A method and system for fast decision of aeroglow gravity wave image of ground base

Abstract

The present application relates to a kind of ground airglow gravity wave image fast determination method and system, method includes: collecting real-time ground airglow image;After the real-time ground airglow image is preprocessed, input layered transfer learning model, output gravity wave existence probability, and save wave image, wherein, the layered transfer learning model is obtained by segment training to ground airglow image dataset, the layered transfer learning model includes pre-training EfficientNetB3 model, deformable convolution layer and classification head.The present application solves the problem that ground airglow imager is difficult to automatically distinguish gravity wave in short-term deployment in real time, reduces artificial expert discrimination workload, while avoiding the problem of overfitting low accuracy of traditional machine learning method in short time small amount of data training.Reach the purpose of rapid deployment, efficient identification.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of space physics observation and computer vision, and in particular to a method and system for rapid interpretation of ground-based airglow gravity wave images. Background Technology

[0002] Atmospheric gravity waves (AGWs) are periodic wave phenomena in the Earth's atmosphere caused by the combined effects of gravity and buoyancy. Their physical essence can be described as follows: when an air mass deviates from its equilibrium position due to disturbances such as orographic lifting, convective bursts, or jet shearing, coupled shear waves (horizontal wavelength 10-300 km) and longitudinal waves (vertical wavelength 2-20 km) are generated under the combined influence of the Coriolis force and static stability. According to the dispersion relation equation:

[0003]

[0004] Where w is the natural frequency, N is the buoyancy frequency, and k is the buoyancy frequency. h With k z These represent the horizontal and vertical wavenumbers, respectively, with H being the atmospheric elevation. The wavenumbers increase exponentially in the vertical direction (amplitude growth factor e). z / 2H When these waves propagate to the mesopause (approximately 87 km in altitude), energy dissipation causes periodic disturbances in the airglow radiation intensity. Typical gravity wave parameters include: horizontal phase velocity of 20-150 m / s, period of 5-60 minutes, and horizontal wavelength of 10-300 km. These waves influence meso-level atmospheric circulation through momentum deposition, with energy fluxes reaching 10-100 mW / m². 2 .

[0005] Airglow imaging, a crucial tool for studying the dynamics of the middle and upper atmosphere, captures radiative disturbances induced by atmospheric gravity waves in the airglow layer, providing key observational data for revealing the propagation mechanism and energy transport processes of gravity waves. Atmospheric gravity waves, as the core carriers of atmospheric energy transfer, have a horizontal wavelength range of 10-300 km. Their momentum deposition between the mesosphere and the lower thermosphere directly influences atmospheric circulation stability and space weather evolution. Accurately identifying the characteristic parameters of gravity waves is of great significance for studying inter-layer atmospheric coupling and improving the accuracy of numerical weather prediction.

[0006] Traditional gravity wave detection primarily relies on ground-based all-sky atmospheric glow imager networks and spaceborne low-light imaging equipment. Ground-based observations extract wave signals through differential and geometric correction of adjacent images, but are limited by the field of view of a single station and cloud obstruction, making it difficult to obtain complete large-scale gravity wave waveforms. While spaceborne imaging technology offers global coverage, its insufficient temporal resolution for single-pass data leads to large errors in period parameter inversion. Existing machine learning methods, such as the improved YOLOv5s using a bidirectional weighted BiFPN module to increase recognition accuracy to 75.8%, are prone to overfitting under small sample conditions. Other models, such as Faster R-CNN, suffer from excessive parameter counts, making them unsuitable for real-time monitoring requirements of ground-based equipment. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for rapid interpretation of ground-based airglow gravity wave images, solving the problem that ground-based airglow imagers are difficult to automatically and in real-time identify gravity waves during short-term deployment.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] A fast interpretation method for ground-based airglow gravity wave images includes:

[0010] Acquire real-time ground airglow images;

[0011] After preprocessing the real-time ground-based airglow image, it is input into a hierarchical transfer learning model, which outputs the probability of gravity wave presence and saves the wave-containing image. The hierarchical transfer learning model is obtained by segmenting the ground-based airglow image dataset. The hierarchical transfer learning model includes a pre-trained EfficientNetB3 model, deformable convolutional layers, and a classification head.

