A Remote Monitoring Method and Related Devices for a Regulatory Meteorological Business System with Low Bandwidth

Abstract

The present invention relates to the field of data processing, and specifically provides a remote monitoring method and related device for a regulatory meteorological service system with low bandwidth. After obtaining a meteorological image to be detected, it is determined whether the change area of the meteorological image to be detected compared with the previous meteorological image is greater than a change threshold; alternatively, it is detected whether the meteorological image to be detected contains an abnormal detection result. When the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold, the change area is encoded, and the change area encoding result is sent to the receiving end; alternatively, when the meteorological image to be detected contains an abnormal detection result, the abnormal detection result is encoded, and the abnormal encoding result is sent to the receiving end, so that the receiving end can synthesize a synthetic meteorological image based on the change area encoding result or the abnormal encoding result and the previous meteorological image. The present invention can complete remote monitoring and protection of the meteorological service system with low bandwidth, low overhead, and high efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a remote monitoring method and related device for a regulatory meteorological business system with low bandwidth. Background Art

[0002] In the operation and management of a meteorological business system, remote monitoring plays a crucial role. It can obtain meteorological information in real time, promptly detect meteorological anomalies, and provide key support for meteorological disaster warning, meteorological resource management, etc. However, there are many deficiencies in the existing remote monitoring methods for meteorological business systems. On the one hand, in terms of data transmission, since the amount of meteorological image data is usually large, traditional transmission methods require a large amount of bandwidth resources. This will lead to problems such as slow data transmission, high latency, and even data loss in a low-bandwidth environment, seriously affecting the real-time performance and reliability of remote monitoring. On the other hand, in terms of abnormal meteorological detection, the existing detection methods often lack pertinence and efficiency. It is difficult to accurately focus on specific types of meteorological anomalies for detection, and the accuracy and reliability of the detection results need to be improved, which cannot well meet the requirements of the meteorological business system for timely and accurate detection of abnormal meteorology. In addition, some monitoring systems lack effective means in resource management and traffic control, and cannot achieve efficient and stable remote monitoring under limited bandwidth conditions, resulting in problems such as slow system response speed and poor stability, thereby affecting the normal operation of the meteorological business system and the timeliness and effectiveness of meteorological disaster warning. Therefore, there is an urgent need for a remote monitoring method for a meteorological business system that can achieve efficient remote monitoring in a low-bandwidth environment, accurately detect abnormal meteorological conditions, and effectively manage network resources. Summary of the Invention

[0003] In view of this, the present invention provides a remote monitoring method and related device for a regulatory meteorological business system with low bandwidth. The technical solution of the present invention is realized as follows:

[0004] On the one hand, the present invention provides a remote monitoring method for a regulatory meteorological service system with low bandwidth, which is applied to a remote monitoring device. The remote monitoring device is communicatively connected to a receiving end. The method includes: acquiring a meteorological image to be detected; determining whether a change area of the meteorological image to be detected compared with the previous meteorological image is greater than a change threshold; or detecting whether the meteorological image to be detected contains an abnormal detection result; when the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold, encoding the change area to obtain a change area encoding result, and sending the change area encoding result to the receiving end so that the receiving end synthesizes a synthetic meteorological image based on the change area encoding result and the previous meteorological image; or when the meteorological image to be detected contains an abnormal detection result, encoding the abnormal detection result to obtain an abnormal encoding result, and sending the abnormal encoding result to the receiving end so that the receiving end synthesizes a synthetic meteorological image based on the abnormal encoding result and the previous meteorological image.

[0005] On the other hand, the present invention provides a remote monitoring device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, the steps in the above method are implemented.

[0006] The present invention can complete a low-bandwidth, low-overhead, and efficient remote monitoring and protection meteorological service system, which only transmits the changed area or abnormal meteorological data, significantly reducing the bandwidth occupancy. At the same time, by combining the token bucket algorithm and the leaky bucket algorithm, the data transmission rate is effectively controlled to ensure efficient and stable remote monitoring under limited bandwidth and ensure the integrity and reliability of the data. It realizes efficient remote monitoring under low bandwidth and is applicable to the remote operation and management of various meteorological service systems. During the training process of the abnormal meteorological detection network, when detecting abnormal meteorology, the target detection branch can guide the abnormal detection in the meteorological training image, and the abnormal meteorological detection of the meteorological training image can focus on the expected detection direction to promote the detection of abnormal information in the meteorological training image. Then, determine the target prior label among at least one prior label matched with the meteorological training image, and the detection branch corresponding to the target prior label is the same as the target detection branch. Based on the abnormal meteorological detection result of the meteorological training image under the target detection branch and the abnormal meteorological detection result under the target detection branch, the meteorological image detection network can be trained so that the meteorological image detection network has the ability to detect abnormal meteorology under the target detection branch. The debugged meteorological image detection network is used to detect abnormal meteorology in meteorological images under the corresponding detection branch. Compared with the network without the target detection branch, the abnormal detection effect of the debugged meteorological image detection network is better. In addition, according to at least one prior label matched with the meteorological training image, the meteorological image detection network can be debugged according to multiple target detection branches, so that the meteorological image detection network can have the ability to detect multiple abnormal meteorological conditions in one debugging, saving the computing power dependence of debugging. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 FIG. is a schematic implementation flowchart of a method for remotely monitoring a low-bandwidth regulatory meteorological service system provided by an embodiment of the present invention.

[0008] Figure 2 FIG. is a schematic hardware entity diagram of a remote monitoring device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] An embodiment of the present invention provides a method for remotely monitoring a low-bandwidth regulatory meteorological service system, and this method can be executed by a processor of a remote monitoring device. The remote monitoring device is communicatively connected to at least one receiving end. Among them, the remote monitoring device can refer to devices with data processing capabilities such as servers, laptops, tablets, desktop computers, and mobile devices. The receiving end can also refer to devices with data processing capabilities such as servers, laptops, tablets, desktop computers, and mobile devices.

[0010] Figure 1Schematic diagram of the implementation process of a remote monitoring method for a low - bandwidth regulatory meteorological business system provided by an embodiment of the present invention, as Figure 1 shown. The method includes:

[0011] Step S100: Obtain the meteorological image to be detected.

[0012] Step S200: Determine whether the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold; or, detect whether the meteorological image to be detected contains an abnormal detection result.

[0013] In step S100, the remote monitoring device obtains the meteorological image to be detected. The meteorological image to be detected refers to the meteorological - related image that needs to be detected and analyzed currently. These images contain various meteorological information, such as cloud distribution, precipitation area, temperature change, etc., which are crucial for the monitoring of the meteorological business system. The remote monitoring device can obtain the meteorological image to be detected through various means, such as directly obtaining the image data captured or monitored in real - time from meteorological monitoring devices such as meteorological satellites and meteorological radars.

[0014] In step S200, the remote monitoring device determines whether the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold, or detects whether the meteorological image to be detected contains an abnormal detection result. Here, the previous meteorological image refers to the meteorological image obtained at the adjacent time point before obtaining the meteorological image to be detected. By comparing the differences between these two images, it can be judged whether the meteorological situation has changed significantly. The change threshold is a preset standard value used to measure the degree of image change. For example, when the change area exceeds a certain proportion (such as 10%) of the total area of the image, it is considered that the change area is greater than the change threshold.

[0015] In order to determine whether the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold, the remote monitoring device can adopt methods of image matching and difference analysis. Specifically, the previous meteorological image and the meteorological image to be detected can be compared at the pixel level, calculate the difference value of each pixel point, and then count the number of pixel points whose difference values exceed a certain threshold to determine the size of the change area. For example, the mean square error (MSE) method can be used to calculate the difference between the two images. The formula is: ; where m and n are the number of rows and columns of the image respectively, and I 1(i,j) and I 2(i,j) are the pixel values of the previous meteorological image and the meteorological image to be detected at the (i, j) - th pixel point respectively. The larger the MSE value calculated, the greater the difference between the two images. When the MSE value exceeds the preset change threshold, it can be judged that the change area of the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold.

[0016] In addition to determining the changed area of the image, the remote monitoring device also needs to detect whether the meteorological image to be detected contains abnormal detection results. The abnormal detection results refer to the situations in the meteorological image that do not conform to normal meteorological laws or characteristics, such as sudden heavy precipitation, abnormal high-temperature areas, etc. To detect whether the meteorological image to be detected contains abnormal detection results, the remote monitoring device can adopt machine learning or deep learning methods, and use the trained meteorological image detection network to analyze and identify the image. The meteorological image detection network can be debugged and optimized through a large number of meteorological training images to improve its detection ability for abnormal meteorological conditions. During the detection process, the remote monitoring device first obtains the meteorological image to be detected and branch prompt information. The branch prompt information is used to indicate the detection branch of the meteorological image to be detected and the corresponding number of branches of the detection branch. For example, the detection branches can include different meteorological abnormal types such as heavy precipitation detection, high-temperature detection, typhoon detection, etc., and each detection branch corresponds to a specific detection task. The remote monitoring device obtains the branch meteorological images matching the detection branches with the corresponding number of branches from at least one candidate meteorological image according to the branch prompt information, and generates a target meteorological image based on the branch meteorological images and the meteorological image to be detected. Then, through the meteorological image detection network that has been debugged, according to the detection branch indicated by the branch prompt information, the target meteorological image is subjected to abnormal meteorological detection, and the abnormal detection result of the target meteorological image under the detection branch is obtained. Taking heavy precipitation detection as an example, the remote monitoring device obtains the meteorological image to be detected and the prompt information of the heavy precipitation detection branch, obtains the branch meteorological images related to heavy precipitation detection from the candidate meteorological images, and synthesizes it with the meteorological image to be detected to form a target meteorological image. Then, the trained meteorological image detection network is used to analyze the target meteorological image to determine whether there is a heavy precipitation area in the image. If a heavy precipitation area is detected, it is considered that the meteorological image to be detected contains abnormal detection results.

[0017] Step S300: When the changed area of the meteorological image to be detected is larger than the change threshold compared with the previous meteorological image, encode the changed area to obtain a changed area encoding result, and send the changed area encoding result to the receiving end so that the receiving end can synthesize a synthetic meteorological image based on the changed area encoding result and the previous meteorological image.

[0018] In the remote monitoring method for a low-bandwidth regulatory meteorological service system, step S300 is a key operation performed by the remote monitoring device when the changed area of the meteorological image to be detected is larger than the change threshold compared with the previous meteorological image. The aim is to efficiently transmit the changed area information to reduce bandwidth occupancy while ensuring that the receiving end can synthesize an accurate meteorological image. When the remote monitoring device determines that the changed area of the meteorological image to be detected is larger than the change threshold compared with the previous meteorological image, it indicates that the meteorological situation has changed significantly. At this time, it is necessary to convey this changed information to the receiving end in a timely and accurate manner.

[0019] The changing area refers to the part of the meteorological image to be detected that has differences compared with the previous meteorological image. These differences may reflect the dynamic changes of meteorological elements, such as the movement of clouds, the expansion of precipitation areas, etc. Encoding is the process of converting the information of the changing area to make it more suitable for transmission in a low-bandwidth environment. The remote monitoring device encodes the changing area to obtain the encoding result of the changing area. This encoding process can adopt various technical means, such as lossless compression encoding and lossy compression encoding. Lossless compression encoding can reduce the data volume without losing image information. Feasible lossless compression algorithms include Huffman coding, run-length encoding, etc.

