A meteorological image transmission method and device based on Beidou communication

Abstract

The present invention provides a meteorological image transmission method and device based on Beidou communication, which relates to the field of artificial intelligence technology. The method includes: obtaining a meteorological detection video containing i meteorological image frames (i>1); extracting specific image descriptors of each meteorological image frame to describe the image frame; generating an overall video descriptor describing the meteorological detection video according to the specific image descriptors of the first j (j is a non-zero natural number less than i) meteorological image frames in the meteorological detection video; combining the overall video descriptor with the specific image descriptor of the i-th meteorological image frame to generate a target image descriptor for this frame; obtaining a meteorological observation result of the i-th meteorological image frame according to the target image descriptor. Embodiments of the present invention can improve the clarity of the process of obtaining meteorological observation results and enhance the credibility of obtaining meteorological observation results.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular, to a meteorological image transmission method and device based on Beidou communication. Background Art

[0002] Meteorological monitoring is crucial for many fields such as meteorological research, disaster warning, and agricultural production. Traditional meteorological monitoring mainly relies on meteorological sensors to obtain data such as temperature, humidity, and air pressure, lacking intuitive meteorological condition image information. In recent years, with the development of image acquisition technology, meteorological image monitoring has gradually emerged, but the transmission and analysis of meteorological images face many challenges.

[0003] On the one hand, meteorological image data is large. In the case of limited communication resources, such as in remote areas or harsh environments, traditional communication methods are difficult to meet the requirements of fast and stable data transmission. On the other hand, existing meteorological image analysis methods have deficiencies in aspects such as image feature extraction and the accuracy of meteorological condition recognition, and cannot efficiently and accurately obtain valuable meteorological observation results from meteorological images. In addition, there is a lack of an effective method that can combine the overall features of meteorological videos and the features of single-frame images for analysis, resulting in the comprehensiveness and credibility of meteorological image analysis needing to be improved. Therefore, there is an urgent need for a meteorological image transmission method and device based on Beidou communication to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to provide a meteorological image transmission method and device based on Beidou communication. The embodiments of the present invention are implemented as follows:

[0005] In a first aspect, an embodiment of the present invention provides a meteorological image transmission method based on Beidou communication. The method includes: collecting a meteorological detection video, where the meteorological detection video includes i meteorological image frames, and the number of meteorological image frames is greater than 1; extracting specific image descriptors respectively matched with the i meteorological image frames in the meteorological detection video, where the specific image descriptors are used to describe the meteorological image frames; obtaining an overall video descriptor based on the specific image descriptors respectively matched with the first j meteorological image frames in the meteorological detection video, where the overall video descriptor is used to describe the meteorological detection video, and j is a non-zero natural number less than i; obtaining a target image descriptor matched with the i-th meteorological image frame based on the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame; performing label assignment on the i-th meteorological image frame according to the target image descriptor matched with the i-th meteorological image frame, and determining the meteorological observation result matched with the i-th meteorological image frame; transmitting the meteorological observation result and the meteorological detection video to a receiving-end database based on the Beidou short message communication protocol.

[0006] On the other hand, an embodiment of the present invention provides a meteorological image transmission device, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, the steps in the above-mentioned method are implemented.

[0007] According to the specific image descriptors respectively matched by the first j meteorological image frames, the present invention determines the overall video descriptor matched by the meteorological detection video, and then obtains the target image descriptor matched by the i-th meteorological image frame according to the overall video descriptor and the specific image descriptor matched by the i-th meteorological image frame. This can make the target image descriptor matched by the i-th meteorological image frame not only reflect the characteristics of the i-th meteorological image frame, but also reflect the influence of the first j meteorological image frames on the i-th meteorological image frame, as well as the global characteristics of the i-th meteorological image frame in the meteorological detection video, thereby improving the clarity and accuracy of the target image descriptor, and further enhancing the understandability of the meteorological image frame classification process and the credibility of obtaining meteorological observation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic flowchart of the implementation of a meteorological image transmission method based on Beidou communication provided by an embodiment of the present invention.

[0009] Figure 2 It is a schematic diagram of the hardware entity of a meteorological image transmission device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] Figure 1 It is a schematic flowchart of the implementation of a meteorological image transmission method based on Beidou communication provided by an embodiment of the present invention. As Figure 1 shown, the method includes:

[0011] Step 100: Collect a meteorological detection video, where the meteorological detection video includes i meteorological image frames, and the number of meteorological image frames is greater than 1.

[0012] The meteorological detection video is obtained by a meteorological image transmission device through a specific video acquisition device, and this video can reflect the dynamic changes of meteorological conditions over a period of time. In the meteorological monitoring scenario, meteorological conditions change continuously over time. For example, the movement of clouds in the sky, the change of wind direction, the appearance and disappearance of precipitation, etc. These changes will be reflected in the meteorological detection video. The video acquisition device is usually installed on platforms such as meteorological monitoring stations, meteorological satellites, or unmanned aerial vehicles to ensure that the meteorological conditions can be monitored comprehensively and from multiple angles; A meteorological image frame is a static picture of the meteorological detection video at a certain moment, similar to each frame in a movie. The meteorological image frame contains rich meteorological information, such as the shape, color, and thickness of clouds, the transparency of the atmosphere, and the presence of meteorological phenomena such as precipitation and lightning. These information are extracted and processed through image analysis technology, so as to provide a basis for meteorological observation and forecasting. The number of image frames contained in the meteorological detection video is represented by i, and i is a natural number greater than 1. This indicates that the meteorological detection video is not a single image, but consists of multiple consecutive image frames. These image frames are arranged in chronological order to form a dynamic video picture. In order to ensure the quality of the collected meteorological detection video, the meteorological image transmission device also needs to preprocess the video. The preprocessing includes operations such as denoising and enhancement to remove the noise and interference in the video and improve the clarity and contrast of the image frame. The denoising operation uses methods such as mean filtering, median filtering, and Gaussian filtering, and these methods effectively remove salt-and-pepper noise, Gaussian noise, etc. in the video. The enhancement operation uses methods such as histogram equalization and adaptive histogram equalization, and these methods improve the brightness and contrast of the image frame, making the meteorological information more clearly visible.

[0013] The formula for mean filtering is: , where, is the pixel value of the filtered image at (x,y), f(x+i,y+j) is the pixel value of the original image at (x+i,y+j), M N is the size of the filtering window. The formula for median filtering is: , where, g(x,y) is the pixel value of the filtered image at (x,y), f(x+i,y+j) is the pixel value of the original image at (x+i,y+j), M N is the size of the filtering window, indicating the median operation. The formula for histogram equalization is: , where, is the gray level of the equalized image, is the gray level of the original image, L is the number of gray levels of the image, is the probability density function of the gray level of the original image.

[0014] Step 200: extracting specific image descriptors that match each of i meteorological image frames in the meteorological detection video, wherein the specific image descriptors are used to describe the meteorological image frames.

[0015] Meteorological image frames are the basic units in meteorological detection videos, similar to each frame in a movie, and contain rich meteorological information, such as the color of the sky, the shape of the clouds, the conditions of precipitation, etc. Meteorological image frames at different times will show different characteristics, which reflect the dynamic changes of meteorological conditions. For example, on a sunny day, the meteorological image frame shows a blue sky and a small amount of white clouds; before a rainstorm, the image frame will show heavy dark clouds and a gloomy sky.

[0016] A specific image descriptor is a digital representation of a meteorological image frame, which converts the complex information in the image frame into a set of representative feature vectors. These feature vectors describe the color distribution, texture features, shape information, etc. of the image frame. For example, the color descriptor uses the mean and standard deviation in the RGB color space to represent the overall color features of the image frame; the texture descriptor extracts the texture information of the image by calculating the gray-level co-occurrence matrix of the image frame; the shape descriptor uses the contour, perimeter, area and other parameters of the object in the image frame to describe the shape features of the object.

[0017] In order to extract specific image descriptors of meteorological image frames, meteorological image transmission equipment uses a variety of technical means. The first is color feature extraction. The meteorological image transmission equipment converts the meteorological image frame into different color spaces, such as RGB, HSV, Lab, etc., and then calculates the statistical features of each channel in the color space, such as mean, standard deviation, histogram, etc. For example, in the RGB color space, the meteorological image transmission equipment calculates the values ​​of the red, green, and blue channels of each pixel, and then counts the mean and standard deviation of these three channels in the entire image frame as the color descriptor.

[0018] Before extracting specific image descriptors, the meteorological image transmission device can preprocess the meteorological image frame to improve the accuracy and stability of feature extraction. Preprocessing operations include denoising, normalization, enhancement, etc. The denoising operation removes noise interference in the image frame, such as salt and pepper noise, Gaussian noise, etc. The meteorological image transmission device uses mean filtering, median filtering, Gaussian filtering and other methods for denoising. The normalization operation adjusts the pixel values ​​of the image frame to a uniform range to eliminate the brightness and contrast differences between different image frames. The meteorological image transmission device uses linear normalization, histogram equalization and other methods for normalization. The enhancement operation improves the clarity and contrast of the image frame and highlights the important features in the image frame. The meteorological image transmission device uses contrast enhancement, sharpening and other methods for enhancement.

[0019] Step 300: Obtain an overall video descriptor based on the specific image descriptors respectively matched by the first j meteorological image frames in the meteorological detection video. The overall video descriptor is used to describe the meteorological detection video, where j is a non-zero natural number less than i.

[0020] A specific image descriptor is a feature quantization representation of a single meteorological image frame, containing various feature information such as color, texture, shape, etc. For example, for a meteorological image frame containing cumulonimbus clouds, its specific image descriptor will reflect features such as darker cloud color, rough texture, and irregular shape. The overall video descriptor, on the basis of these specific image descriptors, comprehensively considers the relationships and changes among the first j image frames, thereby generating a descriptor that can represent the overall features of this video segment.

[0021] To obtain the overall video descriptor, the meteorological image transmission device adopts various technical means. An optional method is based on weighted averaging. According to the importance or the degree of influence on the whole of each image frame, different weights are assigned to its specific image descriptor, and then weighted summation is performed. This method assumes that the contribution of each image frame to the overall video features is different, and important image frames should occupy a larger proportion in the overall description.

[0022] For example, let the specific image descriptors of the first j meteorological image frames be , and the corresponding weights be , and satisfy , then the overall video descriptor is calculated by the following formula:

[0023] In practical applications, the weights are determined by various factors. For example, according to the time sequence of the image frames, the image frames closer to the current moment have a greater impact on the overall video features, so higher weights are assigned. Also according to the content features of the image frames, such as the image frames containing important meteorological phenomena (such as heavy rain, typhoon, etc.) are assigned higher weights.

[0024] Another technical means is the method based on principal component analysis (PCA). Principal component analysis is a data dimensionality reduction and feature extraction method. It converts the high-dimensional specific image descriptor data into low-dimensional principal components while retaining the main information of the data. The meteorological image transmission device combines the specific image descriptors of the first j meteorological image frames into a matrix, and then performs principal component analysis on this matrix to extract the principal components as the overall video descriptor.

