pixel - level cloud probability map is generated, and an automatic and accurate detection of thick clouds, thin clouds, and cloud edges is achieved based on an adaptive threshold and an adaptive optimization strategy of distance weighting.

[0007] According to one aspect of the present specification, there is provided a thick cloud - thin cloud detection method based on a multi - scale deep model, including:

[0008] Step 1: Training multiple scene - level models with different scales based on a cloud detection dataset to generate multiple scene - level cloud probability maps and scene - level binary cloud masks with different scales;

[0009] Step 2: Generating pixel - level cloud probability maps for thick clouds and thin clouds respectively based on multiple scene - level cloud probability maps with different scales and the dark pixel features of the image to be cloud - detected;

[0010] Step 3: Combining the scene - level binary cloud masks, using the dark pixel gradient to fuse the pixel - level cloud probability maps for thick clouds and thin clouds, and then generating a pixel - level binary cloud mask by combining an adaptive threshold and distance weighting to achieve thick cloud - thin cloud integrated detection.

[0011] Further, in the above - mentioned Step 1, generating multiple scene - level cloud probability maps and scene - level binary cloud masks with different scales includes:

[0012] Using the trained scene - level model to perform block processing on the images in the cloud detection dataset, and determining whether each image block contains clouds;

[0013] If the image block contains clouds, marking all the pixels included in the image block as clouds to obtain a scene - level cloud probability map and a scene - level binary cloud mask.

[0014] Further, in the above - mentioned Step 2, generating pixel - level cloud probability maps for thick clouds and thin clouds includes:

[0015] Calculating the dark pixel features using the reflectance of the blue and green bands in the image to be cloud - detected;

[0016] Multiplying the multiple scene - level cloud probability maps with different scales by the dark pixel features pixel - by - pixel to obtain the pixel - level cloud probability map for thick clouds;

[0017] Performing singular value decomposition on the dark pixel features to obtain a low - rank approximation of the dark pixel features;

[0018] Multiply the scene-level cloud probability maps at multiple different scales pixel by pixel with the low-rank approximation of the dark pixel features to obtain the pixel-level cloud probability map of thin clouds.

[0019] Further, in step 3, combining the scene-level binary cloud mask, and using the dark pixel gradient to fuse the pixel-level cloud probability maps for thick and thin clouds, includes:

[0020] Calculate the gradient magnitude of the dark pixel features using the Sobel operator to segment the binary edge mask of the image to be cloud detected;

[0021] Take the intersection of the generated scene-level binary cloud masks at different scales as the cloud coverage area, and at the same time clarify the proportion of the binary edge mask in the cloud coverage area, and fuse the pixel-level cloud probability maps for thick and thin clouds to obtain the fused pixel-level cloud probability map.

[0022] Further, generating the pixel-level binary cloud mask in step 3 includes:

[0023] Segment the fused pixel-level cloud probability map through an adaptive threshold to generate an initial binary cloud mask;

[0024] Adopt distance-weighted adaptive optimization to enhance the fused pixel-level cloud probability map;

[0025] Combined with the initial binary cloud mask, re-segment the enhanced fused pixel-level cloud probability map through an adaptive threshold to generate the final pixel-level binary cloud mask.

[0026] Further, the extraction of the adaptive threshold includes:

[0027] Automatically extract the adaptive threshold of the image to be cloud detected by statistically analyzing the differences in the coverage areas of the scene-level binary cloud masks at different scales.

[0028] Further, the expression of the distance-weighted adaptive optimization is specifically:

[0029]

[0030] In the formula, is the enhanced pixel-level cloud probability map, is the fused pixel-level cloud probability map, is the adaptive distance constant, is the distance from the pixel at the i-th row and j-th column of the image to its nearest cloud block, is the thick cloud probability, is the thin cloud probability.