[0012] Optionally, obtaining the dataset includes: acquiring raw ground-based airglow images, performing data preprocessing and data augmentation on the raw ground-based airglow images, and obtaining the dataset.

[0013] Optionally, data preprocessing and data augmentation of the original ground-based airglow image include: geometric correction and time alignment, three-level adaptive median filtering, multi-scale fusion, and dynamic difference preprocessing of the original ground-based airglow image;

[0014] The preprocessed ground airglow image is randomly horizontally flipped, its brightness is perturbed, and its contrast is adjusted to obtain the dataset.

[0015] Optionally, segmented training of the ground-based airglow image dataset includes:

[0016] The initial hierarchical transfer learning model is trained using the dataset through a freeze phase and a fine-tuning phase to obtain the hierarchical transfer learning model. In the freeze phase, the parameters of the pre-trained EfficientNetB3 model are frozen, and the deformable convolutional layer and the classification head are trained using the AdamW optimizer. In the fine-tuning phase, the remaining first 200 layers of the pre-trained EfficientNetB3 model are unfrozen, and the unfrozen portion of EfficientNetB3, the deformable convolutional layer, and the classification head are fine-tuned using a cosine annealing learning rate scheduling method and a hierarchical weight decay strategy.

[0017] Optionally, after preprocessing the real-time ground airglow image, it is input into a hierarchical transfer learning model, and the output gravity wave existence probability includes:

[0018] The preprocessed real-time ground airglow image is input into the pre-trained EfficientNetB3 model for feature extraction to obtain the feature map.

[0019] The feature map is input into the deformable convolution to predict the spatial offset through the convolutional layer. The spatial offset is then added and convolved element-wise with the preprocessed real-time ground-based airglow image to output the enhanced features.

[0020] The enhanced features are input into the classification head, and the probability of the existence of the gravity wave is output.

[0021] Optionally, the classification head includes: Dropout, Swish activation layer and batch normalization layer, and output layer.

[0022] This invention also provides a fast interpretation system for ground-based airglow gravity wave images, comprising:

[0023] Real-time data acquisition module, used to acquire real-time ground airglow images;

[0024] The real-time recognition output module is used to preprocess the real-time ground-based airglow image and input it into the hierarchical transfer learning model, output the probability of gravity wave presence, and save the wave-containing image. The hierarchical transfer learning model is obtained by segmenting the ground-based airglow image dataset. The hierarchical transfer learning model includes a pre-trained EfficientNetB3 model, deformable convolutional layers, and a classification head.

[0025] Optionally, the system further includes:

[0026] The data receiving and preprocessing module is used to acquire raw foundation airglow images and perform data preprocessing on the raw foundation airglow images;

[0027] The data augmentation module is used to randomly flip the preprocessed ground-based airglow image horizontally, perturb its brightness, and adjust its contrast to obtain the dataset.

[0028] The model building and training module is used to obtain an initial hierarchical transfer learning model by using pre-trained EfficientNetB3, deformable convolutional layers and classification heads, and to train the initial hierarchical transfer learning model using the dataset to perform a freeze phase and a fine-tuning phase, thereby obtaining the hierarchical transfer learning model.

[0029] The beneficial effects of the present invention are as follows: (1) Data augmentation strategy can expand the dataset to meet the generalization requirements of small sample training; (2) By constructing a hierarchical transfer learning model, the amount of data required for training can be significantly reduced while ensuring accuracy, and the complexity of the model can be simplified; (3) The shape of the receptive field can be dynamically adjusted by deformable convolutional layers to adapt to the bending deformation of gravity waves; (4) Training resources can be saved and the training speed can be accelerated by two-stage optimized training.