[0025] Specifically, let the matrix composed of the specific image descriptors of the first j meteorological image frames be X, and its dimension be , where m is the dimension of the specific image descriptor. Through principal component analysis, the covariance matrix of matrix X is obtained , and then solve the eigenvalues and eigenvectors of the covariance matrix C. Select the eigenvectors corresponding to the top p largest eigenvalues to form the projection matrix P, and project the matrix X onto the projection matrix P to obtain the dimensionality-reduced matrix . Each column of the matrix Y is the dimensionality-reduced specific image descriptor, and by comprehensively processing these dimensionality-reduced descriptors, such as calculating the mean, the overall video descriptor can be obtained.

[0026] In addition to the weighted average and principal component analysis methods, the meteorological image transmission device also uses deep learning methods to obtain the overall video descriptor. For example, using models such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), the specific image descriptors of the first j meteorological image frames are sequentially input into the network, and the network will automatically learn the sequential relationships and features between the image frames and finally output an overall video descriptor. This method can better capture the dynamic changes and long-term dependencies in the video.

[0027] When determining the first j meteorological image frames, the meteorological image transmission device needs to make a selection according to the actual situation. The value of j is determined by factors such as the length of the meteorological detection video, the change frequency, and the required analysis accuracy. If the video length is short and the changes are rapid, a relatively large value of j is taken to fully consider the features of more image frames; if the video length is long and the changes are relatively slow, a relatively small value of j is taken to reduce the computational amount.

[0028] The overall video descriptor has important application values in meteorological monitoring and analysis. It is used for the classification and recognition of meteorological videos. By comparing the overall video descriptors of different meteorological videos, it can be determined whether they belong to the same type of meteorological phenomenon. For example, by comparing the overall video descriptors of videos containing different meteorological scenes such as sunny days, cloudy days, and heavy rains, these videos can be accurately classified. In addition, the overall video descriptor is also used for trend analysis of meteorological changes. By analyzing the changes in the overall video descriptors at different time periods, the development trend of meteorological conditions can be predicted.

[0029] After obtaining the overall video descriptor, the meteorological image transmission device stores the overall video descriptor in the database and associates it with the corresponding meteorological detection video for subsequent query and use. At the same time, to ensure the security and credibility of the data, the meteorological image transmission device also needs to back up and encrypt the stored data.

[0030] Step 400: Obtain the target image descriptor corresponding to the i-th meteorological image frame based on the overall video descriptor and the specific image descriptor matching the i-th meteorological image frame;

[0031] To obtain the target image descriptor matching the \(i\)th meteorological image frame, the meteorological image transmission device adopts a variety of technical means. An optional method is feature fusion, that is, directly splicing or weighted combining the overall video descriptor and the specific image descriptor matching the \(i\)th meteorological image frame. Direct splicing is to connect the two descriptors in sequence to form a longer feature vector. Weighted combination is to assign different weights to the overall video descriptor and the specific image descriptor matching the \(i\)th meteorological image frame according to their importance, and then perform weighted summation. If the features of the overall video are more important for the analysis of the current image frame, appropriately increase the weight of the overall video descriptor; conversely, if the features of the current image frame itself are more critical, then increase the weight of the specific image descriptor matching the \(i\)th meteorological image frame.

[0032] Another technical means is the deep learning-based method, such as using a neural network model to fuse the overall video descriptor and the specific image descriptor matching the \(i\)th meteorological image frame. Build a multi-layer perceptron (MLP) model, take the overall video descriptor and the specific image descriptor matching the \(i\)th meteorological image frame as inputs, and through the training and learning of the model, automatically find the optimal combination method between the two descriptors and output the target image descriptor. During the training process, use a large amount of meteorological image frame data and corresponding labels, and continuously adjust the parameters of the model through the backpropagation algorithm to improve the performance of the model.

[0033] Step 500: According to the target image descriptor matching the \(i\)th meteorological image frame, perform label assignment on the \(i\)th meteorological image frame to determine the meteorological observation result matching the \(i\)th meteorological image frame;

[0034] When the meteorological image transmission device performs label assignment, it adopts a variety of technical means. One of the optional methods is the classifier-based method. A classifier is a model that can divide the input feature vector into different categories according to its internal classification rules and trained parameters. The meteorological image transmission device uses a pre-trained classifier, takes the target image descriptor matching the \(i\)th meteorological image frame as the input, and the classifier outputs the meteorological category label to which the image frame belongs according to its internal classification rules and trained parameters. Optional classifiers include support vector machine (SVM), decision tree, random forest, etc.

[0035] Taking the support vector machine as an example, the support vector machine finds an optimal hyperplane in the feature space to separate samples of different categories. Let the target image descriptor be \(x\) and its corresponding meteorological category label be \(y\), (assuming a binary classification problem). The goal is to solve the following optimization problem:

[0036] ;

[0037] where \(w\) is the normal vector of the hyperplane and \(b\) is the bias term, are slack variables used to handle misclassified samples, and \(C\) is the penalty parameter used to balance the model complexity and classification error. By solving the above optimization problem, the optimal \(w\) and \(b\) are obtained. For the new target image descriptor \(x\), according to to determine its belonging class.

[0038] In practical applications, to improve the accuracy of label assignment, the meteorological image transmission device adopts the method of multi-classifier fusion. That is, multiple different types of classifiers are used to classify the target image descriptors of the \(i\)-th meteorological image frame, and then comprehensive judgment is made according to the output results of each classifier. For example, using the voting method, after each classifier classifies the image frame, the number of votes obtained by each class is counted, and the class with the most votes is the final meteorological observation result.

[0039] Suppose that in the \(i\)-th meteorological image frame obtained by the meteorological image transmission device, the sky shows a large area of grayish-white clouds, and the target image descriptor reflects that the color of the image is relatively dull, the texture is relatively uniform, and the cloud shape is relatively regular. The meteorological image transmission device uses a pre-trained support vector machine classifier, inputs the target image descriptor into the classifier, and the classifier judges that the meteorological observation result corresponding to this image frame is "cloudy" according to the parameters and rules obtained from its training.

[0040] Step 600: Transmit the meteorological observation result and the meteorological detection video to the receiving-end database based on the Beidou short message communication protocol.

[0041] When the meteorological image transmission device performs data transmission based on the Beidou short message communication protocol, it first needs to process the meteorological observation result and the meteorological detection video. Since the bandwidth of Beidou short message communication is limited and the amount of data transmitted at one time cannot be too large, the meteorological image transmission device needs to compress the meteorological detection video to reduce the amount of data. Common video compression methods include H.264, H.265, etc. These methods reduce the amount of video data by removing redundant information in the video, such as temporal redundancy and spatial redundancy. For example, for a meteorological detection video with a duration of 10 minutes and a resolution of 1080P, the original data volume reaches hundreds of megabytes, and after being compressed by H.264, the data volume will be reduced to dozens of megabytes or even smaller.

[0042] For the meteorological observation result, the meteorological image transmission device needs to convert it into a format suitable for Beidou short message communication. Usually, the meteorological observation result is encoded and represented as a sequence of numbers or characters. For example, "01" is used to represent sunny, "02" is used to represent cloudy, "03" is used to represent rainfall, etc. This reduces the length of the data and improves the transmission efficiency.

[0043] When transmitting data, the meteorological image transmission device needs to follow the regulations of the Beidou short message communication protocol. The Beidou short message communication is divided into two modes: RDSS (Radio Determination Satellite Service) and RNSS (Radio Navigation Satellite Service). The meteorological image transmission device needs to select a suitable mode according to the actual situation. In the RDSS mode, the meteorological image transmission device sends short messages to the ground control center through Beidou satellites, and the ground control center then forwards the short messages to the receiving end; in the RNSS mode, the meteorological image transmission device communicates directly with Beidou satellites to realize the sending and receiving of short messages.

[0044] When the meteorological image transmission device sends data, it needs to divide the meteorological observation results and the data of the meteorological detection video into several short message data packets, and the length of each data packet cannot exceed the maximum length specified by the Beidou short message communication protocol. For example, the maximum length of a single Beidou short message communication is 120 Chinese characters (about 240 bytes), and the meteorological image transmission device needs to divide the meteorological observation results and the compressed meteorological detection video data according to this length. At the same time, in order to ensure the integrity and accuracy of the data, the meteorological image transmission device needs to add information such as sequence numbers and check codes to each data packet. The sequence number is used to identify the order of the data packets, and the check code is used to detect whether an error occurs during the transmission of the data packets.

[0045] After receiving the short message data packets sent by the meteorological image transmission device, the receiving end database needs to perform unpacking and verification operations. The receiving end first arranges the data packets in the correct order according to the sequence numbers, and then uses the check code to verify each data packet. If the verification result shows that the data packet has no error, it combines them into the complete meteorological observation results and meteorological detection video data; if the verification result shows that the data packet has an error, it needs to send a retransmission request to the meteorological image transmission device, asking the meteorological image transmission device to resend the data packet.

[0046] To improve the credibility of data transmission, the meteorological image transmission device adopts a retransmission mechanism and an acknowledgment mechanism. The retransmission mechanism means that when the meteorological image transmission device sends a data packet, if it does not receive an acknowledgment message from the receiving end within a certain time, it is considered that the data packet transmission fails and needs to be resent. The acknowledgment mechanism means that after receiving a data packet and verifying it without error, the receiving end sends an acknowledgment message to the meteorological image transmission device, indicating that the data packet has been successfully received.

[0047] In the data transmission link, follow the Beidou short message communication protocol to ensure the reliability and compatibility of data transmission. The protocol stipulates the formats of the data header, data body, end identifier, check code, etc., as follows:

[0048] 1. The data packet header contains information such as protocol type (e.g., CCTCQ), Beidou receiver device ID, frequency point, inbound confirmation application, and coding type.

[0049] 2. The data body has different format definitions according to different data types (meteorological observation data, meteorological picture data, lost packet retransmission command, etc.). For example, the data body of meteorological observation data is a calculated meteorological data string, and the data body of meteorological picture data contains information such as time, camera number, number of data packets, etc., as well as the hexadecimal number of the picture data.

[0050] 3. The end identifier is fixed as " ", and the checksum is calculated by the checksum method. From the start identifier of the data to the end identifier (excluding the two characters), the checksum of the remaining characters is accumulated in ASCII code. The accumulated value is encoded in unsigned decimal, and the high - order overflow is taken as the low two digits.

[0051] Since the amount of data transmitted each time in Beidou short message communication is limited (the communication length range of this transmission protocol is 1 - 1000 Chinese characters, that is, at most 2000 bytes are transmitted at a time), for larger data (such as meteorological picture data), packet - splitting processing is required.

[0052] When packet - splitting, information related to the number of data packets, such as the total number of data packets and the current number of data packets, is added to the data body so that the receiving end can correctly splice the data. The data packet header contains information such as protocol type (fixed as CCTCQ), Beidou receiver device ID (used to identify the device receiving the data), frequency point (the value is 2), inbound confirmation application (set whether confirmation is required according to actual needs), coding type (such as using compression code, etc. to improve transmission efficiency), etc.

