On Solar Filament Detection Techniques: From Manual to Intelligent

Abstract

Solar filaments (and their limb counterparts, prominences) are critical tracers of the Sun’s magnetic topology and key precursors to coronal mass ejections (CMEs). Precise identification and continuous tracking of these features are essential for understanding solar eruptive mechanisms and improving space weather forecasting. This systematic review evaluates the evolution of automated detection methodologies, addressing the challenge of processing the exponentially growing volume of high-resolution solar observations. We identify deep learning architectures, particularly U-Net variants and Mask R-CNN, as the most promising current paradigms. Compared to traditional image processing, these data-driven models demonstrate superior robustness against noise and variable observing conditions, achieving high-precision segmentation (>90% accuracy) with sub-second inference speeds. This leap in computational efficiency and accuracy directly facilitates real-time operational monitoring and enables large-scale statistical analysis of filament evolution across solar cycles. We conclude that future breakthroughs lie in developing physics-informed AI and standardized benchmarks to bridge the gap between pixel-level segmentation and physical interpretation, ultimately creating detection systems that are both operationally reliable and scientifically meaningful.

1. Introduction

Research on the historical evolution of solar filament detection methods, including image processing techniques and machine learning applications, has emerged as a critical area of inquiry due to its significance in understanding solar activity and forecasting space weather. Early efforts in solar filament detection relied heavily on manual feature recognition, which was labor-intensive and limited by the volume of observational data [1,2]. A pivotal transition began with the systematic, long-term manual cartography of solar features in the “Synoptic Maps” by the Observatoire de Paris-Meudon, which itself embarked on a crucial digitization and semi-automated synthesis of these records from the 1990s onward [3]. This transition was crucial, as subsequent advancements in telescope technology and digital imaging dramatically increased the volume and resolution of solar observations, thereby necessitating fully automated detection methods to efficiently process the resulting vast datasets [4,5,6,7,8,9,10,11]. For example, based on multi-band H$\alpha$ observations from the Flare Monitoring Telescope (FMT) at Hida Observatory [12], Liu et al. [13] developed an automatic detection system for dynamic dark features such as surges, aiming to replace inefficient, time-consuming and subjective manual visual identification. Automated techniques have thus evolved from basic image enhancement and thresholding to sophisticated algorithms incorporating morphological operations and machine learning classifiers, reflecting a trajectory of increasing complexity and accuracy [14,15,16]. The practical importance of these developments is underscored by the role of solar filaments in predicting geomagnetic storms and coronal mass ejections, which have profound impacts on Earth’s technological infrastructure [17,18].

Despite significant progress, challenges remain in achieving fully automated, accurate, and real-time solar filament detection. The problem lies in the variability of filament appearance, image quality, and the need to distinguish filaments from other solar features such as sunspots [19,20,21,22]. Existing literature reveals a knowledge gap in integrating robust image processing with advanced machine learning techniques to handle diverse datasets from multiple observatories without reliance on empirical thresholds [17,23,24]. Controversies persist regarding the optimal balance between traditional image processing methods and emerging deep learning approaches, with some studies emphasizing morphological operations and others advocating neural networks for improved segmentation and classification [25,26,27]. The consequences of this gap include limitations in the timely and accurate monitoring of solar filaments, which impedes effective space weather forecasting [28].

To provide a concrete illustration of modern automated filament detection, Figure 1 presents a segmentation example obtained by applying a U-Net architecture to CHASE (Chinese H$\alpha$ Solar Explorer) observations. This deep learning-based method effectively delineates filament structures, exemplifying the advanced techniques that have emerged alongside traditional image processing approaches.

The conceptual framework of this review is based on the interplay among solar filament characteristics, image processing methodologies, and machine learning techniques. Solar filaments are typically defined as elongated, cool plasma structures observable in H$\alpha$ full-disk images, and their detection relies on image segmentation and feature extraction [20,29]. Machine learning models, including support vector machines and deep neural networks, learn from labeled data to classify and track filaments, thereby improving detection accuracy and automation [4,30]. This framework supports a systematic review of the evolution of filament detection methods, aiming to identify effective strategies and remaining challenges [31].

This review critically examines the historical development of solar filament detection methods, with an emphasis on the integration of image processing and machine learning. By synthesising existing approaches, highlighting key advances, and identifying gaps, it aims to inform future research and operational applications. Addressing these gaps will help improve automated detection systems that are essential for space weather forecasting [5,18].

