Construction method of fog recognition model and fog recognition method and device

Abstract

The invention provides a construction method of a fog recognition model, and a fog recognition method and device. The construction of the model comprises the following steps: respectively acquiring ground observation station data and satellite observation data; determining corresponding indication information based on the ground observation station data; based on satellite observation data, determining corresponding embedding features; generating a prediction fog mask based on the indication information and the embedded feature; and training the initial fog recognition model according to the real fog mask and the predicted fog mask, thereby obtaining a trained fog recognition model. The ground observation station data and the satellite observation data to be recognized are recognized by using the trained fog recognition model, and the fog recognition efficiency can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a method for constructing a fog recognition model, a fog recognition method and a fog recognition device. Background Art

[0002] Fog is an important atmospheric phenomenon that is highly related to the ecosystem and environment. The two necessary prerequisites for its formation are the presence of fog condensation nuclei and water in the atmosphere. Fog occurring near the ground will reduce visibility, deteriorate air quality, and have a serious impact on transportation, including aviation, navigation, and highway transportation. In addition, heavy fog also has a significant impact on power facilities, human health, agricultural production, military operations, etc. With the advancement of satellite remote sensing technology, fog detection related work based on remote sensing images has been widely studied in recent years. Researching fog detection methods based on satellite remote sensing technology is of great significance for disaster prevention and mitigation, and the sustainable development of the national economy and society.

[0003] Traditional fog detection methods usually use threshold technology to determine whether fog exists by setting thresholds according to the different pixel intensities of medium-high clouds, low clouds, and fog. However, these methods not only rely on empirical threshold selection, but also have low computational efficiency. In recent years, with the rapid development of artificial intelligence technology, semantic segmentation methods based on deep learning have gradually been applied to the field of fog detection. Convolutional neural network (CNN) and visual transformer (ViT), as the two mainstream methods of deep learning, have performed well in many fields. However, these methods are highly dependent on large-scale, accurately annotated pixel-level data. Since the generation process of fog is complex and susceptible to environmental changes, the accurate annotation of fog areas is costly and requires a large amount of pixel-level label data.

[0004] In order to solve the labeling problem, some weakly supervised and unsupervised learning methods have gradually attracted attention in recent years. These methods reduce the labeling cost by generating pseudo labels using point labels, box labels or rough masks. However, although these methods reduce the dependence on precise labeling, there is still a problem of low recognition efficiency in the process of combining multi-source data (satellite remote sensing data, ground meteorological observation data (such as temperature, humidity and visibility), radar data, laser radar (LIDAR) data and numerical weather forecast model prediction data) to identify fog. Summary of the invention

[0005] The present invention provides a method for constructing a fog recognition model, a fog recognition method and a device, so as to solve the defect of low fog recognition efficiency in the prior art.

[0006] The present invention provides a method for constructing a fog recognition model, comprising:

[0007] Obtain ground observation station data and satellite observation data respectively;

[0008] Based on the ground observation station data, determining corresponding indication information;

[0009] Based on the satellite observation data, determining the corresponding embedded features;

[0010] generating a predicted fog mask based on the indication information and the embedded features;

[0011] The initial fog recognition model is trained according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

[0012] According to a method for constructing a fog recognition model provided by the present invention,

[0013] The determining, based on the ground observation station data, the corresponding indication information includes:

[0014] Firstly, based on the ground observation station data, the corresponding initial indication information is determined;

[0015] Extracting grayscale images of the second and thirteenth channels of the satellite observation data, and generating a filtering mask based on the grayscale images of the second and thirteenth channels;

[0016] The filter mask is used to filter out incorrect indication information in the initial indication, and finally the correct indication information of the ground observation station data is obtained.

[0017] According to a method for constructing a fog recognition model provided by the present invention,

[0018] The determining, based on the satellite observation data, the corresponding embedded features includes:

[0019] Extracting the second, third and thirteenth channel grayscale images of the satellite observation data, and merging the second, third and thirteenth channel grayscale images to obtain a pseudo-color image;

[0020] Extracting features from the pseudo-color image to obtain high-frequency features;

[0021] Performing feature extraction on the pseudo-color image to obtain a first embedded feature;

[0022] The first embedded feature and the high-frequency feature are sequentially fused to obtain the embedded feature.

