Dust storm recognition method based on stationary satellite remote sensing data and ground-based cloud image

Abstract

The application discloses a sandstorm identification method based on stationary satellite remote sensing data and ground-based cloud images, and relates to the technical field of meteorological monitoring. The method comprises the following steps: step S10, spatial distortion correction is performed on a stationary satellite cloud image, multi-channel remote sensing data of a specified area is acquired, and standardization processing is performed; step S20, the cloud coverage area and the sandstorm area are identified by using the processed multi-channel satellite remote sensing data; and step S30, according to the identification results of the cloud coverage area and the sandstorm area, the satellite cloud image row and column numbers of the cloud coverage area are acquired, and the sandstorm area under the cloud condition is identified by using the all-sky ground-based cloud image. The application identifies the sandstorm weather of the sandstorm and the sandstorm under the cloud based on the deep learning method, and by fusing the channel data features and the time sequence features, the correlation between the fused features and the sandstorm weather is mined, so that the sandstorm identification precision of the cloud layer and the sand-dust layer boundary area based on the multi-channel satellite remote sensing data is improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring technology, and in particular to a method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images. Background Technology

[0002] Against the backdrop of carbon peaking and carbon neutrality, new power systems dominated by new energy sources exhibit stronger "weather coupling" and "system vulnerability." In recent years, due to the intensification of global climate and environmental change caused by greenhouse gas emissions, the frequency of extreme weather natural disasters has increased globally, seriously threatening the safe operation of power systems. As the proportion of new energy power generation, such as wind and solar power, continues to rise in new power systems, extreme weather events like sandstorms can hinder the output of these new energy sources, leading to power shortages and significantly increasing the risk of accidents. Currently, geostationary meteorological satellites can achieve 24-hour continuous observation, featuring a wide observation range and high observation frequency, and can identify sandstorm weather through multi-channel remote sensing data. Therefore, identifying extreme weather natural disasters such as sandstorms is of great significance for improving the emergency supply capacity of power systems.

[0003] Current technologies utilize multi-channel satellite remote sensing data to mine the characteristics of different channels and extract multi-channel data features to identify large-scale dust storms. However, when clouds are above a dust storm, or in areas where cloud signals are strong while dust layer signals are weak, the visible light and infrared channels of geostationary satellites alone cannot penetrate the clouds to identify dust storms below the clouds. Existing technologies using geostationary satellite data alone can only identify dust storms individually, making it difficult to identify dust storms in situations where there is extensive cloud cover above photovoltaic power stations or in the boundary areas between clouds and dust layers. Summary of the Invention

[0004] The problem this invention aims to solve is to provide a dust storm identification method based on geostationary satellite remote sensing data and ground-based cloud images. Utilizing multi-channel geostationary satellite remote sensing data and all-sky ground-based cloud images from photovoltaic power plants, the method identifies dust storms and sub-cloud dust storms based on deep learning. Furthermore, by fusing channel data features and temporal features, the method mines the correlation between the fused features and dust storm weather, which helps improve the accuracy of dust storm identification in cloud and dust layer boundary areas based on multi-channel satellite remote sensing data.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images, comprising the following steps: Step S10, performing spatial distortion correction on the geostationary satellite cloud image, acquiring multi-channel satellite remote sensing data of a specified area and performing standardization processing; Step S20, using the processed multi-channel satellite remote sensing data to identify cloud-covered areas and dust storm areas, obtaining dust storm areas under cloudless conditions and dust storm areas at cloud boundaries; Step S30, based on the identification results of cloud-covered areas and dust storm areas, obtaining the satellite cloud image row and column numbers of cloud-covered areas, and using all-sky ground-based cloud images to identify dust storm areas under cloud conditions.

[0006] Furthermore, step S10 includes: step S101, selecting effective channel data based on the characteristics of multi-channel geostationary satellite remote sensing data; step S102, reconstructing the cloud image according to the latitude and longitude information and the cloud image row and column number conversion method to achieve spatial distortion correction, obtaining the correspondence between image pixels and actual latitude and longitude, and then extracting multi-channel satellite remote sensing data of a specified area; step S103, standardizing the multi-channel satellite remote sensing data of the specified area to remove the influence of the solar altitude angle on the visible light three-channel data.

