Satellite image fog automatic identification method and device based on machine deep learning

Abstract

The present invention discloses a method and device for automatically identifying heavy fog in satellite images based on machine deep learning, including obtaining high-resolution atmospheric cloud and fog satellite data, and preprocessing the satellite data to obtain image data, wherein fog areas in the image data are annotated; constructing training samples based on the image data to obtain positive sample pairs and negative sample pairs; constructing a machine deep learning model based on the positive sample pairs and negative sample pairs and training; wherein, in the training process, extracting fog area features and non-fog area features based on the backbone network of the model; reconstructing the fog area based on the fog area features and non-fog area features; comparing the annotated fog area with the reconstructed fog area, and training the model with the goal of minimizing the difference between the two. Applying the contrastive learning mechanism to the deep learning model can make fuller use of fog area samples, allowing the neural network to better learn the distinguishing features of fog areas and other non-fog areas, and achieve a higher fog area identification accuracy.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and device for automatically identifying heavy fog in satellite images based on machine deep learning. Background Art

[0002] Fog is an atmospheric phenomenon, which refers to a layer of misty gaseous water between the ground and the clouds. It is formed by the condensation of water vapor into water droplets or ice crystals, and usually appears in the low-altitude layer close to the ground, resulting in reduced visibility. Fog is one of the most serious weather disasters that threaten traffic safety, and is also an important factor affecting public travel and social order.

[0003] Heavy fog monitoring and forecasting have always been important aspects of the meteorological department's disaster prevention and mitigation services. However, since heavy fog is significantly affected by weather systems of different scales, as well as topography and underlying surfaces, and is characterized by suddenness, locality, and rapid changes, it is difficult to monitor and forecast heavy fog.

[0004] Satellite remote sensing can achieve all-weather and full-coverage observation. In order to make up for the insufficient temporal and spatial accuracy of ground-based observations of heavy fog, it is urgent to develop heavy fog identification technology based on high-resolution satellites.

[0005] With the continuous advancement of artificial intelligence technology, the method of fog area identification has also shifted from the traditional method of relying on sparse meteorological stations and buoy monitoring in coastal or offshore areas to a method based on the combination of satellite image analysis and machine learning technology. At present, the method of identifying fog areas based on satellite image analysis is still mainly based on the threshold method. The threshold is set by the brightness temperature difference of different channel data in the satellite image data, so as to realize the fog area identification. However, this method has the problem that the threshold value will change dynamically with time, season, region and other factors. In addition, since fog often appears regionally, this method does not identify the fog area based on the integrity of the image, so this method has problems such as low robustness in fog area identification. At the same time, this method requires a lot of manpower for threshold calculation, has high professional knowledge requirements for practitioners, and requires huge costs for talent training.

[0006] With the continuous development of satellite aerospace technology and people's attention to meteorological issues, more and more meteorological satellites are launched. And these satellites are always working, steadily transmitting satellite image data of the earth back to the earth. Therefore, it is easy to obtain a large amount of satellite image data of foggy areas in multiple regions and ranges. However, these data lack annotations of foggy areas. In supervised deep learning, labeled image data is very important. But manual annotation of data takes a lot of time and consumes a lot of labor costs.

[0007] In summary, in the related technologies, using deep learning methods for foggy area recognition requires a large amount of labeled data, and the data labeling cost is higher. Since the number of foggy area samples is relatively small, the effect of using the same proportion for foggy area recognition learning will be reduced during the learning process, and the utilization rate of foggy area samples is not high. Summary of the invention

[0008] The present application provides a method and device for automatically identifying heavy fog in satellite images based on machine deep learning to solve the problems existing in the related technology.

[0009] In a first aspect, the present invention provides a method for automatic identification of heavy fog in satellite images based on machine deep learning, comprising: obtaining high-resolution atmospheric cloud and fog satellite data, and preprocessing the satellite data to obtain image data, wherein fog areas in the image data are marked; constructing training samples based on the image data to obtain positive sample pairs and negative sample pairs; constructing a machine deep learning model based on the positive sample pairs and negative sample pairs and performing training; wherein, during the training process, fog area features and non-fog area features are extracted based on a backbone network of the model; fog areas are reconstructed based on the fog area features and non-fog area features; and the marked fog areas are compared with the reconstructed fog areas, and the machine deep learning model is trained with the goal of minimizing the difference between the two.

