An improved U-Net-based space-based airglow imaging plasma bubble semantic segmentation method and system

Abstract

The application belongs to the field of machine learning and space physical data processing, and relates to a space-based airglow imaging plasma bubble semantic segmentation method and system based on an improved U-Net, which solves the problems of difficulty in standardizing discrete peak electron density data and automatic segmentation of plasma bubble at the pixel level. The method fuses scanning data of the northern and southern hemispheres within an observation period, generates a two-dimensional gray image through latitude and longitude filtering, gridding interpolation and gray scale normalization; trains a U-Net model adopting hierarchical differentiated Dropout2d with a labeled image, and combines Dice and weighted BCE loss and data enhancement aiming at elastic deformation, dark strip blur and partial loss; inputs the to-be-tested data after preprocessing into the model to obtain a plasma bubble probability graph and a segmentation mask, and removes small-area false detection through connected domain analysis. The application is used for automatic identification and morphological analysis of plasma bubbles in space-based airglow imaging data.

Description

Technical Field

[0001] This invention pertains to the application research of machine learning technology in the field of space physics, specifically involving a semantic segmentation method and system for space-based airglow imaging plasma bubbles based on an improved U-Net. Background Technology

[0002] Equatorial Plasma Bubbles (EPBs) are large-scale plasma depletion structures that occur in the low-latitude F region of the ionosphere after sunset. Rayleigh-Taylor instability is considered the primary mechanism for their formation. These irregularities have significantly lower electron density than the background ionosphere, appearing as darker bubble-like structures in optical instrument data. EPBs cause rapid fluctuations in the amplitude and phase of radio signals passing through them, significantly impacting satellite communications and Global Navigation Satellite Systems (GNSS), and are considered one of the most important space weather features in the ionosphere.

[0003] The space-based far-ultraviolet airglow imaging observation platform operates in geosynchronous orbit and carries a far-ultraviolet imaging spectrometer, enabling continuous imaging observations of the entire Earth's disk. Its peak electron density data products provide the global-scale distribution of peak electron density in the F2 layer of the ionosphere, providing an unprecedented data foundation for global-scale plasma bubble research. Unlike traditional single-station ground-based all-sky airglow imagers that can only cover a limited area, this space-based observation platform can provide continuous observations covering a vast longitude sector.

[0004] However, the peak electron density data from space-based airglow imaging is not a standard image format that can be directly used for image segmentation. Instead, it is stored as discrete latitude-longitude electron density data points. Converting this unstructured discrete observation data into a standardized two-dimensional image suitable for deep learning model input while preserving the morphological features of the plasma bubble is a technical problem that has not yet been fully solved. Specific challenges include: (1) each observation cycle includes two scans of the Northern and Southern Hemispheres, requiring data fusion; (2) the data contains a large number of missing NaN values ​​and invalid observation points; (3) the original data points are spatially unevenly distributed, requiring effective interpolation strategies; and (4) maintaining the morphological integrity and spatial correspondence of the plasma bubble during the image generation process.

[0005] Existing EPB feature extraction methods mainly rely on manual operation, that is, identifying plasma bubble structures in airglow images by eye and manually drawing their outlines. This method is extremely inefficient when processing massive amounts of observation data. Considering that a single station in my country's Meridian Project generates approximately 1TB of data per year per channel, a single person would need to work continuously for tens of hours just for a rough screening. Furthermore, the results of manual identification are highly dependent on the operator's experience, resulting in unavoidable subjectivity. In addition, some researchers have attempted to automate extraction using traditional computer vision techniques such as threshold segmentation and edge detection. However, due to the blurred boundaries of plasma bubbles, low contrast with the background ionosphere, and complex and varied morphologies (elongated, wide, bifurcated, etc.), traditional methods have poor robustness and struggle to achieve stable and reliable detection results in practical applications.

[0006] In recent years, deep learning methods have made significant progress in the field of image analysis, but no mature solution has yet been formed for plasma bubble detection in space-based airglow imaging. The main technical bottlenecks are: (1) Space-based airglow imaging data needs to be specially preprocessed before it can be converted into standardized image input acceptable to deep learning models. Existing technical solutions lack this key step; (2) The cost of obtaining pixel-level labeled data for space-based airglow imaging is extremely high, and the number of available labeled samples is very limited. Existing semantic segmentation models mostly use large-scale backbone networks with a large number of parameters, which are prone to overfitting under small sample conditions; (3) Some solutions only focus on the presence or absence of image-level classification, and cannot achieve accurate pixel-level segmentation and plasma bubble morphology contour extraction; (4) Existing methods have not fully considered the unique physical characteristics of plasma bubbles - variable morphology, blurred boundaries, and low contrast with the background - and lack targeted data augmentation strategies. At the same time, no special loss function has been designed for the extreme class imbalance problem that plasma bubble pixel ratio is usually low.

