A ground time sequence observation image contrail cloud recognition method and system

Abstract

The application discloses a ground time sequence observation image contrail cloud recognition method and system, belongs to the technical field of image processing, obtains the video frame image of the current contrail cloud, and the corresponding shooting time and coordinate information; according to the video frame image and the shooting time, the time label of the video frame image is analyzed, the time feature vector is constructed, and the time feature map of the video frame image is extracted; according to the video frame image and the coordinate information, the channel attention and the spatial attention of the video frame image are obtained, and the spatial feature map of the video frame image is extracted; according to the time feature map and the spatial feature map, the time feature and the spatial feature tensor are spliced in the channel dimension, and a joint feature map is generated; the joint feature map is classified and recognized, and the category of the contrail cloud is obtained. The method can realize accurate identification and classification of the contrail cloud from the generation to the diffusion stage.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for identifying wake clouds in ground time-series observation images. Background Technology

[0002] Contrails are fine cloud bands of ice crystals formed when water vapor emitted by high-altitude aircraft rapidly condenses in low-temperature environments. They typically exhibit a distinct linear morphology and are a common visible feature of aircraft operating at high altitudes. Contrails not only affect the stealth capability of aircraft trajectories but can also interfere with meteorological monitoring results.

[0003] Currently, there are two main methods for monitoring contrails: one is remote sensing identification based on high-resolution satellite imagery, and the other is fixed-point observation using ground-based visible light cameras. Whether using remote sensing or ground-based images, the core processing method relies on image recognition algorithms in computer vision to automatically identify and classify contrails. The mainstream approach primarily employs deep learning techniques, including target detection and segmentation frameworks such as ResNet, YOLO, and UNet. These methods have demonstrated good performance in traditional cloud recognition, smoke detection, and trajectory extraction scenarios, and are therefore widely used in the automatic identification of contrails. In practical applications, researchers typically construct image datasets containing contrails and train and test these models to achieve the detection and segmentation of contrail targets.

[0004] While existing technologies have achieved automatic contrail identification to some extent, the following technical shortcomings and limitations remain, including: Lack of ability to model the temporal characteristics of contrails. The morphology of contrails changes significantly over time: initially, they are elongated with clear edges; in the middle stage, they gradually spread and widen; in the late stage, they may even merge with natural clouds, becoming indistinguishable. Most current mainstream recognition algorithms are trained based on the spatial features of static images, lacking the ability to model temporal labels and evolutionary states, leading to a significant decrease in accuracy when identifying early and late-stage contrails. Insufficient timeliness and realism of existing datasets. Most publicly available contrail image datasets are derived from aerial remote sensing or simulated images. Although the images are high-resolution, they lack real-time labels and the environmental conditions are relatively uniform, failing to reflect the complex lighting, background cloud interference, and other issues under real ground observation conditions. This results in poor generalization ability and insufficient recognition stability when the model is deployed in the field. Lighting and weather interference significantly affect recognition performance. Ground-based imagery is significantly affected by factors such as sunlight angle, cloud brightness, and background skylight. Traditional image recognition models, when faced with low-contrast or high-dynamic lighting conditions, easily misidentify contrails as natural clouds, aircraft trails, or other linear image elements, reducing recognition accuracy. Furthermore, there is a lack of end-to-end acquisition-annotation-recognition systems. Current research focuses primarily on improving model algorithms, but rarely addresses the optimization of the complete data acquisition and annotation process. The absence of comprehensive solutions for time-series evolution and low-cost deployment limits the practical application and real-time performance of this technology.

[0005] In summary, since the shape of contrail clouds changes significantly over time, existing technologies cannot reflect the complex lighting and background cloud interference under real ground observation conditions when acquiring images. This results in poor generalization ability of the model when deployed in the field, thus affecting the recognition accuracy. Summary of the Invention

[0006] To address the problems existing in the aforementioned fields, this invention proposes a method and system for identifying contrail clouds from ground-based time-series observation images. By acquiring image information with time labels, in the feature extraction stage, a time feature map is extracted by constructing a time feature vector to enhance the temporal perception capability of the recognition model. Global average pooling and max pooling operations are used to generate weight maps of channel attention and spatial attention features. The intermediate features of the weight maps are enhanced in both the channel and spatial dimensions to extract the spatial feature map, thereby increasing the model's attention to the main region of the contrail cloud. The time and spatial feature maps are then fused to achieve accurate identification and classification of contrail clouds from their generation to diffusion stages.