[0010] Optionally, the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the machine deep learning model based on a loss function:

[0011] ;

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] in, is the loss function used in the final training, is the cross entropy loss function with hinge loss, is the contrastive loss function; is the cross entropy loss function, is the hinge loss function, , are two hyperparameters that represent the relative influence of hinge loss and contrast loss in the final loss; is the true label, is the predicted value; represents the anchor point sample, Representation and Similar samples are positive samples. Representation and Dissimilar samples are negative samples. Indicated in The prediction results of anchor samples, For positive samples The prediction results, For negative samples The prediction results, is the margin hyperparameter, is the pruning function.

[0017] Optionally, constructing training samples based on image data includes: randomly sampling foggy area images in different satellite images containing foggy areas, and using the obtained sampled samples as positive sample pairs; randomly sampling foggy area images in satellite images containing foggy areas, and randomly sampling non-fog area images in satellite images not containing foggy areas, and using the sampled foggy area images and non-fog area images as negative sample pairs.

[0018] Optionally, the method further includes: acquiring a satellite image to be identified; inputting the satellite image to be identified into the machine deep learning model, and automatically outputting a fog area identification result.

[0019] In a second aspect, the present invention provides a satellite image heavy fog automatic identification device based on machine deep learning, comprising a preprocessing unit, configured to obtain high-resolution atmospheric cloud and fog satellite data, and preprocess the satellite data to obtain image data, wherein the fog area in the image data is marked; a sample construction unit, configured to construct training samples based on the image data to obtain positive sample pairs and negative sample pairs; a training unit, configured to construct and train a machine deep learning model based on the positive sample pairs and negative sample pairs; wherein, during the training process, fog area features and non-fog area features are extracted based on the backbone network of the model; the fog area is reconstructed based on the fog area features and non-fog area features; the marked fog area and the reconstructed fog area are compared, and the machine deep learning model is trained with the goal of minimizing the difference between the two.

[0020] Optionally, the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the model based on a loss function:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;;

[0026] in, is the loss function used in the final training, is the cross entropy loss function with hinge loss, is the contrastive loss function; is the cross entropy loss function, is the hinge loss function, , are two hyperparameters, representing the relative influence of hinge loss and contrast loss in the final loss; y is the true label, is the predicted value; represents the anchor point sample, Representation and Similar samples are positive samples. Representation and Dissimilar samples are negative samples. Indicated in The prediction results of anchor samples, For positive samples The prediction results, For negative samples The prediction results, is the margin hyperparameter, is the pruning function.

[0027] Optionally, constructing training samples based on image data includes: randomly sampling foggy area images in different satellite images containing foggy areas, and using the obtained sampled samples as positive sample pairs; randomly sampling foggy area images in satellite images containing foggy areas, and randomly sampling non-fog area images in satellite images without foggy areas, and using the sampled foggy area images and non-fog area images as negative sample pairs.

[0028] Optionally, the method further includes: acquiring a satellite image to be identified; inputting the satellite image to be identified into the machine deep learning model, and automatically outputting a fog area identification result.

[0029] In a third aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of the first aspects is implemented.

[0030] In a fourth aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method provided in the first aspect when executing the program.