A comprehensive literature survey was conducted, selecting studies from early manual techniques to recent deep-learning models. The analytical framework includes comparative evaluations of algorithmic performance and methodological evolution. The findings are organised chronologically and thematically to elucidate the progression and current state of the field [2,27].

A comprehensive summary of individual studies reviewed in this work is provided in Appendix A (

Table A1. Descriptive Summary of Solar Filament Detection Studies.

Study	Methodology/Approach	Key Findings/Performance	Computational & Efficiency Aspects	Robustness Handling	Datasets & Validation
Mouradian [3]	Semi-automated digitization & synthesis of hand-drawn synoptic maps	Established long-term digital record & synthesis methods for filaments	Human-in-the-loop process; digitization enables computational analysis	Relies on expert human decision, minimizing false detections	Meudon synoptic map series (1919+). Became a benchmark dataset for validation
Gao et al. [32]	Thresholding & region-growing	Automated detection of filament disappearance;alerts on disappearance events	Processes 1 image/minute; suitable for near-real-time alert	Uses “4-pixel” & “40-pixel” methods to handle noise and fragmentation	Tested on BBSO full-disk H$\alpha$ images
Zharkova and Schetinin [25]	Early ANN technique for filament recognition	Effective recognition on varied brightness backgrounds	Computationally light due to simple network	Handles atmospheric distortions well	Tested on Meudon Observatory H$\alpha$ spectroheliograms
Zharkova and Schetinin [63]	Artificial Neural Network with linear background approximation and sliding window technique	Effectively recognizes filaments on variable brightness backgrounds; trained on one fragment	Computationally light network; faster than region-growing, slightly slower than mosaic methods	Adapts to variable intensity backgrounds; no explicit noise correction needed	Meudon Observatory H$\alpha$ spectroheliograms; visually validated against mosaic & region-growing methods
Shih and Kowalski [39]	Advanced local thresholding or global thresholding; directional morphological filtering with linear structuring elements	Excellent detection of large filaments (100%); good detection of small filaments (90.4%); higher false positives than some methods	Faster than region-growing methods; suitable for parallel implementation; efficient directional filtering	Removes sunspots by brightness thresholding; robust to Gaussian and blur noise	Tested on BBSO H$\alpha$ full-disk images; validated against manual markings
Fuller and Aboudarham [1]	Region growing with seed selection for detection	Very good correspondence with manual digitization	Efficient for four months of data	Image cleaning reduces defects and false detections	Tested on Meudon Observatory H$\alpha$ spectroheliograms
Bernasconi et al. [15]	Automated detection, spine/barb extraction, chirality classification, and synodic tracking	72% accuracy compared to manual chirality identification	Efficient for daily tracking over long periods	Unbiased detection reduces operator subjectivity	Tested on BBSO H$\alpha$ images spanning multiple years
Fuller et al. [24]	Cleaning and region growing for filament recognition	Good agreement with manual digitization	Efficient segmentation and skeleton pruning	Corrects instrument defects and noise	Used in EGSO project for solar features
Qahwaji and Colak [19]	Hybrid system: adaptive intensity filtering, modified region growing, and NN verification	FAR reduced to 4% for filaments with NN verification; fast detection suitable for operational use	Fast detection (<3 s/image) suitable for operational use	Limb darkening correction and intensity filtering applied	Tested on Meudon H$\alpha$ & Ca II K3 images; validated against synoptic maps
Liu et al. [13]	Multi-band H$\alpha$ detection; local median thresholding (109 regions, di = 0.9); seed clustering; multi-wavelength classification	89% manual event recall; 10× more detections; precise start/end times; surge/filament/network classification	1 s/frame preprocessing; 1–3 h/day	Data quality filtering; limb-darkening correction; sunspot exclusion	FMT Hida Observatory; H$\alpha$ center, $\pm 0.8$Å, white light; 7 days (1997–2001); validated against manual event catalogs