[0053] During the transmission process, the data acquisition terminal sends the packed data to the Beidou satellite, and the Beidou satellite then forwards the data to the data processing center. If the receiving end (data processing center) finds that data is lost (judged by the data packet header information and the checksum), it sends a lost packet retransmission command to the data acquisition terminal, and the data acquisition terminal re - transmits the lost data packets according to the command.

[0054] In the data receiving end, there are a data receiving and parsing module, a device control instruction generation module, a data storage module, and a data analysis module. The data receiving and parsing module receives the data transmitted by the Beidou short message communication module and parses it according to the protocol format. For meteorological observation data, it is directly stored and further analyzed; for meteorological picture data, splicing and decompression operations are first performed. The received packet data is spliced in order, and then the data is converted back to a binary stream to obtain a directly readable picture. The device control instruction generation module generates camera control commands (such as rotation direction and angle), time calibration commands (accurate to seconds), upload interval, and delayed upload commands, etc., according to user requirements or preset rules. These commands are packaged according to the Beidou short message communication protocol and sent to the data acquisition terminal through the Beidou short message communication module. The data storage module is used to classify and store meteorological observation data and pictures, and establish a meteorological database for subsequent query, analysis, and statistics. The data analysis module uses algorithms such as data mining and machine learning to deeply analyze meteorological observation data, such as predicting meteorological change trends and establishing meteorological disaster warning models.

[0055] As an implementation manner, extracting the specific image descriptors respectively matched by the i meteorological image frames in the meteorological detection video includes:

[0056] Step 210: Perform signal purification on the meteorological detection video to generate a purified meteorological detection video;

[0057] Step 220: Extract the mapping features respectively matched by the i meteorological image frames according to the purified meteorological detection video;

[0058] Step 230: Extract the specific image descriptors respectively matched by the i meteorological image frames according to the mapping features respectively matched by the i meteorological image frames.

[0059] In step 210, the meteorological image transmission device performs signal purification on the meteorological detection video to generate a purified meteorological detection video. The purpose of denoising is to remove unnecessary interference information in the video, make the image frames clearer, and facilitate subsequent feature extraction. Optional noise types include salt-and-pepper noise, Gaussian noise, etc. Salt-and-pepper noise appears as randomly distributed black and white dots in the image, while Gaussian noise makes the image appear blurred and hazy. To remove these noises, the meteorological image transmission device uses various filtering methods, such as mean filtering, median filtering, and Gaussian filtering. Gaussian filtering is a linear smoothing filtering method based on the Gaussian function. It assigns different weights according to the distance between the pixels in the neighborhood and the central pixel. The pixels closer to the central pixel have larger weights, and vice versa. The calculation formula of Gaussian filtering is:

[0060] ;

[0061] Among them is the standard deviation of the Gaussian function, which controls the smoothness of the filtering.

[0062] In step 220, the meteorological image transmission device extracts the mapping features respectively matched with i meteorological image frames based on the purified meteorological detection video. The mapping features can reflect the key feature information of the meteorological image frames, which helps subsequent operations such as classifying and identifying the image frames. The meteorological image transmission device extracts mapping features from multiple aspects, such as color features, texture features, and shape features, etc. Color feature is one of the basic features of an image, which intuitively reflects the overall color distribution of the meteorological image frame. The meteorological image transmission device converts the image frame into different color spaces, such as RGB, HSV, Lab, etc., and then calculates the statistical features of each channel in the color space, such as mean, standard deviation, histogram, etc.

[0063] Texture features reflect the distribution law and variation of gray values in the image, which are used to distinguish different types of meteorological phenomena, such as the texture of clouds, the texture of the ground, etc. The meteorological image transmission device uses methods such as gray-level co-occurrence matrix and local binary pattern (LBP) to extract the texture information of the image frame. The gray-level co-occurrence matrix is a matrix that describes the spatial relationship between gray levels in the image, which reflects features such as the texture fineness and direction of the image. The meteorological image transmission device calculates the gray-level co-occurrence matrix at different distances and angles, and then extracts the statistical features of the matrix, such as contrast, correlation, energy, and homogeneity, etc., as texture descriptors. The local binary pattern generates a binary pattern by comparing the gray value of the central pixel with that of its neighboring pixels, and then counts the occurrence frequency of different binary patterns as texture descriptors. Shape features are used to describe the shape of objects in the meteorological image frame, such as the shape of clouds, the shape of the rainfall area, etc. The meteorological image transmission device uses edge detection algorithms, such as the Canny edge detection algorithm, to extract the edge information of the objects in the image frame, and then calculates parameters such as the contour, perimeter, and area of the objects as shape descriptors. In addition, the meteorological image transmission device also uses methods such as moment features and Fourier descriptors to describe the shape features of objects.

[0064] In step 230, the meteorological image transmission device extracts the specific image descriptors respectively matched with the i meteorological image frames according to the mapping features respectively matched with the i meteorological image frames. The specific image descriptor is a more comprehensive and representative description of the meteorological image frame. It synthesizes various feature information of the image frame and can better reflect the essential features of the image frame. The meteorological image transmission device uses the method of feature fusion to integrate mapping features such as color features, texture features, and shape features to generate specific image descriptors. An optional feature fusion method is direct splicing, that is, connecting different types of mapping features in sequence to form a longer feature vector. For example, the color descriptor, texture descriptor, and shape descriptor are connected in sequence to obtain a comprehensive feature vector as the specific image descriptor. Another method is weighted combination. According to the importance of different types of mapping features, different weights are assigned to them, and then weighted summation is performed. In the meteorological image classification task, since the color feature has a greater impact on classification, the weight of the color feature is appropriately increased. In addition, the meteorological image transmission device also uses machine learning algorithms such as principal component analysis (PCA) and linear discriminant analysis (LDA) to reduce the dimension and select features of the mapping features, remove redundant information, and improve the quality and efficiency of the specific image descriptor. The meteorological image transmission device combines the mapping features of the i meteorological image frames into a matrix, and then performs principal component analysis on the matrix to extract the principal components as the specific image descriptor. Linear discriminant analysis is a supervised feature extraction method. It maximizes the distance between different classes and minimizes the distance within the same class to find the optimal feature projection direction, thereby achieving feature dimension reduction and classification.

[0065] As an implementation manner, the meteorological image frame includes pixel data respectively matched with pixel points;

[0066] The extracting the mapping features respectively matched with the i meteorological image frames according to the purified meteorological detection video includes:

[0067] Step 221: According to the purified meteorological detection video, collect the target pixel points respectively matched with the i meteorological image frames, where the target pixel point is the pixel point with the largest pixel data among the multiple pixel points matched with the meteorological image frame;

[0068] Step 222: For each meteorological image frame, taking the target pixel point matched with the meteorological image frame as a reference point, collect X pixel points along the first diffusion region, and then collect Y pixel points along the second diffusion region to generate a pixel point cluster matched with the meteorological image frame;

[0069] Step 223: Extract the mapping features matched with the meteorological image frame according to the pixel data matched with each pixel point in the pixel point cluster matched with the meteorological image frame.

[0070] In step 221, the target pixel represents a position with significant features in the image. Pixel data usually refers to the grayscale value or color value of the pixel. In a grayscale image, the pixel data is the grayscale value of the pixel, generally ranging from 0 to 255, and the larger the value, the brighter the pixel. In a color image, the pixel data is the value of the three RGB channels. The meteorological image transmission device traverses each pixel in the meteorological image frame, compares the magnitudes of their pixel data, and finds the pixel with the largest pixel data as the target pixel. For example, in a purified meteorological image frame, there are some brighter areas in the sky due to reasons such as light reflection. During the traversal process by the meteorological image transmission device, it will be found that the grayscale value (assuming it is a grayscale image) of a pixel located in the sky area is the largest among all pixels, so this pixel is determined as the target pixel. In actual operation, the meteorological image transmission device uses a loop statement to traverse the pixel matrix of the image frame row by row and column by column, while recording the current maximum pixel data and its corresponding pixel position. Let the pixel matrix of the image frame be P(m,n), where m represents the number of rows and n represents the number of columns. During the traversal process, the meteorological image transmission device will compare the values of P(i,j) (i ranges from 0 to m - 1, j ranges from 0 to n - 1). Assume that the currently recorded maximum pixel data is , and the corresponding pixel position is . Initially = P(0,0), =(0,0). During the traversal process, if P(i,j)> , then update = P(i,j), =(i,j). After the traversal ends, is the position of the target pixel.

[0071] In step 222, the first diffusion region and the second diffusion region are specific regions defined based on the target pixel. Their shapes and ranges are set according to actual requirements. For example, the first diffusion region is a small square region centered on the target pixel with a side length of a, and X pixels are collected within this region. The second diffusion region is a larger square region that expands outward based on the first diffusion region with a side length of b (b > a), and Y pixels are collected within this region. The meteorological image transmission device collects pixels in a certain order, such as from left to right and from top to bottom. Taking the target pixel as an example, when collecting pixels in the first diffusion region, the meteorological image transmission device starts from Start by collecting pixel points in row-major order until X pixel points are collected. When collecting pixel points in the second diffusion region, start from the upper left corner of this region and collect Y pixel points in a certain order. These collected pixel points form a pixel point cluster, which contains the local information around the target pixel point and can reflect the image features of this region.

[0072] In step 223, the mapping feature is a comprehensive representation of the pixel data in the pixel point cluster, which can reflect the local features of the meteorological image frame around the target pixel point. The meteorological image transmission device uses multiple methods to extract the mapping feature. One optional method is to calculate the statistical features of the pixel point cluster, such as mean, standard deviation, variance, etc. The mean reflects the average brightness or color value of the pixel point cluster, and the standard deviation reflects the degree of dispersion of the pixel data. Let the pixel data of the pixel points in the pixel point cluster be , then the mean of the pixel point cluster is , and the standard deviation is . These statistical features are part of the mapping feature. In addition to the statistical features, the meteorological image transmission device also uses histograms to extract the mapping feature.

[0073] In practical applications, for example, if there are multiple significant regions in the meteorological image frame, the meteorological image transmission device selects multiple target pixel points, respectively collects the diffusion regions and extracts the mapping features, and then fuses these mapping features to obtain more comprehensive image features. At the same time, the meteorological image transmission device needs to consider the issues of computational efficiency and storage cost to avoid consuming too many resources during the process of extracting the mapping feature.

[0074] When the meteorological image transmission device executes steps 221 - 223, it also combines image enhancement techniques to further improve the quality of the mapping feature. For example, before collecting the target pixel point, perform contrast enhancement processing on the purified meteorological image frame to make the details in the image clearer, so as to more accurately find the target pixel point. Contrast enhancement uses methods such as histogram equalization. Histogram equalization adjusts the histogram of the image to make the gray distribution of the image more uniform, thereby improving the contrast of the image.