[0030] This achieves high-precision, low-latency real-time identification with few labeled samples. The model is simple and easy to implement, and is suitable for real-time rapid judgment of gravity wave events in short-range ground-based airglow imaging equipment. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of a method for fast interpretation of ground-based airglow gravity wave images according to an embodiment of the present invention;

[0033] Figure 2 This is a comparison chart of data enhancement in an embodiment of the present invention;

[0034] Figure 3 This is a diagram of the hierarchical transfer learning model architecture according to an embodiment of the present invention;

[0035] Figure 4 This is a diagram of the deformable convolutional layer structure according to an embodiment of the present invention;

[0036] Figure 5 The figures show the loss and accuracy curves during model training in this embodiment of the invention, where (a) is the loss curve and (b) is the accuracy curve.

[0037] Figure 6 This is a diagram showing the fast judgment result of the gravity wave output in an embodiment of the present invention. Detailed Implementation

[0038] 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, and 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.

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1:

[0041] like Figure 1 As shown, this embodiment provides a method for fast interpretation of ground-based airglow gravity wave images, including:

[0042] Acquire real-time ground airglow images;

[0043] After preprocessing the real-time ground-based airglow image, it is input into a hierarchical transfer learning model, which outputs the probability of gravity wave presence and saves the wave-containing image. The hierarchical transfer learning model is obtained by segmented training on a model constructed based on EfficientNetB3 and MobileNetV2 and bound to a deformable convolutional layer of the ground-based airglow image dataset. The hierarchical transfer learning model includes a pre-trained EfficientNetB3 model, a deformable convolutional layer, and a classification head.

[0044] Specifically, the real-time acquired ground-based airglow images are denoised by star points and dynamically differencing, then resized to 300x300 and input into the trained network. The `model.predict()` function is called, returning the probability of category 1 (no wave). The final output is the probability of gravity wave presence (1-prediction), and the wave-containing image is saved. Figure 6 As shown.

[0045] Furthermore, obtaining the dataset includes: acquiring raw ground airglow images, performing data preprocessing and data augmentation on the raw ground airglow images, and obtaining the dataset.

[0046] Furthermore, data preprocessing and data augmentation of the original ground-based airglow image include: geometric correction and time alignment of the original ground-based airglow image, three-level adaptive median filtering, multi-scale fusion, and dynamic difference preprocessing.

[0047] The preprocessed ground airglow image was randomly horizontally flipped, its brightness was perturbed, and its contrast was adjusted to obtain the dataset.

[0048] Specifically, it receives the raw PNG image sequence (1024x1024 pixels resolution, 800km field of view coverage diameter) acquired by the ground-based airglow imager and performs geometric correction and time alignment.

[0049] After eliminating star point noise using a three-level adaptive median filter, multi-scale fusion is performed to generate the signal. Then, dynamic difference is performed to generate a denoised image I. diff ;

[0050] For I diff After scaling, random horizontal flipping, brightness perturbation, and contrast adjustment are applied to generate an enhanced dataset.

[0051] Noise is eliminated using a three-stage adaptive median filter, the mathematical expression of which is:

[0052] Mk(x, y) = median i，j∈[-k，k] (I(x+i, y+j)), k∈{9, 15, 21};

[0053] The selection of the filter window k is based on:

[0054] Define a threshold to remove speckle noise, expressed as: Threshold k =μ k +3σ k ;

[0055] Multi-scale fusion is performed based on the filtering results at different scales, as expressed by:

[0056]

[0057] Dynamic difference processing is performed on adjacent images, using the following formula:

[0058]

[0059] Where Δs = 0.781 km / pixel is the spatial resolution of a single pixel, k is the size of the filter window, and λ h For the wavelength of the gravitational wave acting on the target (10-35km), μ k , σ k The filter residuals R are respectively k =|IM k The mean and standard deviation within a k×k window, w = [0.4, 0.3, 0.3] correspond to the weight coefficients of the three-level filter, and the mask, respectively.