[0031] According to one aspect of the present specification, there is provided a thick cloud-thin cloud detection system based on a multi-scale deep model, including:

[0032] A multi-scale scene-level model module is used to train scene-level models of multiple different scales based on a cloud detection dataset, and generate scene-level cloud probability maps and scene-level binary cloud masks of multiple different scales.

[0033] A cloud probability map generation module is used to generate pixel-level cloud probability maps for thick clouds and thin clouds respectively based on scene-level cloud probability maps of multiple different scales and the dark pixel features of the image to be cloud-detected.

[0034] A thick-cloud and thin-cloud integrated detection module is used to combine the scene-level binary cloud mask, and fuse the pixel-level cloud probability maps for thick clouds and thin clouds using the dark pixel gradient, and then generate a pixel-level binary cloud mask by combining an adaptive threshold and distance weighting to achieve thick-cloud and thin-cloud integrated detection.

[0035] According to one aspect of the present specification, there is provided an electronic device including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the thick-cloud and thin-cloud detection method based on a multi-scale depth model are implemented.

[0036] According to one aspect of the present specification, there is provided a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the thick-cloud and thin-cloud detection method based on a multi-scale depth model are implemented.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. The present invention extracts the optical features of thick clouds and thin clouds respectively, and fuses them with multi-scale scene-level cloud probability maps using a differential weight strategy, so as to generate pixel-level cloud probability maps. Compared with the single-feature cloud detection method, it more accurately reflects the distribution characteristics of thick clouds and thin clouds in the image, and realizes high-precision detection of thick clouds and thin clouds while significantly reducing labor costs.

[0039] 2. The present invention realizes the fusion of thick-cloud probability maps and thin-cloud probability maps based on the difference in the gradient amplitude of the dark pixel features in the thick-cloud and thin-cloud regions of the image, in combination with the multi-scale scene-level binary cloud mask.

[0040] 3. By analyzing the differences in the areas covered by the multi-scale scene-level cloud masks, the present invention avoids relying on manually setting segmentation thresholds, adaptively extracts the segmentation thresholds of the image, and generates a binary cloud mask by combining an adaptive optimization strategy of distance weighting, realizing fully automatic and accurate detection of thick clouds, thin clouds and cloud edges. Description of the Drawings

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0042] Figure 1 is the method flowchart of the embodiment of the present invention;

[0043] Figure 2 is the structural diagram of the scene-level deep learning cloud detection model of the embodiment of the present invention. Specific embodiments

[0044] It should be noted that: without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0046] As Figure 1 shown, the embodiment of the present invention provides a thick cloud-thin cloud detection method based on a multi-scale deep model, including: Step 1, training multiple scene-level models with different scales based on a cloud detection dataset and generating multiple scene-level cloud probability maps and scene-level binary cloud masks with different scales; Step 2, generating a pixel-level cloud probability map, and respectively generating pixel-level cloud probability maps for thick clouds and thin clouds by using the scene-level cloud probability maps with different scales and the dark pixel features of the image; Step 3, generating a binary cloud mask, fusing the pixel-level cloud probability maps of thick clouds and thin clouds by using the dark pixel gradient, and generating a pixel-level binary cloud mask in combination with an adaptive threshold and distance weighting optimization.

[0047] Specifically, the embodiment of the present invention also provides an implementation manner of the scene-level cloud probability maps and scene-level binary cloud masks with multiple different scales in Step 1. First, use the existing multi-scale scene-level cloud detection dataset to train deep learning models with different scales. As Figure 2 shown, the used scene-level deep learning cloud detection model is mainly divided into three parts: a feature extractor, a model backbone, and a classifier. The feature extractor is composed of a convolutional layer and a max pooling layer, and can preliminarily extract the features of the input image. The model backbone is composed of 4 groups of residual blocks and can extract multi-level features of the image. The classifier is composed of an average pooling layer and a fully connected layer, and can map the features extracted by the model to cloud probabilities.