[0007] Further, step S20 includes: step S201, based on the multi-channel satellite remote sensing data of the specified area processed in step S10, constructing multi-time, multi-channel input attributes for all pixels at the pixel level to establish a pixel dataset; step S202, obtaining the latitude and longitude ranges of high-confidence dust storms and clouds through historical weather forecast information and cloud detection products, and converting them into satellite cloud image row and column numbers of dust storm and cloud ranges; step S203, synthesizing the visible light channel data of satellite remote sensing in the multi-channel satellite remote sensing data of the specified area into a true-color image, based on the obtained... The satellite cloud image row and column numbers of the dust storm and cloud range are obtained and compared with the true color image to classify and label the image data in the image data set, forming a dataset of various types of pixels, thereby preparing the training set and test set; in step S204, a combined model of convolutional neural network and long short-term memory neural network is established. Using the sequential connection method of the serial structure, the deep features are extracted by the convolutional neural network and input into the LSTM model for classification. The multi-class pixels in the training set are encoded with one-hot codes to realize the identification of dust storm area and cloud-covered area.

[0008] Furthermore, the method for constructing the multi-time, multi-channel input attributes includes: processing multi-channel data C of a single pixel at a single time. t ={C1,C2,…,C n} Expanded to continuous time-series multi-channel data C={C t-k ,…C t ,…,C t+k} as input attribute;

[0009] The arrangement of the multi-channel data at continuous time points is illustrated below:

[0010] .

[0011] Furthermore, metadata classifications include dust pixels, cloud pixels, and clear, cloudless pixels.

[0012] Further, step S30 includes: Step S301, based on the identification results of cloud-covered areas and dust storm areas from multi-channel satellite remote sensing data, checking whether there are dust areas around the cloud-covered areas. If there are no dust areas, it is determined to be dust storm without clouds. If there are dust areas in adjacent areas, the latitude and longitude information is calculated based on the row and column numbers of the satellite cloud image of that area; Step S302, based on step S20, obtaining the full-sky ground-based cloud image of the corresponding location and time of the dust storm weather and the full-sky ground-based cloud image of the weather without dust storms, extracting R, G, and B values ​​respectively, and using the R, G, and B values ​​and their combinations as multi-channel data, constructing multi-time, multi-channel input attributes for all pixels in units of pixels; Step S303 Step S303: Manually label pixels by comparing them with ground-based cloud images. Divide the channel data into three categories: dust pixels during sandstorms, cloud pixels during non-sandstorms, and sky pixels during non-sandstorms, forming datasets for each category, thus preparing training and testing sets. Step S304: Establish a combined model of convolutional neural network and long short-term memory neural network. Using a sequential connection method with a cascaded structure, extract deep features through the convolutional neural network and input the deep features into the LSTM model for classification. Encode the pixels of multiple categories in the training set with one-hot encoding to achieve pixel recognition. Based on the recognition results, identify sandstorm weather by different pixel counts in the all-sky ground-based cloud image.

[0013] Furthermore, by using R, G, B values ​​and their combinations as multi-channel data, the multi-time, multi-channel input attributes of all pixels are constructed on a pixel-by-pixel basis, including the following steps:

[0014] Extract the R, G, B values, red-blue ratio RBR, and green-blue ratio GBG as channel data C. t '={C R C G C B C RBR C GBG},

[0015] in, , ;

[0016] Establish a multi-channel dataset, which combines the channel data of a single pixel at a single time point C. t '={CR C G C B C RBR C GBG} Expanded to continuous time-series channel data C'={C t-k ',…C t ',…,C t+k '}, as the input attribute of the pixel at that moment,

[0017] The continuous time channel data is illustrated below:

[0018] .