[0007] Therefore, there is an urgent need for an automatic plasma bubble detection method that includes a complete preprocessing workflow for peak electron density data in space-based airglow imaging and can achieve pixel-level accurate segmentation under limited labeled sample conditions. Summary of the Invention

[0008] I. Technical problems to be solved

[0009] The purpose of this invention is to overcome the following shortcomings in the prior art: (1) lack of a complete preprocessing scheme for converting discrete peak electron density data of space-based airglow imaging into standardized images; (2) lack of a pixel-level semantic segmentation method for plasma bubbles in space-based airglow imaging data; (3) existing semantic segmentation models have a large number of parameters and are prone to overfitting under limited labeled samples; (4) do not fully consider the physical characteristics of plasma bubbles, such as variable morphology, blurred boundaries, and low contrast with the background, and lack corresponding data augmentation strategies; (5) do not design a special loss function for the extreme class imbalance caused by the low pixel ratio of plasma bubbles. Therefore, this invention provides a semantic segmentation method and system for plasma bubbles in space-based airglow imaging based on an improved U-Net.

[0010] II. Technical Solution

[0011] To address the aforementioned technical problems, this invention provides a semantic segmentation method and system for space-based gasglow imaging plasma bubbles based on an improved U-Net, comprising the following steps:

[0012] Step S1 involves acquiring peak electron density data from space-based far-ultraviolet gasglow imaging, fusing and preprocessing the Northern and Southern Hemisphere scan data for each observation period using gridded interpolation to generate a standardized two-dimensional grayscale image. This includes the following sub-steps:

[0013] S1.1 Load APEX latitude, APEX longitude, and peak electron density data for each observation period from the space-based airglow imaging L2 peak electron density data product. Each observation period consists of two scans of the Earth disk, covering the Northern and Southern Hemispheres respectively, and the data is stored in a matrix format across multiple time frames.

[0014] S1.2 Filter the data for the Northern and Southern Hemispheres based on geographic latitude. For the Northern Hemisphere data, only retain the APEX latitude, APEX longitude, and peak electron density values ​​for pixels with a latitude greater than -0.5 degrees. For the Southern Hemisphere data, only retain the corresponding data for pixels with a latitude less than 0.5 degrees. Simultaneously, remove all missing NaN values ​​from the peak electron density and latitude / longitude data.

[0015] S1.3 merges the filtered APEX latitude, APEX longitude, and peak electron density data of the Northern and Southern Hemispheres into a global data array, forming a global discrete data point set covering the entire observation area.

[0016] S1.4, the global data is cropped according to a preset geographical latitude range. Since the equatorial plasma bubble mainly occurs in the low latitude and equatorial regions, in a preferred embodiment, only data points with APEX latitudes between -25 degrees and 30 degrees are retained, while irrelevant data from mid- and high-latitude regions are removed.

[0017] S1.5 subtracts 360 degrees from the portion of APEX longitude values ​​greater than 180 degrees, and uniformly converts the entire longitude range to a continuous longitude representation from -180 degrees to 180 degrees, avoiding data discontinuity issues caused by longitude crossing the 180-degree line.

[0018] S1.6, the discrete peak electron density data points are interpolated onto a regular grid using the griddata interpolation algorithm (V4 method in a preferred embodiment). The grid resolution is 0.5 degrees multiplied by 0.5 degrees. The interpolation process generates a continuous two-dimensional matrix of peak electron density covering the target region.

[0019] S1.7 converts the interpolated peak electron density two-dimensional matrix into floating-point values ​​in the range of [0,1] through gray-level normalization, and then into an 8-bit grayscale image (0-255 integer values) as the standardized input for subsequent deep learning models.

[0020] Step S2 involves manually annotating the preprocessed grayscale image at the pixel level, labeling pixels in the plasma bubble region as 1 and pixels in the background region as 0, thus obtaining 0-1 binary mask labels. The open-source annotation software LabelMe is used to draw polygons along the edges of the dark areas of the plasma bubble, and a annotation conversion script converts the JSON format annotation file into a PNG format binary mask image. The annotation data is then randomly divided into training, validation, and test sets according to a set ratio.

