Meteorological data acquisition system based on meteorological disaster early warning

Abstract

The invention discloses a meteorological data acquisition system based on meteorological disaster early warning, and relates to the field of meteorological data processing, historical disaster image data and time labels of a target area are acquired, and a training sample set is constructed; training a depth image discrimination model based on the sample set, wherein the depth image discrimination model is used for identifying a disaster cloud cluster form in the real-time image; acquiring image data streams uploaded by a plurality of image acquisition nodes and inputting the image data streams into the model to generate a preliminary judgment result; calculating the disaster similarity based on a judgment result, constructing an image risk thermodynamic diagram, and screening out high-risk nodes; performing sampling frequency scheduling on the high-risk nodes to improve the image updating frequency; further extracting multi-dimensional features such as a wind field, a cloud system form and a brightness temperature gradient, and generating a feature vector set; according to the invention, efficient sensing and dynamic early warning of small-scale and high-locality meteorological disasters can be realized, and the method has the advantages of high response speed and flexible deployment.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing, and specifically to a meteorological data acquisition system for meteorological disaster early warning. Background Technology

[0002] With the increasing frequency of extreme weather events, such as localized tornadoes, severe convective weather, sudden hail, and short-duration heavy rainfall, small-scale meteorological disasters are characterized by high frequency, localization, and suddenness. Traditional meteorological data acquisition methods have revealed significant shortcomings in terms of spatial resolution, temporal accuracy, and disaster response delays. Existing systems mostly rely on large-scale radar, fixed ground meteorological stations, and satellite remote sensing imagery for data acquisition. While these methods are suitable for macro-level meteorological monitoring, they have low recognition rates for small-scale meteorological changes in sparsely distributed or topographically complex areas (such as mountainous areas, canyons, and urban-rural transition zones), making it difficult to capture potential disaster signs in a timely manner.

[0003] For example, micro tornadoes that suddenly appear in urban-rural fringe areas often go undetected and cause serious damage because they cannot be captured by traditional observation equipment. Especially in densely populated areas (such as open-air concerts and farmers' markets), sudden weather events can develop rapidly within minutes, and existing data acquisition systems lack response mechanisms to support high-frequency data collection and intelligent processing in a short period of time, resulting in a serious deficiency in early warning capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a meteorological data acquisition system for meteorological disaster early warning, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a meteorological data acquisition system for meteorological disaster early warning, comprising: Acquisition module: Acquires historical meteorological disaster image data sequences and corresponding time labels for the target area, and constructs a training sample set T; Model training module: Based on the sample set T, a deep image discrimination model M for small-scale meteorological disaster feature recognition is constructed and trained to identify the morphology of potential disaster cloud clusters from real-time image data; Data acquisition module: acquires multiple image acquisition nodes deployed within the target area. Uploaded real-time image data stream The image data streams are then input into model M to obtain preliminary discrimination results. where n is the total number of images; Image analysis module: based on various discrimination results Based on the disaster similarity values, a regional image risk heatmap R is constructed, and areas with heatmap values ​​greater than a set threshold are selected. G is a set of high-risk image nodes. Frequency scheduling module: Schedules the acquisition frequencies of image acquisition nodes in set G, changing their image acquisition frequencies from the initial frequency. Increase to weighted sampling frequency ; Feature extraction module: The updated high-frequency image stream is input into model M again, and a set of feature vectors containing wind field, cloud morphology and brightness temperature gradient is generated by combining the corresponding image metadata.

[0006] Preferably, the acquisition module includes: The system calls upon a multi-source historical image database to extract satellite cloud images and ground-based real-time images of severe weather within a target area over a historical period, according to a preset time window. Preprocess the extracted image; The cloud types, morphological boundaries, and brightness temperature features in the image are labeled to form a structured image feature set; The structured image features are matched with the corresponding historical disaster occurrence times to generate a training sample set T with time labels.

[0007] Preferably, the model training module includes: The initial network structure of the deep image discrimination model M is constructed based on the training sample set T. The network structure includes a multi-scale feature extraction sub-network consisting of convolutional layers, pooling layers and skip connections, which is used to extract local texture and overall morphological features of clouds. The model M is trained in a supervised manner using structured image features from the sample set T as input data and corresponding disaster time labels as supervision signals. The trained model M is validated, and its hyperparameters are optimized through cross-validation.

