Aurora form unsupervised classification method based on all-sky multi-expert model

Abstract

An aurora form unsupervised classification method based on an all-sky multi-expert model comprises the steps that 1) the all-sky multi-expert model refines aurora form characteristics from different angles by adopting a mode of cooperative work of a plurality of expert networks in order to solve the problem of complex content in an all-sky view, and the model characterization capability is enhanced; 2) for the problem of aurora local structure diversity, dynamic convolution and dynamic snakelike convolution are introduced into an all-sky multi-expert model to construct an attention module, so that accurate representation of the aurora local structure is realized; 3) adopting a rapid hierarchical clustering algorithm, and obtaining a complete hierarchical structure representing different granularity levels of the aurora form by constructing an adjacent matrix; and 4) introducing an abnormal value screening strategy into the hierarchical clustering algorithm to improve the reliability of a clustering result. The method provided by the invention can efficiently capture the morphological characteristics of the all-sky aurora image, and explores a natural classification mode of the aurora morphology under an unsupervised condition.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, relates to feature extraction of all-sky auroral morphology and a fast hierarchical clustering algorithm, and specifically relates to an unsupervised classification method of auroral morphology based on an all-sky multi-expert model. Background Art

[0002] Unsupervised learning is an important method in machine learning. Its core idea is to perform statistical and similarity analysis on unlabeled data without pre-defining target variables to discover underlying structures and patterns in the data. Cluster analysis, a fundamental task in unsupervised learning, aims to divide data samples into clusters, so that similar samples are grouped together and dissimilar samples are grouped together. Its core goal is to provide a global perspective for subsequent analysis by discovering the inherent structure of the data. Traditional clustering methods assume that the data has been converted into a vector form suitable for clustering through feature engineering, but face significant challenges when processing high-dimensional data such as images. With the development of deep learning, a large number of deep learning-based image clustering methods have emerged. A direct and effective approach is to first learn a high-level abstract representation of the data through deep neural networks, solving the problem of feature extraction from unstructured data. The extracted features are then input into the clustering algorithm to obtain the final result.

[0003] During the feature extraction phase, contrastive learning, with its core principle of narrowing the distance between positive pairs (similar instances) while simultaneously pushing the distance between negative pairs (dissimilar instances), can learn physically consistent and discriminative feature representations under unsupervised conditions, making it an important tool for improving clustering performance. A representative approach is the MOCO model proposed by He Kaiming's team in 2020. MOCO reduces the work of contrastive learning to a dictionary lookup task, maintaining a dynamic dictionary to store negative samples. The samples in the dictionary are dynamically updated in a queue, with old samples gradually replaced by new ones. During training, the dictionary size can be much larger than the batch size, providing richer contrast information for the learning process. Furthermore, to maintain feature consistency in the dictionary, MOCO employs a momentum update mechanism. Specifically, MOCO uses an online encoder and a momentum encoder to extract query and key features, respectively. The parameters of the momentum encoder are gradually updated through the momentum update mechanism and remain consistent with those of the online encoder. This helps capture more stable feature representations.

[0004] In the image field, two different clustering methods are commonly used during the clustering phase: partition-based clustering and hierarchical clustering. The core idea of the partition-based clustering approach is to partition the dataset into k clusters, each containing at least one example, and the number of clusters k must be specified in advance. The hierarchical clustering approach focuses more on the hierarchical relationships between the data, partitioning clusters by constructing a hierarchical structure. Specifically, these approaches can be categorized as bottom-up (agglomerative) and top-down (divisive). Compared to partition-based clustering, hierarchical clustering does not require a pre-specified k value and can automatically discover clusters of arbitrary shapes. This saves resources and time while also interpreting data at different granularity levels through a hierarchical structure. For example, when using a bottom-up agglomerative hierarchical clustering approach, the fine-grained partitions at the bottom retain higher purity, while the coarse-grained partitions at the top reflect higher-level semantics, both of which are important for measuring similarity between instances. Therefore, hierarchical clustering is more suitable for our original goal of exploring the natural classification of auroral images based on the full-sky auroral dataset itself, without relying on artificial prior knowledge.

[0005] The method described above is a straightforward and effective staged training approach. First, positive and negative samples are constructed through data augmentation. The encoder is trained using a contrastive loss without any manual annotation, resulting in similar instances being closer together and dissimilar instances being further apart in the embedding space. An appropriate clustering algorithm is then selected based on the data characteristics to aggregate image features in the feature space. This method can capture more discriminative feature representations in the image through contrastive learning, thereby improving clustering effectiveness. However, the final results are highly dependent on the quality of the feature representation.