[0020] After obtaining the encoding result of the changing area, the remote monitoring device sends it to the receiving end. After receiving the encoding result of the changing area, the receiving end needs to synthesize it with the previous meteorological image to obtain the synthesized meteorological image. The synthesis process is to accurately integrate the changing information represented by the encoding result of the changing area into the previous meteorological image to reflect the current meteorological conditions. For example, assume that a certain area in the previous meteorological image shows clear sky, while clouds appear in this area in the meteorological image to be detected. After the remote monitoring device detects this changing area, it encodes and sends it to the receiving end. The receiving end adds the cloud information represented by the encoding result to the corresponding position of the previous meteorological image, thus obtaining the synthesized meteorological image containing the latest meteorological information.

[0021] When transmitting the encoding result of the changing area, in order to ensure stable and efficient data transmission in a low-bandwidth environment, the remote monitoring device uses the token bucket algorithm or the leaky bucket algorithm to control the bandwidth. The token bucket algorithm allows a certain amount of burst traffic during communication, but controls the average bandwidth by controlling the long-term average rate. When the traffic exceeds the limit, there are no tokens in the bucket, and the data packets will be discarded or delayed. For example, the token bucket adds tokens to the bucket at a fixed rate, such as adding 100 tokens per second, and each data packet needs to consume 1 token. When there is data to be sent, the remote monitoring device checks whether there are enough tokens in the token bucket. If there are, the data can be sent immediately and the corresponding tokens are consumed; if the bucket is full, the extra tokens will be discarded; if the bucket is empty, the data will need to wait until there are tokens available. This can control the average bandwidth while ensuring a certain ability to handle burst traffic. The leaky bucket algorithm strictly controls the bandwidth according to the specified traffic. Any traffic exceeding the bucket capacity will be immediately discarded. Data flows into the bucket at an irregular rate, and data flows out of the bucket at a constant rate, such as 1000 bytes per second. When the data inflow rate exceeds the output rate of the leaky bucket, the data will accumulate in the bucket. Since the bucket capacity is limited, if the maximum capacity is reached, new data will be discarded to prevent network congestion, thus achieving the goal of precise bandwidth control.

[0022] To ensure the accuracy and integrity of the coding results for the changed areas, the remote monitoring device can also adopt error detection and correction techniques during the coding and transmission processes. For example, parity bits or cyclic redundancy check (CRC) codes can be added to the coding results. After receiving the coding results, the receiving end can verify the accuracy of the data by calculating the parity bits or CRC codes. If an error is found in the data, the remote monitoring device can be requested to resend that part of the data. Additionally, the remote monitoring device can also adopt a retransmission mechanism to retransmit the data packets that are lost or damaged during the transmission process to ensure that the receiving end can receive the coding results of the changed areas completely.

[0023] Step S400: When the meteorological image to be detected contains an anomaly detection result, encode the anomaly detection result to obtain an anomaly coding result, and send the anomaly coding result to the receiving end so that the receiving end can synthesize based on the anomaly coding result and the previous meteorological image to obtain a synthesized meteorological image.

[0024] The remote monitoring device encodes the anomaly detection result to obtain an anomaly coding result. This encoding process converts the anomaly detection result into a data form that is more suitable for transmission in a low-bandwidth network environment. For example, various coding techniques can be used, such as run-length coding, arithmetic coding, etc. Run-length coding is a simple and effective lossless coding method that represents consecutive repeated data with a count value and the value of the data.

[0025] After obtaining the anomaly coding result, the remote monitoring device sends it to the receiving end. After receiving the anomaly coding result, the receiving end will perform a synthesis operation in combination with the previous meteorological image to obtain a synthesized meteorological image. The synthesis process is to accurately add the anomaly information represented by the anomaly coding result to the corresponding position of the previous meteorological image to reflect the actual situation of the current weather. For example, if the previous meteorological image shows that the weather in a certain area is normal, and the meteorological image to be detected detects an anomaly such as a heavy rain cloud cluster in that area, the remote monitoring device encodes the anomaly information and sends it to the receiving end, and the receiving end adds the heavy rain cloud cluster information represented by the coding to the corresponding area of the previous meteorological image to obtain a synthesized meteorological image containing the latest anomaly information.

[0026] As an implementation manner, in step S200, determining whether the changed area of the meteorological image to be detected is larger than the change threshold compared with the previous meteorological image includes:

[0027] Step S201: Perform image segmentation on the previous meteorological image and the meteorological image to be detected respectively to obtain the respective image segmentation sets corresponding to the previous meteorological image and the meteorological image to be detected. The image segments in the image segmentation set of the previous meteorological image and the image segmentation set of the meteorological image to be detected correspond one by one;

[0028] Step S202: Calculate the differences between two corresponding image patches respectively (such as calculating the MSE value) to obtain multiple patch difference results;

[0029] Step S203: Average the multiple patch difference results to obtain an average difference result. If the average difference result is greater than the change threshold, it is determined that the change region of the meteorological image to be detected is larger than the change threshold compared with the previous meteorological image;

[0030] Step S204: When sending the change region coding result or the abnormal coding result to the receiving end, control the bandwidth using the token bucket algorithm or the leaky bucket algorithm.

[0031] In step S201, the remote monitoring device performs image patching on the previous meteorological image and the meteorological image to be detected respectively, obtaining the image patch sets corresponding to the previous meteorological image and the meteorological image to be detected respectively, and the image patches in the image patch set of the previous meteorological image and the image patch set of the meteorological image to be detected correspond one by one. Image patching is a technical means of dividing an image into multiple small regions, aiming to simplify the image processing and analysis process. In the processing of meteorological images, since meteorological images are usually large and contain rich information, directly comparing the entire image will consume a large amount of computing resources and time. After image patching, each small block can be processed and compared independently, thereby improving the processing efficiency. For example, for a meteorological image with a resolution of 1024×1024 pixels, the remote monitoring device can divide it into small blocks of 64×64 pixels, so that the entire image is divided into (1024÷64)×(1024÷64)=256 small blocks. During the division process, the previous meteorological image and the meteorological image to be detected will be patched according to the same rules to ensure that the image patches in their image patch sets correspond one by one, facilitating subsequent difference calculation.

[0032] Step S202 requires the remote monitoring device to calculate the differences between two corresponding image patches respectively to obtain multiple patch difference results. Multiple methods can be used for this difference calculation, and it is feasible to calculate the mean square error (MSE) value. Taking a simple example to illustrate, assume there are two 3×3 pixel image patches. The pixel values of the image patch of the previous meteorological image are [10, 20, 30; 40, 50, 60; 70, 80, 90] respectively, and the pixel values of the corresponding image patch of the meteorological image to be detected are [12, 22, 32; 42, 52, 62; 72, 82, 92] respectively. According to the mean square error formula, first calculate the square of the difference of each pixel point, such as for the first pixel point (12 - 10) 2= 4, then sum the squares of the differences of all pixel points and divide by the total number of pixels 3×3 = 9 to obtain the mean square error value. The larger this mean square error value is, the greater the difference between the two image blocks. In addition to the mean square error, other difference calculation methods can also be used, such as the peak signal-to-noise ratio (PSNR), etc. The calculation formula of PSNR is: ; where, MAX I is the maximum possible value of the image pixel value. For an 8-bit image, MAX I = 255. The larger the PSNR value is, the more similar the two image blocks are and the smaller the difference is.

[0033] Step S203 is that the remote monitoring device averages the multiple block difference results to obtain the mean difference result, and determines whether the mean difference result is greater than the change threshold. The change threshold is a preset standard value used to measure the overall change degree of the meteorological image. By averaging the difference results of all blocks, an index that can reflect the overall change of the entire image can be obtained. For example, assume that after dividing a meteorological image into blocks, 100 block difference results are obtained. Add these 100 results and divide by 100 to obtain the mean difference result. If this mean difference result is greater than the preset change threshold, such as the change threshold is set to 10 and the mean difference result is 15, then the remote monitoring device determines that the changed area of the meteorological image to be detected compared to the previous meteorological image is greater than the change threshold. This means that the meteorological image has changed significantly as a whole, which may indicate a change in the meteorological situation, such as the movement of clouds, the expansion of the precipitation area, etc.

[0034] In step S204, when sending the changed area coding result to the receiving end, or sending the abnormal coding result to the receiving end, the remote monitoring device uses the token bucket algorithm or the leaky bucket algorithm to control the bandwidth.

[0035] As an implementation manner, in step S200, detecting whether the meteorological image to be detected contains an abnormal detection result includes:

[0036] Step S210: Obtain the meteorological image to be detected and the branch hint information. The branch hint information is used to indicate the detection branch of the meteorological image to be detected and the number of branches corresponding to the detection branch;

[0037] Step S220: Obtain the branch meteorological images that match the detection branches with the corresponding number of branches from at least one candidate meteorological image according to the branch hint information, and generate the target meteorological image based on the branch meteorological images and the meteorological image to be detected;

[0038] Step S230: Using the weather image detection network completed through debugging, based on the detection branch indicated by the branch hint information, perform abnormal weather detection on the target weather image to obtain the abnormal detection result of the target weather image under the detection branch.

[0039] In step S210, the remote monitoring device acquires the weather image to be detected and the branch hint information. The branch hint information is used to indicate the detection branch of the weather image to be detected and the number of branches corresponding to this detection branch. The detection branch refers to different categories or directions for abnormal detection of weather images, and each detection branch corresponds to a specific abnormal weather condition or feature. For example, the detection branches can include heavy precipitation detection, high temperature detection, typhoon detection, etc. The number of branches represents the number of branches that need to be detected currently. Suppose the remote monitoring device wants to detect the weather images of a certain area, and the branch hint information indicates that the detection branches are heavy precipitation detection and high temperature detection, and the number of branches is 2. The remote monitoring device can obtain the weather image to be detected and the branch hint information by establishing a connection with a weather database or other data sources, or by reading relevant data from local storage.

[0040] In step S220, the remote monitoring device acquires the branch weather images that match the detection branches with the corresponding number of branches from at least one candidate weather image according to the branch hint information, and generates the target weather image based on the branch weather images and the weather image to be detected. The candidate weather images are a series of pre-stored weather images, which contain the feature information under various different weather conditions. The remote monitoring device filters out the branch weather images related to these detection branches from the candidate weather images according to the detection branches in the branch hint information. For example, in the above example of heavy precipitation detection and high temperature detection, the remote monitoring device finds the images related to heavy precipitation and high temperature from the candidate weather images as the branch weather images. Then, these branch weather images and the weather image to be detected are combined to generate the target weather image. There are various combination methods. For example, the branch weather images can be superimposed on the weather image to be detected as reference information, or their features can be fused to enhance the feature information related to the detection branches in the target weather image.

[0041] In step S230, the remote monitoring device uses the meteorological image detection network that has been debugged to perform abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information, and obtains the abnormal detection result of the target meteorological image under this detection branch. The meteorological image detection network is obtained by debugging a large number of meteorological training images and can accurately identify abnormal conditions in meteorological images. During the detection process, the detection branch plays a guiding role, enabling the meteorological image detection network to focus on specific abnormal meteorological features for detection. For example, for the heavy precipitation detection branch, the meteorological image detection network will focus on features related to precipitation in the target meteorological image, such as the density and color of clouds, to determine whether there is an abnormal heavy precipitation situation.