Qu et al. [14]	Image processing with SVM classification	High precision in filament and spine detection	Moderate speed; suitable for real-time applications	Uses image enhancement to reduce noise and limb effects	Applied on BBSO H$\alpha$ full-disk images
Zharkova and Schetinin [45]	ANN with linear/parabolic background approximation; sliding window technique	∼82.5% accuracy with parabolic fit; outperforms linear fit and earlier mosaic/region-growing methods	Efficient once trained; faster than region-growing; suitable for processing large archives	Adapts to limb darkening and intensity variations without explicit noise correction	Tested on Meudon H$\alpha$ spectroheliograms; compared against region-growing and mosaic methods
Zharkova et al. [50]	Review of filament detection methods (chain linking, region growing); introduces ANN technique with linear/parabolic background approximation	ANN achieves ∼82.5% accuracy for filaments on variable backgrounds; effective without explicit noise correction	Notes computational demands of large archives; ANN efficient once trained	ANN adapts to variable intensity backgrounds; preprocessing (e.g., shape correction)	Techniques tested on Meudon H$\alpha$ data; compared against region-growing results
Shih et al. [4]	SIDE for filament enhancement, adaptive thresholding, and SVM for filament/sunspot discrimination	∼95% detection rate for large filaments; first SVM use for sunspot removal; automated disappearance detection	Efficient processing suitable for continuous monitoring	Robust to low-contrast filaments and limb darkening via SIDE & SVM	Tested on BBSO H$\alpha$ full-disk images; validated against manual detection
Qahwaji and Colak [64]	Local contrast enhancement and adaptive intensity filtering to identify filament and NN for verification	Neural network verification reduces false positives	Image processing and feature-based NN verification enables 1 s/image	Intensity filtering and local detection windows to eliminate noise	Tested on Meudon H$\alpha$ data; compared against BASS manual annotations
Aboudarham et al. [65]	Integrated automatic detection with filament tracking for synoptic map synthesis	Achieved major progress towards fully automatic synoptic map generation	Aims for a fully automated processing pipeline	Image cleaning (limb fitting, intensity normalization) reduces defects	Tested on Meudon H$\alpha$ spectroheliograms; compared with manual maps
Aboudarham et al. [40]	Laplacian enhancement, region growing, morphological closing; Carrington map-based tracking	∼90% detection accuracy; enables tracking across rotations; handles filament splitting/reappearance	Fully automated pipeline; suitable for synoptic map generation and long-term tracking	Corrects limb darkening, transparency variations, and instrumental defects	Meudon H$\alpha$ spectroheliograms; validated against manual synoptic maps
Atoum et al. [66]	Adaptive local thresholding (ALT) using dual-window sliding; intensity- and range-based segmentation	Lower false acceptance rate (FAR: ∼9%) compared to ALT&V (19%); produces well-defined filaments with reduced noise	Efficient sliding window processing; suitable for daily image analysis	Adaptive to local intensity variations; handles background non-uniformity via local statistics	Meudon H$\alpha$ images; validated against manual synoptic maps; FAR used as performance metric
Joshi et al. [28]	Intensity and size thresholding with preprocessing (fragment grouping/foreshortening correction)	81% detection rate across 10 events; capable of real-time eruption monitoring	∼1 min/image processing time; suitable for real-time systems	Limb/foreshortening correction; sunspot removal; multi-observatory robustness	Tested on BBSO, MLSO, KSO, USO data; validated against manual inspection
Yuan et al. [20]	Polynomial luminance correction, adaptive thresholding, morphological operations, and graph-based skeleton extraction	Accuracy >96% for filament number and >99% for area; outperforms previous methods	Efficient cascading Hough transform reduces memory consumption	Robust across different observatories; removes sunspots by shape	Tested on 4 observatories (BBSO, KANZ, OACT, YNAO); validated against manual annotations
Andrijauskas and Gradvohl [49]	Hybrid OpenMP-MPI parallelization; two methods: FDD (diffusion filter + graph-based) and MD (morphological operators)	FDD: ∼80% pixel accuracy, 94% filament detection rate; MD: ∼58% accuracy, 66% detection rate	326× speed-up for FDD, 54× for MD on 5-node cluster; supports high-throughput processing of large images	Diffusion filter preserves edges; handles filament continuity across MPI segments	BBSO H$\alpha$ images; validated against manually annotated ground truth images