[0007] To address the aforementioned technical problems, this invention discloses a method for identifying wake clouds in ground-based time-series observation images, comprising the following steps:

[0008] Acquire the current video frame image of the contrail cloud, along with its corresponding capture time and coordinate information;

[0009] Based on the video frame images and the shooting time, a temporal feature map of the wake cloud video frame images is extracted by constructing a temporal feature vector;

[0010] Based on the video frame image and coordinate information, the channel attention and spatial attention of the video frame image are obtained. The weight map of the channel attention and spatial attention features is generated by global average pooling and max pooling operations. The intermediate features of the weight map are enhanced in the channel dimension and spatial dimension respectively, and the spatial feature map of the video frame image is extracted.

[0011] Based on the temporal and spatial feature maps, the temporal and spatial feature tensors are concatenated along the channel dimension to generate a joint feature map.

[0012] The joint feature map is classified and identified to obtain the category of the wake cloud.

[0013] Preferably, the category of the contrail cloud is identified by an improved YOLOv8 network model, which specifically includes: using the YOLOv8-tiny model as the detection skeleton structure, including Backbone, Neck, Head and time-series label modules;

[0014] An attention mechanism CBAM is inserted into the c2f of the preceding feature extraction layer and the Conv of the following feature extraction layer in the backbone. This mechanism applies channel attention and spatial attention to the intermediate feature maps and performs spatial feature extraction by jointly modeling channel attention and spatial attention. This guides the automatic focus on the wake cloud region during the feature extraction stage.

[0015] The time tag module is used to define image file names based on video frame images and shooting time, parse time tags from image file names, convert time category information into discrete vectors through one-hot encoding, and then map it into a 64-dimensional time feature vector through a fully connected layer;

[0016] In Neck, the 64-dimensional temporal feature vector and spatial features are fused using the concat method to generate a joint feature map;

[0017] The joint feature map is classified and identified using the Head.

[0018] Preferably, the improved YOLOv8 network model introduces a temporal consistency loss function during training, adding a similarity constraint term between the prediction results of the same target in different time frames to the original YOLO loss in the Head of the YOLOv8-tiny model.

[0019] The similarity constraint is:

[0020] ;

[0021] in, This represents the predicted probability vector of the same target at different times; The time adjacency weighting function is represented by the function. This is an adjustment factor.

[0022] Preferably, it further includes:

[0023] Based on the capture time and state of each acquired contrail image, a one-to-one correspondence is established to construct a contrail evolution state labeling system.

[0024] Based on the wake cloud evolution state labeling system and the coordinate information corresponding to each image, target classification is performed on the video frame images to obtain the label image for each category.

[0025] The construction of the wake cloud evolution state labeling system specifically includes:

[0026] By analyzing the evolution of contrails, and combining the shooting time, the target was divided into three states: newly generated (0 min), generated about 30 minutes ago (30 min), and generated about 1 hour ago (60 min). Non-contrails were added as a control group.

[0027] Preferably, the classification and identification of the joint feature map further includes outputting the target location, evolutionary state category, and confidence information of the wake cloud.

[0028] Preferably, acquiring the video frame image of the current contrail specifically includes:

[0029] Select multiple points in the shooting area, set a fixed elevation angle for the camera, and capture video frame images of the contrails in the current state by taking timed shots;

[0030] By performing edge computing on video frame images of contrails, linear structural features in the images are detected, invalid images without significant high-altitude trajectories are excluded, and images that pass the initial quality screening are obtained and labeled.

[0031] Preferably, the coordinate information is GPS positioning coordinates.

[0032] Preferably, it also includes a ground-based temporal observation image wake cloud recognition system, comprising:

[0033] The image acquisition module is used to acquire video frame images of the current contrail cloud, as well as its corresponding shooting time and coordinate information;

[0034] The feature extraction module is used to extract the temporal feature map of the wake cloud video frame image by constructing a temporal feature vector based on the video frame image and the shooting time; and to obtain the channel attention and spatial attention of the video frame image based on the video frame image and coordinate information, generate a weight map of the channel attention and spatial attention features through global average pooling and max pooling operations, and enhance the intermediate features of the weight map in the channel dimension and spatial dimension respectively to extract the spatial feature map of the video frame image.

[0035] The feature fusion module is used to concatenate the temporal and spatial feature tensors along the channel dimension based on the temporal and spatial feature maps to generate a joint feature map.