[0031] The present invention discloses a method for automatically identifying heavy fog in satellite images based on machine deep learning, comprising obtaining high-resolution atmospheric cloud and fog satellite data, and preprocessing the satellite data to obtain image data, wherein fog areas in the image data are annotated; constructing training samples based on the image data to obtain positive sample pairs and negative sample pairs; constructing a machine deep learning model based on the positive and negative sample pairs and training; wherein, in the training process, extracting fog area features and non-fog area features based on the backbone network of the model; reconstructing the fog area based on the fog area features and non-fog area features; comparing the annotated fog area with the reconstructed fog area, and training the model with the goal of minimizing the difference between the two. By applying the contrastive learning mechanism to the deep learning network, the neural network can better learn the distinguishing features between the fog area and other non-fog areas, making the model's results of fog area recognition more stable and reliable, and making more full use of fog area samples. Thereby achieving a higher fog area recognition accuracy. At the same time, it reduces the demand for data in deep learning, improves the ability to learn the similarities of fog area samples, and thus improves the generalization ability of the model, which helps to achieve fog area monitoring with a larger range and full spatial and temporal coverage. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0033] Figure 1 This application provides a flow chart of a method for automatically identifying heavy fog in satellite images based on machine deep learning;

[0034] Figure 2 This is a logical illustration of a method for automatically identifying heavy fog in satellite images based on machine deep learning for this application;

[0035] Figure 3 Schematic diagram of the construction process of positive sample pairs and negative sample pairs in this application;

[0036] Figure 4 A method corresponding to the present application Figure 1 Schematic diagram of an electronic device. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0038] The following is combined with Figure 1 The following is an example of a method for automatically identifying fog in satellite images based on machine deep learning. The method includes the following steps:

[0039] Step 101: Acquire high-resolution atmospheric cloud and fog satellite data, and pre-process the satellite data to obtain image data, wherein fog areas in the image data are marked.

[0040] In this embodiment, L1b data of atmospheric clouds and fog can be obtained from the meteorological satellite website (L1b refers to the data set obtained by the satellite imager, which refers to the original data that has been screened for outliers and has not been processed otherwise, and the data is in .nc format.) The L1b data is clipped, projected, and visualized according to the defined fog area identification area to construct image data that can be used to train the network model. The fog area is labeled according to the image data. The labeling method is to use labeling software for labeling, which is not limited here. After the labeling is completed, positive sample pairs and negative sample pairs can be constructed.

[0041] Step 102: construct training samples based on the image data to obtain positive sample pairs and negative sample pairs.

[0042] In this embodiment, when constructing a positive sample pair, foggy area images are randomly sampled from different satellite images containing foggy areas, and the obtained sampled samples are used as positive sample pairs; when constructing a negative sample pair, foggy area images are randomly sampled from satellite images containing foggy areas, and non-fog area images are randomly sampled from satellite images not containing foggy areas, and the sampled foggy area images and non-fog area images are used as negative sample pairs.

[0043] In this optional implementation, after the data is labeled, a visualized satellite image and a labeled fog area mask labeling result are obtained. Then, based on the obtained results, a sample pair is constructed according to the contrastive learning method. Different from the traditional contrastive learning positive sample pair and negative sample pair construction method used in unsupervised deep learning, the construction method of this embodiment makes the feature distinction between the fog area and other areas greater.

[0044] Different from traditional contrastive learning, in fog area identification, the method of constructing contrastive learning positive and negative sample pairs is as follows: randomly sampling fog area images from different satellite images containing fog areas as positive sample pairs, and randomly sampling fog area images from satellite images containing fog areas and non-fog area images from satellite images without fog areas as negative sample pairs.

[0045] In this optional implementation, the purpose of constructing contrastive learning positive sample pairs and negative sample pairs is to use limited data to fully learn the feature differences between foggy areas and non-foggy areas, and to increase the difference between the two. At the same time, since there are relatively few image samples containing foggy areas compared to non-foggy areas, samples of the same foggy area are constructed as positive sample pairs, making full use of limited foggy area data to learn more common foggy area features and reduce the difference between samples of the same foggy area.

[0046] Step 103: Based on the positive sample pairs and the negative sample pairs, a deep learning model is constructed and trained, wherein during the training process, fog area features and non-fog area features are extracted based on the backbone network of the model; the fog area is reconstructed based on the fog area features and non-fog area features; the marked fog area and the reconstructed fog area are compared, and the model is trained with the goal of minimizing the difference between the two.