[0019] Furthermore, based on the recognition results, the method for identifying dust storm weather by different pixel counts in the all-sky ground-based cloud image is as follows: through... Determine the probability of whether it is a sandstorm, where i sand i cloud andi sky These represent the number of dust pixels, cloud pixels, and sky pixels in a single full-sky ground-based cloud image.

[0020] The beneficial effects of the present invention are: (1) The dust storm weather identification method based on geostationary satellite remote sensing data and all-sky ground-based cloud map of the present invention realizes the identification of dust storm weather under cloudless conditions. By integrating the features and time series features of multi-channel remote sensing data, the correlation between the fused features and dust storm weather is fully explored. The multi-channel satellite remote sensing data is standardized, which helps to improve the identification accuracy at the boundary between the cloud layer and the dust storm area; (2) The dust storm weather identification method based on geostationary satellite remote sensing data and all-sky ground-based cloud map of the present invention uses the identification results of cloud-covered areas based on multi-channel satellite remote sensing data, and uses the latitude and longitude and satellite cloud map row and column number conversion calculation method to realize the latitude and longitude positioning of the pixel points in the cloud-covered area. The dust storm weather is identified by the all-sky ground-based cloud map of the cloud-covered area, and the identification of dust storms under the clouds is realized.

[0021] The present invention will now be described in detail with reference to the accompanying drawings. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the dust storm identification method based on geostationary satellite remote sensing data and ground-based cloud images of the present invention.

[0023] Figure 2 (a) and (b) are satellite cloud images of China before and after spatial distortion correction according to the present invention, respectively;

[0024] Figure 3 (a) and 3(b) are satellite cloud images before and after the visible light channel standardization processing in the geostationary satellite remote sensing data in this invention, respectively;

[0025] Figure 4 This is a satellite cloud image synthesized from three visible light channels for a specified area in this invention;

[0026] Figure 5 This is a schematic diagram illustrating the identification results of dust storm area pixels based on multi-channel satellite remote sensing data in this invention.

[0027] Figure 6 This invention describes the process for identifying dust storm weather based on geostationary satellite remote sensing data and all-sky ground-based cloud images. Detailed Implementation

[0028] See appendix Figure 1 and 6 This invention provides a method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images, comprising the following steps.

[0029] Step S10: Perform spatial distortion correction on the geostationary satellite cloud image, acquire multi-channel satellite remote sensing data of the specified area, and perform standardization processing.

[0030] Step S10 specifically includes the following steps.

[0031] Step S101: Based on the characteristics of the satellite remote sensing data of each channel of the FY-4A satellite and the purpose of each channel, select the effective channel data, which includes six types of channel data: visible light, short, medium and long-wave infrared, and middle and upper water vapor, totaling 14 channel data.

[0032] Step S102, as follows Figure 2 As shown in (a) and 2(b), based on the latitude and longitude information and the cloud map row and column number conversion method, the cloud map is recombined to realize spatial distortion correction and obtain the correspondence between image pixels and actual latitude and longitude. The specific calculation is as follows.

[0033] Convert geographic latitude and longitude to radian units using the following formula, and then convert geographic latitude and longitude to geocentric latitude and longitude:

[0034] (1)

[0035] (2)

[0036] in lon Geographical longitude, lat Geographic latitude; ea The long radius of the Earth ( ea=6378.137km ), eb The shortest radius of the Earth ( eb=6359.7523km ); λ e Longitude of the Earth's core φ e It is the geocentric latitude.

[0037] Calculate using geocentric latitude and longitude r e The formula is as follows:

[0038] (3)

[0039] use r e Find r 1 , r 2 , r 3 The formula is as follows:

[0040] (4)

[0041] (5)

[0042] (6)

[0043] in, h The distance from the Earth's center to the satellite's center of mass ( h=42164km ), λ D This is the longitude of the satellite's nadir point.

[0044] use r 1 , r 2 , r 3 Find r n , x , y The formula is as follows:

[0045] (7)

[0046] (8)

[0047] (9).