[0021] Step S3: Construct a U-Net semantic segmentation model with hierarchical regularization. The model employs a four-level encoder-decoder symmetric structure. Each convolutional block includes BatchNorm batch normalization and Dropout2d spatial regularization, with the Dropout rate of the decoder layer being higher than that of the corresponding encoder layer. The convolutional layers are set to bias=False to cooperate with BatchNorm. The U-Net architecture itself exhibits excellent performance under small sample conditions. Its encoder-decoder symmetric structure and skip connection design can fully utilize pixel-level information in limited labeled samples to achieve accurate localization of plasma bubble morphology.

[0022] The specific structure of the model is as follows: The encoder part consists of four levels: Level 1 (channels 1-32), Level 2 (channels 32-64), Level 3 (channels 64-128), and Level 4 (channels 128-256). Each level contains two 3x3 convolutional layers, followed by a BatchNorm2d layer, a ReLU activation layer, and a Dropout2d layer. The Dropout rate for each level of the encoder is 0.25. Downsampling is performed between levels using 2x2 max pooling. The bottleneck layer contains two 3x3 convolutional layers with channels 256-512, and the Dropout rate is 0.25. The decoder part consists of four levels of upsampling. Each level includes upsampling using 2x2 transposed convolutions, concatenation of skip connection features from the corresponding encoding level along the channel dimension, and a convolutional block with regularization. The fourth decoding stage has 512 to 256 channels with a Dropout rate of 0.35; the third decoding stage has 256 to 128 channels with a Dropout rate of 0.35; the second decoding stage has 128 to 64 channels with a Dropout rate of 0.25; and the first decoding stage has 64 to 32 channels with a Dropout rate of 0.25. The dropout rate of the deep decoder (0.35) is higher than that of the corresponding layer of the encoder and the shallow decoder (0.25). This design is based on the fact that the deep decoder receives the richest feature fusion information, has the highest risk of overfitting, and requires stronger regularization constraints. The output layer contains a 1x1 convolution that maps 32 channels to 2 channels (corresponding to the background category and the plasma bubble category, respectively), followed by a layer with a Dropout2d rate of 0.1. The model weights are initialized using the Kaiming normal distribution (mode is fan_out, nonlinearity is ReLU), and the weights of the BatchNorm layer are initialized to 1, while the biases are initialized to 0. The total number of model parameters is approximately 31M.

[0023] Step S4: Using the Dice and weighted BCE combined loss function, along with a domain-specific data augmentation strategy, the AdamW optimizer is used to train the semantic segmentation model. During the training process, a ReduceLROnPlateau learning rate scheduling strategy based on validation loss and an early stopping mechanism are adopted to obtain the trained plasma bubble semantic segmentation model.

[0024] The formula for calculating the combined loss function of Dice and weighted BCE is: L equals alpha multiplied by (1 minus Dice) plus (1 minus alpha) multiplied by BCE_pos. Where Dice equals 2 multiplied by sigma(p_i multiplied by t_i) plus epsilon divided by sigma(p_i) plus sigma(t_i) plus epsilon, where p_i is the model's predicted probability that the i-th pixel belongs to the plasma bubble category (after sigmoid activation), t_i is the true label of the i-th pixel, and epsilon is a smoothing coefficient of 1e-6. BCE_pos is a binary cross-entropy loss with positive class weights, and the positive class weight pos_weight is 1.5. This design allows the Dice loss to optimize the segmentation result from a macroscopic perspective of region overlap, while the weighted BCE loss optimizes from a microscopic perspective of pixel-by-pixel classification accuracy. The two complement each other, effectively addressing the class imbalance problem caused by the extremely low proportion of plasma bubble pixels. alpha is the weight coefficient of the Dice loss, set to 0.6.

[0025] The domain-specific data augmentation strategies include: (a) Geometric transformation enhancement—random horizontal flip, random vertical flip, random 90-degree rotation, and translational scaling rotation (shift_limit=0.1, scale_limit=0.2, rotate_limit=30 degrees); (b) Elastic deformation enhancement—using ElasticTransform with parameters alpha=1, sigma=50, and alpha_affine=50 to apply random elastic deformation to the image, simulating the natural deformation of the plasma bubble morphology; (c) Dark bar blur simulation—for the masked area, the original image and a Gaussian blur image with a random kernel size of 3, 5, or 7 are weighted and mixed with random intensity in the range of 0.3 to 0.7 to simulate the edge blurring effect of the plasma bubble; (d) Dark bar partial missing simulation—a circular removal area is generated at a random position within the masked area with a probability of 0.2, and the pixels within the removal area are marked as background to simulate the situation where the plasma bubble structure is incomplete or occluded. It also includes random brightness and contrast adjustment, adaptive histogram equalization (CLAHE), Gaussian noise, ISO noise simulation, and photoelectric transformation enhancements such as mesh / optical distortion. Through these enhancement strategies, the diversity of training samples can be effectively expanded using only a limited number of original labeled images (e.g., hundreds), significantly improving the model's generalization ability.