[0060] like Figure 2 As shown, data augmentation operations include:

[0061] Randomly flip the initial image;

[0062] The brightness perturbation of the flipped image is mathematically defined as follows:

[0063] I′=clip(I+Δ,0,255),Δ=αI+β;

[0064] To adjust the contrast of an image after brightness adjustment, the expression is:

[0065] I′=clip(α·I,0,255);

[0066] Where α and β are perturbation factors, and the clip() function constrains the pixel value to the effective range [0, 255].

[0067] Further segmented training on the ground-based airglow image dataset includes:

[0068] The initial hierarchical transfer learning model is trained using the dataset through a freeze phase and a fine-tuning phase to obtain the hierarchical transfer learning model. In the freeze phase, the parameters of the pre-trained EfficientNetB3 model are frozen, and the deformable convolutional layer and the classification head are trained using the AdamW optimizer. In the fine-tuning phase, the remaining first 200 layers of the pre-trained EfficientNetB3 model are unfrozen, and the unfrozen portion of EfficientNetB3, the deformable convolutional layer, and the classification head are fine-tuned using a cosine annealing learning rate scheduling method and a hierarchical weight decay strategy.

[0069] Furthermore, after preprocessing the real-time ground airglow image, it is input into a hierarchical transfer learning model, and the output gravity wave existence probability includes:

[0070] The preprocessed real-time ground airglow image is input into the pre-trained EfficientNetB3 model for feature extraction to obtain the feature map.

[0071] The feature map is input into the deformable convolution to predict the spatial offset through the convolutional layer. The spatial offset is then added and convolved element-wise with the preprocessed real-time ground-based airglow image to output the enhanced features.

[0072] The enhanced features are input into the classification head, and the probability of the existence of the gravity wave is output.

[0073] Specifically, such as Figure 3 As shown, a pre-trained EfficientNetB3 outer deformable convolutional layer is used as the backbone network, the image resolution is adjusted to the input size (300x300), some parameters are frozen, and a fully connected layer is output.

[0074] The custom category header includes Dropout, Swish activation layer (L2 regularization), batch normalization layer, and output layer.

[0075] Freeze phase: Freeze all parameters of EfficientNetB3, train deformable convolutional layers and classification head using AdamW optimizer, and save training resources with early stopping mechanism;

[0076] Fine-tuning phase: Unfreeze the portion of EfficientNetB3 excluding the first 200 layers, and fine-tune the unfrozen portion of EfficientNetB3, deformable convolutional layers, and classification head using cosine annealing learning rate and hierarchical weight decay strategies.

[0077] The pre-trained EfficientNetB3 (input size 300×300) is used, whose basic building block is the MBConv module. This module inherits the core design of the inverse residual structure of MobileNetV2 and improves the activation function while introducing an attention mechanism. It includes an up-dimensional convolutional layer (1×1 convolution with 6 times the number of channels), a depthwise separable convolutional layer, and a down-dimensional convolutional layer (1×1 convolution with 1 compressed channel). The parameters of the first 200 layers are frozen (specifically the weights of the up-dimensional and down-dimensional convolutional layers in the 1st to 16th inverse residual blocks).

[0078] Deformable convolutional layers are deployed at the output of the pre-trained EfficientNetB3 backbone network. Their structure comprises three consecutive operations: first, a 3×3 convolutional layer predicts the spatial offset of the input feature map (the number of output channels strictly matches the 1536 channels of the input feature map); then, the predicted offset is added element-wise to the original feature map to achieve feature deformation; finally, a standard 3×3 convolution operation is performed on the deformed feature map to generate a 512-channel enhanced feature output, forming an adaptive gravity waveform feature enhancement module. The network is then connected to a fully connected classification layer after global average pooling. Specifically, for example... Figure 4 As shown.