[0006] Compared to other types of data, all-sky auroral images have complex dynamic structures and varied morphologies. Samples within the same category often differ significantly in morphology, and the boundaries between samples in different categories are blurred due to transitional morphologies, making it difficult to effectively capture semantic similarity in high-dimensional space. While existing empirical classification methods can describe the different morphologies of auroras to a certain extent, they still rely on the researcher's visual judgment and a small amount of high-quality observational data. This can lead to overlooking potential morphological structural features in the data, and the classification mechanism lacks universality. Therefore, we use cluster analysis to explore natural classification patterns of auroral morphology under unsupervised conditions. In the feature extraction stage, we employ multi-expert collaboration to refine auroral morphological features from different perspectives, aiming to explore data-adaptive classification standards. Summary of the Invention

[0007] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide an unsupervised classification method for auroral morphology based on a full-sky multi-expert model, so as to solve the problem of insufficient key feature extraction caused by the complexity of auroral data and the problem that the existing aurora classification mechanism relying on empirical classification ignores potential information and lacks universality.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] An unsupervised classification method for auroral morphology based on a full-sky multi-expert model includes the following steps:

[0010] Step 1: Preprocess the aurora image:

[0011] Step 1.1: Perform intensity scaling and dark current subtraction on the original aurora image with a size of 512×512 and a brightness value in the dynamic range of [0,17000];

[0012] Step 1.2: Create a circular mask with a radius of 220 pixels, with the zenith as the center. Multiply the mask with the original image to construct the full sky domain.

[0013] Step 1.3: Perform image cropping and rotation on the full-sky aurora image to obtain an image with a size of 440 × 440 pixels and a grayscale value in the dynamic range of [0, 4000] after preprocessing and data enhancement.

[0014] Step 2: Introduce the full-sky multi-expert model as an encoder; this model uses multiple experts to work together to enhance the perception of full-sky content;

[0015] Step 2.1: Input the image after preprocessing and data enhancement in step 1 into the full-sky multi-expert model, and pass the obtained shallow features through the full-sky channel attention module;

[0016] Step 2.2: The full-sky channel attention module performs full-sky maximum pooling and full-sky maximum pooling on the input features in the spatial dimension, and compresses the features into ;

[0017] Step 2.3: Input the results of the full-sky max pooling and the full-sky max pooling into the parameter-sharing fully connected layer respectively, generate two attention vectors and add them together;

[0018] Step 2.4: Normalize the added vector to obtain the final channel attention weight, and multiply it with the initial input feature to obtain the enhanced feature;

[0019] Step 3: Get the final feature map:

[0020] Step 3.1: Input the enhanced features obtained in step 2 into the spatial attention module. The spatial attention mechanism performs full-sky maximum pooling and average pooling on the features in the channel dimension, and compresses the feature map into ;

[0021] Step 3.2: After the features are spliced on the channel, a dynamic snake convolution operation is performed to obtain the spatial attention weight, and the convolution kernel direction is dynamically adjusted according to the local morphology of the features;

[0022] Step 3.3, multiply the weight by the initial feature map at the time of input to obtain the final feature map and input it to the next residual layer to continue extracting features;

[0023] Step 4: Pass the features obtained by the last residual layer of the encoder through an average pooling layer, and then input the obtained feature information into multiple expert networks. The gating network sparsely activates some expert networks. The final output is the weighted sum of the outputs of all expert networks, which can refine the global information from different angles and improve the model representation ability.

[0024] Step 5: Perform cluster analysis on the features extracted in step 4 using a fast hierarchical clustering algorithm;

[0025] Step 5.1: Treat each input image feature as a data point, obtain the nearest neighbor information of the data point using the fast nearest neighbor method, and connect the nearest neighbor points with the help of the adjacency matrix to obtain the first-level clustering result;

[0026] Step 5.2: Then use the threshold strategy to filter out outliers and average the features in each cluster to obtain a mean vector as the cluster prototype, and then use the same method to obtain the next layer of clustering results.

[0027] Furthermore, the multi-expert network includes multiple expert networks with independent parameters. and a gating network that outputs a sparse vector .

[0028] Furthermore, the step (2) of China Railway's full-sky channel attention is performed according to the following steps:

[0029] 1) First, the enhanced view passes through a convolutional layer and full-sky maximum pooling, and then passes through the first residual layer to obtain shallow features ;

[0030] 2) The shallow features obtained In the spatial dimension, full sky domain maximum pooling and full sky domain average pooling are performed respectively to compress the features into , and then input the pooling results into the parameter-sharing fully connected layer to obtain two attention vectors and add them together. The result of the addition is passed through a sigmoid activation function to generate the final channel attention weight, which is combined with the initial feature After multiplication, the enhanced features are obtained , implemented as follows:

[0031] ;

[0032] ;

[0033] in, Represents the input features The channel attention weights obtained after applying the full-sky channel attention, represents pixel-by-pixel multiplication, It represents the result obtained after all-sky channel attention.

[0034] Furthermore, the spatial attention module in step (3) is implemented as follows:

[0035] 1) The features obtained in step (2) Perform channel dimension maximum pooling and average pooling operations respectively to compress the features into , then concatenate the pooled results on the channel and extract local features through a dynamic snake convolution layer (DSConv); DSConv is implemented as follows:

[0036] Given a standard Convolution kernel, the center coordinate is :

[0037] ;

[0038] Introducing deformation offset , ensuring that the convolution kernel conforms to the linear morphological structure;

[0039] From the center coordinates Initially, the position away from the center depends on the position of the previous grid; for Axis direction, the change is expressed as:

[0040] ;

[0041] for Axis direction, the change is expressed as:

[0042] ;

[0043] 2) The result after the dynamic snake convolution layer is passed through a sigmoid activation function to generate the final spatial attention weight; the weight is combined with the feature After multiplication, the enhanced features are obtained ; Implementation is as follows:

[0044] ;

[0045] ;

[0046] in, Represents the input feature map The spatial attention weights obtained after applying the spatial attention mechanism, represents pixel-by-pixel multiplication, Represents the result obtained after the spatial attention mechanism.