[0026] The training configuration includes: the optimizer is AdamW, the initial learning rate is lr=1e-4, and the weight decay is weight_decay=1e-4; the learning rate scheduling uses ReduceLROnPlateau, the monitoring and validation loss is min, the factor is 0.5, and the patience is 5; the gradient clipping is max_norm=1.0; the patience value of the early stopping mechanism is 15 epochs (if the validation Dice coefficient does not show a new optimal value within 15 consecutive epochs, then training is terminated); the batch size is 2 to 8 depending on the hardware configuration.

[0027] In one specific embodiment, training was performed using approximately 480 labeled images, achieving an optimal validation Dice coefficient of 0.7660 and a validation loss of 0.2379. The total training time was approximately 48 minutes (single GPU, 100 epochs).

[0028] Step S5: After preprocessing the peak electron density data of the space-based airglow imaging to be detected into a grayscale image as described in step S1, the image is input into the trained plasma bubble semantic segmentation model. The model outputs a 2-channel logits tensor. Channel 1 (plasma bubble category) is activated by Sigmoid to obtain a probability map. The probability map is binarized using a threshold of 0.5 to obtain an initial segmentation mask. Connectivity analysis is performed on the binarized mask to filter out isolated detection regions with an area less than 50 pixels to eliminate false detections caused by noise. For the retained connected components, the area, centroid coordinates, and average confidence of each region are calculated.

[0029] III. Beneficial Effects

[0030] (1) For the first time, a complete preprocessing scheme for peak electron density data in space-based airglow imaging was provided, solving the technical conversion problem from discrete scan data to standardized deep learning input images. Through key steps such as data fusion between the Northern and Southern Hemispheres, latitude and longitude filtering, griddata interpolation, and grayscale normalization, the scheme efficiently transforms unstructured satellite observation data into standardized images that retain the morphological characteristics of plasma bubbles, filling the gap in this aspect of existing technology.

[0031] (2) Deep learning semantic segmentation technology was successfully applied to the detection of plasma bubbles in space-based airglow imaging, avoiding the inefficiency and subjectivity of traditional manual identification methods when dealing with massive amounts of observation data, and realizing a technological leap from ground-based to space-based and from manual identification to automated pixel-level segmentation.

[0032] (3) Using U-Net as the basic architecture, its encoder-decoder symmetric structure and skip connection design are naturally suitable for pixel-level accurate segmentation under small sample conditions. Only a few hundred labeled images are needed to complete effective training, which greatly reduces the cost of data labeling.

[0033] (4) A hierarchical differential regularization strategy was constructed. Through the encoder-decoder differential Dropout2d design (0.25 for each level of the encoder, 0.25 for the bottleneck layer, 0.25 for the shallow decoder, and 0.35 for the deep decoder), stronger regularization was applied to the deep decoder with the richest network feature fusion, which effectively suppressed the risk of overfitting under small sample conditions, while maintaining the model's ability to extract plasma bubble morphology features. Combined with Kaiming initialization and bias=False convolution design, the stability and convergence speed of deep network training were ensured.

[0034] (5) A combined loss function of Dice and weighted BCE was designed. The Dice loss optimizes the segmentation results from the perspective of region overlap and is insensitive to class imbalance; the weighted BCE loss enhances the classification attention to plasma bubble pixels through positive class weights (pos_weight=1.5) and optimizes from the perspective of pixel-by-pixel accuracy. The two complement each other and effectively solve the training bias problem caused by the usually very low proportion of plasma bubble pixels.

[0035] (6) Domain-specific data augmentation strategies for the physical properties of plasma bubbles are proposed, including elastic deformation enhancement (simulating the natural deformation of plasma bubbles under different ionospheric conditions), dark stripe blurring simulation (simulating the gradual effect of edge contrast), and dark stripe partial missing simulation (simulating the situation of incomplete structure or occlusion), which significantly improves the robustness of the model in detecting dark and weak plasma bubbles with varied morphology and blurred boundaries. Attached Figure Description

[0036] Figure 1 This is a flowchart of a space-based airglow imaging plasma bubble semantic segmentation method and system based on an improved U-Net.