[0048] The formula for determining the row and column numbers corresponding to the latitude and longitude of a specified area is as follows:

[0049] (10)

[0050] (11)

[0051] in, COFF column offset ( COFF=1373.5 ), CFACFor column scaling factor ( CFAC=10233137 ), LOFF column offset ( LOFF=1373.5 ), LFAC For column scaling factor ( LFAC=10233137 ).

[0052] Using the above calculation method, the specific coordinates of the target area on a satellite cloud image can be calculated when the latitude and longitude of the target area are known. For example, Xinjiang is a region prone to sandstorms. The latitude and longitude range of the study area is longitude: 75°E to 95°E, latitude: 36°N to 48°N. Based on the latitude and longitude of each point in the matrix, the corresponding pixel in the satellite cloud image is found. The calculated row number is 70.4794-296.6124, and the column number is 766.3107-1203.5452. Based on these row and column numbers, multi-channel satellite remote sensing data of size 226×437×14 is extracted, thereby realizing the extraction of multi-channel satellite remote sensing data of a specified area.

[0053] Data from the three visible light channels are important features for identifying dust storms. Due to the influence of the solar altitude angle, the values ​​show a general difference between morning and evening and a larger difference at noon. Therefore, standardization processing is required for data quality control.

[0054] Step S103, as follows Figure 3 As shown in (a) and 3(b), the influence of the solar elevation angle on the visible light three-channel data is removed by formula (12) to eliminate the intraday difference in the pixel values ​​of the visible light cloud map.

[0055] (12)

[0056] Where i and j represent the position of the pixel in the cloud image, (i, j) represents the pixel in the i-th row and j-th column of the cloud image; v is the cloud image set number; w is the number of cloud images in the cloud image set; K is the solar constant; ρ is the albedo; α is the solar altitude angle; A is an empirical parameter that is closely related to the location of the satellite, but is currently mostly set by human experience. To correct for intraday variability in the pixel value of the cloud image, the pixel at each pixel location in the cloud image... This constitutes a preliminary cloud map after standardization.

[0057] After the above standardization process, the data from each channel have different numerical ranges as identification features. Therefore, normalization is used to reduce the value range of each channel data to 0-255. The normalization formula is as follows.

[0058] (13)

[0059] in, and For satellite remote sensingi The original and normalized values ​​of the channel, and These are the maximum and minimum values ​​in the sample.

[0060] Step S20, see Appendix Figure 5 Multi-channel satellite remote sensing data is used to identify pixels in cloud-covered areas and dust storm areas, and to identify dust storm areas under cloudless conditions and dust storm areas at cloud boundaries.

[0061] Step S20 specifically includes the following steps.

[0062] Step S201: Based on the multi-channel satellite remote sensing data of the specified area processed in step S10, construct the multi-time, multi-channel input attributes of all pixels in units of pixels, and establish a pixel dataset.

[0063] Because dust storms are continuous processes, meaning satellite remote sensing data exhibits continuity over a certain timeframe, the multi-channel data of a single pixel at a single moment is analyzed. t ={C1,C2,…,C n} Expanded to continuous time-series multi-channel data C={C t-k ,…C t ,…,C t+k} as input attributes.

[0064] The following diagram illustrates the arrangement of multi-channel data across consecutive time points:

[0065] (14).

[0066] Each pixel has n Satellite remote sensing data from each channel, taking data from consecutive time points to form... n×(2k+1) Two-dimensional data.

[0067] To identify dust pixels, data C from n channels is taken. t ={C1,C2,…,C n As the input feature for a single pixel, considering that the occurrence of a sandstorm is continuous and the channel values ​​of a single pixel do not change abruptly, channel data C={C} from consecutive time points are introduced. t-k ,…C t ,…,C t+k The value of k determines the range of consecutive time points to be introduced. The time resolution of satellite remote sensing data is 15 minutes. If k=2, then four consecutive time points are introduced, that is, data within a half-hour range before and after. Sandstorms generally last for about two hours, so k is rounded down and does not exceed 4.