[0023] According to a method for constructing a fog recognition model provided by the present invention, the fusing of the first embedded feature and the high-frequency feature in sequence to obtain the embedded feature includes:

[0024] Fusing the first embedded feature and the high-frequency feature to obtain a first fused feature;

[0025] Extracting an embedded feature of the first fused feature to obtain a second embedded feature;

[0026] Fusing the second embedded feature and the high-frequency feature to obtain a second fused feature;

[0027] Extracting the embedded features of the second fused features again to obtain a third embedded feature;

[0028] The high-frequency features are continuously fused with the embedded features extracted last time to finally obtain the embedded features.

[0029] According to a method for constructing a fog recognition model provided by the present invention, the ground observation station data includes: humidity, visibility, dew point temperature, and weather meteorological code; the satellite observation data is selected from a latitude range of 30°N to 40°N and a longitude range of 110°E to 120°E, with an interval of hours, and the satellite observation data is selected from a range of 0:00 to 8:00.

[0030] The present invention also provides a fog identification method, comprising:

[0031] Respectively obtain the ground observation station data and satellite observation data to be identified;

[0032] The ground observation station data and satellite observation data to be identified are input into a trained fog recognition model, and based on the fog recognition model, the fog area is output.

[0033] The present invention also provides a device for a method of constructing a fog recognition model, comprising:

[0034] An acquisition unit, used for respectively acquiring ground observation station data and satellite observation data;

[0035] A determination unit, configured to determine corresponding indication information based on the ground observation station data;

[0036] The determining unit is further used to determine the corresponding embedding features based on the satellite observation data;

[0037] A generating unit, configured to generate a predicted fog mask based on the indication information and the embedded features;

[0038] The training unit is used to train the initial fog recognition model according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

[0039] The present invention also provides a device for a fog identification method, comprising:

[0040] An acquisition unit, used to respectively acquire ground observation station data and satellite observation data to be identified;

[0041] The identification unit is used to input the ground observation station data and satellite observation data to be identified into a trained fog identification model, and identify the fog based on the fog identification model.

[0042] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements a method for constructing a fog recognition model as described above or implements the fog recognition method as described above.

[0043] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for constructing the fog recognition model as described in any one of the above or the fog recognition method as described above is implemented.

[0044] The method for constructing a fog recognition model, the method for fog recognition, and the device provided by the present invention effectively improve the efficiency and accuracy of fog recognition by using a trained fog recognition model to identify the area where the fog is located. The method for constructing a fog recognition model provided by the present invention extracts the indication information of the ground observation station data through a prompt encoder, extracts the embedded features of the satellite observation data through an image encoder, and finally generates a predicted fog mask through a mask decoder using the indication information and the embedded features. It automatically generates a segmentation mask, quickly segments multiple objects, and avoids the tedious work of labeling one by one. Therefore, the fog recognition model constructed by the present invention greatly simplifies the labeling process of fog images and effectively improves the efficiency of model construction; at the same time, the acquisition of ground observation station data and satellite observation data is combined to generate a predicted fog mask, which also effectively improves the recognition accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a structural block diagram of the SAM model in the prior art;

[0046] Figure 2 A flow chart of the method for constructing a fog recognition model provided by the present invention;

[0047] Figure 3(a) is a prompt map obtained using ground observation station data;

[0048] Figure 3(b) shows the filter mask obtained using satellite observation data;

[0049] FIG3(c) is a prompt image obtained by filtering FIG3(a) using a filter mask;

[0050] Figure 4 Schematic diagram of the fusion process of fusion features;

[0051] Figure 5(a) is a schematic diagram of a pseudo-color image;

[0052] Figure 5(b) is a feature map after filtering out high-frequency components after DCT transformation;