[0037] Figure 2 This is a detailed schematic diagram of the GOLD space-based airglow imaging L2 peak electron density (Nmax) data preprocessing process in step S1 of the present invention.

[0038] Figure 3 This is a schematic diagram of the network structure of the U-Net semantic segmentation model with hierarchical regularization of the present invention.

[0039] Figure 4 This is a schematic diagram illustrating the effect of the domain-specific data augmentation strategy in this invention.

[0040] Figure 5 This is a training curve diagram of the Regularized U-Net model of this invention, including the training loss, validation loss, Dice coefficient, and learning rate change curves.

[0041] Figure 6This is a schematic diagram of the plasma bubble semantic segmentation system of the present invention.

[0042] Figure 7 This is an example image of the plasma bubble semantic segmentation results based on GOLD space-based airglow imaging data, including a four-column comparison of the original image, the true mask, the predicted probability map, and the binary segmentation results. Detailed Implementation

[0043] The technical solutions provided by the present invention will be further illustrated below with reference to the embodiments.

[0044] The inventors of this invention noted that the U-Net architecture performs exceptionally well in small-sample semantic segmentation tasks such as biomedical image segmentation, and its encoder-decoder symmetric structure and skip connection design are particularly suitable for segmentation tasks requiring precise localization. The morphology extraction of plasma bubbles in space-based airglow imaging images shares similar characteristics with medical image segmentation—variable target morphologies, blurred boundaries, and a limited number of available labeled samples. Therefore, this invention features proprietary improvements to the U-Net architecture, constructing a Regularized U-Net model with hierarchical regularization. Simultaneously, the inventors designed a specific preprocessing pipeline to bridge the format differences between the raw observation data and the deep learning model, specifically addressing the unique storage format of peak electron density data from space-based airglow imaging (discrete latitude and longitude-electron density data points).

[0045] Example 1: A semantic segmentation method and system for space-based gasglow imaging plasma bubbles based on an improved U-Net

[0046] This embodiment provides a semantic segmentation method and system for space-based airglow imaging plasma bubbles based on an improved U-Net, including the following steps:

[0047] Step 1: Preprocessing of peak electron density data for space-based gasglow imaging

[0048] L2 peak electron density data products were acquired from a space-based observation data platform. This data product is stored in MAT file format, with each file corresponding to one observation period. Each observation period contains multiple time frames, and each time frame contains scan data from both the Northern and Southern Hemispheres. Each frame of data is a discrete set of latitude / longitude-electron density data points. The preprocessing workflow is as follows:

[0049] (1) Data loading and hemisphere separation: For each observation period, two adjacent frames are extracted as a pair of Northern and Southern Hemisphere scan data. The frame index corresponding to the Northern and Southern Hemispheres is determined based on the hemisphere tag. APEX latitude (APEXla), APEX longitude (APEXlon), and peak electron density data matrices are extracted from the corresponding frames.

[0050] (2) Geographic latitude filtering: Remove pixels with a geographic latitude less than -0.5 degrees from the Northern Hemisphere data; remove pixels with a geographic latitude greater than 0.5 degrees from the Southern Hemisphere data. This step ensures that only valid data of the actual coverage area of ​​each hemisphere is retained.

[0051] (3) NaN value cleaning: Data points with peak electron density of NaN, data points with APEX latitude of NaN, and data points with APEX longitude of NaN are removed in sequence to ensure that all data points used for interpolation are valid values.

[0052] (4) Data fusion between the Northern and Southern Hemispheres: The filtered APEX latitude, APEX longitude and peak electron density data of the Northern and Southern Hemispheres are spliced ​​together along the horizontal direction to form a set of discrete data points covering the target area of ​​the world.

[0053] (5) Geographic latitude clipping: Considering that equatorial plasma bubbles mainly occur in low latitudes and the equatorial region, the target latitude range is set to 30 degrees north latitude to 25 degrees south latitude (i.e., APEX latitude: -25 degrees to 30 degrees). Only data points falling within this range are retained, and irrelevant data from mid- and high latitudes are removed.

[0054] (6) Uniform longitude range: Subtract 360 degrees from the portion of APEX longitude values ​​greater than 180 degrees, and uniformly convert the longitude to a continuous range of -180 degrees to 180 degrees.

