Average Classification Accuracy Research Articles

Online BCI systems: cross-subject motor imagery classification based on weighted time-domain feature extraction methods

Motor imagery electroencephalogram (MI-EEG) is becoming increasingly important. This paper solves the problem of online signal recognition for motor imagery across subjects by finding common features across multiple subjects to improve the generality of the classification model. We analysed the EEG data from left/right-hand motor imagery of eight subjects and proposed a weighted time-domain (WTD) feature extraction method based on a weighted channel screening method. The classification model constructed by combining this feature extraction method with the support vector machine (SVM) classification method was faster in classification and achieved good cross-subject classification accuracy (The average offline classification accuracy was 91.39%). In this paper, an online control system for asynchronous brain-controlled wheelchairs was built with good performance. The online average motor imagery classification accuracy was 81.67%, and the average response time was 1.36s. This method contributes to bringing the online Brain-computer interface (BCI) system out of the laboratory and into wider application.

Rotating Machinery Fault Diagnosis with Limited Multisensor Fusion Samples by Fused Attention-Guided Wasserstein GAN

As vital equipment in modern industry, the health state of rotating machinery influences the production process and equipment safety. However, rotating machinery generally operates in a normal state most of the time, which results in limited fault data, thus greatly constraining the performance of intelligent fault diagnosis methods. To solve this problem, this paper proposes a novel fault diagnosis method for rotating machinery with limited multisensor fusion samples based on the fused attention-guided Wasserstein generative adversarial network (WGAN). Firstly, the dimensionality of collected multisensor data is reduced to three channels by principal component analysis, and then the one-dimensional data of each channel are converted into a two-dimensional pixel matrix, of which the RGB images are obtained by fusing the three-channel two-dimensional images. Subsequently, the limited RGB samples are augmented to obtain sufficient samples utilizing the fused attention-guided WGAN combined with the gradient penalty (FAWGAN-GP) method. Lastly, the augmented samples are applied to train a residual convolutional neural network for fault diagnosis. The effectiveness of the proposed method is demonstrated by two case studies. When training samples per class are 50, 35, 25, and 15 on the KAT-bearing dataset, the average classification accuracy is 99.9%, 99.65%, 99.6%, and 98.7%, respectively. Meanwhile, the methods of multisensor fusion and the fused attention mechanism have an average improvement of 1.51% and 1.09%, respectively, by ablation experiments on the WT gearbox dataset.

Quantifying the accuracy of inter-beat intervals acquired from consumer-grade photoplethysmography wristbands using an electrocardiogram-aided information-based similarity approach

Objective. Although inter-beat intervals (IBI) and the derived heart rate variability (HRV) can be acquired through consumer-grade photoplethysmography (PPG) wristbands and have been applied in a variety of physiological and psychophysiological conditions, their accuracy is still unsatisfactory. Approach. In this study, 30 healthy participants concurrently wore two wristbands (E4 and Honor 5) and a gold-standard electrocardiogram (ECG) device under four conditions: resting, deep breathing with a frequency of 0.17 Hz and 0.1 Hz, and mental stress tasks. To quantitatively validate the accuracy of IBI acquired from PPG wristbands, this study proposed to apply an information-based similarity (IBS) approach to quantify the pattern similarity of the underlying dynamical temporal structures embedded in IBI time series simultaneously recorded using PPG wristbands and the ECG system. The occurrence frequency of basic patterns and their rankings were analyzed to calculate the IBS distance from gold-standard IBI, and to further calculate the signal-to-noise ratio (SNR) of the wristband IBI time series. Main results. The accuracies of both HRV and mental state classification were not satisfactory due to the low SNR in the wristband IBI. However, by rejecting data segments of SNR < 25, the Pearson correlation coefficients between the wristbands' HRV and the gold-standard HRV were increased from 0.542 ± 0.235 to 0.922 ± 0.120 for E4 and from 0.596 ± 0.227 to 0.859 ± 0.145 for Honor 5. The average accuracy of four-class mental state classification increased from 77.3% to 81.9% for E4 and from 79.3% to 83.3% for Honor 5. Significance. Consumer-grade PPG wristbands are acceptable for HR and HRV monitoring when removing low SNR segments. The proposed method can be applied for quantifying the accuracies of IBI and HRV indices acquired via any non-ECG system.