[0053] Figure 5(c) is the feature map returned to RGB space after inverse DCT transformation;

[0054] Figure 6 A schematic diagram of a process for extracting embedded features of a pseudo-color image according to the present invention;

[0055] Figure 7 This is one of the schematic diagrams of the fog identification device provided by the present invention;

[0056] Figure 8 The second schematic diagram of the fog identification device provided by the present invention;

[0057] Fig. 9 A schematic diagram of the structure of a device for constructing a fog recognition model provided by the present invention;

[0058] Fig.10 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The purpose of the present invention is to identify the foggy area based on the use of ground observation station data and FY4A satellite observation data. Due to the distribution constraints of ground observation stations, the actual fog monitoring coverage is incomplete and uneven. Although the use of geostationary meteorological satellite multi-channel observation images and semantic segmentation can identify foggy areas to overcome the limitations of ground observation stations, there are still problems with low recognition efficiency and accuracy. The existing deep learning-based methods are highly dependent on a large number of manually annotated pixel-level labels, which is a major limitation in practical applications: due to the time, conditions and annotation costs of fog formation, it is difficult to obtain a large number of pixel-level labels for training in actual operations, which affects the wide application and accuracy of deep learning methods. In addition, the existing deep learning methods fail to make full use of valuable auxiliary information such as ground observation station data, resulting in insufficient recognition accuracy.

[0061] The present invention uses the SAM model and automatically generates prompts based on ground-based observation stations and multi-channel satellite images to solve the problem of low fog recognition efficiency. This method combines multimodal information to assist image segmentation tasks. Compared with supervised learning that only uses pixel-level labels, it has significant advantages in recognition efficiency and recognition accuracy.

[0062] SAM (Segmen Anything Model) is a general segmentation model that can effectively solve image segmentation tasks in computer vision. It can segment almost any type of object without special training for specific tasks. SAM uses powerful deep learning architectures such as convolutional neural networks (CNN) or visual transformers (ViT) to extract features from images. These encoders are pre-trained with large-scale datasets and can capture the features and structures of various objects in images. The SAM model allows users to specify the objects to be segmented by providing prompts. The prompts can be points, boxes, text, or other forms. Based on these prompts, the model can automatically generate the corresponding segmentation masks. For example, with point prompts, users only need to click one or several points on the object, and the SAM model will automatically generate the segmentation contours of the object based on these points. The SAM model can greatly simplify the image annotation process. In traditional image annotation, users are usually required to manually draw polygons, box selections, or finely annotate each object, which is very time-consuming when dealing with complex backgrounds or objects with rich details. The SAM model can automatically generate segmentation masks and quickly segment multiple objects, avoiding the tedious work of annotating one by one, greatly reducing the workload of manual annotation.

[0063] Figure 1 It is a structural diagram of the SAM model in the prior art, such as Figure 1As shown in FIG. 1 , the SAM model mainly consists of three parts: image encoder, prompt encoder, and mask decoder. SAM uses a large pre-trained encoder based on the Vision Transformer (ViT). The image encoder is responsible for extracting deep feature representations from the image. The visual transformer divides the image into several small blocks and inputs these blocks into the multi-head self-attention mechanism to capture the relationship between different regions in the image. After the image encoder, the image is converted into a high-dimensional feature map, which is used to generate a segmentation mask in subsequent decoding. The prompt encoder is used to process the segmentation prompt information provided by the user, convert this information into a feature representation that the model can understand, and combine it with the image features to guide the segmentation task. The prompts of SAM can be in various forms, such as points, boxes or text prompts. In the embodiment of the present invention, point prompts are used. The mask decoder is one of the core components of SAM. It is responsible for combining image features with prompt features and generating the final segmentation mask. Its job is to segment the corresponding object area in the image based on the visual features extracted by the image encoder and the user guidance provided by the prompt encoder. The mask decoder outputs the object segmentation mask in the image.

[0064] Figure 2 The flowchart of the method for constructing the fog recognition model provided by the present invention is as follows: Figure 2 As shown, the method includes:

[0065] Step 201: Acquire ground observation station data and satellite observation data respectively.