[0055] (7) Grid interpolation: A regular latitude and longitude grid is constructed with a resolution of 0.5 degrees. The V4 method of the griddata interpolation algorithm is used to interpolate the discrete peak electron density data points onto the regular grid to obtain a continuous two-dimensional distribution matrix of peak electron density.

[0056] (8) Grayscale image generation: The mat2gray function is applied to the interpolation matrix to perform grayscale normalization, and the data is linearly mapped to the range of [0,1]. Then, it is converted into an 8-bit grayscale image in the range of 0 to 255 by im2uint8 and saved as a JPG format.

[0057] Step 2: Dataset Creation

[0058] Based on the preprocessed grayscale images and combined with domain expert knowledge, image samples containing plasma bubble structures were selected. Training samples needed to cover plasma bubbles of different spatial scales and morphologies, while also including negative samples without plasma bubbles. Pixel-level manual annotation was performed using LabelMe, with the label being "stripe." After annotation, the images were converted into 0-1 binary PNG mask images. The images were then divided into training, validation, and test sets proportionally.

[0059] Step 3: Building the U-Net semantic segmentation model with hierarchical regularization

[0060] The model employs a four-level encoder-decoder symmetric structure. The encoder's convolutional block structure is as follows: 3x3 convolution (bias=False) to BatchNorm2d to ReLU to Dropout2d (p=0.25) to 3x3 convolution (bias=False) to BatchNorm2d to ReLU. Channel configuration: 1 to 32 to 64 to 128 to 256. Each level is downsampled using 2x2 max pooling. Bottleneck layer: 256 to 512 channels, Dropout rate 0.25. Each level of the decoder is upsampled using 2x2 transposed convolutions and then concatenated with the corresponding encoder level via skip connections. Channel and Dropout rate: 512 to 256 (0.35) to 128 (0.35) to 64 (0.25) to 32 (0.25). Output layer: 1x1 convolution with 32 to 2 channels, followed by Dropout2d (0.1). All convolutional layers have bias=False. Initialize using a Kaiming normal distribution, and initialize BatchNorm weights to 1 and biases to 0. The total number of parameters is approximately 31M.

[0061] Step 4: Model Training

[0062] Data augmentation: geometric transformations (flip, rotate, translate, scale, rotate), elastic deformation (alpha=1, sigma=50, alpha_affine=50, p=0.3), dark bar blur simulation (intensity 0.3 to 0.7), dark bar partial missing simulation (p=0.2), as well as brightness and contrast adjustment, CLAHE, Gaussian / ISO noise, mesh / optical distortion.

[0063] Loss function: Dice + weighted BCE combined loss, alpha=0.6 (Dice accounts for 60%), pos_weight=1.5.

[0064] Training parameters: AdamW optimizer, lr=1e-4, weight_decay=1e-4; ReduceLROnPlateau scheduler (mode=min, factor=0.5, patience=5); gradient clipping max_norm=1.0; early stopping patience=15; batch size 8 (GPU) or 2 (CPU).

[0065] Training results: After training on approximately 480 labeled images for 100 epochs, the optimal validation Dice coefficient was 0.7660, and the validation loss was 0.2379. The learning rate decreased from 1e-4 to 1.5625e-6 in four iterations, and the curve converged well without overfitting.

[0066] Step 5: Model Inference and Post-processing

[0067] The data to be detected is preprocessed according to step 1 to generate a 256x256 grayscale image, which is then input into the model to obtain 2-channel logits. Channel 1 is processed using a Sigmoid function to obtain a probability map, and then binarized with a threshold of 0.5. Connectivity analysis is used to filter out isolated regions with an area less than 50 pixels. The area, centroid coordinates, and average confidence score of the retained regions are calculated.

[0068] Example 2: Semantic Segmentation System for Space-Based Gasglow Imaging Plasma Bubbles Based on Improved U-Net

[0069] This embodiment provides a semantic segmentation system for plasma bubbles in space-based airglow imaging based on an improved U-Net, comprising: a data preprocessing module for acquiring peak electron density data from space-based far-ultraviolet airglow imaging, fusing, filtering by latitude and longitude, performing griddata interpolation, and normalizing grayscale data from the Northern and Southern Hemispheres for each observation period to generate standardized two-dimensional grayscale images; a dataset creation module for pixel-level annotation of the preprocessed grayscale images, generating 0-1 binary mask labels, and dividing them into training, validation, and test sets proportionally; a model building and training module for constructing a U-Net semantic segmentation model with hierarchical regularization, employing a Dice and weighted BCE combined loss function and a domain-specific data augmentation strategy, and training it using the AdamW optimizer in conjunction with the ReduceLROnPlateau scheduling strategy and an early stopping mechanism; and a model inference module for loading the trained model weights, inputting the preprocessed data to be detected into the model for forward inference, outputting a probability map and binary segmentation mask of the plasma bubble region, and performing connected component analysis and area filtering.