[0047] In this embodiment, after constructing the contrast learning positive sample pairs and negative sample pairs and setting the model loss function, the model training is performed to extract the features of the foggy area and the non-foggy area. During the training, a variety of classic deep learning networks can be used as the backbone network (backbone) for feature extraction, such as ViT, ResNet, Mix-ViT, etc. According to the features of the foggy area and the non-foggy area, the foggy area mask image predicted by the feature extraction model is reconstructed and compared with the labeled image to measure the final foggy area identification effect.

[0048] It should be understood that the contrastive learning mechanism for fog areas in this embodiment can be applied to any classic or well-performing neural network as a backbone, and any backbone network can be replaced during model training.

[0049] In the contrastive learning of this embodiment, positive samples and negative samples are distinguished, and more discriminative features are learned by maximizing the similarity between positive samples and minimizing the similarity between negative samples.

[0050] This embodiment can save labor costs, reduce the workload of satellite image data annotation, and at the same time can better utilize existing annotation data, and can better utilize the existing data to extract fog area features.

[0051] As an optional implementation of this embodiment, the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the model based on a loss function:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, is the loss function used in the final training, is the cross entropy loss function with hinge loss, is the contrastive loss function; is the cross entropy loss function, is the hinge loss function, , are two hyperparameters that represent the relative influence of hinge loss and contrast loss in the final loss; is the true label, is the predicted value; represents the anchor point sample, Representation and Similar samples are positive samples. Representation and Dissimilar samples are negative samples. Indicated in The prediction results of anchor samples, For positive samples The prediction results, For negative samples The prediction results, is the margin hyperparameter, is the pruning function.

[0058] In an optional implementation of this embodiment, before training the model, it is necessary to construct the loss function of the model in order to better optimize the model. The ultimate goal of training the model is to increase the difference between the foggy area and the non-foggy area, and reduce the difference between the foggy areas at different times or regions. Therefore, this embodiment introduces contrastive loss to help achieve this result. The goal of contrastive learning is to map similar samples to similar embedding spaces and dissimilar samples to farther spaces. The formula for defining contrastive loss used in this embodiment is as follows:

[0059]

[0060] in, represents the anchor point sample, Representation and Similar samples (positive samples), Representation and Dissimilar samples (negative samples), is a margin hyperparameter, is the pruning function. The intuitive explanation of this loss function is that we hope that the distance between pairs of samples of the same class is smaller than the distance between pairs of samples of different classes, and the difference between them is greater than a predefined margin.

[0061] After constructing the contrast loss, combine the contrast loss with the original loss function suitable for image segmentation in deep learning to construct the overall loss function of the model. The combined formula is as follows:

[0062] ;

[0063] ;

[0064] in, , There are two hyperparameters, which represent the relative influence of hinge loss and contrast loss in the final loss, and are generally in the range of 0.1-1.

[0065] As an optional implementation of this embodiment, the method further includes: acquiring a satellite image to be identified; inputting the satellite image to be identified into the model, and automatically outputting a fog area identification result.

[0066] In an optional implementation of this embodiment, after the model training is completed, the model with the best effect is stored, and this model is used to identify other pre-processed satellite images, ultimately achieving automatic identification of foggy areas.

[0067] refer to Figure 2 The overall logic diagram of this method is given. After obtaining the original data of remote sensing satellites, preprocessing is performed. Random sampling is performed based on the preprocessed data to construct positive sample pairs and negative sample pairs. Then, training is performed under the set contrast loss function, the model loss is calculated and the model is optimized. In this process, feature extraction and mask reconstruction are performed to finally obtain a trained model.

[0068] As an optional implementation method of this embodiment, constructing training samples based on image data includes: randomly sampling fog area images in different satellite images containing fog areas, and using the obtained sampled samples as positive sample pairs; randomly sampling fog area images in satellite images containing fog areas, and randomly sampling non-fog area images in satellite images without fog areas, and using the sampled fog area images and non-fog area images as negative sample pairs.