[0066] Specifically, the ground observation station data is monitored and collected in real time through various meteorological instruments and sensors deployed on the ground. These instruments and sensors can directly measure and record key meteorological elements such as temperature, humidity, air pressure, wind speed, wind direction, and specific parameters such as visibility in the atmosphere. The satellite observation data provided in the embodiment of the present invention is FY-4A satellite observation data. FY-4A satellite observation data can perform three-dimensional monitoring of the atmosphere in a geostationary orbit, and provide continuous monitoring data for land, water, lightning, and space weather. These data are obtained through remote sensing technology, that is, using instruments on satellites to perform long-distance detection and measurement of the earth's atmosphere and surface.

[0067] The data coverage of ground observation stations is relatively small, usually only covering a limited area around the observation station. Resolution: Due to the limited number of observation stations and their uneven distribution, the spatial resolution of ground observation station data may be low, making it difficult to fully reflect the true situation of the atmosphere and the surface. Satellite observation data has a wide coverage range, covering the entire Chinese region and even the world. Resolution: Satellite observation data has high spatial and temporal resolution, and can provide more detailed and comprehensive information on the atmosphere and the surface.

[0068] Data from ground-based observation stations are usually more accurate because the observation instruments and sensors directly measure the physical parameters of the atmosphere and the surface. However, the accuracy of satellite observation data is affected by many factors, including the performance of remote sensing instruments, atmospheric conditions, surface characteristics, and data processing algorithms.

[0069] The satellite observation data provided by the present invention include: observation data with a latitude range of 30°N to 40°N and a longitude range of 110°E to 120°E. And the dates of fog occurrence between 2020 and 2022 were selected according to the fog chronicle. Within a day, the observation data within the range of 0:00 to 8:00 (UTC time) were selected at hour intervals. The data records of the ground observation station include humidity, visibility, dew point temperature, weather meteorological codes and other data related to fog formation.

[0070] Step 202: Determine the corresponding indication information based on the ground observation station data.

[0071] Specifically, an embodiment of the present invention uses a prompt encoder to extract indication information for ground observation station data; the indication information includes: whether there is fog or no fog. When processing ground observation station data, the prompt encoder can design a series of prompt sequences to guide the model to better understand and utilize these data. These prompt sequences are intended to capture the key features of ground observation station data and help the model make more accurate judgments in prediction or analysis tasks. In an embodiment of the present invention, the prompt sequence includes: humidity, visibility, dew point temperature, weather meteorological code, etc. Then, through these prompt sequences, an indication of whether there is fog or no fog is made.

[0072] According to the prompt sequence of the ground observation station data, the embodiment of the present invention defines two types of indication information according to the following formula:

[0073]

[0074] Among them, V represents visibility, H represents humidity, D represents dew point temperature, W represents weather phenomenon code, prompt=1 represents that the area where this observation station is located is "very likely to be foggy", and prompt=0 represents that the area where this observation station is located is "most likely to be foggy".

[0075] Step 203: Determine the corresponding embedded features based on the satellite observation data.

[0076] Specifically, the embodiment of the present invention extracts embedded features of satellite observation data based on an image encoder; the embedded features are used to characterize feature information corresponding to different target objects; the target objects may include at least: cloud layer, surface features; the feature information corresponding to the cloud layer may include at least: cloud type, cloud thickness, cloud top temperature, cloud temperature; the feature information corresponding to the surface features includes surface reflectivity, vegetation coverage, and surface temperature. The image encoder is used to extract deep-level feature representation of the image, which includes a visual transformer and a multi-layer transformer block. First, the image is divided into several small blocks through a visual transformer, and these small blocks are input into a transformer block. The transformer block is used to capture the relationship between different areas in the image, thereby obtaining the embedded features of the image. The key information included in the embedded features includes: texture, shape, color, etc., which provides a basis for subsequent tasks such as image segmentation.

[0077] Step 204: Generate a predicted fog mask based on the indication information and the embedded features.