[0042] To implement steps S210 - S230, the remote monitoring device adopts a series of technical means. When obtaining the meteorological image to be detected and the branch prompt information, data interface technology can be used to connect with meteorological monitoring devices or databases to ensure accurate data transmission. For screening branch meteorological images from candidate meteorological images, an image classification algorithm can be used to classify the candidate meteorological images according to the characteristics of the detection branch and find the matching branch meteorological images. When generating the target meteorological image, an image fusion algorithm can be used to effectively fuse the branch meteorological image and the meteorological image to be detected. In terms of abnormal meteorological detection, the meteorological image detection network can adopt deep learning algorithms, such as convolutional neural networks (CNNs), which can automatically extract feature information in the image and classify and judge abnormal situations.

[0043] As an implementation manner, step S230, using the meteorological image detection network that has been debugged to perform abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information, and obtaining the abnormal detection result of the target meteorological image under the detection branch, includes:

[0044] Step S231: The meteorological image detection network that has been debugged performs abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information to obtain the first abnormal detection result; wherein, the first abnormal detection result includes the element classification corresponding to the maximum support coefficient of each image element in the target meteorological image, and any support coefficient is used to indicate the probability that the element classification of any image element belongs to the corresponding element classification.

[0045] Step S232: Based on the element classification corresponding to the maximum support coefficient of each image element in the target meteorological image, the abnormal detection result of the meteorological image to be detected in the target meteorological image under the detection branch is detected.

[0046] In step S231, the remote monitoring device performs abnormal weather detection on the target weather image according to the detection branch indicated by the branch prompt information through the weather image detection network completed by debugging, and obtains a first abnormal detection result. The first abnormal detection result includes the element classification corresponding to the maximum support coefficient of each image element in the target weather image, where any support coefficient is used to indicate the probability that the element classification of any image element belongs to the corresponding element classification. Here, the image element can be a pixel point, a pixel block, or a region with specific weather characteristics in the target weather image, etc. The element classification is a category predefined according to weather characteristics. For example, under the heavy precipitation detection branch, the element classification can include heavy precipitation regions, light precipitation regions, no precipitation regions, etc. The support coefficient is a value between 0 and 1, which reflects the likelihood of a certain image element belonging to a specific element classification. For example, for a pixel point in the target weather image, its support coefficient for belonging to the heavy precipitation region is 0.8, the support coefficient for belonging to the light precipitation region is 0.1, and the support coefficient for belonging to the no precipitation region is 0.1. Then the maximum support coefficient of this pixel point is 0.8, and the corresponding element classification is the heavy precipitation region.

[0047] To implement this detection process, the weather image detection network extracts and analyzes the features of the target weather image. In actual operation, the weather image detection network can adopt deep learning models such as convolutional neural networks (CNNs). A CNN can automatically learn the feature patterns in the image. For the target weather image, it will extract the feature information of the image through structures such as convolutional layers and pooling layers. Then, in the fully connected layer, the network will perform classification prediction on the extracted features according to the detection branch indicated by the branch prompt information, and calculate the support coefficients of each image element belonging to different element classifications. For example, in a trained CNN model, after inputting the target weather image, the output layer of the model will output the support coefficients of each image element for each element classification.

[0048] In step S232, the remote monitoring device detects and obtains the abnormal detection result of the weather image to be detected in the target weather image under the detection branch based on the element classification corresponding to the maximum support coefficient of each image element in the target weather image. This step comprehensively analyzes and judges the first abnormal detection result obtained in step S231. The remote monitoring device determines whether there is a region that meets the abnormal conditions in the target weather image according to the element classification corresponding to the maximum support coefficient of each image element. For example, under the heavy precipitation detection branch, if the element classification corresponding to the maximum support coefficient of a large number of image elements in the target weather image is the heavy precipitation region, and the distribution and scale of these regions reach the predefined abnormal standard, then the remote monitoring device will judge that there is an abnormality in the weather image to be detected under the heavy precipitation detection branch.

[0049] Taking the typhoon detection branch as an example, in step S231, the meteorological image detection network detects the target meteorological image and calculates the support coefficients of each image element for the classification of elements such as the typhoon eye, typhoon cloud system, and peripheral airflow. Suppose the support coefficient of a certain image element for the typhoon eye is 0.9, and the support coefficients for other element classifications are relatively small. Then, the element classification corresponding to the maximum support coefficient of this image element is the typhoon eye. In step S232, the remote monitoring device counts the distribution and quantity of image elements belonging to element classifications such as the typhoon eye and typhoon cloud system in the target meteorological image. If it is found that there is an obvious typhoon eye area, and there are a large number of image elements belonging to the typhoon cloud system around it, forming a typical typhoon structure, and the scale and intensity of this structure meet the definition criteria of a typhoon, then the remote monitoring device will determine that there is an abnormality in the target meteorological image under the typhoon detection branch, that is, the sign of a typhoon is detected.

[0050] In steps S231 - S232, the meteorological image detection network performs abnormal meteorological detection on the target meteorological image, obtains the element classification corresponding to the maximum support coefficient of each image element, and then based on these classifications, conducts comprehensive analysis and judgment. The remote monitoring device can accurately detect the abnormality of the target meteorological image under the detection branch.

[0051] As another implementation manner, in step S230, the meteorological image detection network that has been debugged performs abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information, and obtains the abnormal detection result of the target meteorological image under the detection branch, including:

[0052] Step S2301: The meteorological image detection network that has been debugged performs abnormal meteorological detection on the target meteorological image according to the target detection branch to obtain a second abnormal detection result; the second abnormal detection result includes the element classifications included in the element classification trajectory corresponding to the maximum trajectory support coefficient; wherein, any element classification trajectory includes the support coefficients of each image element of the target meteorological image for the element classification under the corresponding element classification trajectory, and the conversion support coefficients of adjacent image elements for the corresponding element classification; the trajectory support coefficient of any element classification trajectory is determined based on the support coefficient and conversion support coefficient corresponding to any element classification trajectory.

[0053] Step S2302: Based on the element classifications included in the element classification trajectory corresponding to the maximum trajectory support coefficient, the abnormal detection result of the target meteorological image to be detected under the detection branch is detected.

[0054] In step S2301, the remote monitoring device performs abnormal weather detection on the target weather image based on the target detection branch through the weather image detection network completed by debugging, and obtains a second abnormal detection result. This second abnormal detection result includes the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient. Here, the element classification trajectory refers to a sequence composed of the support coefficient of each image element of the weather training image under the corresponding element classification trajectory and the conversion support coefficient of the corresponding element classification of adjacent image elements. The trajectory support coefficient of any element classification trajectory is determined based on the support coefficient and conversion support coefficient corresponding to the element classification trajectory.

[0055] Specifically, element classification is the division of different feature regions in the weather image. For example, under the heavy precipitation detection branch, element classification can include heavy precipitation regions, light precipitation regions, no precipitation regions, etc. The support coefficient reflects the likelihood of a certain image element belonging to a specific element classification, with a value range between 0 and 1. The larger the value, the higher the likelihood that the image element belongs to this classification. The conversion support coefficient describes the probability of an adjacent image element converting from one element classification to another.

[0056] For a more intuitive understanding, assume a simple weather image composed of 3 adjacent image elements. Under the heavy precipitation detection branch, there may be multiple element classification trajectories. For example, one trajectory is: the first image element belongs to the heavy precipitation region, the second image element belongs to the light precipitation region, and the third image element belongs to the no precipitation region. For this trajectory, each image element has a corresponding support coefficient for its respective classification, and there is also a conversion support coefficient between adjacent image elements. For example, the support coefficient of the first image element belonging to the heavy precipitation region is 0.8, the conversion support coefficient from the heavy precipitation region to the light precipitation region where the second image element belongs is 0.6, the support coefficient of the second image element belonging to the light precipitation region is 0.7, the conversion support coefficient from the light precipitation region to the no precipitation region where the third image element belongs is 0.5, and the support coefficient of the third image element belonging to the no precipitation region is 0.9. Through a certain calculation method (such as multiplying each support coefficient and conversion support coefficient), the trajectory support coefficient of this element classification trajectory can be obtained.

[0057] In actual operation, the meteorological image detection network comprehensively analyzes the target meteorological image and calculates all possible element classification trajectories and their corresponding trajectory support coefficients. This can be achieved with the help of deep learning models, such as recurrent neural networks (RNNs) or their variant long short-term memory networks (LSTMs). These models can handle sequential data and are suitable for analyzing the relationships between adjacent image elements. The model classifies and predicts each image element based on the input target meteorological image and the information of the target detection branch, calculates the support coefficient and the transformed support coefficient, and then obtains the trajectory support coefficients of different element classification trajectories.

[0058] Step S2302 requires the remote monitoring device to detect the anomaly detection result of the meteorological image to be detected in the target meteorological image under the detection branch based on the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient. That is to say, among all the calculated element classification trajectories, find the trajectory with the largest trajectory support coefficient, and the element classification included in this trajectory is an important basis for judging anomalies.

[0059] For example, in the above example of heavy precipitation detection, multiple element classification trajectories and their trajectory support coefficients are calculated. The element classification trajectory corresponding to the maximum trajectory support coefficient shows that there is a large area of heavy precipitation in the image and the distribution conforms to the characteristic pattern of heavy precipitation anomalies. Then the remote monitoring device can judge that there is an anomaly in the meteorological image to be detected under the heavy precipitation detection branch.

[0060] To accurately judge anomalies, the remote monitoring device can combine professional meteorological knowledge and historical data to preset some judgment rules in advance. For example, for heavy precipitation anomalies, it is stipulated that when the area of the heavy precipitation area reaches a certain proportion (such as 15%) and the distribution of the heavy precipitation area shows a specific concentration, it is determined as an anomaly. At the same time, other meteorological factors, such as air pressure and humidity, can also be considered to further verify the reliability of the anomaly detection result.

[0061] In practical applications, to improve the accuracy and efficiency of anomaly detection, the remote monitoring device can preprocess the meteorological image, such as removing noise, enhancing contrast, etc., to improve the image quality and facilitate the meteorological image detection network to better extract features.

[0062] As an implementation method, the meteorological image detection network is obtained through the following steps for debugging:

[0063] Step S10: Obtain meteorological training images and the target detection branches of the meteorological training images; at least one prior label is matched with the meteorological training images, and any prior label is used to indicate the abnormal meteorological detection result of the meteorological training image under the corresponding detection branch;

[0064] Step S20: Perform abnormal weather detection on the meteorological training images according to the target detection branch, and obtain the abnormal weather detection results of the meteorological training images under the target detection branch;

[0065] Step S30: Determine the target prior label among at least one prior label matched by the meteorological training image, and determine the abnormal weather detection results of the meteorological training image under the target detection branch in the target prior label, where the target prior label is the prior label corresponding to the detection branch that is the same as the target detection branch;

[0066] Step S40: Perform network debugging operations on the meteorological image detection network according to the abnormal weather detection results of the meteorological training image under the target detection branch and the abnormal weather detection results under the target detection branch; The meteorological image detection network after debugging is used to perform abnormal weather detection on meteorological images under the corresponding detection branch.