[0047] Furthermore, the multi-expert network in step (4) is implemented as follows:

[0048] 1) The features After the subsequent residual layer and dynamic convolution attention module, the result is passed through an average pooling layer to obtain the feature ;

[0049] 2) First, the gating network learns the correlation between the input features and each expert to obtain a trainable weight matrix ,Multiply the input feature with the weight to obtain the importance score of the current feature to each expert;

[0050] 3) By adding a trainable weight matrix Introducing adjustable Gaussian noise, randomness of the gating mechanism, and then only retaining the top ones with the highest importance scores The scores are converted into probability distributions through a softmax function to provide a basis for the allocation of input features. The implementation is as follows:

[0051] ;

[0052] ;

[0053] ;

[0054] in, represents the output of the gating network, Indicates that the input feature pair The importance score of each expert network;

[0055] 4) According to the gated network obtained in (3), the input features are assigned to the expert network to obtain the output of the final multi-expert network, which is implemented as follows:

[0056] ;

[0057] in, represents the output of the gating network, Indicates the The output of an expert network;

[0058] 5) To prevent a few top experts from being overloaded and other experts from being idle, we add auxiliary loss-balancing experts to utilize the load rate, as follows:

[0059] ;

[0060] ;

[0061] in, Indicates the calculation of importance value, Indicates the calculation of the coefficient of variation, which measures the degree of dispersion of the data. is a scaling factor that can be adjusted manually;

[0062] 6) Linearly add the auxiliary loss and the contrast loss to obtain the total model loss. By continuously optimizing the loss function, the model learns discriminative feature representations, as follows:

[0063] ;

[0064] in, and Set to 0.6 and 0.4 respectively.

[0065] Furthermore, the steps of the fast hierarchical clustering algorithm in step (5) are as follows:

[0066] 1) Treat each feature extracted by the feature enhancement network as a data point, obtain the nearest neighbor information of each sample point through the fast nearest neighbor method, and generate a symmetric sparse adjacency matrix, which is implemented as follows:

[0067] ;

[0068] in: Represents sample points The first neighbor of Represents sample points The first neighboring point of for The first neighbor or sample points and Connect two nodes when they share the same nearest neighbor. Ensure the symmetry of the matrix;

[0069] 2) The adjacency matrix obtained by (1) Get the first-level clustering results, and calculate the average value of the sample points in each cluster as the initial cluster prototype, which represents the typical characteristics of the samples in the cluster;

[0070] 3) Filter out outliers in the clusters. Common filtering strategies include masking the farthest point, masking by threshold, and masking by proportion. A threshold strategy is used based on the data characteristics. The threshold comprehensively considers the standard deviation, IQR, and sample density of the full-sky aurora image. The implementation is as follows:

[0071] Based on the mean times the standard deviation to define the threshold scope:

[0072] ;

[0073] Setting thresholds based on IQR scope:

[0074] ;

[0075] Define an adjustment factor based on sample density When the sample density within the cluster is low, the threshold range is appropriately expanded by adjusting the factor. The final threshold is obtained from the above:

[0076] ;

[0077] 4) Recalculate the mean of the clusters after outlier screening, update the prototype and use it as a new sample point to continue the iterative clustering process. The clustering results obtained at each layer are filtered for outliers. When the number of clusters is less than the specified number of clusters, the algorithm terminates.

[0078] The beneficial effects of the present invention are:

[0079] First, the unsupervised auroral morphology classification method based on the all-sky multi-expert model proposed in this invention combines a multi-expert model with a hierarchical clustering algorithm. It can adaptively mine the potential morphological features in auroral images under unsupervised conditions and explore the natural classification patterns of auroral morphology. The dynamic convolution module and multi-expert network introduced in the feature extraction stage can capture the complex dynamic structure and variable morphology of auroral images, enhancing the model's characterization capabilities. In the clustering stage, a fast hierarchical clustering algorithm capable of screening outliers is used to obtain hierarchical clustering results of auroral data at different granularity levels to explore the potential laws of auroral activity and the spatial distribution characteristics at different scales, which helps to improve the classification mechanism of auroral morphology.

[0080] Second, the present invention introduces a full-sky multi-expert model as an encoder, which uses multi-expert collaboration to enhance perception of full-sky content. The dynamic convolutional attention module introduced in the full-sky multi-expert model uses a hybrid attention mechanism that combines full-sky channel attention and spatial attention. In the spatial attention module, dynamic snake convolution is used to adjust the shape and orientation of the convolution kernel based on the feature map. Compared with traditional attention modules, this module has stronger local feature extraction capabilities and better data adaptability.