[0036] The identification module is used to classify and identify the joint feature map to obtain the category of the wake cloud.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention proposes a ground-based temporal observation image contrail cloud recognition method. This method acquires time-labeled contrail cloud images and, during the feature extraction stage, constructs a temporal feature vector to extract a temporal feature map, thereby enhancing the temporal perception capability of the recognition. Based on video frame images and coordinate information, the method obtains the channel attention and spatial attention of the video frame images. Global average pooling and max pooling operations are used to generate weight maps of the channel attention and spatial attention features. The intermediate features of the weight maps are enhanced in both the channel and spatial dimensions to extract the spatial feature map of the video frame images. This improves the model's focus on the main region of the contrail cloud, suppresses background interference, and enhances the stability of the recognition boundary, maintaining excellent performance even under complex conditions such as strong light and cluttered clouds. The temporal and spatial feature tensors are concatenated along the channel dimension for feature map fusion, generating a joint feature map. This joint feature map not only possesses spatial perception capability but also identifies state differences of the contrail cloud in the temporal evolution dimension. Classification and recognition of the joint feature map enable accurate identification and classification of contrail clouds from their generation to diffusion stages. Attached Figure Description

[0039] Figure 1 This is a flowchart of the ground time-series observation image wake cloud recognition method proposed in this invention;

[0040] Figure 2 An improved YOLOv8 network model architecture was constructed for this invention;

[0041] Figure 3 This is a schematic diagram of the overall structure of the identification system of the present invention;

[0042] Figure 4 The identification results provided in the embodiments of the present invention. Detailed Implementation

[0043] The following will refer to the appendices in the embodiments of the present invention. Figures 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0044] like Figure 1 As shown, this invention proposes a method for identifying contrails in ground-based time-series observation images to improve the accuracy and stability of contrails identification at different stages. The method specifically includes the following steps:

[0045] Acquire the current video frame image of the contrail cloud, along with its corresponding capture time and coordinate information (GPS positioning coordinates);

[0046] Based on the video frame images and the shooting time, the time tags of the video frame images are parsed, and the time feature maps of the video frame images are extracted by constructing time feature vectors.

[0047] Based on video frame images and coordinate information, channel attention and spatial attention features are extracted. A weight map is generated through global average pooling and max pooling operations. The intermediate features of the weight map are enhanced in the channel dimension and spatial dimension respectively, and the spatial feature map of the video frame image is extracted.

[0048] Based on the temporal and spatial feature maps, the temporal and spatial feature tensors are concatenated along the channel dimension to generate a joint feature map.

[0049] The joint feature map is classified and identified to obtain the category of the wake cloud.

[0050] Specifically, classifying and identifying the joint feature map also includes outputting the target location, evolutionary state category, and confidence information of the wake cloud.

[0051] The category of contrail clouds is identified using an improved YOLOv8 network model, which specifically includes: using the YOLOv8-tiny model as the detection skeleton structure, including Backbone, Neck, Head and time-series label modules;

[0052] An attention mechanism CBAM is inserted into the c2f of the preceding feature extraction layer and the Conv of the following feature extraction layer in the backbone. This mechanism applies channel attention and spatial attention to the intermediate feature maps, extracts channel attention and spatial attention features from the wake cloud video frame images, and performs spatial feature extraction by jointly modeling channel attention and spatial attention. This guides the automatic focus on the wake cloud region during the feature extraction stage, improving the recognition ability in complex backgrounds.

[0053] By combining a lightweight neural network model with an attention mechanism, robustness in recognizing complex environments is enhanced. The CBAM module improves the model's focus on the main region of the contrail cloud, suppresses background interference, and improves the stability of the recognition boundary, maintaining excellent performance even under complex conditions such as strong light and cluttered clouds.

[0054] The temporal tagging module (introducing a temporal tagging mechanism) defines image filenames based on video frame images and capture timestamps. It parses time tags from image filenames and converts time category information into discrete vectors using one-hot encoding. These vectors are then mapped to 64-dimensional temporal feature vectors through a fully connected layer. In the Neck, the 64-dimensional temporal feature vectors are fused with spatial features using a concat method to generate a joint feature map, enabling contextual modeling of the temporal evolution state. The Head classifies and identifies the joint feature map.

[0055] The improved YOLOv8 network model enhances its temporal awareness capabilities by introducing a temporal labeling mechanism during training.

[0056] That is, a similarity constraint term between the prediction results of the same target at different time frames is added to the original YOLO loss in the Head of the YOLOv8-tiny model, which is used to classify and identify the joint feature map.

[0057] Specifically, the similarity constraint terms are as follows:

[0058] ;

[0059] in, This represents the predicted probability vector of the same target at different times; The time adjacency weighting function is represented by the function. This is an adjustment factor.

[0060] By establishing a one-to-one correspondence between image capture time and contrail state, and using time state as one of the model input features, the classification accuracy of contrails at different generation stages is significantly improved, solving the problem that traditional models cannot distinguish between "nascent contrails" and "diffuse contrails".

[0061] It also includes establishing a one-to-one correspondence based on the capture time and state of each acquired contrail image, and constructing a contrail evolution state labeling system.