[0067] In step S10, the remote monitoring device obtains the meteorological training images and the target detection branch of the meteorological training images. The meteorological training images are matched with at least one prior label, and any prior label is used to indicate the abnormal weather detection results of the meteorological training image under the corresponding detection branch. The meteorological training images are the basic data for training the meteorological image detection network, and they can come from a large number of meteorological images collected by devices such as meteorological satellites and meteorological radars. The target detection branch refers to a specific type of meteorological abnormality that the meteorological image detection network is expected to focus on and detect, such as heavy precipitation, high temperature, typhoon, etc. The prior label is the abnormal situation of the meteorological training image under different detection branches determined in advance manually or through other reliable methods, and it provides a correct reference standard for the training of the meteorological image detection network. For example, for a meteorological training image, its target detection branch is heavy precipitation detection, and the prior label clearly indicates which areas in the image have heavy precipitation abnormalities. The remote monitoring device can obtain the meteorological training images and target detection branch information from data sources such as meteorological data centers and historical meteorological databases, and these data sources will also provide the corresponding prior labels.

[0068] In step S20, the remote monitoring device performs abnormal weather detection on the meteorological training image according to the target detection branch, and obtains the abnormal weather detection result of the meteorological training image under the target detection branch. This step enables the meteorological image detection network to conduct preliminary analysis and judgment on the meteorological training image. The remote monitoring device inputs the meteorological training image into the meteorological image detection network. According to the requirements of the target detection branch, the network extracts and analyzes the features of the image, and attempts to find possible abnormal situations in the image. For example, under the heavy precipitation detection branch, the meteorological image detection network will focus on features such as the thickness, color, and texture of the clouds in the image, and judge whether there is a heavy precipitation area based on these features. To implement this detection process, the meteorological image detection network can adopt deep learning models such as convolutional neural networks (CNNs). CNNs can automatically extract features in the image, process the image through structures such as convolutional layers and pooling layers, and then classify and predict the extracted features in the fully connected layer, so as to obtain the abnormal weather detection result of the meteorological training image under the target detection branch.

[0069] In step S30, the remote monitoring device determines the target prior label from at least one prior label matched by the meteorological training image, and determines the abnormal weather detection result of the meteorological training image under the target detection branch in the target prior label, where the target prior label is the prior label corresponding to the detection branch that is the same as the target detection branch. Since the meteorological training image may match multiple prior labels, and each prior label corresponds to a different detection branch, it is necessary to find the target prior label that is consistent with the current target detection branch. For example, if the meteorological training image matches the prior labels of three detection branches: heavy precipitation, high temperature, and typhoon, and the current target detection branch is heavy precipitation detection, then the remote monitoring device will select the prior label corresponding to the heavy precipitation detection branch as the target prior label, and obtain the true abnormal weather detection result of the meteorological training image under the heavy precipitation detection branch from it. The purpose of this step is to provide an accurate reference basis for subsequent network debugging.

[0070] In step S40, the remote monitoring device performs network debugging operations on the meteorological image detection network according to the abnormal weather detection result of the meteorological training image under the target detection branch and the abnormal weather detection result under the target detection branch. The meteorological image detection network after debugging is used to perform abnormal weather detection on meteorological images under the corresponding detection branches. Specifically, the remote monitoring device compares the difference between the abnormal weather detection result obtained by the meteorological image detection network and the true abnormal weather detection result indicated by the target prior label, and adjusts the parameters of the meteorological image detection network according to this difference, so that the detection result of the network gradually approaches the true result. For example, the cross-entropy loss function is used to measure the difference between the two.

[0071] In practical applications, in order to improve the training effect of the meteorological image detection network, the remote monitoring device can adopt various strategies. In terms of data acquisition, the diversity of meteorological training images can be increased, including images under different seasons, different regions, and different meteorological conditions, so as to improve the network's adaptability to various meteorological situations. During the training process, the method of batch training can be adopted, dividing the meteorological training images into multiple small batches for training, which can reduce the use of memory and improve the training efficiency at the same time. Regularization techniques, such as L1 and L2 regularization, can also be used to prevent the network from overfitting and improve the generalization ability of the network.

[0072] Taking heavy precipitation detection as an example, the remote monitoring device first obtains a large number of meteorological training images containing heavy precipitation situations from the meteorological database and determines heavy precipitation detection as the target detection branch. Then, these meteorological training images are input into the meteorological image detection network, and the network analyzes the images to obtain the abnormal meteorological detection results under the heavy precipitation detection branch. Next, the remote monitoring device finds out the target prior markers corresponding to the heavy precipitation detection branch from the prior markers matched by the meteorological training images to obtain the real heavy precipitation abnormal situation. Finally, the difference between the network detection result and the real result is compared, and the cross-entropy loss function and the gradient descent algorithm are used to adjust the parameters of the meteorological image detection network. After multiple iterative trainings, the detection ability of the meteorological image detection network for heavy precipitation anomalies will be continuously improved, and it can more accurately identify the heavy precipitation areas in meteorological images.

[0073] As an implementation method, the meteorological image detection network includes a feature information extraction component and an abnormal inference network; the feature information extraction component is used to extract the meteorological image feature vector; the abnormal inference network is used to perform abnormal meteorological detection on the meteorological training images to obtain the corresponding abnormal meteorological detection results.

[0074] The main function of the feature information extraction component is to extract the meteorological image feature vector. The meteorological image feature vector is a vector that can represent the feature information of the meteorological image and the relationship between different image elements. In meteorological images, there is rich information such as cloud morphology, precipitation area distribution, and temperature change characteristics. The role of the feature information extraction component is to convert these complex image information into a vector form that can be processed and analyzed by a computer. Taking the heavy precipitation meteorological image as an example, the feature information extraction component will focus on features such as the thickness, color, and texture of the clouds in the image, as well as information such as the size, position, and boundary of the precipitation area, and quantify and integrate these features and information to form a meteorological image feature vector. Each dimension in this vector corresponds to a certain feature or feature combination of the image, and the features of the meteorological image can be comprehensively described through this vector.

[0075] The characterization information extraction component can adopt various technical means. One method is to use a Convolutional Neural Network (CNN). CNN has powerful feature extraction capabilities. It performs convolution operations on meteorological images through convolutional layers to extract local features in the images, such as edges, textures, etc. The pooling layer is used to reduce the dimensionality of the extracted features, reduce the computational amount, and enhance the robustness of the features. Through multiple layers of convolution and pooling operations, CNN can gradually extract the high-level semantic features of meteorological images and finally combine these features into a meteorological image characterization vector. For example, a typical CNN structure may contain multiple convolutional layers and pooling layers. Each convolutional layer uses different convolutional kernels to extract different types of features. After multiple convolutions and poolings, a fixed-length vector is obtained as the meteorological image characterization vector.

[0076] The anomaly inference network then uses the meteorological image characterization vector extracted by the characterization information extraction component to perform anomaly meteorological detection on meteorological training images and obtain corresponding anomaly meteorological detection results. The task of the anomaly inference network is to judge whether there are abnormal meteorological conditions in the meteorological image based on the meteorological image characterization vector, as well as the type and degree of the anomaly. Continuing with the example of a heavy precipitation meteorological image, the anomaly inference network will judge whether there is a heavy precipitation area in the image based on the input meteorological image characterization vector, as well as information such as the scope and intensity of the heavy precipitation area.

[0077] The anomaly inference network can be implemented using structures such as a fully connected neural network or a recurrent neural network. The fully connected neural network takes the meteorological image characterization vector as input and performs non-linear transformations through multiple fully connected layers, and finally outputs the anomaly meteorological detection results. For example, the output layer of the network can use the softmax function to output the probability of each anomaly category (such as heavy precipitation, high temperature, typhoon, etc.). The category with the highest probability is the predicted anomaly type. The formula for the softmax function is: ; where z j is the input of the j-th neuron in the output layer of the network, and K is the number of anomaly categories.

[0078] The recurrent neural network is suitable for processing meteorological data with sequential relationships, such as a sequence of meteorological images at different time points. It can capture the changing trend of meteorological images over time, thereby performing more accurate anomaly detection. For example, in monitoring the development process of a typhoon, the recurrent neural network can predict the moving direction and intensity changes of the typhoon based on the meteorological image characterization vectors at multiple time points.

[0079] In practical applications, the characterization information extraction component and the anomaly inference network cooperate with each other. The characterization information extraction component provides accurate meteorological image characterization vectors for the anomaly inference network, enabling the anomaly inference network to perform effective anomaly detection based on these vectors. The detection results of the anomaly inference network can be fed back to the characterization information extraction component to help it further optimize the feature extraction methods and parameters, improving the quality of the meteorological image characterization vectors. Through this collaborative work, the meteorological image detection network can continuously improve its ability to detect abnormal meteorological conditions in meteorological training images, providing reliable support for the remote monitoring of meteorological business systems and meteorological disaster warnings.

[0080] As an implementation, the number of target detection branches is at least one, and the number of abnormal meteorological detection results obtained by performing abnormal meteorological detection on meteorological training images according to the target detection branches is the same as the number of target detection branches; the number of detection branches corresponding to the prior labels matched by the meteorological training images is not less than the number of detection branches corresponding to the target detection branches; where, when the number of target detection branches is multiple, the number of obtained abnormal meteorological detection results is also multiple, and the number of detection branches corresponding to the target prior labels determined according to the prior labels matched by the meteorological training images is equal to the number of target detection branches.

[0081] In the remote monitoring of a regulatory meteorological business system with low bandwidth, when the meteorological image detection network performs abnormal meteorological detection on meteorological training images, there are specific corresponding relationships among the setting of the number of target detection branches and the related abnormal meteorological detection results, prior labels, etc. The target detection branch refers to a specific meteorological abnormal type concerned when performing abnormal detection on meteorological training images, and the number of target detection branches can be at least one, which means that the remote monitoring device can either detect only one type of meteorological abnormality or detect multiple types simultaneously. The number of abnormal meteorological detection results obtained by performing abnormal meteorological detection on meteorological training images according to the target detection branches is the same as the number of target detection branches, which reflects a one-to-one correspondence, that is, each target detection branch will correspond to a corresponding abnormal meteorological detection result.

[0082] The number of detection branches corresponding to the prior labels matched by the meteorological training images is not less than the number of detection branches corresponding to the target detection branches. The prior label is the true abnormal situation of the meteorological training image under different detection branches determined in advance, ensuring that there is sufficient reference basis to evaluate the accuracy of the detection results when performing abnormal meteorological detection. When the number of target detection branches is multiple, the number of obtained abnormal meteorological detection results is also multiple, and the number of detection branches corresponding to the target prior labels determined according to the prior labels matched by the meteorological training images is equal to the number of target detection branches, which guarantees that each target detection branch has a corresponding accurate reference standard in the case of multi-branch detection.

[0083] The number of prior labels corresponding to this meteorological training image for matching is at least three. There may also be some other labels for detection branches, but at least it should cover the prior labels for the three target detection branches of heavy precipitation, high temperature, and typhoon. The remote monitoring device determines the target prior labels from these prior labels. The number of detection branches corresponding to these target prior labels is the same as the number of target detection branches, that is, the prior labels for the three detection branches of heavy precipitation, high temperature, and typhoon. By comparing the abnormal meteorological detection results with the true situations indicated by the target prior labels, the remote monitoring device can evaluate the detection accuracy of the meteorological image detection network under each target detection branch and debug and optimize the meteorological image detection network according to the evaluation results.

[0084] In actual operation, the remote monitoring device can adopt the method of multi-task learning to achieve the detection of abnormal meteorology for multiple target detection branches. Multi-task learning allows the meteorological image detection network to process multiple related tasks simultaneously. In this example, it is to detect heavy precipitation, high temperature, and typhoon simultaneously. The network can share a part of the underlying feature extraction layer, and then set independent output layers for each target detection branch to output the detection results of abnormal meteorology for each branch respectively. This can improve the detection efficiency and also utilize the correlation between different detection branches to improve the detection accuracy.