[0075] In addition, when the meteorological image transmission device collects the pixel point clusters and extracts the mapping features, at different scales, the features in the meteorological image frame will change. The meteorological image transmission device uses multi-scale analysis methods, such as Scale-Invariant Feature Transform (SIFT) or Speeded-Up Robust Features (SURF), to extract the mapping features at different scales, and then fuses these features to improve the scale invariance of the features. For rotational invariance, the meteorological image transmission device uses rotation-invariant feature descriptors, such as Local Binary Pattern (LBP) in a circular neighborhood, which can maintain the stability of the features when the image rotates.

[0076] When the meteorological image transmission device executes steps 221-223, it matches the extracted mapping features with known meteorological patterns to further analyze the meteorological conditions. For example, it compares the extracted mapping features with the features of pre-stored meteorological patterns such as cumulonimbus clouds and sunny days. If the matching degree is high, it determines that the meteorological condition corresponding to the meteorological image frame is the corresponding pattern.

[0077] As an implementation, extracting the mapping features matched by the meteorological image frame based on the pixel data matched by each pixel point in the pixel point cluster matched by the meteorological image frame includes:

[0078] Step S10: Cut the pixel data matched by the pixel point cluster to generate a plurality of pixel regions, and the pixel data matched by Z pixel points is included in each pixel region;

[0079] Step S20: For each of the pixel regions, extract the mapping features matched by the pixel region according to the pixel data matched by the pixel region;

[0080] Step S30: Extract the mapping features matched by the meteorological image frame according to the mapping features matched by each of the plurality of pixel regions.

[0081] In step S10, the purpose of cutting the pixel data is to divide the pixel point cluster into several small and relatively independent regions, so as to analyze the pixel data of each region separately, thereby better capturing the local features of the image. The meteorological image transmission device cuts according to certain rules, such as in the order of row priority or column priority. Assuming that the pixel points in the pixel point cluster are arranged in a matrix form, the meteorological image transmission device starts from the upper left corner of the matrix and sequentially selects Z pixel points to form a pixel region until all pixel points are divided into the corresponding pixel regions. For example, if there are 100 pixel points in the pixel point cluster and Z = 10 is set, it is cut into 10 pixel regions. This cutting method is similar to dividing a large image into multiple small sub-images, and each sub-image has certain feature information.

[0082] In step S20, the mapping feature is an abstraction and generalization of the pixel data within the pixel region, which can reflect the feature information such as the color, texture, brightness, etc. of this region. The meteorological image transmission device uses multiple methods to extract the mapping features of the pixel region. An optional method is to calculate the statistical features of the pixel region, such as the mean, standard deviation, variance, etc. The mean reflects the average brightness or color value of the pixel region, and the standard deviation reflects the degree of dispersion of the pixel data.

[0083] In step S30, the mapping features of each pixel region are integrated to obtain the mapping features that can represent the local features of the entire meteorological image frame. The meteorological image transmission device uses multiple methods for integration. One method is weighted average. The weights are determined according to the importance of the pixel region. For example, the pixel regions located in the central area of the image or containing key meteorological features are assigned higher weights. Another method is to use principal component analysis (PCA). The meteorological image transmission device forms a matrix with the mapping features of each pixel region, and then performs PCA analysis on this matrix to extract the principal components as the mapping features for meteorological image frame matching.

[0084] In practical applications, for example, if there are obvious texture features in the meteorological image frame, the meteorological image transmission device uses methods more suitable for texture analysis to extract the mapping features of the pixel region, such as the gray-level co-occurrence matrix, local binary pattern, etc. The gray-level co-occurrence matrix describes the spatial relationship between gray levels within the pixel region. By calculating the statistical features of the gray-level co-occurrence matrix, such as contrast, correlation, energy, and homogeneity, etc., the meteorological image transmission device obtains the texture features of this region. The local binary pattern is to compare the gray value of the central pixel with that of its neighboring pixels to generate a binary pattern, and then count the occurrence frequencies of different binary patterns as the texture descriptor.

[0085] When the meteorological image transmission device cuts the pixel data, it also needs to consider the cutting granularity. If the cutting granularity is too small, the number of pixel points contained in each pixel region is too small to capture sufficient feature information; if the cutting granularity is too large, some local features will be lost. Therefore, the meteorological image transmission device needs to select an appropriate Z value according to factors such as the resolution and feature complexity of the meteorological image frame. For example, for a high-resolution meteorological image frame, since it contains more detailed information, a smaller Z value is selected to better capture local features; for a low-resolution image frame, a larger Z value is selected to reduce the computational amount.

[0086] As an implementation manner, extracting the specific image descriptors respectively matched by the i meteorological image frames according to the mapping features respectively matched by the i meteorological image frames includes:

[0087] Step 231: for each of the meteorological image frames, calculating the matching influence coefficient of each pixel region in the meteorological image frame;

[0088] Step 232: integrating the mapping features matched by each pixel region in the meteorological image frame according to the influence coefficient matched by each pixel region in the meteorological image frame to generate a first integrated feature;

[0089] Step 233: for each pixel region in the meteorological image frame, extract the result of the feature unit alignment multiplication between the feature of the pixel region and the feature of each target pixel region matched by the pixel region; wherein the target pixel region is the next pixel region of the pixel region in the meteorological image frame;

[0090] Step 234: integrating the bitwise multiplication results of the feature units between the pixel region and each target pixel region matched by the pixel region to generate a second integrated feature of the pixel region match;

[0091] Step 235: Integrate the second integrated features matched by each pixel region in the meteorological image frame to generate a third integrated feature;

[0092] Step 236: Integrate the first integrated feature and the third integrated feature to generate a specific image descriptor for matching the meteorological image frame.

[0093] In step 231, the influence coefficient is used to measure the importance of each pixel area in describing the characteristics of the entire meteorological image frame. Different pixel areas contain different meteorological information. For example, in a meteorological image frame containing the sky and the ground, the pixel area of ​​the sky part is more important for judging weather conditions (such as sunny, cloudy), while the pixel area of ​​the ground part also has a certain influence in some cases (such as monitoring the precipitation on the ground). The meteorological image transmission device uses a variety of methods to calculate the influence coefficient. An optional method is based on the variance of the pixel area. The variance reflects the degree of discreteness of the pixel data in the pixel area. The larger the variance, the more drastic the change of the pixel data in the area, and contains more feature information, so a higher influence coefficient is assigned.

[0094] In step 232, the mapping feature is a vector that can reflect the characteristics of the pixel region extracted in the previous step. The purpose of integration is to weight the mapping features of each pixel region according to its influence coefficient to obtain a feature vector that can better represent the entire meteorological image frame. Suppose there are n pixel regions in the meteorological image frame, and the mapping feature of the i-th pixel region is F i , the influence coefficient is , then the first integrated feature Calculated by the following formula: For example, in a meteorological image frame, there are three pixel regions, and their mapping features are respectively =[1, 2, 3], =[4, 5, 6], =[7, 8, 9], and the influence coefficients are respectively = 0.2, , then the first integrated feature: F first =0.2×[1,2,3]+0.3×[4,5,6]+0.5×[7,8,9]=[0.2+1.2+3.5,0.4+1.5+4,0.6+1.8+4.5]=[4.9,5.9,6.9] .

[0095] In step 233, the target pixel region is the next pixel region of this pixel region in the meteorological image frame. Element-wise multiplication of feature units, also known as Hadamard product, refers to multiplying the elements at the corresponding positions of two feature vectors. Let the mapping feature of the current pixel region be F current =[ f 1 , f 2 ,⋯, f m ] , and the mapping feature of the target pixel region be F target =[ g 1 , g 2 ,⋯, g m ] , then the result of their element-wise multiplication of feature units H=[ f 1 × g 1 , f 2 × g 2 ,⋯, f m × g m ] .

[0096] In step 234, the integration method adopts simple summation or weighted summation. Let the results of element-wise multiplication of feature units between the current pixel region and k target pixel regions be respectively , if the simple summation method is adopted, then the second integrated feature matched by this pixel region . For example, the results of element-wise multiplication of feature units between the current pixel region and two target pixel regions are respectively =[1, 2, 3], H 2 =[4,5,6] , then the second integrated feature matched by this pixel region F second-i =[1+4,2+5,3+6]=[5,7,9] 。

[0097] In step 235, similarly, integration is performed by means of weighted summation.

[0098] In step 236, the integration method is to splice or perform weighted summation on two feature vectors. The weights in the weighted summation are adjusted according to specific application scenarios and experimental results. If more attention is paid to the feature information of the pixel region itself, the weight of the first integrated feature is increased; if more attention is paid to the feature correlation information between adjacent pixel regions, the weight of the third integrated feature is increased.

[0099] In practical applications, for example, when calculating the influence coefficient, in addition to the method based on variance, other factors are also considered, such as the position of the pixel region, the specific meteorological features it contains, etc. For pixel regions located at the center of the image or containing key meteorological features (such as rainstorm clouds, typhoon eyes, etc.), a higher influence coefficient is given. When performing feature integration, a more complex integration method is also adopted, such as using a deep learning model for feature fusion. The first integrated feature and the third integrated feature are used as inputs and input into a multi-layer perceptron (MLP) model. Through the training and learning of the model, the optimal feature fusion method is automatically found to generate a more representative specific image descriptor.

[0100] When the meteorological image transmission device calculates the bitwise multiplication of feature units and performs feature integration, in order to improve the calculation efficiency, a parallel calculation method is adopted, and a multi-core processor or a graphics processing unit (GPU) is used for accelerated calculation. The generated specific image descriptor is compressed to reduce the occupancy of storage space.

[0101] In meteorological monitoring and analysis, by comparing the similarity between specific image descriptors of different meteorological image frames, it is judged whether they belong to the same meteorological condition. For example, the cosine similarity is used to calculate the similarity between two specific image descriptors. The higher the similarity, the more similar the meteorological conditions of the two image frames are; at the same time, when it is necessary to find other images similar to a specific meteorological image, fast retrieval is performed according to the specific image descriptor. The meteorological image transmission device stores the specific image descriptors of all meteorological images in the database. When a query image is input, its specific image descriptor is calculated and compared with the descriptors in the database to find images with higher similarity. In addition, by analyzing the changes in the specific image descriptors of meteorological image frames at different time points, the dynamic change trend of meteorological conditions is found.

[0102] As an implementation manner, the signal purification of the meteorological detection video to generate a purified meteorological detection video includes:

[0103] Step 211: Obtain the refined coverage range matching the meteorological detection video. The number of pixel points matching the refined coverage range is W, and W is an odd number;

[0104] Step 212: For the first pixel point in the meteorological detection video, use the first pixel point as the starting point, and collect the pixel data matching each of the W pixel points from within the meteorological detection video to obtain the noisy data to be refined matching the first pixel point;

[0105] Step 213: Determine the robust representative value pixel data within the noisy data to be refined matching the first pixel point as the refined data matching the first pixel point;

[0106] Step 214: Generate the refined meteorological detection video based on the refined data matching each pixel point in the meteorological detection video.