[0062] Based on the wake cloud evolution state labeling system and the coordinate information corresponding to each image, target classification is performed on the video frame images to obtain the label image for each category.

[0063] Constructing a wake cloud evolution state labeling system, specifically including:

[0064] By analyzing the evolution of contrails, and combining the shooting time, the target was divided into three states: newly generated (0 min), generated about 30 minutes ago (30 min), and generated about 1 hour ago (60 min). Non-contrails were added as a control group.

[0065] Obtain video frame images of the wake cloud in the current state, specifically including:

[0066] Select multiple points in the shooting area, set a fixed elevation angle for the camera, and capture video frame images of the contrails in the current state by taking timed shots;

[0067] By performing edge computing on video frame images of contrails, linear structural features in the images are detected, invalid images without significant high-altitude trajectories are excluded, and images that pass the initial quality screening are obtained and labeled.

[0068] This invention also proposes a ground-based temporal observation image wake cloud recognition system, comprising:

[0069] The image acquisition module (image acquisition module) is used to acquire the video frame image of the current trail cloud, as well as its corresponding shooting time and coordinate information.

[0070] Select 5 to 10 typical locations (high-rise buildings, open spaces) within the shooting area and deploy Raspberry Pi 4B devices, each equipped with an HQ Camera Module (12 megapixels) and a GPS coordinate synchronization module. Figure 3 As shown, this system is used for fixed-location aerial image capture. The camera faces north with a fixed elevation angle to avoid accidentally capturing landscapes. One frame is captured every 15 seconds, with a collection period of at least 30 days, operating 24 / 7. The average daily image volume per location is approximately 5760 images; a total of approximately 1,728,000 images (over 30 days) are collected from 10 locations. Edge computing is used to detect linear structural features in the video frames of contrails, eliminating invalid images without significant high-altitude trajectories. Images that pass the initial quality screening are then labeled. After removing images without contrails, approximately 10,000 to 20,000 contrail-related images are retained as candidate samples.

[0071] Images are stored in JPEG format, named with timestamps, and their shooting coordinates are recorded synchronously (for annotation reference and retrospective analysis).

[0072] The feature extraction module is used to parse the time tags of the video frame images of the contrails based on the video frame images and the shooting time, and extract the time feature map of the video frame images of the contrails by constructing a time feature vector. Based on the video frame images and coordinate information of the contrails, the module obtains the channel attention and spatial attention of the video frame images, generates weight maps of the channel attention and spatial attention features through global average pooling and max pooling operations, and enhances the intermediate features of the weight maps in the channel dimension and spatial dimension respectively to extract the spatial feature map of the video frame images.

[0073] Specifically, the temporal and spatial feature maps of the wake cloud are extracted using an improved YOLOv8 network model.

[0074] like Figure 2 As shown, the improved YOLOv8 network model uses the YOLOv8-tiny model as the detection backbone structure. The YOLOv8-tiny model includes: Backbone (CSPDarknet-Tiny, with residual connections and C2f structure), Neck (PAN-FPN structure, used for multi-scale feature fusion) and Detection Head (three output branches, responsible for small, medium and large scale object detection respectively).

[0075] The improved YOLOv8 network model inserts an attention mechanism (CBAM module) into the c2f of the preceding feature extraction layer and the Conv of the following feature extraction layer in the backbone. This mechanism acquires channel attention and spatial attention from video frame images of contrails, combines channel attention and spatial attention, applies channel attention and spatial attention to intermediate feature maps, and performs spatial feature extraction through joint modeling of channel attention and spatial attention. It generates weight maps of channel attention and spatial attention features through global average pooling and max pooling operations, and enhances the intermediate features of the weight maps in both the channel and spatial dimensions to extract the spatial feature map of the image. This guides the model to focus on the contrails region, thereby improving its recognition ability in complex backgrounds.

[0076] A time-series tagging module is introduced, which is a time-series tagging mechanism. The image file name is defined based on the video frame image and the shooting time. The time segment tag (0, 30min, 60min) is parsed from the image file name. One-hot encoding is constructed as a time vector, which is linearly mapped into a 64-dimensional time feature vector. Then, it is fused with the image spatial features in the Neck layer through concat, so as to realize the context modeling of the time-series evolution state.

[0077] Introducing a time consistency loss function: Adding a time consistency constraint term to the original YOLO loss in the Head function. .

[0078] Also includes:

[0079] Based on the capture time and state of each acquired contrail image, a one-to-one correspondence is established to construct a contrail evolution state labeling system.

[0080] Based on the wake cloud evolution state labeling system and the coordinate information corresponding to each image, target classification is performed on the video frame images to obtain the label image for each category.