[0070] Example 3: Comparative Experiment of Different Regularization Strategies

[0071] To verify the effectiveness of the hierarchical differential Dropout strategy, comparisons were made under the same dataset and training configuration: Basic U-Net (without regularization) achieved a Dice of approximately 0.62; adding unified Dropout=0.3 resulted in a Dice of approximately 0.70; using hierarchical Dropout (encoder 0.25 + decoder 0.35) resulted in a Dice of approximately 0.74; the complete solution of this invention (hierarchical Dropout + Kaiming initialization + bias=False + combined loss) achieved a Dice of 0.766. Hierarchical Dropout improves performance by approximately 4 percentage points compared to unified Dropout, and the complete solution improves performance by over 14 percentage points compared to basic U-Net, verifying the synergistic effectiveness of the improved components.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantic segmentation method for plasma bubbles in space-based airglow imaging based on an improved U-Net, comprising: Step S1: Collect peak electron density data from space-based far-ultraviolet gasglow imaging. Perform fusion and gridded interpolation preprocessing on the scanning data from the Northern and Southern Hemispheres for each observation cycle to generate a standardized two-dimensional grayscale image. Step S2: Perform pixel-level manual annotation on the preprocessed grayscale image, labeling pixels in the plasma bubble region as 1 and pixels in the background region as 0, obtaining 0-1 binary mask labels, and dividing the dataset into training, validation, and test sets. Step S3: Construct a U-Net semantic segmentation model with hierarchical regularization. The model adopts a four-level encoder-decoder structure, where each convolutional block includes BatchNorm batch normalization and Dropout2d spatial regularization, and the Dropout rate of the decoder layer is higher than that of the encoder. In step S4, the semantic segmentation model is trained using the AdamW optimizer with a combined loss function of Dice and weighted BCE, along with a domain-specific data augmentation strategy. During training, a ReduceLROnPlateau learning rate scheduling strategy based on validation loss and an early stopping mechanism are used to obtain the trained plasma bubble semantic segmentation model. In step S5, the peak electron density data of the space-based airglow imaging to be detected is preprocessed as described in step S1 to generate a grayscale image, which is then input into the trained plasma bubble semantic segmentation model. The output is a probability map of the plasma bubble region, and the segmentation mask of the plasma bubble is obtained by threshold binarization.

2. The method according to claim 1, characterized in that, The preprocessing described in step S1 includes the following sub-steps: S1.1, loading APEX latitude, APEX longitude, and peak electron density data for each observation period from the L2 peak electron density data product of space-based airglow imaging; S1.2, filtering the data of the Northern and Southern Hemispheres according to the geographical latitude range, retaining only valid pixels with geographical latitude greater than -0.5 degrees in the Northern Hemisphere and geographical latitude less than 0.5 degrees in the Southern Hemisphere, and removing missing NaN values ​​in the peak electron density and latitude / longitude data; S1.3, merge the filtered APEX latitude, APEX longitude and peak electron density data of the Northern and Southern Hemispheres into a global data array respectively; S1.4, trim the global data according to the preset geographical latitude range, retaining the area of ​​APEX latitude between -25 degrees and 30 degrees. S1.5, Subtract 360 degrees from the portion of APEX longitude values ​​greater than 180 degrees, converting them to a longitude range of -180 degrees to 180 degrees; S1.6, Use the griddata interpolation algorithm to interpolate discrete peak electron density data points onto a regular grid, with a grid resolution of 0.5 degrees multiplied by 0.5 degrees; S1.7, Convert the interpolated peak electron density two-dimensional matrix into an 8-bit grayscale image through grayscale normalization, serving as the standardized input for subsequent deep learning models.

3. The method according to claim 1, characterized in that, The U-Net semantic segmentation model with hierarchical regularization described in step S3 is as follows: The encoder part contains four levels: Level 1 (1 to 32 channels), Level 2 (32 to 64 channels), Level 3 (64 to 128 channels), and Level 4 (128 to 256 channels). Each level contains two 3x3 convolutional layers, followed by a BatchNorm2d layer, a ReLU activation layer, and a Dropout2d layer. The Dropout rate for each level of the encoder is 0.