[0107] In step 211, the refined coverage range is a local area used to process pixel points in the meteorological detection video. By operating within this area, noise can be effectively removed. Selecting an odd value for W is to ensure that there is a clear central pixel point in the window, facilitating subsequent processing. For example, when W = 3, the refined coverage range is a 3×3 window, and the central pixel point is located exactly in the middle of the window. The meteorological image transmission device selects an appropriate W value according to the characteristics of the meteorological detection video and the type of noise. If the noise is relatively concentrated in a local area, a smaller W value is selected, such as 3, 5, etc., to more precisely process local noise; if the noise is widely distributed, a larger W value is selected, such as 7, 9, etc., to cover a larger area for denoising processing.

[0108] In step 212, this step is to collect pixel data within the determined purification coverage range. Taking a 3×3 purification coverage range as an example, when the first pixel point is located at the upper left corner of the image frame, the meteorological image transmission device will collect the pixel data of this pixel point and a total of 9 pixel points to its right, below, and bottom right. In actual operation, the meteorological image transmission device needs to ensure that the collected pixel points are within the valid range of the image frame. If boundary pixel points are encountered, special processing is required, such as mirror extension, zero padding, etc. Mirror extension means filling the pixel points outside the boundary with the pixel points at the symmetric positions inside the boundary, and zero padding means filling the pixel points outside the boundary with pixels having a value of 0. Let the pixel matrix of the meteorological detection video be P(m,n), where m represents the number of rows and n represents the number of columns, and the coordinates of the first pixel point be (i,j). Then, within the 3×3 purification coverage range, the collected pixel data is P(i - 1,j - 1), P(i - 1,j), P(i - 1,j + 1), P(i,j - 1), P(i,j), P(i,j + 1), P(i + 1,j - 1), P(i + 1,j), P(i + 1,j + 1) (assuming the boundary processing has been completed). These pixel data form the noisy data to be purified that matches the first pixel point, which contains noise information and requires subsequent processing.

[0109] In step 213, the robust representative value refers to the value that can represent the central tendency of a set of data and is insensitive to outliers. Optional robust representative values include the median. The median is the value located in the middle position after arranging a set of data in ascending or descending order. Taking a 3×3 purification coverage range as an example, the meteorological image transmission device arranges the 9 collected pixel data in ascending order and then takes the middle value as the purified data that matches the first pixel point. For example, if the 9 collected pixel data are [10, 20, 30, 40, 50, 60, 70, 80, 90], after arranging them in order, the middle value is 50, so 50 is the purified data that matches the first pixel point. The advantage of using the median as the robust representative value is that it can effectively resist the interference of noise, especially salt-and-pepper noise. Outliers will have a greater impact on statistics such as the average value, but have a relatively small impact on the median. Therefore, by taking the median, while removing noise, the detailed information of the image is retained.

[0110] In step 214, the meteorological image transmission device repeats the operations of step 212 and step 213 for each pixel point in the meteorological detection video to obtain the purified data matched by each pixel point. Then, these purified data are recombined into a new pixel matrix, and this pixel matrix constitutes the purified meteorological detection video. For example, for an m×n meteorological detection video, the meteorological image transmission device processes each pixel point in sequence to obtain m×n purified data, and arranges these data according to the original positional relationship to obtain the purified pixel matrix. Compared with the original video, the purified meteorological detection video has effectively removed noise, and the image frames are clearer, which is beneficial to subsequent feature extraction and analysis of meteorological image frames.

[0111] In practical applications, for example, when selecting the size of the purification coverage, it is necessary to comprehensively consider the intensity and distribution of noise as well as the detailed information of the image. If the noise intensity is large and the distribution is wide, choosing a larger purification coverage can more effectively remove noise, but it will cause the loss of detailed information of the image; if the noise intensity is small and the distribution is relatively concentrated, choosing a smaller purification coverage can better retain the image details while removing noise. The meteorological image transmission device also selects different calculation methods for robust representative values according to different types of noise. In addition to the median value, other robust statistics are also used, such as the truncated mean. The truncated mean refers to calculating the average value of the remaining data after removing a certain proportion of the maximum and minimum values in a set of data. This method can resist the influence of outliers to a certain extent while comprehensively considering more data information.

[0112] In terms of evaluating the purification effect, the meteorological image transmission device uses some evaluation indicators to measure the quality of the purified meteorological detection video. Optional evaluation indicators include the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM). The peak signal-to-noise ratio is an indicator to measure the degree of image distortion, which reflects the difference between the original image and the purified image. The calculation formula of PSNR is , where is the maximum value of the pixels in the image (for an 8-bit image, = 255), MSE is the mean square error, and the calculation formula is , is the pixel matrix of the original image, is the pixel matrix of the purified image. The higher the PSNR value, the closer the purified image is to the original image, and the better the purification effect. The structural similarity index measures the similarity of the image from three aspects: brightness, contrast, and structure of the image, and it is more in line with the human eye's perception of image quality. The calculation formula of SSIM is , where x and y are the local regions of the original image and the purified image respectively, and are the means of x and y respectively, and are the variances of x and y respectively, is the covariance of x and y, and is a constant set to avoid a zero denominator. The closer the SSIM value is to 1, the more similar the structure of the purified image is to the original image, and the better the purification effect.

[0113] When the meteorological image transmission device executes steps 211 - 214, it also performs comprehensive processing in combination with other denoising methods. For example, first use median filtering to perform preliminary denoising on the meteorological detection video, and then use more complex denoising algorithms, such as wavelet denoising, non-local mean denoising, etc., to further purify the video.

[0114] As an implementation manner, obtaining the overall video descriptor based on the specific image descriptors respectively matched by the first j meteorological image frames in the meteorological detection video includes:

[0115] Step 310: Obtain the influence coefficients respectively matched by the first j meteorological image frames according to the specific image descriptors respectively matched by the first j meteorological image frames;

[0116] Step 320: Adjust the dimensions of the specific image descriptors respectively matched by the first j meteorological image frames to generate the optimized feature vectors respectively matched by the first j meteorological image frames;

[0117] Step 330: Integrate the optimized feature vectors respectively matched by the first j meteorological image frames according to the influence coefficients respectively matched by the first j meteorological image frames to generate the overall video descriptor.

[0118] In step 310, the influence coefficient is used to measure the importance of each meteorological image frame in the process of constructing the overall video descriptor. Different meteorological image frames contain different meteorological information, and their contributions to the overall video features are also different. The meteorological image transmission device uses a variety of methods to determine the influence coefficient. An optional method is based on the time sequence. The closer the meteorological image frame is to the current moment, the greater the influence on the features of the overall video, so a higher weight is given. For example, assume that the first j meteorological image frames are arranged in time sequence, the first image frame is the farthest from the current moment, and the jth image frame is the closest to the current moment. The meteorological image transmission device uses a linearly decreasing weight assignment method. Let the influence coefficient of the ith image frame be , . For example, when j = 3, the influence coefficient of the first image frame , the influence coefficient of the second image frame , the influence coefficient of the third image frame Another method is based on the content features of image frames. For example, image frames containing important meteorological phenomena (such as heavy rain, typhoon, etc.) are given higher weights. The meteorological image transmission device analyzes specific image descriptors to identify image frames containing key meteorological information and assigns a higher influence coefficient to them.

[0119] In step 320, the dimension of the specific image descriptor is relatively high, containing a large amount of feature information, but there are some redundant or unimportant features among them. The purpose of dimension adjustment is to reduce the dimension of the features, remove redundant information, while retaining the main feature information, and improve the computational efficiency and the expressive ability of the features. The meteorological image transmission device uses methods such as principal component analysis (PCA) and linear discriminant analysis (LDA) for dimension adjustment.

[0120] In step 330, the purpose of integration is to weight and combine the optimized feature vectors of each meteorological image frame according to their influence coefficients to obtain an overall descriptor that can represent the features of the first j image frames of the entire meteorological detection video.

[0121] As an implementation manner, it is characterized in that obtaining the target image descriptor matched with the i-th meteorological image frame according to the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame includes:

[0122] Step 410: Fuse the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame to obtain the target image descriptor matched with the i-th meteorological image frame.

[0123] The meteorological image transmission device uses a variety of technical means to achieve the fusion of the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame. One optional method is weighted fusion. Weighted fusion is to assign different weights to the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame according to their importance, and then perform weighted summation. If the features of the overall video are more important for the analysis of the current image frame, it means that the current meteorological situation has a greater correlation with the meteorological trend in the previous period. For example, when judging the meteorological situation at a certain moment during a continuous rainfall process, the previous rainfall trend can provide important reference. At this time, appropriately increase the weight of the overall video descriptor; on the contrary, if the features of the current image frame itself are more critical, such as the sudden appearance of a special meteorological phenomenon (such as a rare optical phenomenon) at a certain moment, then increase the weight of the specific image descriptor matched with the i-th meteorological image frame. Another fusion method is splicing fusion. Splicing fusion is to connect the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame in sequence to form a longer feature vector.

[0124] As another implementation, the meteorological detection video is generated by clipping from the initial meteorological monitoring video; the method further includes:

[0125] Step 10: Clean the first j meteorological image frames from the initial meteorological monitoring video to generate an optimized initial meteorological monitoring video;

[0126] In a meteorological monitoring scenario, there are various problems with the first j meteorological image frames of the initial meteorological monitoring video. For example, when the meteorological monitoring device is just started, it takes a certain amount of time to reach a stable working state, and the image frames collected during this period have problems such as uneven brightness and blurred images. In addition, environmental factors such as transient smoke and dust near the monitoring area also affect the quality of the first j image frames. Suppose meteorological monitoring is carried out in a mountainous area, and when the device is started, there is a gust of mountain wind blowing, raising some dust, which will cause the sky part in the first few image frames to look gray and cannot truly reflect the meteorological conditions at that time. By cleaning these first j image frames affected by interference, the meteorological image transmission device avoids the influence of these inaccurate data on subsequent meteorological analysis.

[0127] To achieve cleaning the first j meteorological image frames from the initial meteorological monitoring video, the meteorological image transmission device first needs to determine the value of j. The value of j is determined based on various factors. One method is based on empirical values. According to past experience in the same or similar meteorological monitoring scenarios, a suitable value of j is determined. For example, in multiple meteorological monitoring of a certain area, it is found that the first 5 image frames after the device is started usually have unstable factors, then j is set to 5. Another method is to determine the value of j by preliminary analysis of the initial meteorological monitoring video. The meteorological image transmission device calculates the characteristic indexes of the first several image frames, such as brightness, contrast, image clarity, etc., and compares them with the corresponding indexes of subsequent image frames. If the indexes of the first several image frames are significantly different from those of subsequent image frames, and this difference will have an adverse impact on meteorological analysis, then these image frames are regarded as objects to be cleaned. For example, the meteorological image transmission device calculates the average brightness value of each image frame. Let the average brightness value of the kth image frame be , by comparing (n is the total number of image frames in the initial meteorological monitoring video), it is found that the average brightness value of the first j image frames fluctuates greatly, while the average brightness value tends to be stable starting from the (j + 1)th image frame, then j is determined as the number of image frames to be cleaned.