[0069] In this optional implementation, the foggy area image is a positive sample, and the non-foggy area image is a negative sample. A single negative sample does not contain a foggy area image, but a negative sample pair contains foggy area and non-foggy area images. That is, one positive sample and one negative sample constitute one negative sample pair.

[0070] Further, refer to Figure 3 When constructing positive and negative sample pairs for contrastive learning, randomly sample foggy images from different satellite images containing foggy areas, including randomly sampling foggy images and randomly sampling foggy images from other images. In deep learning, multiple samples are read and put into the model for training each time the samples are sampled. The other images here refer to the foggy areas in other images obtained during the current random sampling.

[0071] When constructing negative sample pairs, foggy area images and non-foggy area images in other images are randomly sampled.

[0072] This embodiment can achieve higher accuracy by using comparative learning while reducing costs. This overcomes the problem in the prior art that a large amount of work is required for data labeling, a large amount of labor costs are also required in the labeling process, and the labeled data also needs to reach a certain resolution to be accurately labeled, so it is impossible to achieve data labeling with a larger range and full coverage, and it is impossible to achieve fog area monitoring in a larger range.

[0073] In this embodiment, it is very important to obtain obvious and highly distinguishable fog area features using a small amount of data. Contrastive learning can make better use of existing data and fully distinguish the features of fog areas and other non-fog areas in a small amount of data.

[0074] The above is a method provided for one or more embodiments of the present application. Based on the same idea, the present application also provides a corresponding automatic fog identification device for satellite images based on machine deep learning, including a preprocessing unit, configured to obtain atmospheric cloud and fog satellite data, and preprocess the satellite data to obtain image data, wherein the fog area in the image data is marked; a sample construction unit, configured to construct training samples based on the image data to obtain positive sample pairs and negative sample pairs; a training unit, configured to construct and train a machine deep learning model based on the positive sample pairs and negative sample pairs; wherein, during the training process, fog area features and non-fog area features are extracted based on the backbone network of the model; the fog area is reconstructed based on the fog area features and non-fog area features; the marked fog area and the reconstructed fog area are compared, and the machine deep learning model is trained with the goal of minimizing the difference between the two.

[0075] As an optional implementation of this embodiment, the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the model based on a loss function:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in, is the loss function used in the final training, is the cross entropy loss function with hinge loss, is the contrastive loss function; is the cross entropy loss function, is the hinge loss function, , are two hyperparameters that represent the relative influence of hinge loss and contrast loss in the final loss; is the true label, is the predicted value; represents the anchor point sample, Representation and Similar samples are positive samples. Representation and Dissimilar samples are negative samples. Indicated in The prediction results of anchor samples, For positive samples The prediction results, For negative samples The prediction results, is the margin hyperparameter, is the pruning function.

[0082] As an optional implementation method of this embodiment, constructing training samples based on image data includes: randomly sampling fog area images in different satellite images containing fog areas, and using the obtained sampled samples as positive sample pairs; randomly sampling fog area images in satellite images containing fog areas, and randomly sampling non-fog area images in satellite images without fog areas, and using the sampled fog area images and non-fog area images as negative sample pairs.

[0083] As an optional implementation of this embodiment, the trained model obtains a satellite image to be identified;

[0084] The satellite image to be identified is input into the model, and the fog area identification result is automatically output.

[0085] Figure 4 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Figure 4 As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0086] The processor 501 and the memory 502 communicate with each other via the bus 503 ; the processor 501 is used to call program instructions in the memory 502 to execute the methods provided by the above-mentioned method embodiments.

[0087] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above method embodiments.