25. The bottleneck layer contains two 3x3 convolutional layers with 256 to 512 channels, and the Dropout rate is 0.

25. The decoder part contains four levels of upsampling: Level 4 (decoding level 5...). The model has 12 to 256 channels with a Dropout rate of 0.35; the third decoding stage has 256 to 128 channels with a Dropout rate of 0.35; the second decoding stage has 128 to 64 channels with a Dropout rate of 0.25; and the first decoding stage has 64 to 32 channels with a Dropout rate of 0.

25. The output layer contains a 1x1 convolution that maps 32 channels to 2 channels, corresponding to the background category and the plasma bubble category, respectively, followed by a layer with a Dropout2d rate of 0.

1. The weights of the model are initialized using the Kaiming normal distribution, and the weights of the BatchNorm layer are initialized to 1 and the biases are initialized to 0.

4. The method according to claim 1, characterized in that, The formula for calculating the combined loss function of Dice and weighted BCE in step S4 is: L = alpha * (1 - Dice) + (1 - alpha) * BCE_pos, where Dice = (2 * sigma(p_i * t_i) + epsilon) / (sigma(p_i) + sigma(t_i) + epsilon), p_i is the model's predicted probability that the i-th pixel belongs to the plasma bubble category, t_i is the true label of the i-th pixel, and epsilon is the smoothing coefficient, which is 1e-6; BCE_pos is the binary cross-entropy loss with positive class weights, and the positive class weight pos_weight is 1.5; alpha is the weight coefficient of the Dice loss, which ranges from 0.5 to 0.

6.

5. The method according to claim 1, characterized in that, The domain-specific data augmentation strategy described in step S4 includes at least one of the following: (a) Geometric transformation enhancement: random horizontal flip, random vertical flip, random 90-degree rotation, and translation-scaling rotation, wherein the translation-scaling rotation parameters are shift_limit=0.1, scale_limit=0.2, and rotate_limit=30 degrees; (b) Elastic deformation enhancement: using ElasticTransform with parameters alpha=1, sigma=50, and alpha_affine=50 to simulate the natural deformation of the plasma bubble morphology; (c) Dark bar blur simulation: for the masked area, the original image and a Gaussian blur image with a random kernel size of 3, 5, or 7 are weighted and mixed with random intensity in the range of 0.3 to 0.7 to simulate the edge blur effect of the plasma bubble; (d) Dark bar partial missing simulation: a circular removal area is generated at a random position within the masked area with a preset probability, and the pixels within the removal area are marked as background to simulate the case of an incomplete plasma bubble structure.

6. The method according to claim 1, characterized in that, The training process in step S4 further includes: pruning the gradient, with max_norm=1.0; the patience value of the early stopping mechanism is 15 epochs; the parameters of the ReduceLROnPlateau learning rate scheduling strategy are mode=min, factor=0.5, and patience=5; the initial learning rate of the optimizer AdamW is 1e-4, and the weight decay is 1e-4.

7. The method according to claim 1, characterized in that, Step S5 further includes: performing connected component analysis on the binarized mask, filtering out isolated detection regions with an area smaller than a preset threshold, and calculating the coverage, centroid coordinates, and average confidence of the retained regions.

8. A semantic segmentation system for space-based gasglow imaging plasma bubbles based on an improved U-Net, comprising: The data preprocessing module is used to collect peak electron density data from space-based far-ultraviolet gasglow imaging, and to fuse, filter latitude and longitude, perform gridded interpolation and grayscale normalization on the scanning data of the northern and southern hemispheres for each observation period to generate a standardized two-dimensional grayscale image. The dataset creation module is used to perform pixel-level annotation on the preprocessed grayscale images, generate 0-1 binary mask labels, and divide them into training set, validation set and test set; The model building and training module is used to build a U-Net semantic segmentation model with hierarchical regularization. It adopts a combination loss function of Dice and weighted BCE and a domain-specific data augmentation strategy. It is trained with the AdamW optimizer in conjunction with the ReduceLROnPlateau scheduling strategy and early stopping mechanism. The model inference module is used to input the preprocessed peak electron density data of the space-based airglow imaging to be detected into the trained model, and output the probability map of the plasma bubble region and the binary segmentation mask.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images