[0079] Define a custom Swish activation function, with the expression: Swish(x) = x·σ(βx);

[0080] The penalty term for L2 regularization takes the form of:

[0081] Where σ is the Sigmoid function, applied to the weight parameters θ of the fully connected layer. i .

[0082] like Figure 5 (a)- Figure 5 (b) shows the loss and accuracy curves during model training. The two-stage optimization training steps include:

[0083] During the freeze phase, the weight decay of the AdamW optimizer directly affects the parameter update, as expressed by:

[0084]

[0085] During the fine-tuning phase, a cosine annealing learning rate scheduling method is adopted, and the formula is defined as:

[0086] η t =η min +0.5(η max -η min (1+cos(πt / T));

[0087] A hierarchical weight decay strategy is adopted, expressed as: λ l =0.01·e -2l / L ;

[0088] Where η is the learning rate used to control the step size of parameter updates, λ is the weight decay coefficient, and T is the learning rate recovery period. max Let η be the initial learning rate. min is the minimum learning rate, l is the layer number, frozen layers do not participate in the reduction, and L is the total number of fine-tuning layers.

[0089] This method offers a novel approach for rapid interpretation of ground-based airglow gravity wave images, addressing the challenge of automated real-time gravity wave identification in short-term deployments of ground-based airglow imagers. It reduces the workload of human experts in the identification process and avoids the overfitting and low accuracy issues inherent in traditional machine learning methods trained on limited data in a short time. This achieves the goals of rapid deployment and efficient identification.

[0090] Example 2:

[0091] A ground-based airglow gravity wave image fast interpretation system includes:

[0092] Real-time data acquisition module, used to acquire real-time ground airglow images;

[0093] The real-time recognition output module is used to preprocess the real-time ground-based airglow image and input it into the hierarchical transfer learning model, output the probability of gravity wave presence, and save the wave-containing image. The hierarchical transfer learning model is obtained by segmenting the ground-based airglow image dataset. The hierarchical transfer learning model includes a pre-trained EfficientNetB3 model, deformable convolutional layers, and a classification head.

[0094] Specifically, the preprocessed ground airglow image is input into the trained network model. The model outputs the probability value of no gravity wave, and the probability of gravity wave presence is calculated as 1-prediction. When the probability is ≥0.5, it is determined that gravity wave exists and the image is saved. At the same time, the binary classification result (with wave / no wave) and the precise probability value are output.

[0095] The data receiving and preprocessing module is used to acquire raw foundation airglow images and perform data preprocessing on the raw foundation airglow images;

[0096] Specifically, it receives the raw PNG image sequence (1024x1024 pixels resolution, 800km field of view coverage diameter) acquired by the ground-based airglow imager and performs geometric correction and time alignment.

[0097] After eliminating star point noise using a three-level adaptive median filter, multi-scale fusion is performed to generate the signal. Then, dynamic difference is performed to generate a denoised image L. diff .

[0098] A three-level adaptive median filter is used to eliminate both small and large star noise. The 3σ criterion is used to remove star noise that deviates significantly from the background distribution within the window. A multi-scale fusion strategy is employed, using the original image directly in the mask area to avoid gravity wave distortion caused by filtering, and using the results of multi-level filtering in non-star areas to eliminate high-frequency noise. The signal loss of single-level filtering is compensated by fusion weights. Finally, dynamic difference processing is performed on two adjacent processed images within a reasonable time window to ensure the reasonableness of the results.

[0099] The data augmentation module is used to randomly flip the preprocessed ground airglow image horizontally, perturb its brightness, and adjust its contrast to obtain a dataset.

[0100] By defining physical quantities and constraining parameters, we can enhance data diversity while ensuring the physical identifiability of gravity wave characteristics, thus meeting the model generalization requirements under small sample training.