[0008] Preferably, the data acquisition module includes: Multiple image acquisition nodes are deployed at a preset density within the target area. ; Each image acquisition node uploads a real-time image data stream at a fixed time interval t. To edge computing units; The edge computing unit performs preliminary preprocessing on each received image data stream, including compression decoding, image enhancement, and abnormal frame removal. The preprocessed image data is synchronously input into the trained model M, and preliminary discrimination results are output. The results include a probability map and spatial location information of the disaster cloud region in each frame of the image.

[0009] Preferably, the image analysis module includes: For each preliminary judgment result The probability map of disaster cloud clusters output by the system is used to perform regional connectivity analysis, extract suspected disaster areas and calculate their morphological characteristic parameters; Based on the model's output probability mean, area ratio, and edge complexity features, and combined with set weights, a disaster similarity value S is calculated to reflect the degree of similarity between cloud features in the image and historical disaster images. The similarity values ​​of each image acquisition node are bound to the GPS coordinates of the node and mapped to a unified geographic grid of the target area to form a spatial risk distribution matrix; The similarity matrix is ​​spatially expanded using a bilinear interpolation algorithm to construct a continuous regional image risk heat map R, which uses heat color to reflect the disaster potential of different regions.

[0010] Preferably, the frequency scheduling module includes: Obtain the disaster similarity value S of each node in the high-risk image node set G, and classify them into three risk levels: Level 1, Level 2, and Level 3 based on the similarity. Set the corresponding weighted sampling frequency based on the risk level of the node. The According to the proportional relationship, based on the initial sampling frequency The frequency of sampling at Level 1 risk nodes has been increased to [number missing]. Level 2 is Level 3 ; A frequency adjustment command is issued to the nodes in set G to update their acquisition and scheduling parameters so that they operate at the new frequency. Collect image data and upload it.

[0011] Preferably, the feature extraction module includes: The high-frequency image stream obtained after scheduling is re-inputted into the trained deep image discrimination model M to extract the disaster cloud region in the corresponding image frame; By combining image acquisition time and location information, the cloud motion vector field is estimated using the inter-frame optical flow method, and a velocity and direction feature matrix reflecting wind field changes is constructed. Using image segmentation results, cloud boundaries, thickness, and convolution density are extracted; Gradient calculations are performed on the brightness temperature distribution of disaster areas in infrared channel images to form a brightness temperature gradient feature map, which is then fused with wind field and morphological features to generate a set of feature vectors.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention achieves high-timeliness, high-precision, and dynamic perception of small-scale meteorological disasters by constructing a multi-module collaborative system consisting of image acquisition, model training, real-time data collection, image analysis, frequency scheduling, and feature extraction. Compared to existing data acquisition schemes primarily based on radar or fixed weather stations, this invention introduces an image-based disaster identification model and a regional heat map construction mechanism, which can effectively improve the spatial coverage identification capability of disaster cloud clusters, and is particularly suitable for areas with complex terrain and frequent observation blind spots.

[0013] 2. This invention achieves intelligent sampling control of high-risk image nodes through a frequency scheduling mechanism. Combined with multi-dimensional physical features such as wind field estimation, brightness-temperature gradient, and cloud morphology, it forms a structured feature vector set that can be used for disaster level assessment. This scheme not only reduces system resource consumption but also enhances the model's ability to track and dynamically respond to disaster evolution. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the system modules of the present invention.

[0016] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 , 2 As shown in this embodiment, a meteorological data acquisition system for meteorological disaster early warning includes: Acquisition module: Acquires historical meteorological disaster image data sequences and corresponding time labels for the target area, and constructs a training sample set T; Model training module: Based on the sample set T, a deep image discrimination model M for small-scale meteorological disaster feature recognition is constructed and trained to identify the morphology of potential disaster cloud clusters from real-time image data; Data acquisition module: acquires multiple image acquisition nodes deployed within the target area. Uploaded real-time image data stream The image data streams are then input into model M to obtain preliminary discrimination results. where n is the total number of images; Image analysis module: based on various discrimination results Based on the disaster similarity values, a regional image risk heatmap R is constructed, and areas with heatmap values ​​greater than a set threshold are selected. G is a set of high-risk image nodes. Frequency scheduling module: Schedules the acquisition frequencies of image acquisition nodes in set G, changing their image acquisition frequencies from the initial frequency. Increase to weighted sampling frequency ; Feature extraction module: The updated high-frequency image stream is input into model M again, and a set of feature vectors containing wind field, cloud morphology and brightness temperature gradient is generated by combining the corresponding image metadata.