[0081] Specifically, candidate samples are uploaded to the CVAT platform, where annotators perform detailed manual annotations by combining time and image features.

[0082] A contrail cloud evolution state labeling system is constructed, categorized as follows:

[0083] Category 0: Contrails - newly formed (thin, clear lines, consistent with flight paths);

[0084] Category 1: Wheal cloud - generated 30 minutes ago (lines have spread out, outline is blurred);

[0085] Category 2: Contrails - formed within 1 hour (spreads significantly, indistinguishable from natural clouds);

[0086] Category 3: Non-contrail clouds (natural clouds, blue sky background, other cirrus clouds, etc.).

[0087] Each image supports multi-target annotation, using the YOLO standard annotation format: class x_center y_centerwidth height. All coordinate information is normalized to the image resolution (default 1920×1080). The total number of annotated samples is controlled to be no less than 10,000 images, ensuring that there are no less than 2,000 images for each class label.

[0088] The feature fusion module is used to concatenate the temporal and spatial feature tensors along the channel dimension based on the temporal and spatial feature maps to generate a joint feature map.

[0089] The recognition module is used to classify and identify the joint feature map to obtain the category of the wake cloud.

[0090] The acquired video frame images are divided into training, validation, and test sets in a 7:2:1 ratio; the number of training images is controlled at around 7,000; the validation set has 2,000 images; and the test set has 1,000 images, ensuring a balance of labels for each class.

[0091] Training parameter settings include:

[0092] Framework: PyTorch;

[0093] Optimizer: SGD + Cosine Annealing;

[0094] Learning rate: Initially 0.001;

[0095] Batch Size: 16;

[0096] Epochs: 150;

[0097] Transfer learning: Fine-tuning of YOLOv8 pre-trained weights;

[0098] Enhancement strategies: Use image enhancement methods such as Mosaic, Mixup, random brightness, and blurring.

[0099] Evaluation indicators:

[0100] Using mAP 0.5 (mean Average Precision).

[0101] The system integration module of this invention comprises the following sub-modules:

[0102] Image acquisition terminal: The image acquisition service is provided by a Raspberry Pi device, which runs a Python program to acquire, screen and upload images on a regular basis, and supports local caching and network retransmission mechanisms.

[0103] Recognition and inference module: The backend server deploys the improved YOLOv8-tiny model, using ONNX+TensorRT to accelerate deployment, supporting real-time recognition at 30 FPS.

[0104] Control and visualization interface: The control interface is built with Flask and Vue, which supports image preview, recognition result display, wake cloud status statistics and parameter adjustment.

[0105] Data feedback and management: The recognition results are written to the local SQLite database in real time and uploaded to the remote platform synchronously, supporting functions such as trace cloud quantity statistics, generation time distribution analysis, and dynamic evolution trend visualization.

[0106] Example

[0107] To better demonstrate the contrail cloud recognition method proposed in this invention, this embodiment employs a multi-point ground deployment approach, setting up image acquisition terminals at typical airspace observation locations. Each terminal consists of a Raspberry Pi 4B embedded device, equipped with an HQ Camera Module image acquisition module and a GPS positioning module. The image acquisition terminal is pre-installed with image control software, capable of automatically performing airspace shooting tasks in all weather conditions at a preset frequency, ensuring the acquisition of one high-definition image frame every 15 seconds. During the acquisition process, the camera is fixed facing north and maintains a specific elevation angle to avoid interference from ground objects. The system simultaneously records the image generation time and geographic coordinate information. All images are stored in JPEG format and named using timestamps to ensure the integrity and traceability of the image sequence.

[0108] After image acquisition, the device's local program automatically performs initial image quality screening, eliminating nighttime images, severely overexposed images, or invalid images, thereby reducing the burden on the backend recognition. The edge computing module determines whether there are obvious linear structures or high-altitude trajectory features in the images. Image data that passes the initial screening is uploaded to the server and stored in the database via a pre-defined network protocol. The system has a built-in breakpoint resume and caching mechanism, which can automatically resume uploading in the event of a network interruption, ensuring data integrity.

[0109] Uploaded image data is uniformly imported into the CVAT annotation platform, where annotators perform manual annotation based on image content and capture time information. By analyzing the morphological evolution of contrails after generation, annotators categorize image targets into three contrail evolution states: immediately after generation, approximately 30 minutes after generation, and approximately 1 hour after generation. A non-contrail category is also added as a control class. The bounding boxes and category information of targets in each image are uniformly annotated according to the YOLO format, and the coordinates are normalized to adapt to the requirements of neural network training. The number of annotated samples is controlled to be over 10,000 to ensure a balanced distribution of labels for each class, facilitating the construction of a high-quality, time-series-defined training dataset.