[0081] Third, the multi-expert network employed in the all-sky multi-expert model of this invention dynamically activates different sub-models through a gating network to refine global information from different perspectives. Given the high complexity and uncertainty of auroral data, this architecture effectively enhances the model's generalization and representation capabilities while conserving computational resources to a certain extent.

[0082] Fourth, the fast hierarchical clustering algorithm that can filter out outliers adopted by the present invention aggregates samples layer by layer from bottom to top through the adjacency matrix. The algorithm avoids maintaining the direct distance matrix between each data point. While revealing the potential structure of the data through the adjacency relationship, it also significantly reduces the computational complexity, and filters out outliers in the algorithm to improve clustering reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a flow chart for realizing the method of the present invention;

[0084] Figure 2 Schematic diagram of the dynamic convolution attention module in the present invention;

[0085] Figure 3 Schematic diagram of a fast hierarchical clustering algorithm capable of filtering outliers in the present invention;

[0086] Figure 4 This is the confusion matrix result diagram of the present invention and the baseline network in the aurora image classification task, where the left side is the baseline network confusion matrix and the right side is the full-sky multi-expert model confusion matrix;

[0087] Figure 5 This is a t-SNE dimensionality reduction visualization result diagram of data with different numbers of clusters in the hierarchical clustering of the present invention. DETAILED DESCRIPTION

[0088] The present invention will be further described below with reference to the accompanying drawings.

[0089] like Figure 1As shown, the original aurora image is first preprocessed to suppress noise interference introduced by background information, facilitating the subsequent extraction of the key features we need. The preprocessed image is then data-enhanced and input into the full-sky multi-expert model for feature extraction. This model architecture combines multiple expert networks (sub-networks) with a gated network, using the gated network to assign weights to each expert. This allows the model to adaptively select different expert networks to dynamically learn local morphological features in full-sky aurora images, refining global information from multiple perspectives to obtain discriminative feature representations. Furthermore, to more precisely capture auroral morphological features, the present invention introduces a dynamic convolutional attention module to enhance the network's ability to represent auroral morphological features. After the feature extraction stage, the obtained features are input into the hierarchical clustering algorithm. Taking into account the problem of insufficient capture of semantic similarity caused by traditional distance measurement in the clustering process and the impact of outliers on the accuracy of the results, the present invention adopts a fast hierarchical clustering algorithm based on the nearest neighbor matrix, emphasizing the importance of semantic relationships, and screening outliers during the algorithm iteration process. After the screening is completed, the average value of the current cluster is recalculated to update the prototype, avoiding the situation where outliers cause the prototype to deviate from the cluster center.

[0090] The details are as follows:

[0091] Step 1. The original image is preprocessed, including intensity scaling, dark current subtraction, masking and cropping, and image rotation. The resulting image size is 440×440 pixels, the circular mask radius is 220 pixels, and the grayscale value is within the dynamic range of [0, 4000]. The preprocessed image is then subjected to two enhancement methods, strong and weak enhancement, to obtain an enhanced view. A full-sky multi-expert model is used as the encoder to extract image features. Following the residual layer, the encoder uses channel attention and spatial attention with dynamic snake convolution to jointly enhance the model's local feature representation of the full-sky aurora image.

[0092] The dynamic convolution attention module in the all-sky multi-expert model of this embodiment is specifically implemented as follows:

[0093] (1.1) First, the enhanced view passes through a convolutional layer and full-sky maximum pooling, and then passes through the first residual layer to obtain shallow features ;

[0094] (1.2) The shallow features obtained In the spatial dimension, full sky domain maximum pooling and full sky domain average pooling are performed respectively to compress the features into , and then input the pooling results into the parameter-sharing fully connected layer to obtain two attention vectors and add them together. The result of the addition is passed through a sigmoid activation function to generate the final channel attention weight. After multiplication, the enhanced features are obtained The implementation is as follows:

[0095] ;

[0096] ;

[0097] in, Represents the input features The channel attention weights obtained after applying the full-sky channel attention, represents pixel-by-pixel multiplication, It represents the result obtained after all-sky channel attention.

[0098] (1.3) The features obtained in step (1.2) Perform channel dimension maximum pooling and average pooling operations respectively to compress the features into , then concatenate the pooled results on the channel and pass them through a dynamic snake convolution layer (DSConv) to extract local features. DSConv is implemented as follows:

[0099] (1.3.1) Given a standard Convolution kernel, the center coordinate is :

[0100] ;

[0101] (1.3.2) Introducing deformation offset , ensuring that the convolution kernel conforms to the linear morphological structure;

[0102] (1.3.3) From the center coordinates Initially, the position away from the center depends on the position of the previous grid. Axis direction, the change is expressed as:

[0103] ;

[0104] for Axis direction, the change is expressed as:

[0105] ;

[0106] (1.4) The result after the dynamic snake convolution layer is passed through a sigmoid activation function to generate the final spatial attention weight. After multiplication, the enhanced features are obtained The implementation is as follows:

[0107] ;

[0108] ;

[0109] in, Represents the input feature map The spatial attention weights obtained after applying the spatial attention mechanism, represents pixel-by-pixel multiplication, Represents the result obtained after the spatial attention mechanism.