[0101] The model building and training module is used to obtain an initial hierarchical transfer learning model by using pre-trained EfficientNetB3, deformable convolutions, and classification heads. The initial hierarchical transfer learning model is then trained using a dataset through a freeze phase and a fine-tuning phase to obtain the hierarchical transfer learning model.

[0102] A pre-trained EfficientNetB3 outer deformable convolutional layer is used as the backbone network. The image resolution is adjusted to the input size (300x300), some parameters are frozen, and a fully connected layer is output.

[0103] The pre-trained EfficientNetB3 architecture (input size 300×300) is adopted. Its basic building block is the MBConv module, which is an improvement of the MobileNetV2 inverse residual block. While retaining the core structure of 1×1 up-dimensional convolution (channel expansion by 6 times), depthwise separable convolution and 1×1 down-dimensional convolution (linear activation), the Swish activation function and SE channel attention mechanism are introduced; the parameters of the first 200 layers are frozen (specifically the weights of the up-dimensional and down-dimensional convolution layers in the 1st to 16th MBConv modules).

[0104] Deformable convolutional layers are deployed at the output of the pre-trained EfficientNetB3 backbone network. Its structure includes three consecutive operations: First, the spatial offset of the input feature map is predicted through a 3×3 convolutional layer (the number of output channels strictly matches the 1536 channels of the input feature map). Then, the predicted offset is added element-wise to the original feature map to achieve feature deformation. Finally, a standard 3×3 convolution operation is performed on the deformed feature map to generate a 512-channel enhanced feature output, forming an adaptive gravity waveform feature enhancement module. The network ends with a global average pooling layer to output a feature vector, which is then connected to a fully connected classification layer.

[0105] Custom category headers include Dropout, Swish activation layer (L2 regularization), and batch normalization layer;

[0106] Freeze phase: Freeze all parameters of EfficientNetB3, train deformable convolutional layers and classification head using AdamW optimizer, and save training resources with early stopping mechanism;

[0107] Fine-tuning phase: Unfreeze the portion of EfficientNetB3 excluding the first 200 layers, and fine-tune the unfrozen portion of EfficientNetB3, deformable convolutional layers, and classification head using cosine annealing learning rate and hierarchical weight decay strategies.

[0108] The backbone network uses a pre-trained EfficientNetB3, and optimizes the model's depth, width, and resolution through a composite scaling method to achieve efficient and accurate performance. Deformable convolution dynamically adjusts the receptive field shape to adapt to the bending deformation of gravitational waves. A freezing strategy is used to reduce redundant training with small samples, improve feature processing speed, optimize computational resources, and suppress overfitting. A custom Swish activation function is used to adaptively adjust the feature scale, improve the activation rate, and ensure recognition accuracy. L2 regularization and Dropout constraints are used to suppress abnormal weights and improve model robustness.

[0109] The AdamW optimizer is used to decouple weight decay, which can improve gradient stability and prevent overfitting. Cosine annealing learning rate scheduling can accelerate convergence and increase generalization ability, and escape local optima. A hierarchical weight decay mechanism is used to apply an exponential decay strategy to the thawing layer during the fine-tuning stage. The shallow layer has strong decay to constrain overfitting of low-order features, while the deep layer has weak decay to preserve high-order semantic features.

[0110] In this embodiment, preprocessing of the received images improves the reliability and usability of the data, meeting the needs of subsequent operations. Data augmentation strategies expand the dataset, satisfying the generalization requirements of small-sample training. Hierarchical transfer learning model structure freezing, fine-tuning coupling, and hierarchical training strategies significantly reduce the amount of data required for training while maintaining accuracy, simplifying model complexity. Two-stage optimized training saves training resources and accelerates training speed. This achieves high-precision, low-latency real-time recognition under conditions of limited sample annotation. The model is simple and easy to implement, suitable for real-time rapid judgment of gravity wave events in short-range ground-based airglow imaging equipment.