Atoum [17]	Fully automated detection and tracking with NN merging	High confidence in filament disappearance detection	Faster than previous empirical methods	Adaptive thresholding improves noise handling	Uses Heliographic Carrington coordinates for tracking
Bonnin et al. [56]	Co-rotational frame tracking with skeleton curve-matching	∼90% tracking accuracy for filaments	Efficient for large datasets; ∼30 s per Carrington rotation	Stable in co-rotating frame	Applied on Meudon H$\alpha$ spectroheliograms
Peng et al. [2]	Early automated methods including ANN and region growing	Moderate accuracy; ANN improves detection near solar limb	Region growing is time-consuming; ANN faster	Sensitive to background cleaning and limb proximity	Tested on limited datasets, mainly H$\alpha$ images
Hao et al. [31]	Canny edge detection & morphological operators for detection; differential rotation for tracing	Quantified latitudinal migration speeds in Cycle 23; ∼85% detection accuracy	Efficient processing (∼1 s/image for detection)	Preprocessing removes limb darkening; sunspot filtering by geometry	Tested on MLSO H$\alpha$ images; validated against manual detection
Riegler et al. [67]	Variational multi-label segmentation for simultaneous flare and filament detection	Accurate classification vs. expert annotations; F-score: 0.89 (filaments), 0.92 (flares)	Real-time alert potential with preprocessing steps	Image normalization, registration, and structural bandpass filter	Tested on KSO H$\alpha$ image sequences; validated against NOAA annotations
Atoum and Ali [29]	Geometrical approach for spine extraction	More accurate and longer spine detection	Computationally less complex and faster	Tracks filament backbone precisely	Suitable for real-time tracking
Schuh et al. [54]	Comparative evaluation leveraging AAFDCC metadata to train a general Trainable Feature Recognition (TFR) module	Achieved >82.7% filament detection accuracy; confirmed reliability of AAFDCC reporting	Efficient with certain classifiers (e.g., J48, RF) for large-scale data	labeling schemes to mitigate label noise	Trained and validated on BBSO H$\alpha$ images using AAFDCC module metadata as benchmark
Atoum and Ali [23]	Image processing + neural network for merging broken filaments; uses spatial & orientation features as NN inputs	92% true-positive rate in merging fragments;reduces false merging compared to earlier empirical methods	Efficient once NN is trained; avoids constant thresholds	Adaptive to fragment distance & orientation; handles fragmentation caused by imaging artifacts	Applied on BBSO H$\alpha$ images; validated against manual merging
Atoum [58]	Context-based sliding window with adaptive thresholding	Adapts window size to local intensity variability; handles both high- and low-contrast filaments	Adaptive window reduces unnecessary computation	Robust to uneven illumination and intensity variations; no fixed structuring elements	Tested on BBSO and Meudon H $\alpha$ images
Hao et al. [5]	Canny edge detection for segmentation, and morphological operations for feature extraction	Generated filament “butterfly diagram” from nearly 3 solar cycles	Efficient and versatile method applied to multi-cycle data	Limb-darkening correction; sunspot exclusion	Applied to full-disk H$\alpha$ data (primarily BBSO)
Ahmadzadeh et al. [30]	Mask R-CNN deep neural network framework	Competitive performance using Mask R-CNN; scalable with more data	Scalable deep learning framework	Handles false positives and negatives well	Trained on BBSO H$\alpha$ images with HEK metadata
Salasa and Arymurthy [44]	Mask R-CNN deep learning for filament detection	Achieves 96% precision without intensity normalization	Fast detection at 0.42 s per image	Robust to intensity variations without normalization	Applied on LAPAN H$\alpha$ images
Zhu et al. [26]	Improved U-Net with dropout layers & nearest-neighbor upsampling	Average TPR: 0.9145, DSC: 0.8944; effective for large filaments	GPU training: ∼49 min; CPU: ∼57 h; suitable for small datasets	Handles noise and artifacts; sensitive to uneven illumination	Dataset from BBSO/FDH A H$\alpha$ images