[0019] To achieve effective identification of severe weather within the target area and model training, the acquisition module collects and constructs a training sample set T through the following specific steps: The system calls upon a multi-source historical image database to extract satellite cloud images and ground-based real-time images of severe weather within a target area over a historical period, according to a preset time window. By setting a time window (e.g., June to September each year is the peak period for severe convection), the system accesses a multi-source historical image database, including data interfaces from the National Meteorological Satellite Center, ground monitoring image storage systems, and UAV aerial photography databases. For the target area, the system limits the image capture range by geographic coordinates and filters data by setting hazardous weather keywords (such as "thunderstorm," "hail," "tornado," and "short-term heavy rainfall").

[0020] The time window can be dynamically adjusted based on historical meteorological disaster records of the target area and can be configured as continuous or non-continuous time periods. The data is uniformly converted to JPEG or TIFF format via API calls, with a resolution of at least 1024×768 pixels to ensure the clarity requirements of subsequent image processing.

[0021] The extracted images are first subjected to image quality verification and standardization. The preprocessing steps include: Image denoising: A Gaussian filter is used for smoothing to remove random noise caused by sensor errors or transmission. Contrast Enhancement: Histogram equalization algorithm is used to enhance image contrast, making cloud boundaries clearer; Size normalization: Images from different sources are scaled to a fixed size (e.g., 256×256 pixels) to ensure consistent input dimensions for subsequent models; Image correction and registration: Spatial alignment of ground images and satellite images is performed using feature-point-based image registration algorithms (such as SIFT or ORB algorithms) to ensure consistent spatial dimensions.

[0022] The cloud types, morphological boundaries, and brightness temperature features in the image are labeled to form a structured image feature set, including: A pre-trained image segmentation model, U-Net, is used to perform semantic segmentation of cloud regions in each image and automatically identify cloud boundaries. The model is pre-trained on publicly available remote sensing image sets (such as SEVIR) and fine-tuned using local samples.

[0023] Specific feature extraction includes: Cloud type identification: Based on cloud thickness, texture, and brightness level, it is identified as cumulonimbus, stratus, cirrus, etc., and the classification output is performed using Softmax. Morphological boundary extraction: Generate boundary contours based on segmentation results, and calculate morphological indicators such as area, perimeter, and roundness; Brightness temperature feature extraction: The pixel brightness in the infrared image is converted into a brightness temperature value. The brightness temperature is calculated from the sensor radiometer value. A linear conversion model is used to map the pixel value to the corresponding brightness temperature (in K).

[0024] All extracted features are constructed into a structured data table, with each image corresponding to a set of feature vectors in the format: Ci={cloud type, boundary coordinate set, area, perimeter, maximum brightness temperature, average brightness temperature}; where Ci is the cloud feature set of the i-th image.

[0025] Structured image features are matched with the corresponding historical disaster occurrence times to generate a time-labeled training sample set T, including: By comparing the image capture timestamps with event timestamps in the meteorological disaster record database, image samples are associated with specific disaster events that have occurred. Disaster event records are sourced from local meteorological bureaus, emergency management departments, and news reports, and are automatically extracted via web crawlers and manually verified.

[0026] If the time interval between the image capture time and the occurrence of a known disaster event is less than a set threshold Δt (e.g., 30 minutes), the image is marked as a "disaster-related" sample and assigned a corresponding label, such as "severe convection" or "short-term heavy rainfall"; otherwise, it is marked as a "background sample".

[0027] The final training sample set T contains image feature vectors C, time labels T, and corresponding disaster types L. The data format is as follows: T={(C1,T1,L1),(C2,T2,L2),...,(Cn,Tn,Ln)}, where n is the total number of images. This training set will be used as the training input for the subsequent deep image discrimination model M.