[0110] The wake cloud recognition system provided in this embodiment consists of an image acquisition module, an image annotation module, an algorithm improvement module, a model training and testing module, and a system integration module. The system integration module includes a recognition inference module, a control and visualization interface, and a data feedback and management module as well as a dynamic evolution trend analysis module. The image acquisition module runs a Python program on a Raspberry Pi device to implement timed acquisition. It reads NMEA data from the GPS module via a serial port to achieve coordinate synchronization, binds timestamps and latitude / longitude to each image frame and stores them in a local cache queue. An LRU strategy is used to manage a cache space with a capacity of 100 frames. In the event of a network interruption, an exponential backoff retransmission mechanism is activated to prioritize sending the latest data. The recognition and inference module deploys an improved YOLOv8 network model, with YOLOv8-tiny as the backbone network. A CBAM attention mechanism is added to the end of the backbone network to enhance the ability to extract trail cloud features. A temporal feature fusion module is introduced to capture dynamic changes by processing three consecutive frames of feature maps through LSTM units. A temporal consistency loss function is added to the loss function to address abrupt changes in recognition results between adjacent frames. After the improved YOLOv8 network model is converted from PyTorch to ONNX, it is optimized by FP16 quantization and layer fusion through the TensorRT engine to achieve real-time inference of 30 FPS with a latency of ≤33ms.

[0111] The control and visualization interface is built using Flask and Vue, defining a four-level classification standard for contrail cloud status: Level 0 indicates no clouds, Level 1 corresponds to a single area less than 50% of the image region, Level 2 is a single area exceeding 50%, and Level 3 represents the simultaneous presence of multiple contrail clouds. The interface integrates a dynamic evolution trend analysis module, using exponential smoothing based on SQLite time-series data to predict changes in the number of contrail clouds. It also combines an OpenGL engine to generate contrail diffusion heatmaps and supports 1-hour sliding window analysis. The data feedback and management system adopts a dual SQLite database architecture. The terminal database caches original images and GPS metadata and automatically cleans up old data older than 24 hours, while the platform database permanently stores recognition results and statistical indicators. After network recovery, incremental synchronization of cached data during network outages is achieved via HTTPS protocol.

[0112] The model training process employs a 7:2:1 dataset partitioning ratio, utilizing a Mosaic+MixUp augmentation strategy combined with random HSV jitter and ±15-degree affine transformation to enhance data diversity. Transfer learning and parameter optimization techniques are incorporated into the training process. The evaluation system includes core metrics such as mAP@0.5 / 0.95, recall, precision, and inference speed to ensure the model meets performance requirements in real-world deployments. All modules collaborate through standardized interfaces, forming a complete closed-loop system from data acquisition and real-time analysis to visualization.

[0113] This embodiment uses the YOLOv8-tiny model as the skeleton recognition network model and makes targeted improvements to the network structure.

[0114] First, an attention mechanism (CBAM module) is introduced into the backbone network. By jointly modeling channel attention and spatial attention, the neural network is guided to automatically focus on the contrail region during the feature extraction stage, reducing the influence of background interference factors such as natural clouds and sunlight scattering. After the feature map is output, this module generates a weight map through global average pooling and max pooling operations, and enhances the intermediate features of the weight map in both the channel and spatial dimensions, thereby improving the model's sensitivity to the contrail region and the accuracy of boundary judgment.

[0115] Secondly, a time-series labeling module is introduced in the feature extraction stage. The system parses the time tags according to the image file name through the time-series labeling mechanism, and converts the time category information into a discrete vector through one-hot encoding, and then maps it into a 64-dimensional time feature vector through a fully connected layer.

[0116] In the feature fusion stage, the Neck layer is fused with the image spatial features through concat. The 64-dimensional temporal feature vector and the image spatial feature tensor are concatenated along the channel dimension to generate a joint feature map. This enables the model to not only have spatial perception capabilities, but also to identify the state differences of the wake cloud in the temporal evolution dimension.

[0117] To enhance the model's recognition consistency over time, a temporal consistency loss function is introduced. This loss function adds a similarity constraint term between predictions of the same target in different time frames to the original target detection loss, thereby encouraging the model to maintain semantic coherence and stage rationality in its predictions during the wake cloud evolution process.