Liu et al. [47]	Multiple improved U-Net variants with ASPP module; compared with CGAN	dilation-122436 achieves Jac: 0.6320, MCC: 0.7829, F1: 0.7710 on high-quality images; CGAN better for low-quality images	Improved U-Nets train faster (≈1.2h) vs. CGAN (80 h); suitable for limited data	ASPP improves high-quality image segmentation; CGAN more robust to poor image quality	Trained on 3000 pairs from HSOS & BBSO H$\alpha$ images; validated against manual ground truth
Guo et al. [52]	Improved CondInst instance segmentation with ResNet-C/D/v2 backbone + Matrix NMS	Mean precision: 90.83%, recall: 83.88%, AP: 82.86%, F1: 87.22%; effectively detects fragmented filaments	Training: 30 h on 1×RTX 2080; inference: 0.458 s/image	Robust to uneven brightness, limb darkening, low contrast; generalizes across observatories	BBSO H $\alpha$ images; 12,519 filaments; tested on SMART & HSOS data
Priyadarshi et al. [57]	K-means clustering for optimal RGB thresholding on hand-drawn suncharts	Effective extraction of filament parameters; revealed poleward migration and tilt sign dominance	Processing of scanned drawings is computationally light	Handles variability in hand-drawn data and discolorations	Unique dataset from Kodaikanal Solar Observatory; cross-verified with H$\alpha$ plates
Wu et al. [68]	AA-UNet with axial attention for segmentation	Achieves F1-score of 0.77, robust to noise	Efficient segmentation with attention blocks	Handles uneven image quality effectively	Tested on solar image test sets
You and Shang [69]	Improved DeepLab V3+ deep learning model	F1-score improved by over 2% compared to original	Optimized convolutions reduce incomplete detections	Enhanced edge optimization for small filaments	Dataset from BBSO H$\alpha$ images
Bandyopadhyay and Pant [70]	Active Contours Without Edges (ACWE) with preprocessing (inpainting, log transform, sharpening) and postprocessing	Max TPR: 0.9075; accuracy: 99.63%; outperforms Otsu and K-means in boundary preservation	Slower execution (∼29.8 s/image); no training required	Robust to noise; preserves boundaries; sensitive to local minima	BBSO H$\alpha$ images; validated against manual ground truth
Diercke et al. [27]	Semi-supervised deep learning combining YOLOv5 and U-Net	92% accuracy in filament segmentation	Semi-supervised learning reduces annotation cost	Generalizable to multiple H$\alpha$ filtergrams	Applied on ChroTel, GONG, and KSO data
Jiang and Li [41]	Attention U2-Net; integrates attention gates into RSU blocks for enhanced feature selection	F1-score: 0.83; outperforms standard U-Net variants; effective for multi-scale filaments	Trained on an NVIDIA Quadro RTX4000 GPU; batch size of 2 over 150 epochs. (No inference time or comparative efficiency metrics reported)	Attention mechanism improves robustness to noise and complex backgrounds	CHASE/H$\alpha$ images (2023); 600 images with labels from clustering & post-processing
Shang et al. [42]	Hybrid CNN-Transformer model with multi-scale residual blocks, deformable large kernel attention, and res path	F1: 91.19%, Precision: 91.50%, Recall: 90.89%; superior anti-interference to sunspots and noise	Batch size: 4; Transformer layers: 12; image size: 512×512; data augmentation used	Enhanced global feature extraction; robust to sunspots and background noise	Dataset from BBSO H$\alpha$ images; 2700 augmented training samples
Zheng et al. [18]	U-Net and CSRT tracking for filament dynamics	Validated for accurate filament identification	Automated processing suitable for large data	Robust velocity inversion and spine extraction	Uses CHASE/HIS full-disk spectroscopic data
Zhu et al. [43]	Flat U-Net ultralightweight model for segmentation	Precision  0.93, DSC up to 0.82 with attention blocks	Highly efficient with reduced parameters	Performs well despite image quality variations	Dataset from BBSO H$\alpha$ images
Seth et al. [22]	YOLOv8-obb (oriented bounding boxes); self-validation via statistical moments (entropy, skew, etc.) and Tamura texture features (GLCM).	On FITS: P = 0.788, R = 0.863, mAP@0.5 = 0.874. Entropy best differentiates plages, sunspots, filaments, off-limb.	500 epochs, batch size 8, LR 0.01. Data augmentation: grayscale, HSV, gamma, flips.	Gamma augmentation for intensity variations; manual rescaling (×8) on real SUIT; statistical/Tamura self-validation for unlabeled data.	Mock SUIT from IRIS Mg II k (2013–2023); 217→3056 augmented JPG (train), 123 FITS (val), 92 real SUIT FITS (test).