[0028] To achieve automatic identification and discrimination of small-scale meteorological disaster characteristics, the model training module constructs a deep image discrimination model M and optimizes it based on the training sample set T, specifically including: The initial network structure of the deep image discrimination model M is constructed based on the training sample set T. The network structure includes a multi-scale feature extraction sub-network composed of convolutional layers, pooling layers, and skip connections, which is used to extract local texture and overall morphological features of clouds. Specifically, it includes: In this step, the depth image discrimination model M is constructed using an improved U-Net network structure, which includes two parts: an encoder (downsampling path) and a decoder (upsampling path). High-resolution features are transferred through skip connections.

[0029] Convolutional layer: used to extract low-level features (such as edges and textures) from the input image. It uses a 3×3 convolutional kernel for feature mapping and ReLU (corrected linear unit) activation function to prevent gradient vanishing. Pooling layer: In the encoding path, a 2×2 max pooling layer is added after every two convolutional layers to reduce spatial dimensionality and retain key features; Skip connection: Each layer of the decoder receives the high-resolution feature map output by the corresponding encoder, and improves the localization capability and boundary recognition effect through the stitching operation; Input and output: The model input is a preprocessed 256×256 pixel image, and the output is a disaster probability heatmap of the same size as the input (each pixel value represents the probability of belonging to a disaster cloud).

[0030] The multi-scale structure ensures the model's ability to comprehensively extract local details (such as cloud textures) and global shapes (such as cloud boundaries).

[0031] Using structured image features from sample set T as input data and corresponding disaster time labels as supervision signals, supervised training of model M is performed, specifically including: The sample set T consists of structured image features and time labels. Each sample contains an input image and its corresponding label image. The value of each pixel in the label image indicates whether it belongs to a disaster-related cloud cluster area, which constitutes the supervision signal for model training.

[0032] The training process employs supervised learning, and the specific methods are as follows: Loss function: The weighted cross-entropy loss function is adopted, which assigns higher weights (such as 1.5) to pixels in the disaster area and a weight of 1 to the background area, thereby enhancing the model's ability to learn disaster features; Optimization algorithm: The Adam optimizer is used for parameter updates, with the initial learning rate set to 0.001 and dynamically adjusted based on the performance of the validation set during training; Batch training method: Training is performed with a batch size of 16, and the network parameters are updated in each iteration. The network gradually converges through the backpropagation algorithm.

[0033] Training typically involves 100 iterations and takes approximately 2 hours, with the exact time depending on the size of the sample set.

[0034] The trained model M is validated and its hyperparameters are optimized through cross-validation, specifically including: After training, model M is validated to ensure it has good generalization ability under different image sources and weather conditions. The specific steps are as follows: Validation set partitioning: 20% of the data in the sample set is randomly selected as the validation set, and the remaining 80% is used for training to ensure the independence of the validation samples; Evaluation metrics: Precision, Recall, F1-score, and IoU (Intersection over Union) are used as evaluation metrics. Models with an IoU below 0.5 are considered to have unsatisfactory performance. Cross-validation: A 5-fold cross-validation method is used to alternate between training and validation on the training data, and the results of each round are averaged to avoid the model from overfitting to specific samples; Based on the validation results, key hyperparameters, including learning rate, kernel size, and weight decay coefficient, were adjusted to determine the optimal model version for subsequent real-time image recognition deployment.

[0035] The final trained model M can receive image input in real time and output a probability map of disaster clouds.

[0036] To achieve continuous and efficient acquisition and model discrimination input of real-time meteorological image data within the target area, the data acquisition module adopts an architecture combining multi-point deployment, edge computing, and model integration recognition, specifically including: Multiple image acquisition nodes are deployed at a preset density within the target area. Specifically, it includes: Image acquisition nodes are deployed in zones within the target area according to geographical features and risk levels. The nodes... For independently installed image sensing devices, each node includes: High-definition visible light camera with a resolution of no less than 1280×720 pixels; Low-power wireless communication modules, such as LoRa, NB-IoT, or LTE modules, are used for image data uploading; Built-in GPS positioning module provides accurate geographic coordinate information (accuracy not less than ±10 meters). The power supply can be either solar power or mains power, ensuring that the node operates 24 / 7.

[0037] The preset density is dynamically determined based on the area and monitoring accuracy requirements. For example, 2-3 nodes are deployed per square kilometer in sensitive urban areas, while one node is deployed per 5 square kilometers in rural areas. Node deployment information is uniformly recorded in the management platform for subsequent image data attribution and synchronization processing.