[0110] (1.5) The characteristics After the subsequent residual layer and dynamic convolution attention module, the features are further extracted, and finally the features are obtained through an average pooling layer. ;

[0111] Step 2. Features The global information is refined from different perspectives through a multi-expert network. The network consists of multiple expert networks with independent parameters. The core concept is to assign input features to different expert sub-networks, then dynamically assign weights to each expert's output through a gating network, perform a weighted sum, and obtain the final result. This ensures that each expert handles the task they are best suited for, thereby improving the model's representational capabilities. The expert network can adopt any machine learning or deep learning model. In this invention, the expert network adopts a feed-forward neural network (FFN) structure. The specific implementation steps are as follows:

[0112] (2.1) First, learn the input features through the gated network The correlation between each expert and a trainable weight matrix ,Multiply the input feature with the weight to obtain the importance score of the current feature to each expert;

[0113] (2.2) By adding a trainable weight matrix Introducing adjustable Gaussian noise to enhance the randomness of the gating mechanism, and then only retaining the top The scores are converted into probability distributions through a softmax function to provide a basis for the distribution of input features. The implementation is as follows:

[0114] ;

[0115] ;

[0116] ;

[0117] in, represents the output of the gating network, Indicates that the input feature pair The importance score of an expert network.

[0118] (2.3) Based on the gating network obtained in (2.2), the input features are assigned to the expert network to obtain the output of the final multi-expert network. The implementation is as follows:

[0119] ;

[0120] in, represents the output of the gating network, Indicates the The output of an expert network.

[0121] (2.4) To prevent a few top experts from being overloaded and other experts from being idle, we add auxiliary losses to balance the expert utilization load rate. This is achieved as follows:

[0122] ;

[0123] ;

[0124] in, Indicates the calculation of importance value, Indicates the calculation of the coefficient of variation, which measures the degree of dispersion of the data. is a scaling factor that can be adjusted manually.

[0125] (2.5) The auxiliary loss and the contrast loss are linearly added to obtain the total model loss. By continuously optimizing the loss function, the model learns discriminative feature representations. The implementation is as follows:

[0126] ;

[0127] in, and Set to 0.6 and 0.4 respectively.

[0128] Step 3. Input the image features extracted by the all-sky multi-expert model into the fast hierarchical clustering algorithm, which relies on the indirect relationship of data points rather than the direct distance. While revealing the potential relationship of the data through the adjacency matrix, it also significantly reduces the computational complexity. Specifically, the algorithm regards each input feature image as a data point, obtains the nearest neighbor information of the data point through the fast approximate nearest neighbor method, and connects the nearest neighbor points based on this information to obtain the first layer of clustering results. In the subsequent iterative process, the features in each cluster are averaged to obtain a mean vector as the prototype of the cluster, and the outliers are filtered out to update the prototype, and then the same method is used to obtain the next layer of clustering results. The specific implementation steps are as follows:

[0129] (3.1) Obtain the nearest neighbor information of each sample point through the fast nearest neighbor method and generate a symmetric sparse adjacency matrix. The implementation is as follows:

[0130] ;

[0131] in: Represents sample points The first neighbor of Represents sample points The first neighboring point.

[0132] The formula shows that when the sample point for The first neighbor or sample points and Connect two nodes when they share the same nearest neighbor. Ensure the symmetry of the matrix.

[0133] (3.2) The adjacency matrix obtained by (3.1) Get the first-level clustering results, and calculate the average value of the sample points in each cluster as the initial cluster prototype, which represents the typical characteristics of the samples in the cluster;

[0134] (3.3) Filter outliers from the clusters. Common filtering strategies include masking the farthest point, masking by threshold, and masking by proportion. This paper adopts a threshold strategy based on data characteristics. The threshold comprehensively considers the standard deviation, IQR, and sample density of the full-sky aurora image. The implementation is as follows:

[0135] Based on the mean times the standard deviation to define the threshold scope:

[0136] ;

[0137] Setting thresholds based on IQR scope:

[0138] ;

[0139] Define an adjustment factor based on sample density , when the sample density within the cluster is low, the threshold range is appropriately expanded by adjusting the factor. In summary, the final threshold is:

[0140] ;

[0141] (3.4) Recalculate the mean of the clusters after outlier screening, update the prototype and use the prototype as the new sample point to continue the iterative clustering process. The clustering results obtained at each layer are screened for outliers. When the number of clusters is less than the specified number of clusters, the algorithm terminates.

[0142] Example:

[0143] The dataset used was obtained from winter observations from 2003-2004, using the All Sky Imaging (ASI) instrument at the Yellow River Station in the Arctic, China (geographic coordinates 78.92°N, 11.93°E, corrected magnetic latitude 76.24°N, MLT = UT+3.1h). 45,000 images were selected from these winter observations for the experiment.