[0111] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for fast interpretation of ground-based airglow gravity wave images, characterized in that, include: Acquire real-time ground airglow images; After preprocessing the real-time ground-based airglow image, it is input into a hierarchical transfer learning model, which outputs the probability of gravity wave presence and saves the wave-containing image. The hierarchical transfer learning model is obtained by segmenting the ground-based airglow image dataset and includes a pre-trained EfficientNetB3 model, deformable convolutional layers, and a classification head. Segmented training of the ground-based airglow image dataset includes: The initial hierarchical transfer learning model is trained using the dataset through a freeze phase and a fine-tuning phase to obtain the hierarchical transfer learning model. In the freeze phase, the parameters of the pre-trained EfficientNetB3 model are frozen, and the deformable convolutional layer and the classification head are trained using the AdamW optimizer. In the fine-tuning phase, the remaining first 200 layers of the pre-trained EfficientNetB3 model are unfrozen, and the unfrozen portion of EfficientNetB3, the deformable convolutional layer, and the classification head are fine-tuned using a cosine annealing learning rate scheduling method and a hierarchical weight decay strategy.

2. The method for rapid interpretation of ground-based airglow gravity wave images according to claim 1, characterized in that, Obtaining the dataset includes: acquiring raw ground-based airglow images, performing data preprocessing and data augmentation on the raw ground-based airglow images, and obtaining the dataset.

3. The method for rapid interpretation of ground-based airglow gravity wave images according to claim 2, characterized in that, Data preprocessing and data augmentation of the original ground-based airglow image include: geometric correction and time alignment, three-level adaptive median filtering, multi-scale fusion, and dynamic difference preprocessing. The preprocessed ground airglow image is randomly horizontally flipped, its brightness is perturbed, and its contrast is adjusted to obtain the dataset.

4. The method for rapid interpretation of ground-based airglow gravity wave images according to claim 1, characterized in that, After preprocessing the real-time ground airglow image, it is input into a hierarchical transfer learning model, and the output gravity wave existence probability includes: The preprocessed real-time ground airglow image is input into the pre-trained EfficientNetB3 model for feature extraction to obtain the feature map. The feature map is input into the deformable convolution to predict the spatial offset through the convolutional layer. The spatial offset is then added and convolved element-wise with the preprocessed real-time ground-based airglow image to output the enhanced features. The enhanced features are input into the classification head, and the probability of the existence of the gravity wave is output.

5. The method for rapid interpretation of ground-based airglow gravity wave images according to claim 1, characterized in that, The classification head includes: Dropout, Swish activation layer, batch normalization layer, and output layer.

6. A rapid interpretation system for ground-based airglow gravity wave images, implemented according to any one of claims 1-5, characterized in that, include: Real-time data acquisition module, used to acquire real-time ground airglow images; The real-time recognition output module is used to preprocess the real-time ground-based airglow image and input it into the hierarchical transfer learning model, output the probability of gravity wave presence, and save the wave-containing image. The hierarchical transfer learning model is obtained by segmenting the ground-based airglow image dataset. The hierarchical transfer learning model includes a pre-trained EfficientNetB3 model, deformable convolutional layers, and a classification head.

7. The fast judgment system for ground-based airglow gravity wave images according to claim 6, characterized in that, The system also includes: The data receiving and preprocessing module is used to acquire raw foundation airglow images and perform data preprocessing on the raw foundation airglow images; The data augmentation module is used to randomly flip the preprocessed ground-based airglow image horizontally, perturb its brightness, and adjust its contrast to obtain the dataset. The model building and training module is used to obtain an initial hierarchical transfer learning model by using pre-trained EfficientNetB3, deformable convolutional layers and classification heads, and to train the initial hierarchical transfer learning model using the dataset to perform a freeze phase and a fine-tuning phase, thereby obtaining the hierarchical transfer learning model.

Images