[0068] Step S202: Obtain the latitude and longitude range of high-confidence dust storms and clouds through historical weather forecast information and cloud detection products, and convert it into the row and column numbers of satellite cloud images of dust storm and cloud range.

[0069] Taking Xinjiang as an example, the timing of dust storm occurrences is obtained by using historical forecasts of dust storms in Xinjiang from the national meteorological website and based on the satellite cloud image processing method in step S10. t Satellite remote sensing data C t ={C1,C2,…,C n The cloud detection products at the corresponding time are obtained through the Fengyun Satellite Remote Sensing Data Service Network. This information provides the latitude and longitude range of the high-confidence sandstorm and cloud layer. The latitude and longitude conversion method in step S102 can then be used to convert the latitude range of the specified area to... lat 1 - lat 2 Longitude range is lon 1 -lon 2 Convert to satellite cloud image row and column numbers for a specified area ( l 1 -l 2 , c 1 -c 2 ).

[0070] Step S203: Combine the satellite remote sensing visible light channel data processed in step S103 into a true-color image, such as... Figure 4 As shown, dust storms and cloud areas can be displayed in true-color images. Based on the cloud map row and column numbers of the dust storm and cloud range obtained in step S201 and compared with the true-color images, the pixels are manually labeled. The multi-channel remote sensing data is divided into three categories: dust pixels, cloud pixels, and clear sky pixels, forming datasets of each type of pixel, thereby preparing training and test sets.

[0071] Step S204: Establish a combined model of a convolutional neural network and a long short-term memory neural network. The one-dimensional convolutional neural network includes a first convolutional layer, a first pooling layer, and a second convolutional layer connected sequentially. Deep features are extracted through this convolutional neural network. The first convolutional layer has a filter size of 10 and a filter dimension of 2, the second convolutional layer has a filter size of 5 and a filter dimension of 2, and the pooling layer is a max-pooling layer. Using this sequential connection method, the deep features are input into an LSTM model for classification, with dust pixels, cloud pixels, and other pixels as classification labels.

[0072] Using the dataset constructed in step S203, the pixels in the training set are encoded using one-hot encoding, resulting in class labels of 100, 010, and 001. The constructed 1DCNN-LSTM network outputs three-dimensional data, which is then processed by a softmax activation function to produce probability values ​​between 0 and 1. The maximum probability is set to 1, and the rest are set to 0, thus enabling the recognition of dust and cloud pixels.

[0073] Step S30: Based on the identification results of cloud-covered areas and dust storm areas, obtain the satellite cloud image row and column numbers of the cloud-covered areas, and use the all-sky ground-based cloud image to identify the dust storm areas under cloud conditions.

[0074] Step S30 specifically includes the following steps.

[0075] Step S301: Based on the identification results of cloud-covered areas and dust storm areas from multi-channel satellite remote sensing data, check whether there are dust areas around the cloud-covered areas. If there are no dust areas, it is determined to be dust without clouds. If there are dust areas in adjacent areas, calculate the latitude and longitude information based on the row and column numbers of the satellite cloud image of that area. The calculation method is as follows (obtain the all-sky ground-based cloud image of the area based on the latitude and longitude information).

[0076] Calculate using the known row and column numbers of the satellite cloud image x , y The formula is as follows:

[0077] (15)

[0078] (16).

[0079] use x , y, ea, eb, h Seeking S 1 、S 2 、S 3 、S xy The formula is as follows:

[0080] (17)

[0081] (18)

[0082] (19)

[0083] (20).

[0084] use S 1、S 2 、S 3 、S xy That is, to obtain latitude and longitude, the formula is as follows:

[0085] (twenty one)

[0086] (twenty two).

[0087] Based on step S20, obtain the corresponding location and time of the sandstorm. t The full-sky ground-based cloud image and the full-sky ground-based cloud image without sandstorm weather were extracted respectively. R , G , B Value, red-blue ratio RBR and green-blue ratio GBG As channel data C t ={C R C G C B C RBR C GBG},in

[0088] (twenty three)

[0089] (twenty four).