), which details methodologies, performance metrics, and validation approaches across the historical development of solar filament detection techniques.

2. Historical Evolution of Solar Filament Detection Techniques

The historical progression of solar filament detection methods reveals a clear technological trajectory from labor-intensive manual approaches to increasingly sophisticated automated systems. This evolution can be broadly categorized into three distinct phases, each characterized by fundamental shifts in methodology, accuracy, and computational efficiency.

2.1. Manual and Semi-Automated Methods (1990s–2002)

The earliest efforts in automated filament detection emerged from the need to process growing volumes of solar observational data. Foundational work by Mouradian [3] established long-term digital records through semi-automated digitization of hand-drawn synoptic maps from the Observatoire de Paris-Meudon, creating benchmark datasets for validation. These methods maintained human oversight to minimize false detections but represented a crucial first step toward automation.

Early fully automated approaches employed basic image processing techniques. Gao et al. [32] implemented thresholding and region-growing algorithms for filament disappearance detection, achieving processing rates of one image per minute—suitable for near-real-time alerts. These pioneering methods addressed fundamental challenges like noise and fragmentation through simple heuristics such as “4-pixel” and “40-pixel” connectivity rules, though they remained limited by sensitivity to image quality variations and background non-uniformity.

2.2. Image Processing and Early Machine Learning (2003–2010)

This period witnessed significant methodological diversification and the initial integration of machine learning. Some general introductions to pattern and feature detection in digital images can be found in the textbooks by Gonzalez and Woods [33], Jähne [34], Pao [35], Woods [36], Mallat [37] and Ripley [38]. Advanced image processing techniques emerged, with Shih and Kowalski [39] demonstrating excellent detection of large filaments (100%) using directional morphological filtering with linear structuring elements, while Fuller and Aboudarham [1] achieved very good correspondence with manual digitization through region growing with seed selection.Targeting dynamic H$\alpha$ dark features such as active filaments and chromospheric surge jets, Liu et al. [13] employed local median-based dynamic thresholding to identify dark points and integrated multi-wavelength information for event classification. Their method successfully identified 89% of manually recorded events and determined the start and end times of events more accurately.

The integration of machine learning classifiers marked a critical advancement. Zharkova and Schetinin [25] pioneered artificial neural networks (ANNs) for filament recognition on variable brightness backgrounds, achieving approximately 82.5% accuracy with parabolic background approximation. Subsequent work introduced support vector machines (SVMs) for filament classification and sunspot removal, with Shih et al. [4] reporting ∼95% detection rates for large filaments. Hybrid systems combining image processing with neural network verification, such as that by Qahwaji and Colak [19], reduced false acceptance rates to 4% while maintaining fast processing times (<3 s per image).

Automated tracking and characterization capabilities also advanced significantly during this period. Bernasconi et al. [15] developed automated spine extraction and chirality classification with 72% accuracy compared to manual identification, while Aboudarham et al. [40] achieved ∼90% detection accuracy through Laplacian enhancement and region growing, enabling fully automated synoptic map generation. A representative operational example of such traditional pipelines is the Automatic Solar Synoptic Analyzer (ASSA) developed by the Korean Space Weather Center, which integrates morphological directional filtering for filament detection [39] and provides real-time online products (ASSA webpage: https://spaceweather.kasa.go.kr/assa accessed on 2 June 2026).

2.3. Deep Learning Era (2011–Present)

The advent of deep learning architectures represents a paradigm shift in solar filament detection, enabling end-to-end learning with minimal hand-crafted feature engineering. Initial approaches focused on adapting existing architectures, with improved U-Net variants achieving average true positive rates of 0.9145 and Dice similarity coefficients of 0.8944 [26].

A typical U-Net architecture designed for filament segmentation is illustrated in Figure 2. As shown, the model takes H$\alpha$ line-center images as input and produces a binary segmentation mask of the same spatial dimensions. The encoder path captures contextual information through successive down-sampling, while the decoder path restores spatial resolution via up-sampling and skip connections. This symmetric design enables precise localization of filament boundaries, which is a key reason why U-Net and its variants have become dominant in this domain [18].

Building on this foundational design, subsequent works have introduced attention mechanisms, transformer modules, and lightweight modifications to further improve accuracy and efficiency [41,42,43].

Instance segmentation models like Mask R-CNN demonstrated remarkable performance, with Salasa and Arymurthy [44] reporting 96% precision without intensity normalization and inference times of 0.42 s per image. These approaches showed enhanced robustness to noise and background interference while eliminating the need for extensive preprocessing.