[0078] Specifically, the embodiment of the present invention generates a predicted fog mask based on the indication information and the embedded features based on the mask decoder. The embodiment of the present invention inputs the final prompt (1=fog, 0=no fog) output by the prompt encoder into the mask decoder and combines it with the embedded features obtained by the image encoder to generate a predicted fog mask, that is, the fog area predicted by the model.

[0079] Step 205: Train the initial fog recognition model according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

[0080] Specifically, the embodiment of the present invention uses a loss function to train the initial fog recognition model according to the real fog mask and the predicted fog mask, thereby obtaining a trained fog recognition model. The loss function calculates the error between the real fog mask and the predicted fog mask, and transmits the error to each layer model of the model through back propagation. The model adjusts the parameters according to the error, so that the next prediction is more accurate. After training, the value of the loss function gradually decreases, the segmentation performance of the model gradually improves, and finally reaches a higher prediction accuracy.

[0081] The method for constructing a fog recognition model provided by the present invention extracts the indication information of the ground observation station data through the prompt encoder, extracts the embedded features of the satellite observation data through the image encoder, and finally generates a predicted fog mask by using the indication information and the embedded features through the mask decoder. It automatically generates a segmentation mask, quickly segments multiple objects, and avoids the tedious work of labeling one by one. Therefore, the fog recognition model constructed by the present invention greatly simplifies the labeling process of fog images and effectively improves the efficiency of model construction; at the same time, the predicted fog mask is generated by combining the acquisition of ground observation station data and satellite observation data, which also effectively improves the recognition accuracy of the model.

[0082] Furthermore, based on the ground observation station data, the corresponding indication information is determined to specifically include:

[0083] Firstly, based on the ground observation station data, the corresponding initial indication information is determined; the grayscale images of the second and thirteenth channels of the satellite observation data are extracted, and a filtering mask is generated based on the grayscale images of the second and thirteenth channels; the incorrect indication information in the initial indication is filtered out by using the filtering mask, and finally the correct indication information of the ground observation station data is obtained.

[0084] Specifically, in actual images, the results of ground observation stations often conflict with image features, and the most common situation is that high clouds block fog. That is, the observation station can detect the situation of "very likely to have fog", but due to the obstruction of high clouds, it is difficult to identify fog on satellite images (there is fog blocked by high clouds, and satellite images show the image characteristics of high clouds). The present invention visualizes all fourteen channels of FY4A. The visualization shows that in the high channel, that is, the channel with a larger visible light band, the high cloud area is obviously black, which is different from the white-gray color of other areas. The points where prompt=1 fall on these two types of areas cannot be used as a criterion for guiding the model to identify fog even if fog areas are actually formed. Therefore, for satellite data at each moment, the present invention automatically extracts the grayscale images of the second and thirteenth channels, and generates a filter mask after simple binarization and merging. Filter incorrect prompts in the initial indication information according to the filter mask. FIG3(a) is a prompt image obtained using ground observation station data; FIG3(b) is a filter mask obtained using satellite observation data; FIG3(c) is a prompt image obtained by filtering FIG3(a) using the filter mask; As shown in FIG3(a)-(b), the present invention uses the filter mask to filter the generated initial indication information, and the green dots in FIG3(a) represent positive points (prompt=1), the white dots represent negative points (prompt=0), and the blue dots represent filtered points (points filtered out by the filter mask). FIG3(b) shows the filter mask obtained after processing the satellite data, and FIG3(c) shows the correct indication information obtained after filtering.

[0085] The method for constructing a fog recognition model provided by the present invention generates a filter mask by using the grayscale images of the second and thirteenth channels of satellite observation data, and uses the filter mask to filter out incorrect indication information in the initial indication, so that the indication information extracted by using the prompt encoder has a higher accuracy, thereby further improving the recognition accuracy of the model.