Seizure prediction based on improved vision transformer model for EEG channel optimization

Epileptic seizures are unpredictable events caused by abnormal discharges of a patient’s brain cells. Extensive research has been conducted to develop seizure prediction algorithms based on long-term continuous electroencephalogram (EEG) signals. This paper describes a patient-specific seizure prediction method that can serve as a basis for the design of lightweight, wearable and effective seizure-prediction devices. We aim to achieve two objectives using this method. The first aim is to extract robust feature representations from multichannel EEG signals, and the second aim is to reduce the number of channels used for prediction by selecting an optimal set of channels from multichannel EEG signals while ensuring good prediction performance. We design a seizure-prediction algorithm based on a vision transformer (ViT) model. The algorithm selects channels that play a key role in seizure prediction from 22 channels of EEG signals. First, we perform a time-frequency analysis of processed time-series signals to obtain EEG spectrograms. We then segment the spectrograms of multiple channels into many non-overlapping patches of the same size, which are input into the channel selection layer of the proposed model, named Sel-JPM-ViT, enabling it to select channels. Application of the Sel-JPM-ViT model to the Boston Children’s Hospital–Massachusetts Institute of Technology scalp EEG dataset yields results using only three to six channels of EEG signals that are slightly better that the results obtained using 22 channels of EEG signals. Overall, the Sel-JPM-ViT model exhibits an average classification accuracy of 93.65%, an average sensitivity of 94.70% and an average specificity of 92.78%.

Deep transfer learning model for brain tumor segmentation and classification using UNet and chopped VGGNet

Brain tumors (BT) are a leading cause of cancer-related mortality worldwide, underscoring the critical need for early and precise detection to improve patient survival rates. Computer-aided diagnosis (CAD) plays a pivotal role in early BT detection by providing medical experts with valuable information image analysis. Various researchers have developed distinct methodologies, drawing from both machine and deep learning approaches. ML relies on manual feature analysis, which entails a time-intensive procedure of selecting an optimal feature extractor and necessitates domain experts with a deep understanding of feature selection. Conversely, deep learning methods exhibit superior performance compared to ML owing to their end-to-end, automated, high-level, and robust attribute mining capabilities. In this study introduced an innovative two-stage framework designed for the automatic classification of BT. In the initial stage, utilize U-Net models to conduct BT segmentation as part of the pre-processing step. Subsequently, in the second stage, utilize the improved BT images as input for a transfer learning-based model known as visual geometry group neural network (VGGNet), which excels in BT classification. The experimental analysis shows that the proposed approach has reported the average classification accuracy as 98.6%, 98.76%, and 99.45% for Meningioma, Glioma, and Pituitary BTs, respectively.

Multi-channel high-order network representation learning research.

The existing network representation learning algorithms mainly model the relationship between network nodes based on the structural features of the network, or use text features, hierarchical features and other external attributes to realize the network joint representation learning. Capturing global features of the network allows the obtained node vectors to retain more comprehensive feature information during training, thereby enhancing the quality of embeddings. In order to preserve the global structural features of the network in the training results, we employed a multi-channel learning approach to perform high-order feature modeling on the network. We proposed a novel algorithm for multi-channel high-order network representation learning, referred to as the Multi-Channel High-Order Network Representation (MHNR) algorithm. This algorithm initially constructs high-order network features from the original network structure, thereby transforming the single-channel network representation learning process into a multi-channel high-order network representation learning process. Then, for each single-channel network representation learning process, the novel graph assimilation mechanism is introduced in the algorithm, so as to realize the high-order network structure modeling mechanism in the single-channel network representation learning. Finally, the algorithm integrates the multi-channel and single-channel mechanism of high-order network structure joint modeling, realizing the efficient use of network structure features and sufficient modeling. Experimental results show that the node classification performance of the proposed MHNR algorithm reaches a good order on Citeseer, Cora, and DBLP data, and its node classification performance is better than that of the comparison algorithm used in this paper. In addition, when the vector length is optimized, the average classification accuracy of nodes of the proposed algorithm is up to 12.24% higher than that of the DeepWalk algorithm. Therefore, the node classification performance of the proposed algorithm can reach the current optimal order only based on the structural features of the network under the condition of no external feature supplementary modeling.