[0090] Establish a multi-channel dataset, based on the method described in step S203, and process the channel data C of a single pixel at a single time. t '={C R C G C B C RBR C GBG} Expanded to continuous time-series channel data C'={C t-k ',…C t ',…,C t+k '} serves as the input attribute for the pixel at that moment.

[0091] The continuous time-time channel data is illustrated below:

[0092] (25).

[0093] Each pixel has 5 Data from each channel, taken from data at consecutive time points. 5×(2k+1) Two-dimensional data.

[0094] To identify dust pixels, data from 5 channels C was collected. t '={CR C G C B C RBR C GBG As the input feature for a single pixel, considering that the occurrence of a sandstorm is continuous and the channel values ​​of a single pixel do not change abruptly, channel data C'={C} from consecutive time points are introduced. t-k ',…C t ',…,C t+k The value of k determines the range of consecutive time points to be introduced. The time resolution of satellite remote sensing data is 15 minutes. If k=2, then four consecutive time points are introduced, that is, data within a half-hour range before and after. Sandstorms generally last for about two hours, so k is rounded down and does not exceed 4.

[0095] Step S303: Manually label the pixels by comparing them with the ground-based cloud image. Divide the channel data into three categories: dust pixels under dust storm weather, cloud pixels under no dust storm weather, and sky pixels under no dust storm weather. This forms a dataset of each type of pixel, thereby preparing the training set and the test set.

[0096] Step S304: A combined model of convolutional neural network and long short-term memory neural network is established. Using a sequential connection method with a cascaded structure, features are extracted through the convolutional neural network and input into the LSTM model for classification. When constructing the ground-based cloud image metadata dataset, each pixel uses 5 channels of data. Since the data sequence is relatively short in this dimension, the convolutional neural network only requires two sequentially connected layers, including a convolutional layer and a pooling layer. The convolutional layer has a filter size of 10 and a filter dimension of 2, and the pooling layer is a max-pooling layer. One-hot encoding is performed on the various classes of pixels in the training set to achieve pixel recognition. Based on the recognition results, sandstorm weather is identified by the different number of pixels in the all-sky ground-based cloud image.

[0097] Based on the above identification results, if the number of dust pixels in a full-sky ground-based cloud image is greater than or equal to the sum of cloud pixels and sky pixels, it is determined to be a dust storm; otherwise, it is not a dust storm, as shown in the following formula.

[0098] (26)