[0144] During the feature extraction phase, to validate the superiority of our proposed all-sky multi-expert model in extracting auroral morphological features, we compared it with the classic residual networks ResNet18 (baseline) and ResNet34. To further verify the effectiveness of each component in the feature extraction phase, we conducted ablation experiments to analyze in detail the impact of two key operations, the dynamic convolutional attention module and the multi-expert network, on model performance. Furthermore, to investigate the impact of varying the number of experts in the multi-expert network on experimental results, we conducted ablation experiments with varying numbers of experts in the network to explore the optimal number of experts for the model. During the image clustering phase, we quantitatively analyzed the clustering results of hierarchical clustering at levels two to ten and visualized the results, revealing hidden features and patterns in auroral images and contributing to the improvement of auroral classification systems.

[0145] Through the above experimental design, we hope to rigorously demonstrate the effectiveness of our proposed approach compared to traditional methods, as well as the contribution of each component to the overall model performance. We will carefully analyze the experimental results and use visualization methods to intuitively demonstrate the advantages of our approach.

[0146] Experiment 1: Backbone network comparison to verify the superiority of the present invention

[0147] To validate the effectiveness of the proposed all-sky multi-expert model for feature extraction, we designed a classification task using 10,000 auroral images (5,000 of arc-shaped auroras and 5,000 of other auroral shapes). We compared the proposed all-sky multi-expert model with a classic skeleton network using four classification evaluation metrics: accuracy (ACC), precision, recall, and F1 score. Specifically, we first trained the encoder using 45,000 selected auroral images. The trained network was then used to extract features from the 10,000 images used in the classification task and to train the classification head. The comparison results are shown in Table 1.

[0148] Table 1 Backbone network comparison experiment results:

[0149]

[0150] As shown in the backbone network comparison experiments in Table 1, the all-sky multi-expert model outperforms traditional networks across all evaluation metrics. Specifically, the model achieves 97.80% accuracy, a 5.15% improvement over the baseline ResNet-18 model. It also significantly outperforms ResNet-18 in precision, recall, and F1 score. Compared to ResNet-34, the model achieves improvements of over 2% across all metrics. These experimental data demonstrate that our proposed all-sky multi-expert model can more effectively extract and utilize auroral image feature information than traditional methods, leading to superior results in subsequent tasks.

[0151] In order to more intuitively demonstrate the superiority of the model, we plotted the confusion matrix of the baseline network and the proposed model on the classification task. The confusion matrix can help us understand the performance of the model on different categories by showing the comparison between the actual category and the predicted category. Figure 4 As shown in the figure, the confusion matrix for the baseline network is on the left, and the confusion matrix for the all-sky multi-expert model is on the right. Using the same classification head, we can see that the proposed model increases the number of successful predictions for images belonging to class 0 from 869 to 971, and the number of successful predictions for images belonging to class 1 from 971 to 985. Overall, when the all-sky multi-expert model is used for feature extraction, the confusion matrix shows a significant increase in the number of correctly classified samples and a significant decrease in the number of misclassified samples. This demonstrates that our proposed network is able to learn the intrinsic structure of the data at a deeper level, capture the key features of auroral images, and generate more discriminative feature representations.

[0152] Experiment 2: Ablation experiment to verify the effectiveness of each component of the all-sky multi-expert model

[0153] To further verify the effectiveness of the two important components introduced in the proposed model, the dynamic convolutional attention module and the multi-expert network, in the feature extraction stage, an ablation experiment was designed to analyze in detail the independent performance of each component and its contribution to the overall performance of the model. The experimental results are shown in Table 2:

[0154] Table 2 Performance comparison results of each operation of the full-sky multi-expert model:

[0155]

[0156] Operation (i) indicates the introduction of the dynamic convolution attention module, and (ii) indicates the introduction of the multi-expert network. Experimental data show that after adding the dynamic convolution module, all classification evaluation indicators of the model are improved compared with the baseline model, especially the accuracy rate is improved by 4.26%, indicating that the dynamic convolution attention module can capture the key local detail features in the aurora image. After adding the multi-expert network, the model performance is further improved, indicating that through the collaborative work of multiple expert networks, the model can more comprehensively extract different features in the aurora image, and give higher weights to key features, further improving the model's representation ability. The all-sky multi-expert model proposed in this invention combines the advantages of the dynamic convolution attention module and the hybrid expert network to achieve the best performance.

[0157] Experiment 3: Comparative experiment, the impact of the number of experts

[0158] The sub-expert networks in the all-sky multi-expert model can capture the complex structure of auroral data from multiple perspectives, thereby improving the model's overall representational capabilities. However, how to effectively utilize individual experts and integrate their outputs remains a key challenge. Specifically, too few experts may lead to insufficient feature extraction and reduced model generalization, while too many experts may oversize the model and cause overfitting. Therefore, to further explore the impact of varying numbers of experts on model performance and identify an optimal number that balances model performance and complexity, we designed experiments to observe model performance using varying numbers of experts. The experimental results are shown in Table 3.