[0085] In addition, in order to ensure the accuracy and integrity of the target prior labels, the remote monitoring device can adopt a method combining manual annotation and automatic annotation. Manual annotation can be carried out by meteorological experts to mark the meteorological training images in detail according to professional knowledge to ensure the accuracy of the labels. Automatic annotation can use existing meteorological data and algorithms to preliminarily annotate a large number of meteorological training images, and then be reviewed and corrected by humans to improve the annotation efficiency.

[0086] By reasonably setting the number of target detection branches and ensuring the corresponding relationships among the abnormal meteorological detection results, prior labels, etc., the remote monitoring device can detect abnormal meteorology in meteorological training images more comprehensively and accurately. This multi-branch detection method can simultaneously meet the monitoring requirements for various meteorological anomaly types and provide richer and more accurate information for the remote monitoring of meteorological business systems and meteorological disaster warnings. In future meteorological monitoring and warning work, with the continuous enrichment of meteorological data and the continuous progress of technology, the method for detecting abnormal meteorology with multiple target detection branches is expected to be further optimized and improved, providing stronger support for the development of the meteorological field.

[0087] As an implementation method, the acquisition methods of meteorological training images include:

[0088] Step S11: Obtain at least one candidate meteorological image and branch hint information; the branch hint information is used to hint the number of branches of the target detection branch.

[0089] Step S12: Determine, according to the branch hint information, a candidate meteorological image that matches the target detection branch with the corresponding number of branches from at least one candidate meteorological image, and use the determined candidate meteorological image as the branch meteorological image of the meteorological training image.

[0090] Step S13: Obtain a sample meteorological image, combine the branch meteorological image and the sample meteorological image to obtain a meteorological training image.

[0091] In step S11, the remote monitoring device obtains at least one candidate meteorological image and branch hint information, and the branch hint information is used to hint the number of branches of the target detection branch. The candidate meteorological images are a series of meteorological images collected in advance, and these images cover various features under different meteorological conditions, such as images of different meteorological scenarios such as sunny, cloudy, rainy, typhoon, etc. The branch hint information provides crucial guidance for subsequent processing, and it tells the remote monitoring device the number of target detection branches to focus on. For example, the branch hint information may indicate that the target detection branches are heavy precipitation detection and high temperature detection, and the number of branches is 2. The remote monitoring device can obtain the candidate meteorological images and branch hint information by establishing connections with data sources such as meteorological data centers and meteorological monitoring devices, or read relevant data from local storage.

[0092] In step S12, the remote monitoring device determines, according to the branch hint information, a candidate meteorological image that matches the target detection branch with the corresponding number of branches from at least one candidate meteorological image, and uses the determined candidate meteorological image as the branch meteorological image of the meteorological training image. This step is to screen out the images related to the target detection branch from numerous candidate meteorological images according to the branch hint information. For example, in the above example of heavy precipitation detection and high temperature detection, the remote monitoring device finds the images related to heavy precipitation and high temperature from the candidate meteorological images as the branch meteorological images. An image classification algorithm can be used to implement this screening process. By extracting and analyzing the features of the candidate meteorological images, they are classified into different target detection branch categories. For example, a convolutional neural network (CNN) is used to extract the features of the candidate meteorological images, and then the images are classified according to the feature model of the target detection branch to find the images that match the heavy precipitation and high temperature detection branches.

[0093] In step S13, the remote monitoring device acquires a sample meteorological image, combines the branch meteorological image and the sample meteorological image to obtain a meteorological training image. The sample meteorological image refers to a basic image containing various meteorological features. It can be a representative meteorological image or a processed comprehensive meteorological image. There are various ways of combination. For example, the branch meteorological image can be superimposed on the sample meteorological image as supplementary information, or their features can be fused. For example, for the branch meteorological image of the heavy precipitation detection branch, the part representing the heavy precipitation features can be superimposed on the corresponding position of the sample meteorological image to enhance the feature information related to heavy precipitation in the sample meteorological image. In terms of feature fusion, methods such as feature splicing or weighted averaging can be used to fuse the features of the branch meteorological image and the sample meteorological image to generate a new meteorological training image.

[0094] Taking typhoon detection as an example, the remote monitoring device first acquires a large number of candidate meteorological images and branch hint information. The hint information indicates that the target detection branch is typhoon detection and the number of branches is 1. Then, the remote monitoring device uses an image classification algorithm to screen out the images related to typhoons from the candidate meteorological images as the branch meteorological images. These images may contain typical features such as typhoon eyes and typhoon cloud systems. Next, the remote monitoring device acquires the sample meteorological image and combines the typhoon-related branch meteorological image with the sample meteorological image. The features of the typhoon eye and typhoon cloud system in the branch meteorological image can be superimposed on the appropriate position of the sample meteorological image, or the features of the two can be fused by weighted averaging, and finally a meteorological training image for typhoon detection training is obtained.

[0095] In practical applications, in order to improve the quality and diversity of the meteorological training image, the remote monitoring device can preprocess the candidate meteorological images, such as removing noise, adjusting brightness and contrast, etc., to improve the clarity and quality of the images. The sample meteorological image can also be enhanced, such as rotation, flipping, scaling, etc., to increase the diversity of the samples and improve the generalization ability of the meteorological image detection network. In addition, in order to ensure that the combination of the branch meteorological image and the sample meteorological image is reasonable and effective, some evaluation indicators can be used to measure the quality of the combined meteorological training image, such as information entropy, correlation, etc. Through these methods, high-quality and diverse meteorological training images can be obtained, providing strong support for the training of the meteorological image detection network, thereby improving the accuracy and reliability of meteorological anomaly detection and better serving the remote monitoring of the meteorological business system and meteorological disaster warning work.

[0096] As an implementation method, in step S20, abnormal meteorological detection is performed on the meteorological training image according to the target detection branch to obtain the abnormal meteorological detection result of the meteorological training image under the target detection branch, including:

[0097] Step S21: Extract the representation information from the meteorological training image according to the target detection branch to obtain the meteorological image representation vector corresponding to the meteorological training image; the meteorological image representation vector is used to represent the feature information of the image elements in the meteorological training image and the involvement relationship between different image elements.

[0098] Step S22: Perform abnormal meteorological detection on the meteorological training image based on the meteorological image representation vector to obtain the abnormal meteorological detection result of the meteorological training image under the target detection branch.

[0099] In step S21, the remote monitoring device extracts the representation information from the meteorological training image according to the target detection branch to obtain the meteorological image representation vector corresponding to the meteorological training image. The meteorological image representation vector is used to represent the feature information of the image elements in the meteorological training image and the involvement relationship between different image elements. The meteorological training image contains rich meteorological information, such as the shape of clouds, the distribution of precipitation areas, the change of temperature, etc., while the meteorological image representation vector is a vector representation obtained by quantifying and integrating these complex information.

[0100] To extract the meteorological image representation vector, the remote monitoring device can use the convolutional neural network (CNN) in deep learning. Through the alternating operations of multiple convolutional layers and pooling layers, the CNN can gradually extract different levels of features of the meteorological training image. For example, in the first convolutional layer, some simple edge and texture features may be extracted; as the number of network layers increases, the subsequent convolutional layers will extract more advanced semantic features, such as the type of clouds, the boundary of precipitation areas, etc.

[0101] In step S22, the remote monitoring device performs abnormal meteorological detection on the meteorological training image based on the meteorological image representation vector to obtain the abnormal meteorological detection result of the meteorological training image under the target detection branch. This step is to use the meteorological image representation vector extracted previously to judge whether there is an abnormal meteorological situation corresponding to the target detection branch in the meteorological training image.

[0102] The remote monitoring device can use a fully connected neural network to implement abnormal meteorological detection. The fully connected neural network takes the meteorological image representation vector as the input and performs non-linear transformation through multiple fully connected layers. Assume that the input meteorological image representation vector is x, the weight matrix of the l-th layer is , and the bias vector is , then the input of the (l + 1)-th layer can be expressed as: ; where x l is the output of the l-th layer. After being processed by a non-linear activation function (such as the ReLU function, ReLU(z) = max(0, z)), the output x l+1 of the (l + 1)-th layer is obtained.

[0103] In abnormal weather detection, the output layer of the fully connected neural network outputs the probabilities of the weather training image belonging to different abnormal categories under the target detection branch. For example, under the heavy precipitation detection branch, the output layer may have two neurons, representing the probabilities of the existence and non-existence of heavy precipitation anomalies respectively. By comparing the magnitudes of these two probabilities, the remote monitoring device can determine whether there is an anomaly in the weather training image under the heavy precipitation detection branch.

[0104] Continuing with the heavy precipitation detection as an example, the remote monitoring device inputs the weather image feature vector obtained in step S21 into the fully connected neural network. The network processes and analyzes the vector, and calculates the probability of the existence of heavy precipitation anomaly in the weather training image based on the feature information and relationships contained in the vector. If this probability exceeds a pre-set threshold (such as 0.5), the remote monitoring device determines that there is an anomaly in the weather training image under the heavy precipitation detection branch; otherwise, it determines that there is no anomaly.

[0105] To improve the accuracy of abnormal weather detection, the remote monitoring device can adopt the cross-entropy loss function when training the fully connected neural network.

[0106] As an implementation, before step S22, which performs abnormal weather detection on the weather training image based on the weather image feature vector and obtains the abnormal weather detection result of the weather training image under the target detection branch, the method further includes:

[0107] Step S2201: Obtain the weather image type corresponding to the weather image to which each image element in the weather training image belongs;

[0108] Step S2202: Generate a masking matrix based on the weather image type; the masking matrix includes multiple masking elements, and the image elements included in the weather training image correspond to the masking elements in the masking matrix;

[0109] Step S2203: Perform a masking operation on the weather image feature vector of the weather training image based on the masking matrix to obtain a masked weather image feature vector; the masked weather image feature vector includes the element feature vectors corresponding to each image element in the sample weather image, and the masked weather image feature vector is used to perform abnormal weather detection on the weather training image.

[0110] In step S2201, the remote monitoring device obtains the meteorological image types corresponding to the meteorological training images for each image element. The meteorological image type is a way to classify meteorological images, and different meteorological image types have different meteorological characteristics and patterns. For example, meteorological image types can be divided into sunny images, cloudy images, precipitation images, typhoon images, etc. In the meteorological training images, each image element may belong to different meteorological image types, and understanding these types helps with subsequent analysis and processing. The remote monitoring device can obtain meteorological image type information in various ways. One method is to use a pre-annotated meteorological image dataset, which labels the corresponding type for each meteorological image. Another method is to use an image classification algorithm. By extracting and analyzing the features of the meteorological training images, they are classified into different meteorological image types. For example, a convolutional neural network (CNN) is used to extract the features of the meteorological training images, and then based on a pre-trained classification model, the meteorological image type to which each image element belongs is determined.

[0111] In step S2202, the remote monitoring device generates a shielding matrix based on the meteorological image types. The shielding matrix includes multiple shielding elements, and the image elements included in the meteorological training images correspond to the shielding elements in the shielding matrix. The role of the shielding matrix is to shield some information that is irrelevant to the current abnormal meteorological detection according to the meteorological image type. For example, when detecting strong precipitation anomalies, if the meteorological image type to which an image element belongs is a sunny image, then this image element is likely to be irrelevant to the strong precipitation anomaly, and the corresponding information can be shielded through the shielding matrix. The specific method for generating the shielding matrix can be designed according to different meteorological image types. A simple method is to assign a shielding coefficient to each meteorological image type. When an image element belongs to a specific meteorological image type, the value of its corresponding shielding element is set to the shielding coefficient of that type. For example, for the sunny image type that is irrelevant to strong precipitation detection, the shielding coefficient can be set to 0, while for the precipitation image type that is related to strong precipitation, the shielding coefficient can be set to 1. In this way, the elements in the shielding matrix can be assigned values according to the meteorological image types of the image elements in the meteorological training images.