[0128] After determining the value of j, the meteorological image transmission device needs to remove the first j meteorological image frames from the initial meteorological monitoring video. This is achieved by operating on the storage structure of the video data. If the initial meteorological monitoring video is stored on a disk or other storage device in the form of a continuous image sequence, the meteorological image transmission device directly skips the first j image files and starts reading and processing data from the (j + 1)-th image file. If the video data is stored in a compressed format, such as an optional video coding format (H.264, H.265, etc.), the meteorological image transmission device needs to decode the video file, locate the positions of the first j image frames, then delete the data of these image frames from the decoded video stream, and finally re-encode the processed video stream to generate an optimized initial meteorological monitoring video.

[0129] In terms of the feature extraction of meteorological images, since the first j problematic image frames are removed, the meteorological image transmission device can more accurately extract the features of meteorological image frames, such as color features, texture features, and shape features. For example, when extracting the color feature of the sky, if the first j image frames have problems with uneven brightness, it will lead to inaccurate extraction of color features. However, the optimized video avoids this situation, making the extracted color feature better reflect the true color of the sky, thus providing a more reliable basis for judging weather conditions (such as sunny, cloudy, overcast, etc.). In terms of the analysis of meteorological change trends, the optimized video more clearly shows the dynamic changes in meteorological conditions.

[0130] Step 20: Obtain the meteorological observation results respectively matched to each remaining meteorological image frame in the optimized initial meteorological monitoring video;

[0131] In Step 20, the meteorological observation result is a summary description of the meteorological conditions reflected by the meteorological image frame, such as different meteorological types like sunny, cloudy, rainfall, snowfall, etc. The optimized initial meteorological monitoring video is the video after removing the first j meteorological image frames, and the remaining meteorological image frames are the remaining meteorological image frames.

[0132] To obtain the meteorological observation results respectively matching the legacy meteorological image frames, the meteorological image transmission device needs to rely on meteorological observation algorithms. Specifically, the meteorological image transmission device uses the algorithm to extract specific image descriptors respectively matching the legacy meteorological image frames. The specific image descriptors are used to describe the characteristics of the meteorological image frames and contain key information of the meteorological image frames, such as color distribution, texture features, etc. For example, in the meteorological image frame on a sunny day, the color of the sky part is mainly blue, and its specific image descriptor will reflect this color feature. Then, the algorithm generates an overall video descriptor based on the specific image descriptors respectively matching the first j legacy meteorological image frames in the optimized initial meteorological monitoring video. The overall video descriptor is used to describe the overall characteristics of the optimized initial meteorological monitoring video. Next, by combining the overall video descriptor and the specific image descriptor matching the current i-th legacy meteorological image frame, the target image descriptor matching the i-th legacy meteorological image frame is obtained. Finally, according to the target image descriptor matching the i-th legacy meteorological image frame, label assignment is performed on the i-th legacy meteorological image frame, so as to determine the meteorological observation result matching the i-th legacy meteorological image frame.

[0133] Step 30: Cluster each of the legacy meteorological image frames according to the meteorological observation results respectively matching them, to obtain a plurality of image frame clusters, and each of the image frame clusters corresponds to one of the meteorological observation results;

[0134] To achieve the clustering of the legacy meteorological image frames, the meteorological image transmission device needs to use the meteorological observation results respectively matching these image frames as the basis for clustering. The meteorological observation result is a key feature of the meteorological image frame, which can reflect the essential attributes of the meteorological condition represented by the image frame. For example, all the legacy meteorological image frames marked as "sunny" have similar characteristics in terms of sky color, light intensity, etc.; while the legacy meteorological image frames marked as "rainy" show similarities in features such as cloud morphology and whether there are rain streaks in the image. Based on these similarities, the meteorological image transmission device classifies the legacy meteorological image frames to form different image frame clusters.

[0135] In actual operation, meteorological image transmission equipment uses a variety of clustering algorithms to complete this task. Among them, the K-means clustering algorithm is one method. The basic idea of ​​this algorithm is to divide data points into K different clusters in an iterative manner so that the sum of the distances from each data point to the centroid of the cluster to which it belongs is minimized. In the scenario of meteorological image frame clustering, the value of K is the number of types of meteorological observation results. For example, if the meteorological observation results include sunny, cloudy, rainy, and snowy, then K is equal to 4. The meteorological image transmission device first randomly initializes K centroids, each of which represents the center of an image frame cluster. Then, for each legacy meteorological image frame, the meteorological image transmission device calculates its distance to each centroid and assigns it to the image frame cluster represented by the nearest centroid. Next, the meteorological image transmission device updates the centroid of each image frame cluster, that is, calculates the average value of the feature vectors of all legacy meteorological image frames in the cluster. Repeat the above process of assigning and updating the centroid until the centroid no longer changes significantly or reaches the preset number of iterations. Assume that the feature vector of the legacy meteorological image frame is , n is the number of legacy meteorological image frames), the centroid is , then the cluster to which each legacy meteorological image frame x_i belongs is determined by the following formula: ,in Represents the norm of a vector, usually the Euclidean norm.

[0136] In addition to the K-means clustering algorithm, the meteorological image transmission device also uses a hierarchical clustering algorithm. Hierarchical clustering algorithms are divided into two types: agglomerative and divisive. The agglomerative hierarchical clustering algorithm starts with each legacy meteorological image frame as a separate cluster, and then gradually merges similar clusters until all legacy meteorological image frames are merged into one cluster or the number of preset clusters is reached. The divisive hierarchical clustering algorithm is the opposite. It starts with all legacy meteorological image frames as a cluster, and then gradually splits the cluster into smaller clusters. In the clustering of meteorological image frames, the meteorological image transmission device selects a suitable hierarchical clustering algorithm according to the characteristics of the meteorological observation results.

[0137] Before clustering, the meteorological image transmission device needs to preprocess the features of the legacy meteorological image frames. Since the features of the meteorological image frames have different scales and ranges, this will affect the performance of the clustering algorithm. Therefore, the meteorological image transmission device normalizes the features and maps the feature values ​​to a uniform range, such as the [0, 1] interval. The normalization methods include minimum-maximum normalization and Z-score normalization. The formula for minimum-maximum normalization is: , where x is the original eigenvalue, and They are the minimum and maximum values of the features respectively, and \(x_{norm}\) is the normalized feature value. The formula for Z-score normalization is: , where is the mean of the features, is the standard deviation of the features.

[0138] The meteorological image transmission device also needs to evaluate the clustering results to ensure the quality of clustering. The evaluation metrics include the silhouette coefficient, Calinski-Harabasz index, etc. The silhouette coefficient comprehensively considers the closeness of data points to their own clusters and the separation from other clusters, and its value range is [-1, 1]. The closer the value is to 1, the better the clustering effect. The Calinski-Harabasz index evaluates the clustering effect by calculating the ratio of the within-cluster variance to the between-cluster variance. The larger the value, the better the clustering effect. The meteorological image transmission device adjusts and optimizes the clustering results according to these evaluation metrics, such as adjusting the parameters of the clustering algorithm or selecting more appropriate features for clustering.

[0139] Step 40: Select one of the legacy meteorological image frames from the multiple image frame clusters respectively multiple times, and fuse them to obtain multiple spliced meteorological images;

[0140] The meteorological image transmission device performs multiple selection and fusion operations to construct diverse spliced meteorological images to cover the feature information under different combinations of meteorological observation results, thereby providing a rich data basis for the subsequent evaluation of the clarification analysis approach. For example, in one fusion, the meteorological image transmission device selects a legacy meteorological image frame from the sunny image frame cluster and a legacy meteorological image frame from the rainfall image frame cluster, and fuses them into a spliced meteorological image containing sunny and rainfall features. Through multiple such operations, spliced meteorological images containing various combinations of meteorological observation results are obtained.

[0141] In actual operation, the meteorological image transmission device uses the following technical means to complete this process. First, for selecting legacy meteorological image frames from multiple image frame clusters, the meteorological image transmission device uses a random selection method. It sets an index range for each image frame cluster, then randomly generates an index within this range, and selects a legacy meteorological image frame from the corresponding image frame cluster according to this index.

[0142] When performing image fusion, the meteorological image transmission device adopts a pixel-based fusion method. This method is to perform weighted combination of the corresponding pixels of the selected multiple legacy meteorological image frames to generate the pixel values of the spliced meteorological image. Suppose the meteorological image transmission device selects two legacy meteorological image frames and , and their pixel values at a certain position (x, y) are and For the fused stitched meteorological image, the pixel value P(x, y) at this position is calculated by the following formula: where is a weight coefficient with a value range between [0, 1], which determines the contribution ratio of the two image frames during the fusion process. When = 0.5, it means that the two image frames have the same contribution during the fusion; when is close to 1, has a greater contribution; when is close to 0, has a greater contribution. The meteorological image transmission device adjusts the value of the weight coefficient according to actual requirements.

[0143] The meteorological image transmission device also adopts a feature-based fusion method. This method first extracts the features of the selected legacy meteorological image frames, such as color features, texture features, etc., and then performs fusion according to the importance of these features. For example, for color features, the meteorological image transmission device calculates the color histograms of the two image frames and then combines them with weights to obtain the color histogram of the stitched meteorological image. For texture features, methods such as gray-level co-occurrence matrix are used for extraction and fusion.

[0144] During multiple selections and fusions, the meteorological image transmission device needs to record the information of the legacy meteorological image frames used in each fusion to avoid fusing image frames corresponding to completely irrelevant meteorological conditions. For example, simply fusing the image frames of sunny days and heavy snow will make the stitched meteorological image meaningless.

[0145] Step 50: For multiple types of clarification analysis approaches, obtain the evaluation values of each of the clarification analysis approaches under the constraint of each of the stitched meteorological images. The evaluation values are used to evaluate the ratio of the promotion contribution value of the first legacy meteorological image frame in the stitched meteorological image to the promotion contribution value of the stitched meteorological image;

[0146] In order for the meteorological image transmission device to obtain the evaluation values of the clarification analysis approaches under the constraint of each stitched meteorological image, it is necessary to process each type of clarification analysis approach and each stitched meteorological image separately. In the embodiments of the present invention, one clarification analysis approach corresponds to one interpretability method. Taking the gradient method as an example, the feature changes of the image are analyzed by calculating the gradient values of each pixel point in the stitched meteorological image. For a stitched meteorological image, the meteorological image transmission device will use the gradient method to calculate its gradient information, and then calculate the promotion contribution values of the first legacy meteorological image frame and the entire stitched meteorological image under the constraint of meteorological observation results respectively. Suppose the promotion contribution value of the first legacy meteorological image frame is , and the promotion contribution value of the stitched meteorological image is , then the evaluation value of the gradient method under the constraint of the spliced meteorological image is calculated by the formula .

[0147] Alternatively, the integrated gradient method is another way of clarity analysis. It calculates the promotion contribution value by integrating the gradient. When the meteorological image transmission device uses the integrated gradient method, it integrates the gradient along the path from the reference image to the spliced meteorological image to obtain the promotion contribution value of each remaining meteorological image frame. The reference image is a blank image or a basic image with specific meteorological characteristics. Similarly, the meteorological image transmission device calculates the promotion contribution values of the first remaining meteorological image frame and the entire spliced meteorological image, and calculates the evaluation value of the integrated gradient method under the constraint of the spliced meteorological image according to the above formula.