[0086] Furthermore, based on the image encoder, the embedded features of the satellite observation data are extracted, including: the image encoder includes: a visual transformer, a high-frequency extractor, and a multi-layer transformer block; firstly, the second, third, and thirteenth channel grayscale images of the satellite observation data are extracted, and the second, third, and thirteenth channel grayscale images are merged to obtain a pseudo-color image; then, the high-frequency extractor is used to extract features of the pseudo-color image to obtain high-frequency features; the visual transformer is used to extract features of the pseudo-color image to obtain a first embedded feature; and the multi-layer transformer block is used to fuse the first embedded feature and the high-frequency feature in sequence to finally obtain the embedded feature.

[0087] Specifically, the embodiment of the present invention improves the existing SAM model, specifically: a high-frequency extractor is added to the image encoder, that is, the image encoder includes a visual transformer, a high-frequency extractor, and a multi-layer transformer block. At the same time, in order to facilitate the model to learn image features, the present invention selects the 2nd (0.65μm), 3rd (0.825μm), and 13th channel (12.0μm) data of the fourteen channels of FY-4A satellite data, and synthesizes them into a pseudo-color image. By inputting the pseudo-color image into the image encoder of the improved SAM model, the deep-level feature representation of the image is extracted. The visual transformer divides the image into several small blocks, and these blocks are input into the transformer block to capture the mutual relationship between different areas in the image to obtain the embedded features of the image. After the image encoder is processed by the present invention, the pseudo-color image is converted into a high-dimensional feature, and this high-dimensional feature will be used to generate a fog segmentation mask in the subsequent decoding process.

[0088] The present invention extracts high-frequency features of pseudo-color images because clouds and fog are difficult to distinguish in pseudo-color images, and feature details such as texture are not obvious in RGB space. In order to distinguish the foggy area, the present method embodiment designs a high-frequency extractor, through which features are extracted from the frequency domain as a supplement to the RGB space features. Then, the embedded features and high-frequency features of the image are used and input into the transformer block to obtain the final embedded features.

[0089] The following formula shows how to fuse image embedding features and high-frequency features using the transformer block:

[0090]

[0091] Among them, MLP is a feature linear layer used to generate specialized tasks. is an embedded feature (an embedded feature is a feature that converts image blocks into high-dimensional vector representations through a visual transformer to capture local and global information of the image), is a high-frequency feature (high-frequency features correspond to feature details such as texture in the image, helping the model to better distinguish between clouds and fog), F i refers to the fused features of each transformer block attached to the model. Figure 4 is a schematic diagram of the fusion process of fusion features, such as Figure 4 As shown in Figure 1, after the embedded features and high-frequency features are fused, they are passed through MLP to generate a feature linear layer for a specific task, and finally through the activation function and MLP to generate the fused feature F i .

[0092] The method for constructing a fog recognition model provided by the present invention utilizes a high-frequency extractor to extract high-frequency features of a pseudo-color image, and utilizes embedded features and high-frequency features of the image to obtain embedded features, thereby effectively improving the model's recognition accuracy for fog.

[0093] The design principle of the high-frequency extractor is introduced below: Taking the pseudo-color image X as input, the high-frequency extractor first transforms it from the RGB domain to the frequency domain using discrete cosine transform (DCT).

[0094] X q =D(X)

[0095] The low-frequency features of the image after DCT transformation are concentrated in the upper left corner, so the present invention adopts a high-pass filter to eliminate the low-frequency components, and then restores the filtered image to the RGB domain through inverse DCT transformation.

[0096] X h =D -1 (F(X q ,α))

[0097] Among them, F represents a high-pass filter, α is a manually designed threshold used to control the low-frequency components to be filtered out, and X h The high-frequency feature is finally obtained. The high-frequency feature extraction process is shown in Figure 5. Figure 5(a) is a schematic diagram of a pseudo-color image, Figure 5(b) is a feature map after filtering out high-frequency components after DCT transformation, and Figure 5(c) is a feature map returned to the RGB space after inverse DCT transformation.