[0038] Each image acquisition node uploads a real-time image data stream at a fixed time interval t. To the edge computing unit, specifically including: Each node periodically collects image frames within the current field of view at a set time interval t (recommended value is 30 seconds, but can be configured to 5~60 seconds in practice), and uploads the image along with the timestamp and GPS location information to the nearest edge computing unit.

[0039] Edge computing units are located at points with good network communication conditions within the region, typically base stations or network relay points, and have local storage and preliminary processing capabilities.

[0040] The edge computing unit performs preliminary preprocessing on each received image data stream, including compression decoding, image enhancement, and abnormal frame removal, specifically including: The edge computing unit performs the following preprocessing steps on the image data stream: Compression Decoding: Decodes JPEG encoded images to restore them to standard RGB format pixel images; Image enhancement: Histogram equalization is used to adjust the image brightness and enhance cloud edge and texture information; Abnormal frame removal: The pixel change rate between the current image and the previous frame is calculated using a differential detection algorithm. When the change rate is less than a set threshold (e.g., 2%) or the overall brightness of the image is abnormal (e.g., a black frame at night), it is determined to be an abnormal frame and removed.

[0041] The preprocessed image data is synchronously input into the trained model M, and preliminary discrimination results are output. The results include a probability map and spatial location information of the disaster cloud region in each frame of the image, specifically including: The preprocessed images are sequentially input into the previously trained deep image discrimination model M, which is based on the U-Net architecture and has been trained and optimized on the historical sample set T. Model M is deployed locally in the edge computing unit, supporting fast inference operations.

[0042] After processing each frame of image, a probability map of the same size as the input image is output. The value of each pixel is a decimal between 0 and 1, representing the probability that the location belongs to a disaster cloud cluster. In addition, the system extracts the contour coordinates of high-probability areas in the image and combines them with the image's original GPS information to map the distribution range of the disaster cloud cluster in the actual geographic space.

[0043] Output Data is sent to the image analysis module via the data bus for regional risk aggregation and heat map construction.

[0044] The image analysis module is used to analyze preliminary judgment results. The system performs disaster similarity calculations and spatial risk modeling on images acquired by each image acquisition node, ultimately outputting a regional image risk heatmap R and a set of high-risk nodes G. This provides a basis for subsequent early warning and acquisition frequency scheduling, specifically including: For each preliminary judgment result Regional connectivity analysis is performed on the disaster cloud probability map output from the database to extract suspected disaster areas and calculate their morphological characteristic parameters, specifically including: Model M outputs a probability map for each frame of image as a single-channel grayscale image, with each pixel value ranging from 0 to 1, representing the probability that the pixel belongs to a disaster cloud cluster.

[0045] In this step, the system uses a set probability threshold (e.g., 0.6) to binarize the probability map, and pixels above this threshold are considered suspected disaster areas. Subsequently, an eight-neighborhood connectivity algorithm is used to cluster consecutive high-probability pixels, extract each connected region, and calculate the following morphological feature parameters for it: area (number of pixels per unit); perimeter; edge complexity = ratio of perimeter to area; aspect ratio and geometric moment index.

[0046] The above parameters are used for subsequent disaster similarity calculations and can filter out small, misidentified interference areas (such as areas less than 50 pixels).

[0047] Based on the model's output probability mean, area proportion, and edge complexity features, and combined with set weights, a disaster similarity value S is calculated to reflect the degree of similarity between cloud features in the image and historical disaster images. Specifically, this includes: The suspected disaster areas extracted from each image correspond to the following three core features: Probability mean (P_mean): The average probability of all pixels within the disaster area; Area ratio (A_ratio): The ratio of the disaster area to the total area of ​​the image; Edge complexity (E_complex): As mentioned earlier, it is the ratio of the perimeter of the outline to its area.

[0048] Let the corresponding weight coefficients be α, β, and γ, respectively, and let the three satisfy the following conditions: (For example, α=0.4, β=0.4, γ=0.2), then the formula for calculating the disaster similarity value S is: The edge complexity is calculated by taking the reciprocal, which results in higher scores for regions with high edge clarity and regular shapes. The value of S ranges from 0 to 1, with values ​​closer to 1 indicating that the image is more likely to correspond to hazardous weather clouds.