[0148] Or, the class activation mapping method is adopted to display the regions in the image that make important contributions to specific meteorological observation results by generating class activation maps. When the meteorological image transmission device uses the class activation mapping method, it obtains a heat map reflecting the contribution degree of each region in the spliced meteorological image to the meteorological observation results. Then, by analyzing and quantifying the regions corresponding to the first remaining meteorological image frame and the entire spliced meteorological image in the heat map, their promotion contribution values are obtained, and then the evaluation value is calculated.

[0149] Or, the gradient class activation mapping method combines the ideas of gradient information and class activation mapping. It considers the dual factors of gradient and class activation when calculating the promotion contribution value. When the meteorological image transmission device uses this method, it first calculates the gradient and class activation information of the spliced meteorological image, and then combines these two kinds of information to calculate the promotion contribution value and the evaluation value.

[0150] In actual operation, first, for each clarity analysis approach, the meteorological image transmission device needs to preprocess the spliced meteorological image, such as normalization, denoising, etc., to improve the accuracy of the analysis. For example, when using the gradient method, normalizing the image makes the gradient calculation more stable. Then, the meteorological image transmission device uses the corresponding algorithm to implement the clarity analysis approach and calculates the promotion contribution value of each remaining meteorological image frame.

[0151] Step 60: Obtain the final evaluation value matching the clarity analysis approach according to the evaluation values of the clarity analysis approach under the constraints of each of the spliced meteorological images;

[0152] The meteorological image transmission device obtains the final evaluation value that matches the clarification analysis approach, in order to comprehensively consider the performance of this analysis approach on multiple spliced meteorological images, so as to more accurately evaluate its applicability. Since different spliced meteorological images have different meteorological characteristics and image structures, a single evaluation value cannot fully reflect the performance of the clarification analysis approach. Therefore, by calculating the final evaluation value, the meteorological image transmission device objectively compares each clarification analysis approach, providing a strong basis for determining the clarification analysis approach suitable for the initial meteorological monitoring video in the follow-up.

[0153] In actual operation, an optional method for calculating the final evaluation value adopted by the meteorological image transmission device is to calculate the average value. Suppose the evaluation values of a certain clarification analysis approach under the constraints of n spliced meteorological images are , then the final evaluation value matched by this clarification analysis approach is calculated by the following formula: . For example, if the evaluation values of the gradient method under the constraints of 5 spliced meteorological images are 0.6, 0.7, 0.5, 0.8, 0.6 respectively, then the final evaluation value of the gradient method is .

[0154] Before the meteorological image transmission device calculates the final evaluation value, since there are outliers in the evaluation values, these outliers will have a great impact on the calculation of the final evaluation value. Therefore, the meteorological image transmission device adopts methods for outlier detection and processing, such as the method based on the standard deviation. The meteorological image transmission device first calculates the mean value and the standard deviation of the evaluation values, and then regards the data points whose difference from the mean value in the evaluation values exceeds k times the standard deviation (k is usually taken as 2 or 3) as outliers, and corrects or eliminates them. For example, if the mean value of the evaluation values is 0.7, the standard deviation is 0.1, and k is taken as 3, then the points with evaluation values less than 0.4 or greater than 1 can be regarded as outliers.

[0155] Step 70: Determine the clarification analysis approach suitable for the initial meteorological monitoring video according to the final evaluation values respectively matched by various clarification analysis approaches.

[0156] When determining the applicable clarity analysis approach, a meteorological image transmission device needs to comprehensively consider the final evaluation values of various clarity analysis approaches. Generally speaking, the higher the final evaluation value, the more accurately the clarity analysis approach can evaluate the ratio of the promotion contribution value of the first remaining meteorological image frame to the promotion contribution value of the spliced meteorological image when analyzing the spliced meteorological image, and thus it is more applicable to the initial meteorological monitoring video. For example, if the final evaluation value of the gradient method is 0.8, the final evaluation value of the integrated gradient method is 0.7, the final evaluation value of the class activation mapping method is 0.6, and the final evaluation value of the gradient class activation mapping method is 0.75, then from the perspective of the final evaluation value, the gradient method is the most suitable clarity analysis approach for the initial meteorological monitoring video.

[0157] In actual operation, a method for a meteorological image transmission device to determine the clarity analysis approach is to directly select the clarity analysis approach with the highest final evaluation value. Suppose the final evaluation values of various clarity analysis approaches are respectively (m is the number of clarity analysis approaches), and the meteorological image transmission device determines the index of the clarity analysis approach with the highest final evaluation value through the formula , so as to select this approach as the applicable analysis method.

[0158] As an implementation manner, in step 50, obtaining the evaluation values of the clarity analysis approach under the constraint of each of the spliced meteorological images includes:

[0159] Step 51: For each remaining meteorological image frame in the spliced meteorological image, obtain the total promotion contribution value between the remaining meteorological image frame and the clarity analysis approach. The total promotion contribution value is the summation result of the promotion contribution values between each pixel data in the remaining meteorological image frame and the clarity analysis approach under the constraint of the meteorological observation result matched by the remaining meteorological image frame;

[0160] Step 52: Sum up the total promotion contribution values between each remaining meteorological image frame in the spliced meteorological image and the clarity analysis approach respectively to obtain a first summation result;

[0161] Step 53: Divide the total promotion contribution value between the remaining meteorological image frame and the clarity analysis approach by the first summation result to collect the evaluation value of the clarity analysis approach under the constraint of the spliced meteorological image.

[0162] In step 51, the meteorological observation result is obtained by assigning labels to the legacy meteorological image frames through a meteorological observation algorithm, representing the meteorological conditions reflected by the image frames, such as sunny, cloudy, rainy, etc. Different clarification analysis approaches calculate the promotion contribution value in different ways. Taking the gradient method as an example, this method analyzes the feature changes of an image by calculating the gradient values of each pixel point in the image. For a legacy meteorological image frame, the meteorological image transmission device calculates the gradient value of each pixel point in the image frame under the constraint of the meteorological observation result, takes these gradient values as the promotion contribution value between the pixel point and the gradient method, and then adds up the promotion contribution values of all pixel points to obtain the total promotion contribution value between the legacy meteorological image frame and the gradient method. Suppose a legacy meteorological image frame has N pixel points, and the promotion contribution value of the i-th pixel point is , then the total promotion contribution value G between the legacy meteorological image frame and the gradient method is calculated by the formula . For example, for a 100×100 pixel legacy meteorological image frame, after the meteorological image transmission device calculates the gradient value of each pixel point, it adds up these 10,000 gradient values to obtain the total promotion contribution value between the image frame and the gradient method.

[0163] In actual operation, the meteorological image transmission device uses the following technical means to calculate the total promotion contribution value. First, for the gradient method, the meteorological image transmission device uses the Sobel operator or the Prewitt operator to calculate the gradient value of the pixel point. The Sobel operator is an edge detection operator that calculates the gradient components of the pixel point in the horizontal and vertical directions by performing a convolution operation on the image, and then combines these two components to obtain the gradient value of the pixel point. The Prewitt operator is similar to the Sobel operator and is also used to calculate the gradient of the pixel point. For the integral gradient method, the meteorological image transmission device needs to integrate the gradient along the path from the reference image to the legacy meteorological image frame. The reference image is a blank image or a basic image with specific meteorological features. The meteorological image transmission device discretizes this path into multiple steps, calculates the gradient value at each step, and then integrates these gradient values to obtain the promotion contribution value of the pixel point. For the class activation mapping method, the meteorological image transmission device generates a heat map reflecting the contribution degree of each region in the image to the meteorological observation result, and then determines the promotion contribution value of the pixel point according to the value of the corresponding pixel point in the heat map. For the gradient class activation mapping method, the meteorological image transmission device combines the ideas of gradient information and class activation mapping, first calculates the gradient and class activation information of the image, and then synthesizes these two types of information to determine the promotion contribution value of the pixel point.

[0164] Step 52 is to calculate the total promotion contribution value between the entire spliced meteorological image and the clarification analysis approach. Suppose a spliced meteorological image consists of M legacy meteorological image frames, and the total promotion contribution value between the j-th legacy meteorological image frame and the clarification analysis approach is , then the first summation result T is calculated by the formula . For example, a spliced meteorological image consists of 3 legacy meteorological image frames, and their total promotion contribution values to the gradient method are 100, 150, and 200 respectively. Then the first summation result is 100 + 150 + 200 = 450.

[0165] In step 53, suppose the total promotion contribution value between the legacy meteorological image frame and the clarification analysis approach is G 1 , and the first summation result is T. Then the evaluation value E of the clarification analysis approach under the constraint of the spliced meteorological image is calculated by the formula . For example, the total promotion contribution value between the legacy meteorological image frame and the gradient method is 100, and the first summation result is 450. Then the evaluation value of the gradient method under the constraint of this spliced meteorological image is .

[0166] The evaluation value reflects the relative importance of the legacy meteorological image frame in the spliced meteorological image. The higher the evaluation value, the greater the promotion contribution of the legacy meteorological image frame to the meteorological observation result represented by the spliced meteorological image; the lower the evaluation value, the relatively smaller the promotion contribution of the legacy meteorological image frame.

[0167] As another implementation manner, the method is completed through a meteorological observation algorithm, and the method further includes:

[0168] Step a: Obtain an algorithm tuning template, where the algorithm tuning template includes i meteorological image frames, and each meteorological image frame is correspondingly labeled with a meteorological observation label, and the number of meteorological image frames is greater than 1.

[0169] The algorithm tuning template is training data for training a meteorological observation algorithm, and there are various ways for a meteorological image transmission device to obtain the algorithm tuning template. An optional way is to collect meteorological image frames from a meteorological monitoring database. The meteorological monitoring database usually stores a large amount of meteorological image data, which are collected through various meteorological monitoring devices, such as meteorological satellites, meteorological radars, ground cameras, etc. The meteorological image transmission device screens out i meteorological image frames from the database according to certain rules, for example, selects image frames at different time periods and different geographical locations to ensure the diversity and representativeness of the algorithm tuning template. During the screening process, the meteorological image transmission device also needs to add corresponding meteorological observation labels to each meteorological image frame, and these labels are manually marked according to the image content and are also automatically generated by an existing reliable meteorological observation system.

[0170] Step b: Extract specific image descriptors respectively matched with i meteorological image frames in the algorithm tuning template through the meteorological observation algorithm, where the specific image descriptors are used to describe the meteorological image frames;

[0171] Step c: Generate an overall video descriptor through the meteorological observation algorithm according to the specific image descriptors respectively matched with the first j meteorological image frames in the algorithm tuning template, where the overall video descriptor is used to describe the algorithm tuning template, and j is a non-zero natural number less than i;

[0172] Step d: Generate a target image descriptor matched with the i-th meteorological image frame according to the overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame;

[0173] Step e: Perform label assignment on the i-th meteorological image frame through the meteorological observation algorithm according to the target image descriptor matched with the i-th meteorological image frame to obtain a predicted meteorological observation result matched with the i-th meteorological image frame;

[0174] Step f: Debug the meteorological observation algorithm according to the predicted meteorological observation result and the meteorological observation label matched with the i-th meteorological image frame to generate a meteorological observation algorithm with accurate tuning, where the meteorological observation algorithm with accurate tuning is used to perform label assignment on meteorological image frames in a meteorological detection video.