[0048] Specifically, in the embodiments of the present invention, scene-level cloud detection samples of different scales are divided into a training set and a test set according to a 7:3 ratio scale by scale. During the training process of the multi-scale model, the reflectance of the visible light band of the images in the training set is used as the model input, and the corresponding manually labeled labels are used as the output, where the labels only mark whether there are clouds in the images. Each scale model is trained only using the samples of its corresponding scale, and the learning rates of all scale models are uniformly set to 10 -4 . Each model is iteratively trained 100 times, and the model with the highest accuracy on the test set of the corresponding scale is selected as the final scene-level cloud detection model for that scale.

[0049] Specifically, in the model application stage, according to the scale requirements of the multi-scale deep model, the image to be detected is segmented into image patches of different scales, and the image patches are input into the deep models of the corresponding scales, so as to generate the cloud probability of each image patch, where the probability values of each image patch are the same. Then, the cloud probabilities of the image patches of each scale are recombined to generate a multi-scale scene-level cloud probability map, and the regions with cloud probabilities greater than 50% are regarded as cloud-covered regions, so as to generate a multi-scale scene-level binary cloud mask.

[0050] Specifically, the embodiments of the present invention also provide an implementation method for generating a pixel-level cloud probability map in step 2. The dark pixel feature of the image is calculated using the reflectance of the blue band and the green band of the image to be detected. The dark pixel feature can be calculated by formula (1):

[0051] (1)

[0052] Wherein, is the dark pixel feature of the pixel at the i-th row and j-th column of the image, is the reflectance of the blue band of the image, is the reflectance of the green band of the image. The larger the dark pixel value of the image pixel, the more likely the pixel is a cloud.

[0053] Specifically, in order to obtain the thick cloud pixel-level cloud probability map of the image, the multi-scale scene-level cloud probability map generated in step 1 is multiplied by the dark pixel feature of the image pixel by pixel. The thick cloud pixel-level cloud probability map of the final image can be expressed as:

[0054] (2)

[0055] Wherein, is the cloud probability of the thick cloud, is scene-level cloud probabilities of different scales, and the scales of the probability maps are arranged from large to small, is the adaptive weight constant of the cloud probability maps of different scales, where the value of the weight constant is .

[0056] Specifically, since thin clouds in the image usually exhibit the characteristic of large-scale distribution, in order to reduce the interference of local details on the extraction of the overall characteristics of thin clouds, singular value decomposition (SVD) is used to decompose the dark pixel features of the image, which are represented as two orthogonal matrices and a diagonal matrix. By retaining the first 20 largest singular values, a low-rank approximation of the dark pixel features is obtained, which can be expressed as:

[0057] (3)

[0058] where, is the low-rank approximation of the dark pixel features of the image, is the first 20 columns of the left singular vector matrix, are the first 20 singular values, is the first 20 rows of the right singular vector matrix.

[0059] Specifically, in this embodiment, the multi-scale scene-level cloud probability map is also multiplied pixel by pixel with the low-rank approximation of the dark pixel features of the image. The thin cloud pixel-level cloud probability map of the final image can be expressed as:

[0060] (4)

[0061] where, is the cloud probability of the thin cloud.

[0062] Specifically, the embodiment of the present invention also provides an implementation method for generating the binary cloud mask in step 3. The Sobel operator is used to calculate the gradient magnitude of the dark pixel features of the image, and the binary edge mask of the image is segmented. The binary edge mask can be expressed as:

[0063] (5)

[0064] (6)

[0065] (7)

[0066] In the formula, is the binary edge mask, is the gradient magnitude of the dark pixel features of the image. The thick cloud edge region usually exhibits a relatively high gradient magnitude, while the thin cloud region exhibits a relatively low gradient magnitude. are the convolution kernels of the Sobel operator in the horizontal and vertical directions respectively, is the adaptive edge segmentation constant. By sorting the gradient magnitude of the dark pixel features of the image from large to small and taking the gradient magnitude at the 15% position after sorting as the segmentation constant, it can be expressed as:

[0067] (8)

[0068] Among them, is the gradient amplitude of the dark pixel feature sorted from large to small, is the total number of pixels of the dark pixel feature.