[0049] The similarity values ​​of each image acquisition node are bound to the node's GPS coordinates and mapped to a unified geographic grid of the target area, forming a spatial risk distribution matrix, which specifically includes: Each image acquisition node Ni uploads its current GPS coordinates (Latitude Lat, Longitude Lon) along with the image. After calculating the corresponding similarity value S, the system maps it to the designated regional geographic grid. The specific method is as follows: Divide the target area into a fixed grid (e.g., each grid is 0.01 degrees × 0.01 degrees, in latitude and longitude units); Each image acquisition node is assigned to its respective grid. The similarity value S of each node in a grid is used as the initial risk factor for that grid. If there are multiple nodes in a grid, their average value is taken.

[0050] This results in a two-dimensional spatial risk distribution matrix, where each cell represents the disaster similarity value corresponding to that geographic grid.

[0051] The similarity matrix is ​​spatially expanded using a bilinear interpolation algorithm to construct a continuous regional image risk heatmap R. This heatmap R uses heat color to reflect the disaster potential of different regions; and regions with heat values ​​greater than a set threshold are selected. The set of high-risk image nodes G includes: To achieve a more continuous and refined spatial representation of the risk map, this step employs bilinear interpolation to interpolate the spatial distribution matrix, filling in the grid regions that were not directly sampled. The algorithm calculates the current grid value using a distance-weighted average, centered on each empty grid and referencing the similarity values ​​of its four adjacent grids, thus achieving a smooth spatial transition.

[0052] After interpolation, the system converts the risk matrix into an image and draws a regional image risk heat map R. The colors in the image, from cold (blue) to hot (red), represent similarity values ​​from low to high, intuitively displaying the distribution of disaster risks.

[0053] Finally, set a risk heat threshold. (e.g., 0.65), filter out all popularity values ​​exceeding The grid area is mapped back to the corresponding image acquisition node number, forming a high-risk image node set G.

[0054] The frequency scheduling module performs hierarchical scheduling based on the disaster similarity of image nodes, dynamically adjusting the image acquisition frequency to achieve more intensive image updates in key areas. This module includes: Obtain the disaster similarity value S for each node in the high-risk image node set G, and classify them into three risk levels (Level 1, Level 2, and Level 3) based on the similarity magnitude. Specifically, this includes: The disaster similarity value S of each image acquisition node is read from the set G output by the image analysis module. S is a real number in the interval [0,1], and the larger the value, the higher the disaster risk.

[0055] Based on the set grading thresholds, the system classifies similarity values ​​into three risk levels, as follows: Level 1 risk: similarity S ≥ 0.8; Level 2 risk: 0.65 ≤ S < 0.8; Level 3 risk: 0.5 ≤ S < 0.65; Nodes with similarity below 0.5 are not included in the scheduling scope. This grading rule can be customized by the user according to the actual scenario, and the grading results are stored in the scheduling index table for use in formulating subsequent sampling strategies.

[0056] Set the corresponding weighted sampling frequency based on the risk level of the node. The According to the proportional relationship, based on the initial sampling frequency The frequency of sampling at Level 1 risk nodes has been increased to [number missing]. Level 2 is Level 3 .

[0057] Set a new sampling frequency for each node based on its current risk level. This increases image acquisition density under high-risk conditions. The sampling frequency is adjusted using a multiplicative weighted strategy, with the specific calculation formula as follows: Level 1 risk node: Secondary risk node: Level 3 risk node: ;in, The default sampling frequency configured for the system (e.g., one frame every 30 seconds). Units and Consistent. The weighting coefficients can be flexibly adjusted according to system load capacity and early warning requirements. The frequency adjustment strategy is encapsulated as an instruction template and stored in the scheduling instruction queue along with parameters such as node ID, coordinates, and risk level, ready for issuance.

[0058] A frequency adjustment command is issued to the nodes in set G to update their acquisition scheduling parameters, enabling them to acquire and upload image data at the new frequency f1. Specifically, this includes: Frequency adjustment commands are sent to each node in set G via an edge control platform or wireless communication module. The command format includes: node unique identifier (ID); current risk level identifier; and new sampling frequency. ; Scheduling effective timestamp. After receiving the instruction, the node sets the local acquisition timer parameter f to . The system will immediately begin image acquisition and uploading at the new frequency. The frequency adjustment status of all nodes will be reported in the system scheduling log for subsequent verification and strategy optimization.