[0099] in, α This indicates the probability of a sandstorm. i sand , i cloud and i sky These represent the number of dust pixels, cloud pixels, and sky pixels in a single full-sky ground-based cloud image.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images, characterized in that, Includes the following steps: Step S10: Perform spatial distortion correction on the geostationary satellite cloud image, acquire multi-channel satellite remote sensing data of the specified area, and perform standardization processing; Step S20: Use the processed multi-channel satellite remote sensing data to identify cloud-covered areas and dust storm areas, and obtain the dust storm areas under cloudless conditions and the dust storm areas at the cloud boundary. Step S30: Based on the identification results of cloud-covered areas and dust storm areas, obtain the satellite cloud image row and column numbers of cloud-covered areas, and use the all-sky ground-based cloud image to identify dust storm areas under cloud conditions. Step S20 includes: Step S201: Based on the multi-channel satellite remote sensing data of the specified area processed in step S10, construct the multi-time and multi-channel input attributes of all pixels in units of pixels to establish a pixel dataset; Step S202: Obtain the latitude and longitude range of high-confidence dust storms and clouds through historical weather forecast information and cloud detection products, and convert it into the row and column numbers of satellite cloud images of dust storm and cloud range; Step S203: Combine the visible light channel data of satellite remote sensing in the multi-channel satellite remote sensing data of the specified area into a true color image. Based on the obtained row and column numbers of the satellite cloud image of the sandstorm and cloud range and compare it with the true color image, classify and label the image data in the image data set to form a dataset of various types of pixels, thereby preparing the training set and the test set. Step S204: Establish a combined model of convolutional neural network and long short-term memory neural network. Using the sequential connection method of the serial structure, extract deep features through convolutional neural network and input the deep features into LSTM model for classification. Encode the multi-class pixels of the training set with one-hot code to realize the identification of sandstorm areas and cloud-covered areas. Step S30 includes: Step S301: Based on the identification results of cloud-covered areas and dust storm areas from multi-channel satellite remote sensing data, check whether there are dust areas around the cloud-covered areas. If there are no dust areas, it is determined to be dust under no clouds. If there are dust areas in adjacent areas, calculate the latitude and longitude information based on the row and column numbers of the satellite cloud image of that area. Step S302: Based on step S20, obtain the full-sky ground-based cloud image of the corresponding location and time during the sandstorm weather and the full-sky ground-based cloud image without sandstorm weather, extract the R, G, and B values ​​respectively, and use the R, G, and B values ​​and their combinations as multi-channel data to construct the multi-time and multi-channel input attributes of all pixels in units of pixels; Step S303: Manually label the pixels by comparing them with the ground-based cloud image. Divide the channel data into three categories: dust pixels under sandstorm weather, cloud pixels without sandstorm weather, and sky pixels without sandstorm weather. This forms a dataset of each type of pixel, thereby preparing the training set and the test set. Step S304: Establish a combined model of convolutional neural network and long short-term memory neural network. Using the sequential connection method of serial structure, extract deep features through convolutional neural network and input the deep features into LSTM model for classification. Encode the multi-class pixels in the training set with one-hot code to realize pixel recognition. Based on the recognition results, identify sandstorm weather by different pixel counts in the all-sky ground-based cloud map.

2. The method for identifying sandstorms based on geostationary satellite remote sensing data and ground-based cloud images according to claim 1, characterized in that, Step S10 includes: Step S101: Select effective channel data based on the characteristics of multi-channel geostationary satellite remote sensing data; Step S102: Reconstruct the cloud map according to the latitude and longitude information and the cloud map row and column number conversion method to achieve spatial distortion correction, obtain the correspondence between image pixels and actual latitude and longitude, and then extract multi-channel satellite remote sensing data of the specified area; Step S103: Standardize the multi-channel satellite remote sensing data of the designated area to remove the influence of the solar elevation angle on the visible light three-channel data.

3. The method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images according to claim 1, characterized in that, The method for constructing the multi-time, multi-channel input attributes includes: C represents the multi-channel data of a single pixel at a single moment. t ={C1,C2,…,C n } Expanded to continuous time-series multi-channel data C={C t-k ,…C t ,…,C t+k } as input attribute; The arrangement of the multi-channel data at continuous time points is illustrated below: 。 4. The method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images according to claim 1, characterized in that, Metadata categories include dust pixels, cloud pixels, and clear, cloudless pixels.

5. The method for identifying sandstorms based on geostationary satellite remote sensing data and ground-based cloud images according to claim 1, characterized in that, Using R, G, B values ​​and their combinations as multi-channel data, the following steps are taken to construct multi-time, multi-channel input attributes for all pixels, on a pixel-by-pixel basis: Extract the R, G, B values, red-blue ratio RBR, and green-blue ratio GBG as channel data C. t '={C R C G C B C RBR C GBG }, in, , ; Establish a multi-channel dataset, which combines the channel data of a single pixel at a single time point C. t '={C R C G C B C RBR C GBG } Expanded to continuous time-series channel data C'={C t-k ',…C t ',…,C t+k '}, as the input attribute of the pixel at that moment, The continuous time channel data is illustrated below: 。 6. The method for identifying dust storms based on geostationary satellite remote sensing data and ground-based cloud images according to claim 1, characterized in that, The method for identifying dust storm weather based on the recognition results and different pixel counts in all-sky ground-based cloud images is as follows: pass Determine the probability of whether it is a sandstorm, where i sand i cloud andi sky These represent the number of dust pixels, cloud pixels, and sky pixels in a single full-sky ground-based cloud image.

Images