Enhanced Knowledge Distillation for Advanced Recognition of Chinese Herbal Medicine.

The identification and classification of traditional Chinese herbal medicines demand significant time and expertise. We propose the dual-teacher supervised decay (DTSD) approach, an enhancement for Chinese herbal medicine recognition utilizing a refined knowledge distillation model. The DTSD method refines output soft labels, adapts attenuation parameters, and employs a dynamic combination loss in the teacher model. Implemented on the lightweight MobileNet_v3 network, the methodology is deployed successfully in a mobile application. Experimental results reveal that incorporating the exponential warmup learning rate reduction strategy during training optimizes the knowledge distillation model, achieving an average classification accuracy of 98.60% for 10 types of Chinese herbal medicine images. The model boasts an average detection time of 0.0172 s per image, with a compressed size of 10 MB. Comparative experiments demonstrate the superior performance of our refined model over DenseNet121, ResNet50_vd, Xception65, and EfficientNetB1. This refined model not only introduces an approach to Chinese herbal medicine image recognition but also provides a practical solution for lightweight models in mobile applications.

A study of annual tree-wise LiDAR intensity patterns of boreal species observed using a hyper-temporal laser scanning time series

This study introduces the annual tree-wise intensity patterns of three boreal tree species, silver birch (Betula pendula Roth.), Scots pine (Pinus sylvestris L.), and Norway spruce (P,icea abies H. Karst.), observed from a long-term hyper-temporal point cloud dataset collected with a permanent laser scanning (LiDAR) station. An annual LiDAR intensity pattern refers to the trend of variations of tree-wise calibrated LiDAR intensity values over the course of a year, primarily linked to species-specific phenological characteristics. Such pattern was discovered from hyper-temporal (76 observations between April 2020 and April 2021) high resolution (0.01 m 3D point spacing at 100 m distance) point cloud observations acquired using a single wavelength (1550 nm) static LiDAR system installed on a climate monitoring tower at the Hyytiälä forest research station in central Finland. A set of experiments was designed to explore the interactions among tree-wise calibrated LiDAR intensity, species, data quality, and observation dates. Thus, to provide practical suggestions for expectable accuracies of tree species classification using diverse types of LiDAR systems and data acquisition strategies. It was revealed that by combining high spatial resolution (up to 100,000 points/m2) with bi-weekly observations (two scans per week except winter period), the average tree species classification accuracy reaches 96.8% for the three boreal tree species in the study area, where the few misclassified cases are attributed to inaccurate individual tree segmentation. This suggests that the annual tree-wise LiDAR intensity patterns, which reflects the tree-wise phenological dynamics, are robust and species-specific. When the spatial resolution was reduced to 5 points/ m2, the average accuracy of species classification among the three studied species reached 85% when using bi-temporal data from spring and autumn. Our analysis further indicates that arranging data acquisition with a two-week variability in late autumn and late spring could yield higher accuracy in species classification. Furthermore, our study revealed the importance of subcanopy information for species classification, specifically in distinguishing different coniferous species. When limited to upper canopy data (as in simulated airborne LiDAR scenarios), achieving a tree species classification accuracy of over 80% required either high point density (e.g. 8000 points/m2) combined with bi-temporal collection in a year, or low point density (5 points/ m2) with a dense time series (e.g., 76 time points within a year), which showcased the possibility to compensate low spatial resolution with high temporal resolution, and vice versa, in capturing species-specific phenological characteristics of trees.