[0112] In step S2203, the remote monitoring device performs a masking operation on the meteorological image representation vector of the meteorological training image based on the masking matrix to obtain the masked meteorological image representation vector. The masked meteorological image representation vector includes the element representation vectors corresponding to each image element in the sample meteorological image, and this vector is used to perform abnormal meteorological detection on the meteorological training image. The masking operation is to multiply the masking matrix and the meteorological image representation vector element by element, so as to filter out the unnecessary information. Through this masking operation, the element values corresponding to the information irrelevant to the current abnormal meteorological detection become 0, while the information related to the detection is retained. For example, in heavy precipitation detection, after the masking operation, the value of the image element corresponding to the sunny day image in the meteorological image representation vector becomes 0, while the value of the image element corresponding to the precipitation image remains unchanged, so that the feature information related to heavy precipitation can be highlighted and the accuracy of abnormal detection can be improved.

[0113] Taking heavy precipitation detection as an example, the remote monitoring device first obtains the meteorological image types corresponding to the meteorological training image for each image element. Assume that the meteorological training image contains image elements in sunny areas, cloudy areas, and precipitation areas. The remote monitoring device uses an image classification algorithm to determine the meteorological image types to which the image elements in different areas belong. Then, a masking matrix is generated based on these meteorological image types. For the image elements in the sunny area, the corresponding masking element value is set to 0; for the image elements in the precipitation area, the corresponding masking element value is set to 1; for the image elements in the cloudy area, according to their correlation with heavy precipitation, the masking element value is set to a value between 0 and 1, such as 0.2. Next, the remote monitoring device multiplies the generated masking matrix and the meteorological image representation vector of the meteorological training image element by element to obtain the masked meteorological image representation vector. In this masked vector, the value of the element representation vector corresponding to the image element in the sunny area becomes 0 and is thus masked, while the values of the element representation vectors corresponding to the image elements in the precipitation area and some cloudy areas with a certain correlation with heavy precipitation are retained. Finally, this masked meteorological image representation vector is used to perform heavy precipitation abnormal detection on the meteorological training image. Since the irrelevant information is masked, the detection process can focus more on the features related to heavy precipitation, improving the accuracy and efficiency of the detection.

[0114] In practical applications, in order to improve the effectiveness of steps S2201 - S2203, the remote monitoring device can adopt some optimization strategies. When obtaining the meteorological image type, a multi-modal classification algorithm based on deep learning can be used, combining the visual features of the image and the features of meteorological data to improve the accuracy of classification. When generating the shielding matrix, the shielding coefficient can be dynamically adjusted according to different target detection branches to meet different abnormal meteorological detection requirements. For example, when detecting typhoons, for some cloudy image types that are somewhat related to the outer cloud system of the typhoon, the shielding coefficient can be appropriately increased to retain more information that may be related to the typhoon. In addition, to ensure the rationality and effectiveness of the shielding matrix, the remote monitoring device can use the method of cross-validation for evaluation. By dividing the meteorological training images into a training set and a validation set, generating the shielding matrix and performing shielding operations on the training set, and then evaluating the effect of abnormal meteorological detection on the validation set. According to the evaluation results, adjust the generation method of the shielding matrix and the shielding coefficient until the best detection effect is achieved.

[0115] As an implementation manner, step S22, based on the meteorological image feature vector, performs abnormal meteorological detection on the meteorological training images to obtain the abnormal meteorological detection results of the meteorological training images under the target detection branch, including:

[0116] Step S221: Based on the meteorological image feature vector, infer the abnormal classification prediction information of each image element in the meteorological training image; the abnormal classification prediction information of any image element represents the probability that the element classification of the corresponding image element belongs to the element classification of the abnormal meteorology under the target detection branch;

[0117] Step S222: Based on the abnormal classification prediction information corresponding to each image element in the meteorological training image, generate the abnormal meteorological detection results of the meteorological training image under the target detection branch.

[0118] In step S221, the remote monitoring device infers the abnormal classification prediction information of each image element in the meteorological training image based on the meteorological image feature vector. The abnormal classification prediction information of any image element represents the probability that the element classification of the corresponding image element belongs to the element classification of the abnormal meteorology under the target detection branch. The meteorological image feature vector is obtained after performing feature extraction on the meteorological training image in the previous steps, and it contains the feature information of the image elements in the meteorological training image and the involved relationships between different image elements. The abnormal classification prediction information is the result of quantifying the possibility of each image element belonging to the abnormal meteorological element classification based on these feature information.

[0119] To implement this inference process, the remote monitoring device can adopt a neural network model, such as a fully connected neural network. The fully connected neural network takes the meteorological image feature vector as the input and performs non-linear transformations through multiple fully connected layers.

[0120] Taking the object detection branch for heavy precipitation detection as an example, assume that the element classification of abnormal weather has three types: heavy precipitation area, light precipitation area, and no precipitation area. The remote monitoring device inputs the meteorological image feature vector of the meteorological training image into a fully connected neural network. After the calculation and processing of the network, the three neurons in the output layer respectively output the probabilities that the image element belongs to the heavy precipitation area, the light precipitation area, and the no precipitation area. For example, for a specific pixel point in the image, the network outputs the probability that it belongs to the heavy precipitation area as 0.8, the probability that it belongs to the light precipitation area as 0.1, and the probability that it belongs to the no precipitation area as 0.1. Then the abnormal classification prediction information of this pixel point is such a set of probability values.

[0121] In step S222, the remote monitoring device generates the abnormal weather detection result of the meteorological training image under the object detection branch based on the abnormal classification prediction information corresponding to each image element in the meteorological training image. This step comprehensively analyzes and judges the abnormal classification prediction information of each image element to determine the abnormal situation of the entire meteorological training image under the object detection branch.

[0122] The remote monitoring device can adopt various methods to generate the abnormal weather detection result. One method is to count the number or proportion of image elements belonging to each abnormal weather element classification, and then judge whether there is an abnormality according to a pre-set threshold.

[0123] More complex methods can also be adopted, combining the spatial relationship and feature correlation between different image elements for judgment. For example, for adjacent image elements, if they both have a high probability of belonging to the heavy precipitation area and these elements form a certain continuous area in space, then it can be more confidently judged that this is a heavy precipitation abnormal area.

[0124] Continuing with the heavy precipitation detection as an example, the remote monitoring device counts the probabilities that all image elements in the meteorological training image belong to the heavy precipitation area. Assume that there are 1000 pixel points in the image, and among them, 200 pixel points have a probability of belonging to the heavy precipitation area greater than 0.8. If the pre-set heavy precipitation abnormal threshold is that the proportion of the number of pixel points reaches 15%, then since 200 / 1000 = 20% > 15%, the remote monitoring device judges that there is an abnormality in the meteorological training image under the heavy precipitation detection branch.

[0125] To improve the accuracy of the abnormal weather detection result, the remote monitoring device can adopt the cross-entropy loss function when training the neural network.

[0126] In practical applications, the effective execution of steps S221 - S222 is crucial for the performance of the meteorological image detection network. Accurate abnormal classification prediction information can provide a basis for generating reliable abnormal meteorological detection results, while a reasonable result generation method can make accurate judgments based on this prediction information. The remote monitoring device can improve the accuracy of abnormal classification prediction information and the reliability of abnormal meteorological detection results by continuously optimizing the structure and parameters of the neural network, as well as increasing the diversity and quantity of training data. At the same time, other meteorological data, such as meteorological observation data, numerical weather prediction data, etc., can be combined to further improve the practicality and effectiveness of abnormal meteorological detection.

[0127] As an implementation, the element classification of abnormal meteorology under the object detection branch includes at least one. The abnormal classification prediction information of any image element includes: the support coefficients of the element classification of the corresponding image element belonging to different element classifications, and any support coefficient is used to indicate the probability that the element classification of the corresponding image element belongs to the corresponding element classification. Based on this, step S222, based on the abnormal classification prediction information corresponding to each image element in the meteorological training image, generates the abnormal meteorological detection result of the meteorological training image under the object detection branch, including:

[0128] Step S2221: Among the support coefficients included in the abnormal classification prediction information of any image element, determine the maximum support coefficient;

[0129] Step S2222: Take the maximum support coefficient and the element classification corresponding to the maximum support coefficient as the target prediction information corresponding to any image element;

[0130] Step S2223: Generate the abnormal meteorological detection result of the meteorological training image under the object detection branch according to the target prediction information of each image element in the meteorological training image.

[0131] In step S2221, the remote monitoring device determines the maximum support coefficient among the support coefficients included in the abnormal classification prediction information of any image element. The support coefficient refers to the probability that the element classification of the corresponding image element belongs to different element classifications, and it is obtained through the abnormal classification prediction of the meteorological training image in the previous steps. For example, under the object detection branch of heavy precipitation detection, the element classification of abnormal meteorology may include heavy precipitation areas, light precipitation areas, and no precipitation areas. For a specific image element in the meteorological training image, its abnormal classification prediction information may show that the support coefficient for this element belonging to the heavy precipitation area is 0.8, the support coefficient for the light precipitation area is 0.1, and the support coefficient for the no precipitation area is 0.1. In this case, the maximum support coefficient is 0.8.

[0132] In step S2222, the remote monitoring device classifies the maximum support coefficient and the element classification corresponding to the maximum support coefficient as the target prediction information corresponding to any image element. Continuing with the above example of heavy precipitation detection, if the maximum support coefficient is 0.8 and the corresponding element classification is the heavy precipitation area, then the target prediction information for this image element is (0.8, heavy precipitation area). This step simplifies the abnormal classification prediction information for each image element, focusing on the most likely element classification and its corresponding support coefficient, providing a clearer basis for generating the overall abnormal detection result of the meteorological training image in the subsequent steps.

[0133] In step S2223, the remote monitoring device generates the abnormal meteorological detection result of the meteorological training image under the target detection branch based on the target prediction information of each image element in the meteorological training image. This step requires comprehensive analysis and judgment of the target prediction information of all image elements. A feasible method is to count the number or proportion of image elements belonging to different element classifications and determine whether there is an abnormality in the meteorological training image according to a preset threshold.

[0134] For example, in heavy precipitation detection, the remote monitoring device counts the number of image elements in the meteorological training image whose target prediction information is (high support coefficient, heavy precipitation area). Suppose there are 10,000 image elements in the meteorological training image, and among them, 1,500 image elements have target prediction information indicating that they belong to the heavy precipitation area and the support coefficient is greater than 0.8. If the preset heavy precipitation anomaly threshold is that the proportion of image elements in the heavy precipitation area reaches 10%, since 1,500 / 10,000 = 15% > 10%, then the remote monitoring device can determine that there is an abnormality in the meteorological training image under the heavy precipitation detection branch.