[0069] Specifically, in the embodiment of the present invention, the intersection of the binary cloud masks of different scales generated in step 1 is taken as the clear cloud coverage area , and by statistically calculating the proportion of the binary edge mask in the clear cloud coverage area, the fusion of the thick cloud probability map and the thin cloud probability map is realized. The pixel-level cloud probability map of image fusion can be expressed as:

[0070] (9)

[0071] (10)

[0072] (11)

[0073] Among them, is the pixel-level cloud probability map after fusion, is the proportion of the cloud area edge, is the binary scene-level cloud mask, is the fusion constant.

[0074] Specifically, in the embodiment of the present invention, to avoid the participation of artificial thresholds and improve the generality of the present invention, by statistically calculating the differences in the coverage areas of the binary scene-level cloud masks at different scales, the adaptive threshold of the image is automatically extracted. The adaptive threshold of different images can be expressed as:

[0075] (12)

[0076] Among them, is the adaptive threshold of the image, is the thick cloud probability, obtained by statistically calculating the probability mean of the coverage area of the maximum-scale scene-level cloud mask, is the thin cloud probability, obtained by statistically calculating the probability mean of the coverage difference area between the maximum-scale scene-level cloud mask and the minimum-scale scene-level cloud mask.

[0077] Specifically, based on the adaptive threshold we segment the pixel-level cloud probability map after image fusion , and generate an initial binary cloud mask, which can be expressed as:

[0078] (13)

[0079] Among them, To initialize the binary cloud mask, since it uses an adaptive threshold instead of an artificially optimal threshold and it is difficult to achieve precise coverage of thin clouds, an adaptive optimization method with distance weighting is adopted for the pixel-level cloud probability map after fusion. It is enhanced and can be expressed as:

[0080] (14)

[0081] Wherein, is the enhanced pixel-level cloud probability map, is the adaptive distance constant, which can be calculated by and is the distance from the pixel at the i-th row and j-th column of the image to its nearest cloud block. Finally, the enhanced pixel-level cloud probability map is segmented again by the adaptive threshold to generate the final binary cloud mask. Specifically, in summary, a thick cloud-thin cloud detection method based on a multi-scale deep model provided by an embodiment of the present invention first trains a multi-scale scene-level cloud detection model using an existing multi-scale scene-level cloud detection sample set, and generates a multi-scale scene-level cloud probability and a binary cloud mask. Then, combining the dark pixel features of thick clouds and thin clouds with the multi-scale scene-level cloud probability, pixel-level cloud probability maps for thick clouds and thin clouds are generated respectively. Finally, the fusion of thick cloud and thin cloud cloud probabilities is realized based on the dark pixel gradient, and a binary pixel-level cloud mask is generated using an adaptive threshold and a distance-weighted optimization method. Considering that existing cloud detection methods rely on artificial thresholds or high-quality sample sets to achieve high-precision pixel-level cloud detection, however, the setting of artificial thresholds and the annotation of high-quality pixel-level samples both require a large amount of manpower. The present invention extracts the spectral features of thick clouds and thin clouds in the image respectively, and combines a multi-scale scene-level cloud detection model to obtain pixel-level cloud detection results. While significantly reducing the labor cost, high-precision detection of thick clouds and thin clouds is achieved.

[0082] Specifically, the embodiment of the present invention takes into account that existing cloud detection methods rely on artificial thresholds or high-quality sample sets to achieve high-precision pixel-level cloud detection. However, the setting of artificial thresholds and the annotation of high-quality pixel-level samples both require a large amount of manpower. The present invention combines the spectral features of thick clouds and thin clouds and a multi-scale deep model to generate cloud probability maps for thick clouds and thin clouds respectively, and uses the gradient magnitude difference of the spectral features of thick clouds and thin clouds to realize the fusion of cloud probability maps. In addition, based on the differences in the covered areas of scene-level cloud masks at different scales, an adaptive threshold of the image is automatically extracted, and a high-precision binary pixel-level cloud mask is obtained by combining a distance-weighted adaptive optimization method.