Event recognition method based on feature synthesizing for a zero-shot intelligent distributed optical fiber sensor

Phase-sensitive optical time domain reflectometer (Φ-OTDR) is an emergent distributed optical sensing system with the advantages of high localization accuracy and high sensitivity. It has been widely used for intrusion identification, pipeline monitoring, under-ground tunnel monitoring, etc. Deep learning-based classification methods work well for Φ-OTDR event recognition tasks with sufficient samples. However, the lack of training data samples is sometimes a serious problem for these data-driven algorithms. This paper proposes a novel feature synthesizing approach to solve this problem. A mixed class approach and a reinforcement learning-based guided training method are proposed to realize high-quality feature synthesis. Experiment results in the task of eight event classifications, including one unknown class, show that the proposed method can achieve an average classification accuracy of 42% for the unknown class and obtain its event type, meanwhile achieving a 74% average overall classification accuracy. This is 29% and 7% higher, respectively, than those of the ordinary instance synthesizing method. Moreover, this is the first time that the Φ-OTDR system can recognize a specific event and tell its event type without collecting its data sample in advance.

Identifying optical microscope images of CVD-grown two-dimensional MoS2 by convolutional neural networks and transfer learning.

In Complementary Metal-Oxide Semiconductor (CMOS) technology, scaling down has been a key strategy to improve chip performance and reduce power losses. However, challenges such as sub-threshold leakage and gate leakage, resulting from short-channel effects, contribute to an increase in distributed static power. Two-dimensional transition metal dichalcogenides (2D TMDs) emerge as potential solutions, serving as channel materials with steep sub-threshold swings and lower power consumption. However, the production and development of these 2-dimensional materials require some time-consuming tasks. In order to employ them in different fields, including chip technology, it is crucial to ensure that their production meets the required standards of quality and uniformity; in this context, deep learning techniques show significant potential. This research introduces a transfer learning-based deep convolutional neural network (CNN) to classify chemical vapor deposition (CVD) grown molybdenum disulfide (MoS2) flakes based on their uniformity or the occurrence of defects affecting electronic properties. Acquiring and labeling a sufficient number of microscope images for CNN training may not be realistic. To address this challenge, artificial images were generated using Fresnel equations to pre-train the CNN. Subsequently, accuracy was improved through fine-tuning with a limited set of real images. The proposed transfer learning-based CNN method significantly improved all measurement metrics with respect to the ordinary CNNs. The initial CNN, trained with limited data and without transfer learning, achieved 68% average accuracy for binary classification. Through transfer learning and artificial images, the same CNN achieved 85% average accuracy, demonstrating an average increase of approximately 17%. While this study specifically focuses on MoS2 structures, the same methodology can be extended to other 2-dimensional materials by simply incorporating their specific parameters when generating artificial images.

Application of Semi-Supervised Learning Model to Coal Sample Classification

As an extremely important energy source, improving the efficiency and accuracy of coal classification is important for industrial production and pollution reduction. Laser-induced breakdown spectroscopy (LIBS) is a new technology for coal classification which has the ability to rapidly analyze coal compared with traditional coal analysis methods. In the practical application of LIBS, a large amount of labeling data is usually required, but it is quite difficult to obtain labeling data in industrial sites. In this paper, to address the problem of insufficient labeled data, a semi-supervised classification model (SGAN) based on adversarial neural network is proposed, which can utilize unlabeled data to improve the classification accuracy. The effects of labeled and unlabeled samples on the classification accuracy of the SGAN model are investigated, and the results show that the number of labeled and unlabeled samples are positively correlated, and the highest average classification accuracy that the model can achieve is 98.5%. In addition, the classification accuracies of SGAN and other models (e.g., CNN, RF) are also compared, and the results show that, with the same number of labeled samples in the three models, SGAN performs better after the number of unlabeled samples reaches a certain level, with an improvement of 0.7% and 2.5% compared to the CNN and RF models, respectively. This study provides new ideas for the application of semi-supervised learning in LIBS.