[0059] The feature extraction module is used for in-depth analysis of disaster cloud regions in high-frequency image streams, extracting multi-dimensional features such as wind field, cloud morphology, and brightness-temperature gradient, ultimately generating a set of feature vectors for input to the intelligent early warning model. Specifically, it includes: The high-frequency image stream obtained after scheduling is re-inputted into the trained deep image discrimination model M to extract the disaster cloud region in the corresponding image frame, specifically including: After the scheduling module adjusts the sampling frequency, the high-frequency image stream uploaded by the image acquisition nodes is re-inputted into the trained deep image discrimination model M. This model is built on the U-Net architecture and has semantic segmentation capabilities. The model performs pixel-by-pixel classification on the input image and outputs a disaster probability map with the same size as the original image. The probability map is binarized according to a set threshold (e.g., 0.6), and regions with probability values ​​greater than the threshold are extracted and considered as disaster cloud regions, forming a cloud mask. The mask is used for region restriction in subsequent feature extraction operations.

[0060] Combining image acquisition time and location information, the cloud motion vector field is estimated using the inter-frame optical flow method, and a velocity and direction feature matrix reflecting wind field changes is constructed, specifically including: For image sequences acquired at adjacent times t and t+1 from the same image acquisition node, inter-frame motion estimation is performed based on an optical flow algorithm. The Farneback dense optical flow method is used to estimate the displacement vector (Δx, Δy) of each pixel from the continuous images, and the calculation result is the cloud motion vector.

[0061] By combining the image time interval Δt and the pixel spacing, the actual wind speed and direction can be further calculated to form a two-dimensional wind field feature matrix. Each element in the matrix includes: horizontal wind speed (unit: m / s); vertical wind speed (if it can be estimated); and wind direction angle (unit: degree). This wind field matrix provides a basis for understanding the evolution trend of disaster cloud clusters and local wind field disturbances.

[0062] Using image segmentation results, cloud boundaries, thickness, and convolution density are extracted, specifically including: Based on the cloud mask image output by the model, morphological processing algorithms are used to extract the cloud boundaries, and the following morphological parameters are further calculated: Boundary length: The contour length is obtained through an edge detection algorithm, and the unit is pixels; Thickness index: The vertical development of cloud clusters is simulated using image grayscale values ​​or brightness temperature gradients, and the thickness is estimated through projection analysis; Convolution density: The image within the masked region is convolved with a Gaussian kernel in two dimensions, and the sum of its response intensities is calculated to reflect the density and structural compactness of the cloud.

[0063] Gradient calculation of brightness temperature distribution in disaster areas of infrared channel images is performed to form a brightness temperature gradient feature map, which is then fused with wind field and morphological features to generate a feature vector set, specifically including: In infrared images, the brightness value of each pixel corresponds linearly to the cloud top brightness temperature. The corresponding disaster mask region is extracted from the infrared image, and the following operations are performed: Gradients are calculated for pixel values ​​(using the Sobel operator) to obtain the brightness temperature change rates in the horizontal and vertical directions. A brightness temperature gradient amplitude map is calculated to represent the intensity of spatial brightness temperature changes, reflecting the activity of cloud development. Statistical features such as the maximum gradient value, average gradient value, and gradient direction variation range are extracted. Finally, the following three types of features are fused: wind field features (e.g., wind speed, wind direction); cloud morphology features (e.g., boundary length, thickness, convolution density); and brightness temperature gradient features (e.g., maximum value, average value). These are integrated into a standardized set of feature vectors.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A meteorological data acquisition system for meteorological disaster early warning, characterized in that: include: Acquisition module: Acquires historical meteorological disaster image data sequences and corresponding time labels for the target area, and constructs a training sample set T; Model training module: Based on the sample set T, a deep image discrimination model M for small-scale meteorological disaster feature recognition is constructed and trained to identify the morphology of potential disaster cloud clusters from real-time image data; Data acquisition module: acquires multiple image acquisition nodes deployed within the target area. Uploaded real-time image data stream The image data streams are then input into model M to obtain preliminary discrimination results. where n is the total number of images; Image analysis module: based on various discrimination results Based on the disaster similarity values, a regional image risk heatmap R is constructed, and areas with heatmap values ​​greater than a set threshold are selected. G is a set of high-risk image nodes. Frequency scheduling module: Schedules the acquisition frequencies of image acquisition nodes in set G, changing their image acquisition frequencies from the initial frequency. Increase to weighted sampling frequency ; Feature extraction module: The updated high-frequency image stream is input into model M again, and a set of feature vectors containing wind field, cloud morphology and brightness temperature gradient is generated by combining the corresponding image metadata.