[0083]

[0084] ​The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, on the basis of the above embodiments, an embodiment of the present invention provides a thick cloud-thin cloud detection system based on a multi-scale depth model, which is used to execute a thick cloud-thin cloud detection method based on a multi-scale depth model in the above method embodiments.

[0085] The system includes: a multi-scale scene-level model module, which is used to train scene-level models of multiple different scales based on a cloud detection data set to generate scene-level cloud probability maps and scene-level binary cloud masks of multiple different scales; a cloud probability map generation module, which is used to generate pixel-level cloud probability maps for thick clouds and thin clouds respectively based on the scene-level cloud probability maps of multiple different scales and the dark pixel features of the image to be cloud-detected; a thick cloud-thin cloud integrated detection module, which is used to combine the scene-level binary cloud mask, and use the dark pixel gradient to fuse the pixel-level cloud probability maps for thick clouds and thin clouds, and then generate a pixel-level binary cloud mask by combining an adaptive threshold and distance weighting to achieve thick cloud-thin cloud integrated detection.

[0086] The thick cloud-thin cloud detection system based on a multi-scale depth model provided by the embodiment of the present invention solves the problem that the existing cloud detection methods are limited by factors such as the complexity of thin cloud features and the quality of training samples. By using several modules, it combines the spectral features and the thick cloud-thin cloud integrated detection technology based on a multi-scale depth model. Based on a multi-scale scene-level depth model, it uses the differential optical features of thick clouds and thin clouds to generate pixel-level cloud probability maps, and realizes the automatic and accurate detection of thick clouds, thin clouds and cloud edges based on an adaptive optimization strategy of an adaptive threshold and distance weighting.

[0087] Based on the same inventive concept as the foregoing embodiments, an embodiment of the present invention also provides an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement a thick cloud-thin cloud detection method based on a multi-scale depth model as proposed in the above embodiments.

[0088] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it overcomes the problem that the existing cloud detection methods are limited by factors such as the complexity of thin cloud features and the quality of training samples, and it is often difficult to achieve accurate detection of thin clouds. It accurately reflects the distribution characteristics of thick clouds and thin clouds in the image, and realizes the high-precision detection of thick clouds and thin clouds while significantly reducing the labor cost.

[0089] The storage medium can be any non-volatile storage device such as a hard disk, a solid-state drive, a flash drive, an optical disc, etc., and is used to store computer program codes and necessary data files. The stored computer programs include: a multi-scale scene-level model module, a cloud probability map generation module, and a thick cloud-thin cloud integrated detection module.

[0090] Finally, it should be pointed out that the above specific embodiments are only relatively representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and there can be many variations. Any simple modification, equivalent change and modification made to the above specific embodiments based on the technical essence of the present invention shall be considered to fall within the protection scope of the present invention.

Claims

1. A thick cloud-thin cloud detection method based on a multi-scale deep model, characterized in that Including: Step 1: Train scene-level models of multiple different scales based on a cloud detection dataset to generate scene-level cloud probability maps and scene-level binary cloud masks of multiple different scales; Step 2: Generate pixel-level cloud probability maps for thick clouds and thin clouds respectively based on the scene-level cloud probability maps of multiple different scales and the dark pixel features of the image to be cloud-detected; Step 3: Combine the scene-level binary cloud masks, use the dark pixel gradient to fuse the pixel-level cloud probability maps for thick clouds and thin clouds, and then generate a pixel-level binary cloud mask by combining an adaptive threshold and distance weighting to achieve integrated detection of thick clouds and thin clouds.