Advancing spinal cord injury care through non-invasive autonomic dysreflexia detection with AI

This paper presents an AI-powered solution for detecting and monitoring Autonomic Dysreflexia (AD) in individuals with spinal cord injuries. Current AD detection methods are limited, lacking non-invasive monitoring systems. We propose a model that combines skin nerve activity (SKNA) signals with a deep neural network (DNN) architecture to overcome this limitation. The DNN is trained on a meticulously curated dataset obtained through controlled colorectal distension, inducing AD events in rats with spinal cord surgery above the T6 level. The proposed system achieves an impressive average classification accuracy of 93.9% ± 2.5%, ensuring accurate AD identification with high precision (95.2% ± 2.1%). It demonstrates a balanced performance with an average F1 score of 94.4% ± 1.8%, indicating a harmonious balance between precision and recall. Additionally, the system exhibits a low average false-negative rate of 4.8% ± 1.6%, minimizing the misclassification of non-AD cases. The robustness and generalizability of the system are validated on unseen data, maintaining high accuracy, F1 score, and a low false-negative rate. This AI-powered solution represents a significant advancement in non-invasive, real-time AD monitoring, with the potential to improve patient outcomes and enhance AD management in individuals with spinal cord injuries. This research contributes a promising solution to the critical healthcare challenge of AD detection and monitoring.

Autoregressive Feature Extraction with Topic Modeling for Aspect-based Sentiment Analysis of Arabic as a Low-resource Language

This paper proposes an approach for aspect-based sentiment analysis of Arabic social data, especially the considerable text corpus generated through communications on X (formerly known as Twitter) for expressing opinions in Arabic-language tweets during the COVID-19 pandemic. The proposed approach examines the performance of several pre-trained predictive and autoregressive language models; namely, Bidirectional Encoder Representations from Transformers (BERT) and XLNet, along with topic modeling algorithms; namely, Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF), for aspect-based sentiment analysis of online Arabic text. In addition, Bidirectional Long Short Term Memory (Bi-LSTM) deep learning model is used to classify the extracted aspects from online reviews. Obtained experimental results indicate that the combined XLNet-NMF model outperforms other implemented state-of-the-art methods through improving the feature extraction of unstructured social media text with achieving values of 0.946 and 0.938, for average sentiment classification accuracy and F-measure, respectively.

MetRoBERTa: Leveraging Traditional Customer Relationship Management Data to Develop a Transit-Topic-Aware Language Model

Transit riders’ feedback provided in ridership surveys, customer relationship management (CRM) channels, and, in more recent times, through social media, is key for transit agencies to better gauge the efficacy of their services and initiatives. Getting a holistic understanding of riders’ experience through the feedback shared in those instruments is often challenging, mostly because of the open-ended, unstructured nature of text feedback. In this paper, we propose leveraging traditional transit CRM feedback to develop and deploy a transit-topic-aware large language model (LLM) capable of classifying open-ended text feedback to relevant transit-specific topics. First, we utilize semi-supervised learning to engineer a training dataset of 11 broad transit topics detected in a corpus of 6 years of customer feedback provided to the Washington Metropolitan Area Transit Authority. We then use this dataset to train and thoroughly evaluate a language model based on the RoBERTa architecture. We compare our LLM—MetRoBERTa—with classical machine learning approaches utilizing keyword-based and lexicon representations. Our model outperforms those methods across all evaluation metrics, providing an average topic classification accuracy of 90%. Finally, we provide a value proposition of this work demonstrating how the language model, alongside additional text processing tools, can be applied to add structure to open-ended text sources of feedback such as Twitter. The framework and results we present provide a pathway for an automated, generalizable approach for ingesting, visualizing, and reporting transit riders’ feedback at scale, enabling agencies to better understand and improve customer experience.