[0135] In addition, the remote monitoring device can also consider the spatial distribution of image elements. If the image elements belonging to the heavy precipitation area form a continuous and relatively large area in space, it can better indicate the existence of heavy precipitation anomaly. On the contrary, if these elements are scattered, further analysis or combination with other meteorological data may be required for judgment. In practical applications, in order to improve the accuracy of the abnormal detection result, the remote monitoring device can adopt some optimization strategies. When determining the maximum support coefficient, a confidence threshold can be set, and only when the maximum support coefficient exceeds this threshold, the corresponding element classification is used as part of the target prediction information. For example, if the confidence threshold is set to 0.7, and the maximum support coefficient of a certain image element is 0.6, then further analysis of the classification of this element may be required or the information of this element can be ignored.

[0136] When generating anomaly detection results, the correlation between different image elements can be combined. For example, adjacent image elements often have similar meteorological characteristics. If their target prediction information all indicates that they belong to the heavy precipitation area, then the credibility of the heavy precipitation anomaly judgment can be increased. The correlation can be evaluated by calculating the consistency of the target prediction information of adjacent image elements. In addition, the remote monitoring device can also dynamically adjust the threshold according to different meteorological scenarios and target detection branches. For example, in some regions or seasons, the criteria for heavy precipitation may be different. At this time, the threshold of the proportion of image elements in the heavy precipitation area can be adjusted according to the actual situation. Taking typhoon detection as an example, the element classification of abnormal meteorology may include the typhoon eye, typhoon cloud system, and peripheral airflow. For each image element in the meteorological training image, the remote monitoring device first determines the maximum support coefficient in its abnormal classification prediction information. Suppose the support coefficient of an image element belonging to the typhoon eye is 0.9, the support coefficient belonging to the typhoon cloud system is 0.05, and the support coefficient belonging to the peripheral airflow is 0.05. Then the maximum support coefficient is 0.9, and the corresponding element classification is the typhoon eye. The target prediction information of this image element is (0.9, typhoon eye).

[0137] Then, the remote monitoring device counts the target prediction information of all image elements in the meteorological training image. If it is found that there are multiple image elements with the target prediction information of (high support coefficient, typhoon eye), and these elements form a relatively concentrated area in space, and there are also a large number of image elements with the target prediction information of (high support coefficient, typhoon cloud system) surrounding them, then it can be judged that there is an anomaly in the meteorological training image under the typhoon detection branch, that is, the sign of a typhoon may be detected.

[0138] As another implementation manner, the element classification of abnormal meteorology under the target detection branch includes at least one. The abnormal classification prediction information is also used to indicate: when the element classification of the corresponding image element belongs to different element classifications, the probability that the element classification of adjacent image elements belongs to different element classifications, and the probability is represented based on the conversion support coefficient of adjacent image elements; the abnormal classification prediction information includes: the conversion support coefficient and the support coefficient that the element classification of the corresponding image element belongs to different element classifications; any support coefficient is used to indicate the probability that the element classification of the corresponding image element belongs to the corresponding element classification. Based on this, step S222, based on the abnormal classification prediction information corresponding to each image element in the meteorological training image, generate the abnormal meteorology detection result of the meteorological training image under the target detection branch, including:

[0139] Step S222A: Generate at least one element classification trajectory corresponding to the meteorological training image according to the abnormal classification prediction information; any element classification trajectory includes the element classification of each image element in the meteorological training image; any element classification trajectory includes the support coefficient of the element classification of each image element in the corresponding element classification trajectory and the conversion support coefficient of adjacent image elements;

[0140] Step S222B: Determine the trajectory support coefficient of any element classification trajectory based on the support coefficient and conversion support coefficient corresponding to any element classification trajectory;

[0141] Step S222C: Determine the element classification trajectory corresponding to the maximum trajectory support coefficient among at least one element classification trajectory, and generate the abnormal meteorological detection result of the meteorological training image under the target detection branch according to the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient.

[0142] In step S222A, the remote monitoring device generates at least one element classification trajectory corresponding to the meteorological training image according to the abnormal classification prediction information. Any element classification trajectory includes the element classification of each image element in the meteorological training image, and at the same time includes the support coefficient of the element classification of each image element in the corresponding element classification trajectory and the conversion support coefficient of adjacent image elements. The abnormal classification prediction information not only indicates the probability (i.e., support coefficient) of each image element belonging to different element classifications, but also reflects the probability (i.e., conversion support coefficient) of adjacent image elements converting from one element classification to another. The element classification trajectory is a sequence formed by arranging the element classifications of all image elements in the meteorological training image in a certain order and combining the corresponding support coefficients and conversion support coefficients.

[0143] Taking the object detection branch for heavy precipitation detection as an example, assume that the element classification of abnormal weather has three types: heavy precipitation area, light precipitation area, and no precipitation area. For three adjacent image elements A, B, and C in a meteorological training image, the abnormal classification prediction information shows that: the support coefficient of image element A belonging to the heavy precipitation area is 0.8, the support coefficient of belonging to the light precipitation area is 0.1, and the support coefficient of belonging to the no precipitation area is 0.1; the support coefficient of image element B belonging to the heavy precipitation area is 0.7, the support coefficient of belonging to the light precipitation area is 0.2, and the support coefficient of belonging to the no precipitation area is 0.1; the support coefficient of image element C belonging to the heavy precipitation area is 0.6, the support coefficient of belonging to the light precipitation area is 0.3, and the support coefficient of belonging to the no precipitation area is 0.1. At the same time, from image element A to image element B, the transition support coefficient from the heavy precipitation area to the heavy precipitation area is 0.8, the transition support coefficient from the heavy precipitation area to the light precipitation area is 0.1, and the transition support coefficient from the heavy precipitation area to the no precipitation area is 0.1; there are also corresponding transition support coefficients from image element B to image element C. Based on this information, an element classification trajectory can be generated: (A: heavy precipitation area, 0.8; B: heavy precipitation area, 0.7; C: heavy precipitation area, 0.6), where each element in the parentheses represents the image element, element classification, and support coefficient in sequence, and the transition support coefficients between adjacent image elements are implied in this trajectory. To generate the element classification trajectory, the remote monitoring device can use the exhaustive method or a search algorithm. The exhaustive method will traverse all possible combinations of element classifications, calculate the support coefficients and transition support coefficients corresponding to each combination, and thus generate all possible element classification trajectories. The search algorithm can search for a better element classification trajectory within a certain search space to reduce the amount of calculation. For example, using the dynamic programming algorithm, by recording intermediate results, repeated calculations can be avoided, and the efficiency of generating the element classification trajectory can be improved.

[0144] In step S222B, the remote monitoring device determines the trajectory support coefficient of any element classification trajectory based on the support coefficient and transition support coefficient corresponding to the element classification trajectory. The trajectory support coefficient is an index to measure the rationality and possibility of the element classification trajectory, which comprehensively considers the support coefficient of each image element and the transition support coefficient between adjacent image elements. A feasible method for calculating the trajectory support coefficient is to multiply the support coefficient of each image element in the element classification trajectory by the transition support coefficient of adjacent image elements.

[0145] Let the element classification trajectory be , where I i represents the i-th image element, C i represents its element classification, s i represents its support coefficient, and the adjacent image element Ii to I i+1 The conversion support coefficient to is t i, i+1 , then the trajectory support coefficient S of this element classification trajectory can be expressed as: ; Continuing with the above example of heavy precipitation detection, for the element classification trajectory (A: heavy precipitation area, 0.8; B: heavy precipitation area, 0.7; C: heavy precipitation area, 0.6), assuming the conversion support coefficient t from A to B A, B = 0.8, and the conversion support coefficient t from B to C B, C = 0.7, then the trajectory support coefficient of this trajectory .

[0146] In step S222C, the remote monitoring device determines the element classification trajectory corresponding to the maximum trajectory support coefficient among at least one element classification trajectory, and generates an abnormal meteorological detection result of the meteorological training image under the target detection branch based on the element classification included in this element classification trajectory. The element classification trajectory corresponding to the maximum trajectory support coefficient represents the most likely element classification combination. By analyzing the distribution of element classifications in this trajectory, it can be determined whether there is an abnormality in the meteorological training image under the target detection branch. For example, in heavy precipitation detection, if the element classification trajectory corresponding to the maximum trajectory support coefficient shows that a large area of image elements in the meteorological training image are classified as heavy precipitation areas and these areas are spatially continuous, then the remote monitoring device can determine that there is an abnormality in the meteorological training image under the heavy precipitation detection branch. The specific judgment criteria can be preset according to the actual situation. For example, the area ratio of the heavy precipitation area reaches a certain threshold (such as 15%), or the heavy precipitation area forms a specific shape or pattern, etc.

[0147] As an implementation manner, in step S40, according to the abnormal meteorological detection result of the meteorological training image under the target detection branch and the abnormal meteorological detection result under the target detection branch, network debugging operations are performed on the meteorological image detection network, including:

[0148] Step S41: Generate a target training cost under the target detection branch according to the error between the abnormal meteorological detection result of the meteorological training image under the target detection branch and the abnormal meteorological detection result under the target detection branch;

[0149] Step S42: Iterate the parameter variables of the meteorological image detection network according to the goal of reducing the target training cost to perform network debugging operations on the meteorological image detection network.

[0150] In step S41, the remote monitoring device generates the target training cost under the target detection branch based on the abnormal meteorological detection results of the meteorological training image under the target detection branch and the error between the abnormal meteorological detection results under the target detection branch. Here, two sets of abnormal meteorological detection results are involved. One set is the results obtained by the meteorological image detection network detecting the meteorological training image, and the other set is the true abnormal meteorological detection results of the meteorological training image under the target detection branch indicated by the target prior label. The error is an index measuring the degree of difference between these two sets of results, and the target training cost is a value used to evaluate the current performance of the network, which reflects the degree of deviation between the network prediction result and the true result.

[0151] To generate the target training cost, the remote monitoring device can use a loss function, such as the cross-entropy loss function for calculation.

[0152] In step S42, the remote monitoring device iterates the parameter variables of the meteorological image detection network according to the goal of reducing the target training cost to perform network debugging operations on the meteorological image detection network. The parameter variables of the meteorological image detection network mainly include the weights and biases in the network, and these parameters determine the calculation method and output results of the network. Iteration means continuously adjusting the parameter variables, and recalculating the target training cost after each adjustment until the target training cost reaches a satisfactory value or meets certain stopping conditions.

[0153] To achieve the iterative adjustment of the parameter variables, the remote monitoring device usually adopts an optimization algorithm, such as the Stochastic Gradient Descent (SGD) algorithm. The basic idea of the SGD algorithm is to update the parameter variables according to the gradient of the target training cost with respect to the parameter variables. Let the parameter variables be , and the target training cost be . Then, at the t-th iteration, the update formula for the parameter variables is: ; where is the learning rate, which controls the step size of the parameter variable update in each iteration, is the gradient of the target training cost at .

[0154] In practical applications, to improve the effect of network debugging, the remote monitoring device can adopt some optimization strategies. The method of learning rate decay can be adopted. As the number of iterations increases, the learning rate is gradually reduced, so that it can converge quickly in the initial stage of training and adjust the parameters more finely in the later stage. The batch normalization technique can also be adopted to normalize the input of each layer in the network, accelerating the convergence speed of the network and improving the stability. In addition, to prevent the network from overfitting, regularization techniques, such as L2 regularization, can be adopted.

[0155] Steps S41 - S42 continuously optimize the network performance by calculating the target training cost and iteratively adjusting the parameters of the meteorological image detection network. In the remote monitoring of a regulatory meteorological service system with low bandwidth, the debugged meteorological image detection network can more accurately detect abnormal meteorological conditions in meteorological images under the target detection branch, providing strong support for the stable operation of the meteorological service system and meteorological disaster early warning.