2. The meteorological data acquisition system for meteorological disaster early warning according to claim 1, characterized in that: The acquisition module includes: The system calls upon a multi-source historical image database to extract satellite cloud images and ground-based real-time images of severe weather within a target area over a historical period, according to a preset time window. Preprocess the extracted image; The cloud types, morphological boundaries, and brightness temperature features in the image are labeled to form a structured image feature set; The structured image features are matched with the corresponding historical disaster occurrence times to generate a training sample set T with time labels.

3. The meteorological data acquisition system for meteorological disaster early warning according to claim 1, characterized in that: The model training module includes: The initial network structure of the deep image discrimination model M is constructed based on the training sample set T. The network structure includes a multi-scale feature extraction sub-network consisting of convolutional layers, pooling layers and skip connections, which is used to extract local texture and overall morphological features of clouds. The model M is trained in a supervised manner using structured image features from the sample set T as input data and corresponding disaster time labels as supervision signals. The trained model M is validated, and its hyperparameters are optimized through cross-validation.

4. A meteorological data acquisition system for meteorological disaster early warning according to claim 3, characterized in that: The data acquisition module includes: Multiple image acquisition nodes are deployed at a preset density within the target area. ; Each image acquisition node uploads a real-time image data stream at a fixed time interval t. To edge computing units; The edge computing unit performs preliminary preprocessing on each received image data stream, including compression decoding, image enhancement, and abnormal frame removal. The preprocessed image data is synchronously input into the trained model M, and preliminary discrimination results are output. The results include a probability map and spatial location information of the disaster cloud region in each frame of the image.

5. A meteorological data acquisition system for meteorological disaster early warning according to claim 4, characterized in that: The image analysis module includes: For each preliminary judgment result The probability map of disaster cloud clusters output by the system is used to perform regional connectivity analysis, extract suspected disaster areas and calculate their morphological characteristic parameters; Based on the model's output probability mean, area ratio, and edge complexity features, and combined with set weights, a disaster similarity value S is calculated to reflect the degree of similarity between cloud features in the image and historical disaster images. The similarity values ​​of each image acquisition node are bound to the GPS coordinates of the node and mapped to a unified geographic grid of the target area to form a spatial risk distribution matrix; The similarity matrix is ​​spatially expanded using a bilinear interpolation algorithm to construct a continuous regional image risk heat map R, which uses heat color to reflect the disaster potential of different regions.

6. A meteorological data acquisition system for meteorological disaster early warning according to claim 1, characterized in that: The frequency scheduling module includes: Obtain the disaster similarity value S of each node in the high-risk image node set G, and classify them into three risk levels: Level 1, Level 2, and Level 3 based on the similarity. Set the corresponding weighted sampling frequency based on the risk level of the node. The According to the proportional relationship, based on the initial sampling frequency The frequency of sampling at Level 1 risk nodes has been increased to [number missing]. Level 2 is Level 3 ; A frequency adjustment command is issued to the nodes in set G to update their acquisition and scheduling parameters, so that they operate at the new frequency. Collect image data and upload it.

7. A meteorological data acquisition system for meteorological disaster early warning according to claim 6, characterized in that: The feature extraction module includes: The high-frequency image stream obtained after scheduling is re-inputted into the trained deep image discrimination model M to extract the disaster cloud region in the corresponding image frame; By combining image acquisition time and location information, the cloud motion vector field is estimated using the inter-frame optical flow method, and a velocity and direction feature matrix reflecting wind field changes is constructed. Using image segmentation results, cloud boundaries, thickness, and convolution density are extracted; Gradient calculations are performed on the brightness temperature distribution of disaster areas in infrared channel images to form a brightness temperature gradient feature map, which is then fused with wind field and morphological features to generate a set of feature vectors.

Images