[0118] During model training, the neural network was constructed using the PyTorch framework, and pre-trained YOLOv8 parameters were loaded using transfer learning. The dataset was divided into training, validation, and test sets to ensure model generalization ability. During training, the system employed various image enhancement techniques, including Mosaic and Mixup enhancements, brightness perturbations, and blurring, to improve the model's robustness under complex lighting and blurred cloud conditions. The loss function consisted of bounding box error, class cross-entropy, and temporal consistency loss. Model optimization employed a cosine annealing strategy to adjust the learning rate and improve convergence efficiency.

[0119] After training, the model is exported in ONNX format and deployed on a backend server. The TensorRT inference engine is used to provide high-concurrency, high-speed real-time recognition inference services. The server provides a RESTful interface for frontend calls, returning wake cloud recognition results immediately after image upload, including target location, evolutionary state category, and confidence information. Recognition results are written to a local database in real time and synchronously transmitted to a remote central platform for subsequent statistical analysis and trend modeling.

[0120] The system integration uses Flask as the backend service framework and Vue to implement the data visualization interface. Users can view image acquisition status, recognition results, and contrail quantity statistics in real time on the interface. The system supports dynamic parameter configuration and model updates, and also provides a visualization module for the temporal distribution and state evolution trend of contrail generation, facilitating long-term monitoring and meteorological correlation analysis by researchers.

[0121] The overall system boasts advantages such as low-cost deployment, high temporal sensitivity, and strong environmental adaptability, making it suitable for contrail cloud recognition tasks under complex weather conditions. It can also be widely applied to various scenarios including aviation stealth assessment, airspace monitoring and early warning, and meteorological data correction. This invention features significant technical innovations in hardware deployment methods, image acquisition strategies, temporal label embedding mechanisms, and temporal consistency loss design, effectively overcoming the technical bottlenecks of traditional methods in contrail cloud evolution recognition, and possessing good practicality and scalability.

[0122] like Figure 4 The figure shows typical output results of the improved YOLOv8 network model after identifying video frame image samples of contrails collected in this implementation. The figure clearly presents the detection boxes, corresponding class labels, and confidence scores of different targets in the image by the improved YOLOv8 network model. Specifically:

[0123] Figure 4 The main targets identified were two types of contrail clouds, Contrail=1 hour and Contrail=Fresh, where:

[0124] Contrail=1hour was identified multiple times with confidence levels of 0.90, 0.80, 0.52, and 0.31, respectively, indicating that such targets are significantly present in the scene and that the improved YOLOV8 network model has high reliability in identifying them.

[0125] Contrail=Fresh appeared most frequently, with confidence levels of 0.80, 0.57, 0.82, 0.62, 0.69, and 0.45, respectively. It is the most important recognition category in the image, but its confidence level fluctuates relatively widely, reflecting the differences in target state or imaging conditions.

[0126] Figure 4 The identification results also included other salient targets. "Sun" was identified with a high confidence score of 0.90, indicating the presence of the sun as a salient feature in the image. "Parasite" was identified with the highest confidence score of 0.93, demonstrating the strong ability of the improved YOLOv8 network model to identify such features and the obvious presence of this target in the image. "Unknown" was a target detected by the improved YOLOv8 network model with a confidence score of 0.62, but its specific category could not be determined.

[0127] In conclusion, Figure 4 The contrail recognition results shown intuitively demonstrate that the improved YOLOv8 network model can effectively identify various key targets in video frames of contrails, particularly performing well for categories such as Contrail=1hour, Contrail=Fresh, Sun, and Parasite. However, the results also include detection boxes in the "unknown" category, reflecting the challenges in real-world recognition scenarios and the difference in confidence levels between the improved YOLOv8 network model and these categories. Figure 4 The density of the bounding boxes also suggests that the contrails and related targets are relatively concentrated in the image.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0129] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for identifying wake clouds in ground-based time-series observation images, characterized in that, Includes the following steps: Acquire the current video frame image of the contrail cloud, along with its corresponding capture time and coordinate information; Based on the video frame images and the shooting time, a temporal feature map of the wake cloud video frame images is extracted by constructing a temporal feature vector; Based on the video frame image and coordinate information, the channel attention and spatial attention of the video frame image are obtained. The weight map of the channel attention and spatial attention features is generated by global average pooling and max pooling operations. The intermediate features of the weight map are enhanced in the channel dimension and spatial dimension respectively, and the spatial feature map of the video frame image is extracted. Based on the temporal and spatial feature maps, the temporal and spatial feature tensors are concatenated along the channel dimension to generate a joint feature map. The joint feature map is classified and identified to obtain the category of the wake cloud; The category of the contrail cloud is identified using an improved YOLOv8 network model, which specifically includes: The YOLOv8-tiny model was used as the detection skeleton structure, including the Backbone, Neck, Head and time label modules; An attention mechanism CBAM is inserted into the c2f of the preceding feature extraction layer and the Conv of the following feature extraction layer in the backbone. This mechanism applies channel attention and spatial attention to the intermediate feature maps and performs spatial feature extraction by jointly modeling channel attention and spatial attention. This guides the automatic focus on the wake cloud region during the feature extraction stage. The time tag module is used to define image file names based on video frame images and shooting time, parse time tags from image file names, convert time category information into discrete vectors through one-hot encoding, and then map it into a 64-dimensional time feature vector through a fully connected layer; In Neck, the 64-dimensional temporal feature vector and spatial features are fused using the concat method to generate a joint feature map; Classification and recognition of joint feature maps are performed using Head; The improved YOLOv8 network model introduces a temporal consistency loss function during training, adding a similarity constraint term between the prediction results of the same target in different time frames to the original YOLO loss in the Head of the YOLOv8-tiny model. The similarity constraint is: ； in, This represents the predicted probability vector of the same target at different times; The time adjacency weighting function is represented by the function. This is an adjustment factor.