[0098] Further, Figure 6 FIG. 1 is a flow chart of extracting embedded features of a pseudo-color image according to the present invention, as shown in FIG. Figure 6As shown, it specifically includes: fusing the first embedded feature and the high-frequency feature to obtain the first fused feature; extracting the embedded feature of the first fused feature to obtain the second embedded feature; fusing the second embedded feature and the high-frequency feature to obtain the second fused feature; extracting the embedded feature of the second fused feature again to obtain the third embedded feature; and continuously fusing the high-frequency feature with the embedded feature extracted last time to finally obtain the embedded feature.

[0099] Specifically, the embodiment of the present invention fuses the first embedded feature and the high-frequency feature to obtain the first fused feature; uses the transformer block to extract the embedded feature of the first fused feature to obtain the second embedded feature; fuses the second embedded feature and the high-frequency feature to obtain the second fused feature; uses the next layer of transformer blocks to extract the embedded feature of the second fused feature to obtain the third embedded feature; and continuously uses the next layer of transformer blocks to fuse the embedded features of the previous layer of transformer blocks and the high-frequency features to finally obtain the embedded feature.

[0100] The method for constructing a fog recognition model provided by the present invention extracts high-frequency features of a pseudo-color image and fuses the extracted high-frequency features with the embedded features extracted from each layer to obtain the final embedded features, and generates a predicted fog mask based on the final embedded features and indication information, thereby further improving the model's recognition accuracy for fog.

[0101] Furthermore, the present invention selects humidity, visibility, dew point temperature and weather code of ground observation station data as extraction objects of prompt information, which can effectively improve the accuracy of prompt information. At the same time, the range of satellite observation data selection is defined as: latitude range is 30°N to 40°N, longitude range is 110°E to 120°E, and satellite observation data within the range of 0:00 to 8:00 is selected at hourly intervals, which can improve the comprehensiveness and accuracy of satellite observation data. By combining the two to identify fog, the fog recognition accuracy can be effectively improved.

[0102] Figure 7 One of the flow charts of the fog recognition method provided by the present invention is as follows: Figure 7 As shown, the method includes:

[0103] Step 701: respectively obtaining ground observation station data and satellite observation data to be identified;

[0104] Step 702: input the ground observation station data and satellite observation data to be identified into the trained fog recognition model, and output the fog area based on the fog recognition model.

[0105] Specifically, Figure 8 The second schematic diagram of the fog identification device provided by the present invention is as follows: Figure 8 As shown in the figure, firstly, the data processing module is used to process the ground observation station data and satellite observation data to be identified, and the corresponding prompt information and embedded features are obtained; then, the prompt information and embedded features are input into the mask decoder through the network training module, and the predicted fog mask is generated through the mask decoder; finally, the test module is used to test the ground observation station data and satellite observation data to be identified, and finally the fog area is output.

[0106] The fog recognition method provided by the present invention effectively improves the fog recognition efficiency and accuracy by using a trained fog recognition model to recognize the fog area.

[0107] The following is a description of a device for constructing a fog recognition model provided by the present invention. The device for constructing a fog recognition model described below and the method for constructing a fog recognition model described above can be referred to each other. Fig. 9 A schematic diagram of the structure of the fog recognition model construction device provided by the present invention is shown in FIG. Fig. 9 As shown, the device comprises:

[0108] An acquisition unit 901 is used to acquire ground observation station data and satellite observation data respectively;

[0109] The determination unit 902 is used to determine the corresponding indication information based on the ground observation station data; the indication information includes: whether there is fog or no fog;

[0110] The determining unit 902 is further used to determine the corresponding embedding features based on the satellite observation data;

[0111] A generating unit 903, configured to generate a predicted fog mask based on the indication information and the embedded features;

[0112] The training unit 904 is used to train the initial fog recognition model according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

[0113] The present invention also provides a device for a fog identification method, comprising:

[0114] An acquisition unit, used to respectively acquire ground observation station data and satellite observation data to be identified;

[0115] The identification unit is used to input the ground observation station data and satellite observation data to be identified into the trained fog identification model, and identify the fog based on the fog identification model.