[0088] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0089] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0090] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0091] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for automatically identifying heavy fog in satellite images based on machine deep learning, characterized in that: include: Acquire atmospheric cloud and fog satellite data, and pre-process the satellite data to obtain image data, wherein fog areas in the image data are marked; Construct training samples based on image data to obtain positive sample pairs and negative sample pairs; Based on the positive sample pairs and the negative sample pairs, a machine deep learning model is constructed and trained; Among them, during the training process, fog area features and non-fog area features are extracted based on the backbone network of the model; the fog area is reconstructed based on the fog area features and non-fog area features; the marked fog area and the reconstructed fog area are compared, and the model is trained with the goal of minimizing the difference between the two; the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the model based on the loss function: L final =L seg +μ2L Contrastive ; L seg =L bce +μ1L hinge ; Among them, L final is the loss function used in the final training, L seg is the cross entropy loss function with hinge loss, L Contrastive is the contrast loss function; L bce is the cross entropy loss function, L hinge is the hinge loss function, μ1 and μ2 are two hyperparameters, indicating the relative influence of hinge loss and contrast loss in the final loss; y is the true label, is the predicted value; x i represents the anchor point sample, x + Represents x i Similar samples are positive samples, x - Represents x i Dissimilar samples are negative samples, f(x i ) means that in x i The prediction result of the anchor sample, f(x + ) means that for the positive sample x + The prediction result, f(x - ) means for negative sample x - The prediction result, α is the margin hyperparameter, [z] + =max(z,0) is the pruning function.

2. The method for automatically identifying heavy fog in satellite images based on machine deep learning according to claim 1 is characterized in that: Building training samples based on image data includes: In different satellite images containing foggy areas, randomly sample the foggy area images, and use the obtained sampled samples as positive sample pairs; Foggy area images are randomly sampled from satellite images containing foggy areas, and non-foggy area images are randomly sampled from satellite images without foggy areas, and the sampled foggy area images and non-foggy area images are used as negative sample pairs.

3. The method for automatically identifying heavy fog in satellite images based on machine deep learning according to claim 1 is characterized in that: The method further comprises: Obtain high-resolution satellite images to be identified; The high-resolution satellite image to be identified is input into the model, and the fog area identification result is automatically output.

4. A satellite image fog automatic identification device based on machine deep learning, characterized in that: include: A preprocessing unit is configured to obtain atmospheric cloud and fog satellite image data and preprocess the satellite image data to obtain image data, wherein fog areas in the image data are marked; Construct training samples based on image data to obtain positive sample pairs and negative sample pairs; Based on the positive sample pairs and the negative sample pairs, a machine deep learning model is constructed and trained; In the training process, the backbone network of the model is used to extract the fog area features and the non-fog area features; the fog area is reconstructed based on the fog area features and the non-fog area features; the marked fog area and the reconstructed fog area are compared, and the model is trained with the goal of minimizing the difference between the two; Among them, the machine deep learning model is trained with the goal of minimizing the difference between the two, including training the model based on the loss function: L final =L seg +μ2L Contrastive ; L seg =L bce +μ1L hinge ; Among them, L final is the loss function used in the final training, L seg is the cross entropy loss function with hinge loss, L Contrastive is the contrast loss function; L bce is the cross entropy loss function, L hinge is the hinge loss function, μ1 and μ2 are two hyperparameters, indicating the relative influence of hinge loss and contrast loss in the final loss; y is the true label, is the predicted value; x i represents the anchor point sample, x + Represents x i Similar samples are positive samples, x - Represents x i Dissimilar samples are negative samples, f(x i ) means that in x i The prediction result of the anchor sample, f(x + ) means that for the positive sample x + The prediction result, f(x - ) means for negative sample x - The prediction result, α is the margin hyperparameter, [z] + =max(z,0) is the pruning function.

5. The automatic satellite image fog identification device based on machine deep learning according to claim 4 is characterized in that: Building training samples based on image data includes: In different satellite images containing foggy areas, randomly sample the foggy area images, and use the obtained sampled samples as positive sample pairs; Foggy area images are randomly sampled from satellite images containing foggy areas, and non-foggy area images are randomly sampled from satellite images without foggy areas, and the sampled foggy area images and non-foggy area images are used as negative sample pairs.

6. The automatic satellite image fog identification device based on machine deep learning according to claim 4 is characterized in that: After the machine deep learning model training is completed, the satellite image to be identified is obtained; based on the model, the fog area identification result is automatically output.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 3 is implemented.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 3 is implemented.

Images