[0159] Table 3 Experimental results of the full-sky multi-expert model with different numbers of experts:

[0160]

[0161] The number of experts was set to 2, 3, 4, and 5 to observe the network's evaluation indicators for the classification task. The results showed that the setting of the number of experts had a significant impact on the experimental results, which means that different numbers of experts have different effectiveness in the feature extraction process. Specifically, when the number of experts was 2 and 3, the model performance improved slightly; when the number of experts was 4, the model performance improved significantly, with the accuracy, precision, and F1 score reaching their peaks; when the number of experts continued to increase to 5, all three indicators decreased slightly. This shows that the impact of the number of experts on model performance is nonlinear, and there is an optimal number of experts that can better balance feature extraction ability and model complexity. In the experiment, when the number of experts was 4, the model's feature extraction ability reached its optimal level, which can bring more accurate results for subsequent tasks.

[0162] Experiment 4: Comparative experiment, quantitative analysis of clustering results

[0163] The above experiments have fully demonstrated the superiority of the proposed all-sky multi-expert model. After extracting valid features, the feature information was clustered using our proposed fast hierarchical clustering algorithm, which can filter out outliers. After obtaining a complete hierarchical structure, we quantitatively analyzed the results from the second level (number of clusters: 2) to the tenth level (number of clusters: 10). The experimental results are shown in Table 4, with the top three best results for each clustering metric indicated in bold. The SC index represents the silhouette coefficient; a larger value indicates greater intracluster compactness; the CH index reflects intracluster compactness and intercluster separation; a larger value indicates better clustering; and the DB index reflects intercluster clustering; a lower value indicates better clustering.

[0164] Table 4 Clustering results of different cluster numbers on all-sky image data:

[0165]

[0166] Experimental data shows that all three metrics reach optimal values when the number of clusters is 3, indicating that the intra-class features are more compact and the inter-class features are more separated. The metrics for 2 clusters are just below the optimal value, indicating that this result can still represent the general distribution of auroral imagery features, but is not very detailed. The CH metric drops significantly when the number of clusters is 4, indicating that the inter-class separation is reduced and there is some overlap between clusters. In summary, the three-class classification better aligns with the distribution of auroral data.

[0167] Experiment 5: Visualization of clustering results

[0168] Intuitively display the pedigree relationship between the data obtained by hierarchical clustering, such as Figure 5As shown, the results demonstrate a clear hierarchy. When the number of clusters is large, the images are divided into finer subcategories, indicating that the fine-grained partitions at the bottom of the tree structure retain a higher level of purity. As the number of clusters gradually decreases, the results reflect the primary category distribution of aurora images, classifying images with visually similar morphological features into the same category, indicating that the coarse-grained partitions at the top reflect higher-level semantics.

[0169] The above simulation analysis proves the correctness and effectiveness of the method proposed in the invention.

[0170] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0171] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. An unsupervised classification method for auroral morphology based on a full-sky multi-expert model, characterized by: The following steps are involved: Step 1: Preprocess the aurora image: Step 1.1: Perform intensity scaling and dark current subtraction on the original aurora image with a size of 512×512 and a brightness value in the dynamic range of [0,17000]; Step 1.2: Create a circular mask with a radius of 220 pixels, with the zenith as the center. Multiply the mask with the original image to construct the full sky domain. Step 1.3: Perform image cropping and rotation on the full-sky aurora image to obtain an image with a size of 440 × 440 pixels and a grayscale value in the dynamic range of [0, 4000] after preprocessing and data enhancement. Step 2: Introduce the full-sky multi-expert model as the encoder; Step 2.1: Input the image after preprocessing and data enhancement in step 1 into the full-sky multi-expert model, and pass the obtained shallow features through the full-sky channel attention module; Step 2.2: The full-sky channel attention module performs full-sky maximum pooling and full-sky maximum pooling on the input features in the spatial dimension, and compresses the features into ; Step 2.3: Input the results of the full-sky max pooling and the full-sky max pooling into the parameter-sharing fully connected layer respectively, generate two attention vectors and add them together; Step 2.4: Normalize the added vector to obtain the final channel attention weight, and multiply it with the initial input feature to obtain the enhanced feature; Step 3: Get the final feature map: Step 3.1: Input the enhanced features obtained in step 2 into the spatial attention module and compress the feature map into ; Step 3.2: After the features are spliced on the channel, a dynamic snake convolution operation is performed to obtain the spatial attention weight, and the convolution kernel direction is dynamically adjusted according to the local morphology of the features; Step 3.3, multiply the weight by the initial feature map at the time of input to obtain the final feature map and input it to the next residual layer to continue extracting features; Step 4: Pass the features obtained by the last residual layer of the encoder through an average pooling layer, and then input the obtained feature information into multiple expert networks. The gating network sparsely activates some expert networks, and the final output is the weighted sum of the outputs of all expert networks; Step 5: Perform cluster analysis on the features extracted in step 4 using a fast hierarchical clustering algorithm; Step 5.1: Treat each input image feature as a data point, obtain the nearest neighbor information of the data point using the fast nearest neighbor method, and connect the nearest neighbor points with the help of the adjacency matrix to obtain the first-level clustering result; Step 5.2: Then use the threshold strategy to filter out outliers and average the features in each cluster to obtain a mean vector as the cluster prototype, and then use the same method to obtain the next layer of clustering results.

2. The unsupervised classification method for auroral morphology based on a full-sky multi-expert model according to claim 1, characterized in that: The multi-expert network includes multiple expert networks with independent parameters and a gating network that outputs a sparse vector .