[0116] Fig.10 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.10As shown, the electronic device may include: a processor 1010, a communication interface 1020, a memory 1030 and a communication bus 1040, wherein the processor 1010, the communication interface 1020 and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 may call the logic instructions in the memory 1030 to execute the fog recognition model construction method or the fog recognition method.

[0117] In addition, the logic instructions in the above-mentioned memory 1030 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0118] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the fog recognition model construction method or the fog recognition method provided by the above methods.

[0119] 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 displayed as units may or may not be physical units, that is, 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. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0120] 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 described in each embodiment or some parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a fog recognition model, characterized in that: include: Obtain ground observation station data and satellite observation data respectively; Based on the ground observation station data, determining corresponding indication information; Based on the satellite observation data, determining the corresponding embedded features; generating a predicted fog mask based on the indication information and the embedded features; The initial fog recognition model is trained according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

2. The method for constructing a fog recognition model according to claim 1, characterized in that: The determining, based on the ground observation station data, the corresponding indication information includes: Firstly, based on the ground observation station data, the corresponding initial indication information is determined; Extracting grayscale images of the second and thirteenth channels of the satellite observation data, and generating a filtering mask based on the grayscale images of the second and thirteenth channels; The filter mask is used to filter out incorrect indication information in the initial indication, and finally the correct indication information of the ground observation station data is obtained.

3. The method for constructing a fog recognition model according to claim 1 or 2, characterized in that: The determining, based on the satellite observation data, the corresponding embedded features includes: Extracting the second, third and thirteenth channel grayscale images of the satellite observation data, and merging the second, third and thirteenth channel grayscale images to obtain a pseudo-color image; Extracting features from the pseudo-color image to obtain high-frequency features; Performing feature extraction on the pseudo-color image to obtain a first embedded feature; The first embedded feature and the high-frequency feature are sequentially fused to obtain the embedded feature.

4. The method for constructing a fog recognition model according to claim 3, characterized in that: The first embedding feature and the high-frequency feature are sequentially fused to obtain the embedding feature, comprising: Fusing the first embedded feature and the high-frequency feature to obtain a first fused feature; Extracting an embedded feature of the first fused feature to obtain a second embedded feature; Fusing the second embedded feature and the high-frequency feature to obtain a second fused feature; Extracting the embedded features of the second fused features again to obtain a third embedded feature; The high-frequency features are continuously fused with the embedded features extracted last time to finally obtain the embedded features.

5. The method for constructing a fog recognition model according to claim 3, characterized in that: The ground observation station data include: humidity, visibility, dew point temperature, weather and meteorological code; the satellite observation data is selected from the latitude range of 30°N to 40°N and the longitude range of 110°E to 120°E, with hourly intervals, and the satellite observation data is selected from the range of 0:00 to 8:

00.

6. A fog identification method, characterized in that: include: Respectively obtain the ground observation station data and satellite observation data to be identified; The ground observation station data and satellite observation data to be identified are input into a trained fog recognition model, and based on the fog recognition model, the fog area is output.

7. A device for executing the method for constructing a fog recognition model according to claim 1, characterized in that: include: An acquisition unit, used for respectively acquiring ground observation station data and satellite observation data; A determination unit, configured to determine corresponding indication information based on the ground observation station data; The determining unit is further used to determine the corresponding embedding features based on the satellite observation data; A generating unit, configured to generate a predicted fog mask based on the indication information and the embedded features; The training unit is used to train the initial fog recognition model according to the real fog mask and the predicted fog mask, so as to obtain a trained fog recognition model.

8. A device for executing the fog identification method according to claim 6, characterized in that: include: An acquisition unit, used to respectively acquire ground observation station data and satellite observation data to be identified; The identification unit is used to input the ground observation station data and satellite observation data to be identified into a trained fog identification model, and identify the fog based on the fog identification model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for constructing a fog recognition model as described in any one of claims 1 to 5 is implemented, or the fog recognition method as described in claim 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for constructing a fog recognition model as described in any one of claims 1 to 5 or the fog recognition method as described in claim 6 is implemented.

Images