[0175] The principles of Steps b to e can refer to the aforementioned Steps 200 to 600 and will not be elaborated here. When the meteorological image transmission device performs algorithm debugging, it calculates the error between the predicted meteorological observation result and the meteorological observation label. An optional error calculation method is to use a loss function. For example, a cross-entropy loss function can be adopted. Assume that there are C categories in the meteorological observation result, and the true label of the i-th meteorological image frame is ( is a C-dimensional one-hot encoded vector, where only the element corresponding to the true category is 1 and the rest are 0), and the predicted result is ( is a C-dimensional vector, and each element represents the probability that the image frame belongs to the corresponding category), then the cross-entropy loss function L is calculated through the formula , where and are respectively the c-th elements of and . For example, if the true label = [1, 0, 0] (indicating sunny) of a certain meteorological image frame, and the predicted result = [0.8, 0.1, 0.1], then the cross-entropy loss . .

[0176] After obtaining the error, the meteorological image transmission device needs to adjust the parameters of the meteorological observation algorithm according to the error. If the meteorological observation algorithm is based on a neural network, the parameter adjustment method is the backpropagation algorithm. The backpropagation algorithm calculates the gradient of the loss function with respect to each parameter in the neural network, and then updates the parameters according to the direction and magnitude of the gradient. Suppose a certain parameter in the neural network is , and the loss function is L, then the update formula for the parameter is , where is the learning rate, which controls the step size of parameter update. For example, if , , and the learning rate = 0.01, then the updated parameter .

[0177] Figure 2 FIG. [FIG. number not provided in the original] is a schematic diagram of the hardware entity of a meteorological image transmission device provided by an embodiment of the present invention. As Figure 2 shown, the hardware entity of the meteorological image transmission device A includes: a processor A1 and a memory A2. Among them, the memory A2 stores a computer program that can run on the processor A1, and when the processor A1 executes the program, it implements the steps in the method of any of the above embodiments.

[0178] The above is only the embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for transmitting meteorological images based on Beidou communication, characterized in that: The method comprises: Collecting a meteorological detection video, wherein the meteorological detection video includes i meteorological image frames, and the number of the meteorological image frames is greater than 1; Performing signal purification on the meteorological detection video to generate a purified meteorological detection video; Extracting mapping features matching each of the i meteorological image frames according to the purified meteorological detection video; For each of the meteorological image frames, calculating the influence coefficient of each pixel region in the meteorological image frame; Integrate the mapping features matched by each pixel region in the meteorological image frame according to the influence coefficient matched by each pixel region in the meteorological image frame to generate a first integrated feature; For each pixel region in the meteorological image frame, extract the result of the feature unit bitwise multiplication between the feature of the pixel region and the feature of each target pixel region matched by the pixel region; wherein the target pixel region is the next pixel region of the pixel region in the meteorological image frame; Integrate the bitwise multiplication results of the feature units between the pixel region and each target pixel region matched by the pixel region to generate a second integrated feature of the pixel region match; Integrate the second integrated features matched by each pixel region in the meteorological image frame to generate a third integrated feature; Integrate the first integrated feature and the third integrated feature to generate a specific image descriptor matched with the meteorological image frame, wherein the specific image descriptor is used to describe the meteorological image frame; According to the specific image descriptors matched by the first j meteorological image frames in the meteorological detection video, an overall video descriptor is obtained, wherein the overall video descriptor is used to describe the meteorological detection video, and j is a non-zero natural number less than i; Obtaining a target image descriptor matched to the i-th meteorological image frame according to the overall video descriptor and the specific image descriptor matched to the i-th meteorological image frame; According to the target image descriptor matched by the i-th meteorological image frame, label assignment is performed on the i-th meteorological image frame to determine the meteorological observation result matched by the i-th meteorological image frame; The meteorological observation results and the meteorological detection video are transmitted to a receiving end database based on the Beidou short message communication protocol.

2. The method for transmitting meteorological images based on Beidou communication as claimed in claim 1, characterized in that: The meteorological image frame includes pixel data of respective matching pixel points; The step of extracting the mapping features matched by each of the i meteorological image frames based on the purified meteorological detection video includes: According to the purified meteorological detection video, target pixel points that are matched by each of the i meteorological image frames are collected, and the target pixel point is the pixel point with the largest pixel data among the multiple pixel points matched by the meteorological image frame; For each of the meteorological image frames, taking the target pixel point matched by the meteorological image frame as a reference point, collecting X pixel points along the first diffusion area, and then collecting Y pixel points along the second diffusion area, to generate a pixel point cluster matched by the meteorological image frame; The mapping features of the meteorological image frame matching are extracted based on the pixel data of each pixel point matched in the pixel point cluster of the meteorological image frame matching.

3. The method for transmitting meteorological images based on Beidou communication as claimed in claim 2, characterized in that: The step of extracting the mapping features of the meteorological image frame matching based on the pixel data of each pixel point matched in the pixel point cluster of the meteorological image frame matching comprises: Cutting the pixel data matched by the pixel point cluster to generate a plurality of pixel regions, each of which contains pixel data matched by Z pixel points; For each of the pixel regions, extracting mapping features of the pixel region matches according to the pixel data matched by the pixel region; The mapping features matched by the meteorological image frame are extracted based on the mapping features matched by the plurality of pixel regions.

4. The method for transmitting meteorological images based on Beidou communication according to claim 1, characterized in that: The step of performing signal purification on the meteorological detection video to generate a purified meteorological detection video includes: Obtaining a purified coverage range matched by the meteorological detection video, wherein the number of pixel points matched by the purified coverage range is W, where W is an even number; For a first pixel point in the meteorological detection video, taking the first pixel point as a starting point, collecting pixel data matched by each of the W pixel points in the meteorological detection video, and obtaining the noisy data to be purified matched by the first pixel point; Determine the robust representative value pixel data in the noisy data to be purified matched by the first pixel point as the purified data matched by the first pixel point; The purified meteorological detection video is generated based on the purified data matched to each pixel point in the meteorological detection video.

5. The method for transmitting meteorological images based on Beidou communication according to claim 1, characterized in that: The method of obtaining the overall video descriptor based on the specific image descriptors that are matched to the first j meteorological image frames in the meteorological detection video comprises: Obtaining influence coefficients of the matches of the first j meteorological image frames according to the specific image descriptors that the first j meteorological image frames respectively match; Performing dimension adjustment on the specific image descriptors that match the first j meteorological image frames, respectively, to generate optimized representation vectors that match the first j meteorological image frames, respectively; According to the influence coefficients of the respective matches of the first j meteorological image frames, the optimized representation vectors of the respective matches of the first j meteorological image frames are integrated to generate the overall video descriptor; The step of obtaining a target image descriptor matched by the i-th meteorological image frame based on the overall video descriptor and the specific image descriptor matched by the i-th meteorological image frame comprises: The overall video descriptor and the specific image descriptor matched with the i-th meteorological image frame are fused to obtain a target image descriptor matched with the i-th meteorological image frame.

6. The method for transmitting meteorological images based on Beidou communication according to claim 1, characterized in that: The meteorological detection video is generated by editing the initial meteorological monitoring video; the method further comprises: Cleaning the first j meteorological image frames in the initial meteorological monitoring video from the initial meteorological monitoring video to generate an optimized initial meteorological monitoring video; Obtaining meteorological observation results that match each of the remaining meteorological image frames in the optimized initial meteorological monitoring video; Clustering each of the legacy meteorological image frames according to the meteorological observation results matched by each of the legacy meteorological image frames to obtain a plurality of image frame clusters, each of the image frame clusters corresponding to one of the meteorological observation results; Selecting one of the legacy meteorological image frames from the plurality of image frame clusters for multiple times and fusing them to obtain a plurality of spliced ​​meteorological images; For multiple types of clearing analysis approaches, obtaining evaluation values ​​of the clearing analysis approaches under the constraints of each of the stitched meteorological images, wherein the evaluation values ​​are used to evaluate the ratio of the promotion contribution value of the first legacy meteorological image frame in the stitched meteorological image to the promotion contribution value of the stitched meteorological image; According to the evaluation value of the clarification analysis pathway under each of the stitched meteorological image constraints, obtaining a final evaluation value of the clarification analysis pathway matching; Determining a clarification analysis approach suitable for the initial meteorological monitoring video according to the final evaluation values ​​matched by the various clarification analysis approaches; The step of obtaining the evaluation value of the clarification analysis approach under the constraints of each of the stitched meteorological images comprises: For each legacy meteorological image frame in the stitched meteorological image, a total amount of promotion contribution values ​​between the legacy meteorological image frame and the clarifying analysis path is obtained, wherein the total amount of promotion contribution values ​​is a summation result of promotion contribution values ​​between each pixel data in the legacy meteorological image frame and the clarifying analysis path under the constraint of the meteorological observation result matched by the legacy meteorological image frame; Adding the total amount of promotion contribution values ​​between each legacy meteorological image frame in the spliced ​​meteorological image and the clarification analysis path to obtain a first summation result; The total amount of the promotion contribution value between the legacy meteorological image frame and the clearing analysis approach is divided by the first summation result, and the evaluation value of the clearing analysis approach under the constraint of the spliced ​​meteorological image is collected.

7. The method for transmitting meteorological images based on Beidou communication according to claim 1, characterized in that: The method is performed by a meteorological observation algorithm, and the method further comprises: Obtain an algorithm tuning template, wherein the algorithm tuning template includes i meteorological image frames, each of the meteorological image frames is marked with a meteorological observation label, and the number of the meteorological image frames is greater than 1; Extracting specific image descriptors that match each of i meteorological image frames in the algorithm adjustment template through the meteorological observation algorithm, wherein the specific image descriptors are used to describe the meteorological image frames; Generate an overall video descriptor by using the meteorological observation algorithm according to the specific image descriptors that match the first j meteorological image frames in the algorithm training template, wherein the overall video descriptor is used to describe the algorithm training template, and j is a non-zero natural number less than i; Generate a target image descriptor for matching the i-th meteorological image frame according to the overall video descriptor and the specific image descriptor for matching the i-th meteorological image frame; Performing label assignment on the i-th meteorological image frame according to the target image descriptor matched by the i-th meteorological image frame through the meteorological observation algorithm to obtain a predicted meteorological observation result matched by the i-th meteorological image frame; The meteorological observation algorithm is debugged based on the predicted meteorological observation results and meteorological observation labels matched by the i-th meteorological image frame to generate a precisely calibrated meteorological observation algorithm, which is used to perform label assignment on meteorological image frames in meteorological detection videos.

8. A meteorological image transmission device based on Beidou communication, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images