3. The unsupervised classification method for auroral morphology based on a full-sky multi-expert model according to claim 1, characterized in that: The step (2) of China Railway's all-sky channel attention is carried out as follows: 1) First, the enhanced view passes through a convolutional layer and full-sky maximum pooling, and then passes through the first residual layer to obtain shallow features ; 2) The shallow features obtained In the spatial dimension, full sky domain maximum pooling and full sky domain average pooling are performed respectively to compress the features into , and then input the pooling results into the parameter-sharing fully connected layer to obtain two attention vectors and add them together. The result of the addition is passed through a sigmoid activation function to generate the final channel attention weight, which is combined with the initial feature After multiplication, the enhanced features are obtained , implemented as follows: ； ； in, Represents the input features The channel attention weights obtained after applying the full-sky channel attention, represents pixel-by-pixel multiplication, Represents the result obtained after full sky channel attention.

4. The unsupervised classification method for auroral morphology based on a full-sky multi-expert model according to claim 1, characterized in that: The spatial attention module in step (3) is implemented as follows: 1) The features obtained in step (2) Perform channel dimension maximum pooling and average pooling operations respectively to compress the features into , then concatenate the pooled results on the channel and extract local features through a dynamic snake convolution layer (DSConv); DSConv is implemented as follows: Given a standard Convolution kernel, the center coordinate is : ； Introducing deformation offset , ensuring that the convolution kernel conforms to the linear morphological structure; From the center coordinates Initially, the position away from the center depends on the position of the previous grid; for Axis direction, the change is expressed as: ； for Axis direction, the change is expressed as: ； 2) The result after the dynamic snake convolution layer is passed through a sigmoid activation function to generate the final spatial attention weight; The weight and feature After multiplication, the enhanced features are obtained ; Implementation is as follows: ； ； in, Represents the input feature map The spatial attention weights obtained after applying the spatial attention mechanism, represents pixel-by-pixel multiplication, Represents the result obtained after the spatial attention mechanism.

5. The unsupervised classification method for auroral morphology based on a full-sky multi-expert model according to claim 1, characterized in that: The multi-expert network in step (4) is implemented as follows: 1) The features After the subsequent residual layer and dynamic convolution attention module, the result is passed through an average pooling layer to obtain the feature ; 2) First, the gating network learns the correlation between the input features and each expert to obtain a trainable weight matrix ,Multiply the input feature with the weight to obtain the importance score of the current feature to each expert; 3) By adding a trainable weight matrix Introducing adjustable Gaussian noise, randomness of the gating mechanism, and then only retaining the top ones with the highest importance scores The scores are converted into probability distributions through a softmax function to provide a basis for the allocation of input features. The implementation is as follows: ； ； ； in, represents the output of the gating network, Indicates that the input feature pair The importance score of each expert network; 4) According to the gated network obtained in (3), the input features are assigned to the expert network to obtain the output of the final multi-expert network, which is implemented as follows: ； in, represents the output of the gating network, Indicates the The output of an expert network; 5) To prevent a few top experts from being overloaded and other experts from being idle, we add auxiliary loss-balancing experts to utilize the load rate, as follows: ； ； in, Indicates the calculation of importance value, Indicates the calculation of the coefficient of variation, which measures the degree of dispersion of the data. is a scaling factor that can be adjusted manually; 6) Linearly add the auxiliary loss and the contrast loss to obtain the total model loss. By continuously optimizing the loss function, the model learns discriminative feature representations, as follows: ； in, and Set to 0.6 and 0.4 respectively.

6. The unsupervised classification method of auroral morphology based on the all-sky multi-expert model according to claim 1, characterized in that: The steps of the fast hierarchical clustering algorithm in step (5) are as follows: 1) Treat each feature extracted by the feature enhancement network as a data point, obtain the nearest neighbor information of each sample point through the fast nearest neighbor method, and generate a symmetric sparse adjacency matrix, which is implemented as follows: ； in: Represents sample points The first neighbor of Represents sample points The first neighboring point of for The first neighbor or sample points and Connect two nodes when they share the same nearest neighbor. Ensure the symmetry of the matrix; 2) The adjacency matrix obtained by (1) Get the first-level clustering results, and calculate the average value of the sample points in each cluster as the initial cluster prototype, which represents the typical characteristics of the samples in the cluster; 3) Filter out outliers in the clusters. Common filtering strategies include masking the farthest point, masking by threshold, and masking by proportion. A threshold strategy is used based on the data characteristics. The threshold comprehensively considers the standard deviation, IQR, and sample density of the full-sky aurora image. The implementation is as follows: Based on the mean times the standard deviation to define the threshold scope: ； Setting thresholds based on IQR scope: ； Define an adjustment factor based on sample density When the sample density within the cluster is low, the threshold range is appropriately expanded by adjusting the factor. The final threshold is obtained from the above: ； 4) Recalculate the mean of the clusters after outlier screening, update the prototype and use it as a new sample point to continue the iterative clustering process. The clustering results obtained at each layer are filtered for outliers. When the number of clusters is less than the specified number of clusters, the algorithm terminates.

Images