Characterization and depth detection of internal delamination defects in CFRP based on line laser scanning infrared thermography

In this article, the shape characterization and depth detection of delamination defects in carbon fiber reinforced polymer (CFRP) sandwich are achieved by detecting delamination defects through line laser infrared thermography. In terms of defect shape characterization, the temperature matrix difference imaging method is proposed to characterize the shape of internal defects, and the defect shape distortion is corrected by the principal component analysis algorithm to optimize the effect of defect shape characterization. In terms of defect depth detection, this study constructs a dataset of cooling curves for defects of different depths and builds an encoder–decoder convolutional neural network structure based on the attention mechanism, which is centered on a streamlined encoder–decoder structure as well as a highly integrated attention gate, which is capable of accelerating the extraction of features and reducing the number of iterations of the algorithm. The results demonstrate that we achieved an average classification accuracy of 88.9% for defects with a depth gradient difference of 0.2 mm, attaining the highest classification accuracy of 95% for a single defect. Furthermore, our approach outperformed several classical classifiers by achieving superior accuracy with fewer iterations. The experimental results demonstrate that the line laser scanning inspection method is effective in the task of detecting internal defects in CFRP materials.

Optical Emission Spectroscopy for the Real-Time Identification of Malignant Breast Tissue

Breast conserving resection with free margins is the gold standard treatment for early breast cancer recommended by guidelines worldwide. Therefore, reliable discrimination between normal and malignant tissue at the resection margins is essential. In this study, normal and abnormal tissue samples from breast cancer patients were characterized ex vivo by optical emission spectroscopy (OES) based on ionized atoms and molecules generated during electrosurgical treatment. The aim of the study was to determine spectroscopic features which are typical for healthy and neoplastic breast tissue allowing for future real-time tissue differentiation and margin assessment during breast cancer surgery. A total of 972 spectra generated by electrosurgical sparking on normal and abnormal tissue were used for support vector classifier (SVC) training. Specific spectroscopic features were selected for the classification of tissues in the included breast cancer patients. The average classification accuracy for all patients was 96.9%. Normal and abnormal breast tissue could be differentiated with a mean sensitivity of 94.8%, a specificity of 99.0%, a positive predictive value (PPV) of 99.1% and a negative predictive value (NPV) of 96.1%. For 66.6% patients all classifications reached 100%. Based on this convincing data, a future clinical application of OES-based tissue differentiation in breast cancer surgery seems to be feasible.

Dynamic Threshold Distribution Domain Adaptation Network: A Cross-Subject Fatigue Recognition Method Based on EEG Signals

Electroencephalogram (EEG) has been widely used in driver fatigue recognition due to its high-resolution and non-invasive characteristics in recent years. However, EEG signals have strong variability and significant individual differences, which put forward extremely high requirements on the generalization ability of the model in practical use. To address these issues, a dynamic threshold distribution domain adaptation network (DTDDAN) is proposed to classify the EEG signals for driver fatigue recognition. In the model, the domain discriminator is introduced to alleviate the marginal distribution difference of different domains, and a new loss, Jensen–Shannon loss (JS loss), is applied to relieve the conditional distribution difference of different domains to learn cross-subject invariant deep features of EEG signals. In addition, we propose a dynamic threshold pseudo label strategy to assign pseudo labels to the samples of the target domain in the training process of the model. This strategy can make sure that the high-quality unlabeled data of the target domain contribute to the model training and the model can fully learn the information of the target domain. With extensive experiments on our self-constructed EEG-based fatigue driving dataset, competitive performance is achieved for the cross-subject classification task, with average classification sensitivity, specificity, and accuracy of 89.8%,93.8% and 92.2%, respectively. And it is shown that our model can effectively extract cross-subject invariant EEG signals deep features by the experimental results.