[0156] Figure 2 The following is a schematic diagram of the hardware entity of a remote monitoring device provided by an embodiment of the present invention, as Figure 2 shown. The hardware entity of the remote monitoring device 1000 includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any of the above embodiments.

Claims

1. A low-bandwidth remote monitoring method for a regulatory meteorological service system, characterized in that: Applied to a remote monitoring device, the remote monitoring device is communicatively connected to a receiving end, and the method comprises: Acquire the meteorological image to be detected; Determine whether the change area of ​​the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold; or, detect whether the meteorological image to be detected contains an abnormal detection result; When a change area of ​​the meteorological image to be detected compared with the previous meteorological image is greater than a change threshold, the change area is encoded to obtain a change area encoding result, and the change area encoding result is sent to a receiving end, so that the receiving end synthesizes the change area encoding result and the previous meteorological image based on the change area encoding result to obtain a synthesized meteorological image; Alternatively, when the meteorological image to be detected includes an abnormal detection result, the abnormal detection result is encoded to obtain an abnormal encoding result, and the abnormal encoding result is sent to the receiving end, so that the receiving end synthesizes the abnormal encoding result and the previous meteorological image based on the abnormal encoding result to obtain a synthesized meteorological image; The step of judging whether a change area of ​​the meteorological image to be detected compared with the previous meteorological image is greater than a change threshold comprises: Performing image segmentation on the previous meteorological image and the meteorological image to be detected respectively, to obtain image segmentation sets corresponding to the previous meteorological image and the meteorological image to be detected respectively, wherein the image segmentation set of the previous meteorological image and the image segments in the image segmentation set of the meteorological image to be detected correspond to each other one by one; Calculate the differences between two image blocks that correspond to each other one by one, and obtain multiple block difference results; Averaging the plurality of block difference results to obtain a mean difference result, and if the mean difference result is greater than the change threshold, determining that the change area of ​​the meteorological image to be detected compared with the previous meteorological image is greater than the change threshold; When sending the changed region encoding result to the receiving end, or sending the abnormal encoding result to the receiving end, a token bucket algorithm or a leaky bucket algorithm is used to control the bandwidth; The detecting whether the meteorological image to be detected contains an abnormal detection result includes: Acquire a meteorological image to be detected and branch prompt information, wherein the branch prompt information is used to indicate a detection branch of the meteorological image to be detected and the number of branches corresponding to the detection branch; Acquire a branch meteorological image matching a detection branch of a corresponding number of branches from at least one candidate meteorological image according to the branch prompt information, and generate a target meteorological image according to the branch meteorological image and the meteorological image to be detected; Through the debugged meteorological image detection network, abnormal meteorological detection is performed on the target meteorological image according to the detection branch indicated by the branch prompt information, and the abnormal detection result of the target meteorological image under the detection branch is obtained.

2. The method according to claim 1, characterized in that The meteorological image detection network completed through debugging performs abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information, and obtains the abnormal detection result of the target meteorological image under the detection branch, including: The meteorological image detection network that has been debugged performs abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information to obtain a first abnormal detection result; wherein the first abnormal detection result includes the element classification corresponding to the maximum support coefficient of each image element in the target meteorological image, and any support coefficient is used to indicate the probability that the element classification of any image element belongs to the corresponding element classification; Based on the element classification corresponding to the maximum support coefficient of each image element in the target meteorological image, detecting and obtaining the abnormality detection result of the meteorological image to be detected in the target meteorological image under the detection branch; Alternatively, the meteorological image detection network completed through debugging performs abnormal meteorological detection on the target meteorological image according to the detection branch indicated by the branch prompt information, and obtains the abnormal detection result of the target meteorological image under the detection branch, including: The meteorological image detection network completed through debugging performs abnormal meteorological detection on the target meteorological image according to the target detection branch to obtain a second abnormal detection result; the second abnormal detection result includes the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient; wherein any element classification trajectory includes the support coefficient of the element classification of each image element of the target meteorological image under the corresponding element classification trajectory, and the conversion support coefficient of the element classification corresponding to the adjacent image elements; the trajectory support coefficient of any element classification trajectory is determined based on the support coefficient and the conversion support coefficient corresponding to the any element classification trajectory; Based on the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient, an abnormality detection result of the to-be-detected meteorological image in the target meteorological image under the detection branch is detected.

3. The method according to claim 2, characterized in that The meteorological image detection network is debugged by the following steps: Acquire a meteorological training image and a target detection branch of the meteorological training image; the meteorological training image is matched with at least one priori mark, and any priori mark is used to indicate an abnormal meteorological detection result of the meteorological training image under a corresponding detection branch; Performing abnormal meteorological detection on the meteorological training image according to the target detection branch to obtain an abnormal meteorological detection result of the meteorological training image under the target detection branch; Determine a target priori marker from at least one priori marker matched by the meteorological training image, and determine an abnormal meteorological detection result of the meteorological training image under the target detection branch from the target priori marker, wherein the target priori marker is a priori marker that corresponds to a detection branch and is the same as the target detection branch; According to the abnormal meteorological detection result of the meteorological training image under the target detection branch and the abnormal meteorological detection result under the target detection branch, a network debugging operation is performed on the meteorological image detection network; the debugged meteorological image detection network is used to perform abnormal meteorological detection of meteorological images under the corresponding detection branch; The number of the target detection branches is at least one, and the number of abnormal meteorological detection results obtained by performing abnormal meteorological detection on the meteorological training image according to the target detection branches is the same as the number of the target detection branches; the number of detection branches corresponding to the priori marks matched by the meteorological training image is not less than the number corresponding to the target detection branches; wherein, when the number of the target detection branches is multiple, the number of abnormal meteorological detection results obtained is also multiple, and the number of detection branches corresponding to the target priori marks determined according to the priori marks matched by the meteorological training image is equal to the number of the target detection branches; the meteorological image detection network includes a representation information extraction component and an abnormal reasoning network; the representation information extraction component is used to extract meteorological image representation vectors; the abnormal reasoning network is used to perform abnormal meteorological detection on the meteorological training image to obtain corresponding abnormal meteorological detection results; The method for obtaining the meteorological training image includes: Acquire at least one candidate meteorological image and branch prompt information; the branch prompt information is used to prompt the number of branches of the target detection branch; Determine, according to the branch prompt information, a candidate meteorological image that matches the target detection branch of the corresponding number of branches in the at least one candidate meteorological image, and use the determined candidate meteorological image as a branch meteorological image of the meteorological training image; A sample meteorological image is obtained, and the branch meteorological image and the sample meteorological image are combined to obtain a meteorological training image.

4. The method according to claim 3, characterized in that The step of performing abnormal meteorological detection on the meteorological training image according to the target detection branch to obtain an abnormal meteorological detection result of the meteorological training image under the target detection branch includes: Extracting representation information of the meteorological training image according to the target detection branch to obtain a meteorological image representation vector corresponding to the meteorological training image; the meteorological image representation vector is used to represent feature information of image elements in the meteorological training image and the relationship between different image elements; Abnormal meteorological detection is performed on the meteorological training image based on the meteorological image characterization vector to obtain an abnormal meteorological detection result of the meteorological training image under the target detection branch.

5. The method according to claim 4, characterized in that Before performing abnormal meteorological detection on the meteorological training image based on the meteorological image representation vector and obtaining the abnormal meteorological detection result of the meteorological training image under the target detection branch, the method further includes: Obtaining the meteorological image type corresponding to the meteorological image to which each image element in the meteorological training image belongs; Generate a shielding matrix based on the meteorological image type; the shielding matrix includes a plurality of shielding elements, and the image elements included in the meteorological training image correspond to the shielding elements in the shielding matrix; Based on the shielding matrix, a shielding operation is performed on the meteorological image representation vector of the meteorological training image to obtain a shielded meteorological image representation vector; the shielded meteorological image representation vector includes an element representation vector corresponding to each image element in the sample meteorological image, and the shielded meteorological image representation vector is used to perform abnormal meteorological detection on the meteorological training image; The performing abnormal meteorological detection on the meteorological training image based on the meteorological image representation vector to obtain the abnormal meteorological detection result of the meteorological training image under the target detection branch includes: Based on the meteorological image representation vector, the abnormal classification prediction information of each image element in the meteorological training image is inferred; the abnormal classification prediction information of any image element represents the probability that the element classification of the corresponding image element belongs to the element classification of abnormal meteorology under the target detection branch; Based on the abnormal classification prediction information corresponding to each image element in the meteorological training image, an abnormal meteorological detection result of the meteorological training image under the target detection branch is generated.

6. The method according to claim 5, characterized in that The element classification of abnormal weather under the target detection branch includes at least one, and the abnormal classification prediction information of any image element includes: the support coefficient of the element classification of the corresponding image element belonging to different element classifications, and any support coefficient is used to indicate the probability that the element classification of the corresponding image element belongs to the corresponding element classification; the abnormal weather detection result of the meteorological training image under the target detection branch is generated based on the abnormal classification prediction information corresponding to each image element in the meteorological training image, including: Determine a maximum support coefficient among the support coefficients included in the abnormal classification prediction information of any image element; classifying the maximum support coefficient and the elements corresponding to the maximum support coefficient as target prediction information corresponding to any one of the image elements; According to the target prediction information of each image element in the meteorological training image, an abnormal meteorological detection result of the meteorological training image under the target detection branch is generated.

7. The method according to claim 5, characterized in that The element classification of abnormal weather under the target detection branch includes at least one, and the abnormal classification prediction information is also used to indicate the probability that the element classification of the corresponding image element belongs to a different element classification, and the probability is expressed based on the conversion support coefficient of the adjacent image element; The abnormal classification prediction information includes: the conversion support coefficient and the support coefficient of the element classification of the corresponding image element belonging to different element classifications; any support coefficient is used to indicate the probability that the element classification of the corresponding image element belongs to the corresponding element classification; the abnormal classification prediction information corresponding to each image element in the meteorological training image is used to generate the abnormal meteorological detection result of the meteorological training image under the target detection branch, including: At least one element classification trajectory corresponding to the meteorological training image is generated according to the abnormal classification prediction information; any element classification trajectory includes the element classification of each image element in the meteorological training image; any element classification trajectory includes the support coefficient of the element classification of each image element in the meteorological training image under the corresponding element classification trajectory, and the conversion support coefficient of the adjacent image elements; Determining a trajectory support coefficient of any element classification trajectory based on a support coefficient and a conversion support coefficient corresponding to any element classification trajectory; An element classification trajectory corresponding to a maximum trajectory support coefficient is determined in the at least one element classification trajectory, and an abnormal meteorological detection result of the meteorological training image under the target detection branch is generated based on the element classification included in the element classification trajectory corresponding to the maximum trajectory support coefficient.

8. The method according to claim 3, characterized in that The network debugging operation is performed on the meteorological image detection network according to the abnormal meteorological detection result of the meteorological training image under the target detection branch and the abnormal meteorological detection result under the target detection branch, including: Generate a target training cost under the target detection branch according to an abnormal meteorological detection result of the meteorological training image under the target detection branch and an error between the abnormal meteorological detection results under the target detection branch; According to the goal of reducing the target training cost, the parameter variables of the meteorological image detection network are iterated to perform a network debugging operation on the meteorological image detection network.

9. A remote monitoring device, 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 8 are implemented.

Images