2. The method for identifying wake clouds in ground-based time-series observation images according to claim 1, characterized in that, Also includes: Based on the capture time and state of each acquired contrail image, a one-to-one correspondence is established to construct a contrail evolution state labeling system. Based on the wake cloud evolution state labeling system and the coordinate information corresponding to each image, target classification is performed on the video frame images to obtain the label image for each category.

3. The method for identifying wake clouds in ground-based time-series observation images according to claim 2, characterized in that, The construction of the wake cloud evolution state labeling system specifically includes: By analyzing the evolution of contrails, and combining the shooting time, the target was divided into three states: newly generated (0 min), generated about 30 minutes ago (30 min), and generated about 1 hour ago (60 min). Non-contrails were added as a control group.

4. The method for identifying wake clouds in ground-based time-series observation images according to claim 1, characterized in that, The classification and recognition of the joint feature map also includes outputting the target location, evolutionary state category, and confidence information of the wake cloud.

5. The method for identifying wake clouds in ground-based time-series observation images according to claim 1, characterized in that, The acquisition of the current wake cloud video frame image specifically includes: Select multiple points in the shooting area, set a fixed elevation angle for the camera, and capture video frame images of the contrails in the current state by taking timed shots; By performing edge computing on video frame images of contrails, linear structural features in the images are detected, invalid images without significant high-altitude trajectories are excluded, and images that pass the initial quality screening are obtained and labeled.

6. The method for identifying wake clouds in ground-based time-series observation images according to claim 1, characterized in that, The coordinate information is GPS positioning coordinates.

7. A ground-based temporal observation image wake cloud recognition system, characterized in that, include: The image acquisition module is used to acquire video frame images of the current trail cloud, as well as its corresponding shooting time and coordinate information; The feature extraction module is used to extract the temporal feature map of the wake cloud video frame image by constructing a temporal feature vector based on the video frame image and the shooting time; and to obtain the channel attention and spatial attention of the video frame image based on the video frame image and coordinate information, generate a weight map of the channel attention and spatial attention features through global average pooling and max pooling operations, and enhance the intermediate features of the weight map in the channel dimension and spatial dimension respectively to extract the spatial feature map of the video frame image. The feature fusion module is used to concatenate the temporal and spatial feature tensors along the channel dimension based on the temporal and spatial feature maps to generate a joint feature map. The identification module is used to classify and identify the joint feature map to obtain the category of the wake cloud; The category of the contrail cloud is identified using an improved YOLOv8 network model, which specifically includes: The YOLOv8-tiny model was used as the detection skeleton structure, including the Backbone, Neck, Head and time label modules; An attention mechanism CBAM is inserted into the c2f of the preceding feature extraction layer and the Conv of the following feature extraction layer in the backbone. This mechanism applies channel attention and spatial attention to the intermediate feature maps and performs spatial feature extraction by jointly modeling channel attention and spatial attention. This guides the automatic focus on the wake cloud region during the feature extraction stage. The time tag module is used to define image file names based on video frame images and shooting time, parse time tags from image file names, convert time category information into discrete vectors through one-hot encoding, and then map it into a 64-dimensional time feature vector through a fully connected layer; In Neck, the 64-dimensional temporal feature vector and spatial features are fused using the concat method to generate a joint feature map; Classification and recognition of joint feature maps are performed using Head; The improved YOLOv8 network model introduces a temporal consistency loss function during training, adding a similarity constraint term between the prediction results of the same target in different time frames to the original YOLO loss in the Head of the YOLOv8-tiny model. The similarity constraint is: ； in, This represents the predicted probability vector of the same target at different times; The time adjacency weighting function is represented by the function. This is an adjustment factor.

Images