2. The thick cloud-thin cloud detection method based on a multi-scale depth model according to claim 1, wherein In the said Step 1, generating scene-level cloud probability maps and scene-level binary cloud masks of multiple different scales includes: Use the trained scene-level model to perform block processing on the images in the cloud detection dataset, and judge whether each image block contains clouds; If the image block contains clouds, mark all the pixels contained in the image block as clouds to obtain the scene-level cloud probability map and the scene-level binary cloud mask.

3. A thick cloud-thin cloud detection method based on a multi-scale depth model according to claim 1, characterized in that In the said Step 2, generating pixel-level cloud probability maps for thick clouds and thin clouds includes: Calculate the dark pixel features using the reflectance of the blue band and the green band in the image to be cloud-detected; Multiply the scene-level cloud probability maps of multiple different scales and the dark pixel features pixel by pixel to obtain the pixel-level cloud probability map for thick clouds; Perform singular value decomposition on the dark pixel features to obtain the low-rank approximation of the dark pixel features; Multiply the scene-level cloud probability maps of multiple different scales and the low-rank approximation of the dark pixel features pixel by pixel to obtain the pixel-level cloud probability map for thin clouds.

4. A thick cloud - thin cloud detection method based on a multi - scale deep model according to claim 1, characterized in that, In the said Step 3, combining the scene-level binary cloud masks and using the dark pixel gradient to fuse the pixel-level cloud probability maps for thick clouds and thin clouds includes: Use the Sobel operator to calculate the gradient magnitude of the dark pixel features and segment the binary edge mask of the image to be cloud-detected; Take the intersection of the generated scene-level binary cloud masks of different scales as the cloud coverage area, and at the same time clarify the proportion of the binary edge mask in the cloud coverage area, and fuse the pixel-level cloud probability maps for thick clouds and thin clouds to obtain the fused pixel-level cloud probability map.

5. A thick cloud-thin cloud detection method based on a multi-scale depth model according to claim 1, characterized in that In the said Step 3, generating the pixel-level binary cloud mask includes: Segment the fused pixel-level cloud probability map through an adaptive threshold to generate an initial binary cloud mask; Perform adaptive optimization with distance weighting to enhance the fused pixel-level cloud probability map; Combine the initial binary cloud mask and re-segment the enhanced fused pixel-level cloud probability map through an adaptive threshold to generate the final pixel-level binary cloud mask.

6. The thick cloud-thin cloud detection method based on a multi-scale depth model according to claim 5, characterized in that The extraction of the said adaptive threshold includes: Automatically extract the adaptive threshold of the image to be cloud-detected by statistically analyzing the differences in the coverage areas of the scene-level binary cloud masks of different scales.

7. A thick cloud-thin cloud detection method based on a multi-scale deep model according to claim 5, characterized in that, The expression of the adaptive optimization with distance weighting is specifically: , In the formula, is the enhanced pixel-level cloud probability map, is the fused pixel-level cloud probability map, is the adaptive distance constant, is the distance from the pixel at the i-th row and j-th column of the image to its nearest cloud block, is the thick cloud probability, is the thin cloud probability.

8. A thick cloud-thin cloud detection system based on a multi-scale deep model, characterized in that, Including: A multi-scale scene-level model module for training scene-level models of multiple different scales based on a cloud detection dataset to generate scene-level cloud probability maps and scene-level binary cloud masks of multiple different scales; A cloud probability map generation module for generating pixel-level cloud probability maps for thick clouds and thin clouds respectively based on the scene-level cloud probability maps of multiple different scales and the dark pixel features of the image to be cloud-detected; The thick cloud - thin cloud integrated detection module is used to combine the scene - level binary cloud mask, adopt the dark pixel gradient to fuse the pixel - level cloud probability maps of thick clouds and thin clouds, and generate a pixel - level binary cloud mask by combining an adaptive threshold and distance - weighted optimization, so as to achieve the integrated detection of thick clouds and thin clouds.

9. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the thick cloud - thin cloud detection method based on the multi - scale depth model according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the thick cloud - thin cloud detection method based on the multi - scale depth model according to any one of claims 1 to 7.

Images