Temporal and spatial variability of dynamic microstate brain network in disorders of consciousness.

Accurately diagnosing patients with the vegetative state (VS) and the minimally conscious state (MCS) reached a misdiagnosis of approximately 40%. A method combined microstate and dynamic functional connectivity (dFC) to study the spatiotemporal variability of the brain in disorders of consciousness (DOC) patients was proposed. Resting-state EEG data were obtained from 16 patients with MCS and 16 patients with VS. Mutual information (MI) was used to assess the EEG connectivity in each microstate. MI-based features with statistical differences were selected as the total feature subset (TFS), then the TFS was utilized to feature selection and fed into the classifier, obtaining the optimal feature subsets (OFS) in each microstate. Subsequently, an OFS-based MI functional connectivity network (MIFCN) was constructed in the cortex. The group-average MI connectivity matrix focused on all channels revealed that all five microstates exhibited stronger information interaction in the MCS when comparing with the VS. While OFS-based MIFCN, which only focused on a few channels, revealed greater MI flow in VS patients than in MCS patients under microstates A, B, C, and E, except for microstate D. Additionally, the average classification accuracy of OFS in the five microstates was 96.2%. Constructing features based on microstates to distinguish between two categories of DOC patients had effectiveness.

A tensor decomposition scheme for EEG-based diagnosis of mild cognitive impairment

Mild Cognitive Impairment (MCI) is the primary stage of acute Alzheimer's disease, and early detection is crucial for the person and those around him. It is difficult to recognize since this mild stage does not have clear clinical signs, and its symptoms are between normal aging and severe dementia. Here, we propose a tensor decomposition-based scheme for automatically diagnosing MCI using Electroencephalogram (EEG) signals. A new projection is proposed, which preserves the spatial information of the electrodes to construct a data tensor. Then, using parallel factor analysis (PARAFAC) tensor decomposition, the features are extracted, and a support vector machine (SVM) is used to discriminate MCI from normal subjects. The proposed scheme was tested on two different datasets. The results showed that the tensor-based method outperformed conventional methods in diagnosing MCI with an average classification accuracy of 93.96% and 78.65% for the first and second datasets, respectively. Therefore, it seems that maintaining the spatial topology of the signals plays a vital role in the processing of EEG signals.

Deep learning approach for identification of H ii regions during reionization in 21-cm observations – II. Foreground contamination

ABSTRACT The upcoming Square Kilometre Array Observatory will produce images of neutral hydrogen distribution during the epoch of reionization by observing the corresponding 21-cm signal. However, the 21-cm signal will be subject to instrumental limitations such as noise and galactic foreground contamination that pose a challenge for accurate detection. In this study, we present the SegU-Net v2 framework, an enhanced version of our convolutional neural network, built to identify neutral and ionized regions in the 21-cm signal contaminated with foreground emission. We trained our neural network on 21-cm image data processed by a foreground removal method based on Principal Component Analysis achieving an average classification accuracy of 71 per cent between redshift z = 7 and 11. We tested SegU-Net v2 against various foreground removal methods, including Gaussian Process Regression, Polynomial Fitting, and Foreground-Wedge Removal. Results show comparable performance, highlighting SegU-Net v2’s independence on these pre-processing methods. Statistical analysis shows that a perfect classification score with ${\rm AUC}=95~{{\ \rm per\ cent}}$ is possible for 8 &amp;lt; z &amp;lt; 10. While the network prediction lacks the ability to correctly identify ionized regions at higher redshift and differentiate well the few remaining neutral regions at lower redshift due to low contrast between 21-cm signal, noise, and foreground residual in images. Moreover, as the photon sources driving reionization are expected to be located inside ionized regions, we show that SegU-Net v2 can be used to correctly identify and measure the volume of isolated bubbles with $V_{\rm ion}\gt (10\, {\rm cMpc})^3$ at z &amp;gt; 9, for follow-up studies with infrared/optical telescopes to detect these sources.