Recent advances incorporate sophisticated architectural innovations. Attention mechanisms, as implemented in Attention U²-Net by Jiang and Li [41], improve feature selection and model interpretability. Hybrid CNN-Transformer models like that of Shang et al. Shang et al. [42] achieve F1-scores of 91.19% through enhanced global feature extraction. Semi-supervised approaches combining YOLOv5 and U-Net by Diercke et al. [27] maintain high accuracy (92%) while reducing annotation costs, addressing a critical limitation in the field.

The development of lightweight architectures represents another significant trend. Models like Flat U-Net by Zhu et al. [43] demonstrate high efficiency with reduced parameters while maintaining competitive performance (precision ∼0.93), enabling deployment on embedded systems and facilitating real-time processing.

This evolutionary progression has yielded substantial improvements in detection performance. Accuracy has advanced from moderate levels (72–85%) in early methods to consistently high performance (90%+) in recent deep learning approaches. Computational efficiency has similarly improved, with processing times evolving from minutes per image in manual methods to sub-second inference in modern systems. The scope of validation has expanded from single-observatory testing to comprehensive multi-instrument evaluations, enhancing operational applicability across diverse observational platforms.

A chronological summary of these methodological developments, from semi-automated digitization to modern deep learning architectures, is presented in

Table 1. Chronological Development of Solar Filament Detection Methods.

Period	Key Developments and Representative Works
1990s–2002	Semi-automated digitization of synoptic maps [3] Thresholding and region-growing for disappearance detection [32] Processing rates: ∼1 image/minute Basic noise handling through connectivity rules
2003–2010	Advanced image processing: directional morphological filtering [39], region growing [1] Multi-band detection of dynamic dark features: adaptive thresholding and region growing [13] Machine learning integration: ANN [25,45,46], SVM [4,14] Hybrid systems with NN verification [19] Automated tracking and characterization [15,40] Accuracy: 72–95%; Processing: 1–3 s/image
2011–Present	Deep learning architectures: U-Net variants [26,47], Mask R-CNN [30,44] Advanced features: attention mechanisms [41], transformer-CNN hybrids [42] Semi-supervised learning [27] Lightweight models: Flat U-Net [43] Accuracy: >90%; Processing: <0.5 s/image

.

3. Performance Comparison of Key Technical Approaches

The evolution of solar filament detection has produced three dominant technical paradigms: traditional image processing, classical machine learning, and deep learning architectures. Each approach exhibits distinct performance characteristics, advantages, and limitations that make them suitable for different application scenarios and operational constraints.

3.1. Traditional Image Processing Methods

Traditional image processing techniques formed the foundation of automated filament detection, relying on carefully designed algorithms for feature extraction and segmentation. These methods typically employed sequential processing pipelines including image enhancement, thresholding, morphological operations, and region growing [1,32,48].

Performance Characteristics:

Early approaches demonstrated moderate accuracy levels, with region growing methods achieving very good correspondence with manual digitization [1] and directional morphological filtering showing excellent detection of large filaments (100%) with 90.4% detection rate for small filaments [39]. Advanced implementations incorporating robust preprocessing achieved higher performance, with polynomial luminance correction and adaptive thresholding reporting >96% accuracy for filament number detection and >99% for area measurement [20]. For dynamic features such as surges, the method achieved a recognition rate of 89% compared with manually created catalogs, while detecting over ten times more manually missed events [13].

Computational Efficiency: Processing speeds varied significantly based on algorithm complexity. Basic thresholding and region growing methods typically required 1–3 s per image [19], while more sophisticated approaches with extensive preprocessing could take up to 1 min per image [28]. Parallel implementations demonstrated substantial speed improvements, with hybrid OpenMP-MPI parallelization achieving 326×acceleration for filter-based diffusion methods [49].

Strengths and Limitations: The primary advantages of traditional methods include computational transparency, minimal training requirements, and well-understood algorithmic behavior. However, these approaches suffer from sensitivity to image quality variations, requirement for manual parameter tuning, and limited adaptability to new datasets without algorithm modification [1,50]. They particularly struggle with fragmented filaments, limb-darkening effects, and distinguishing filaments from other solar features like sunspots [2,28].

3.2. Classical Machine Learning Approaches

The integration of machine learning classifiers represented a significant advancement in filament detection capability, enabling improved feature discrimination and reduced false positive rates through learned decision boundaries.

Performance Characteristics: Artificial neural networks achieved approximately 82.5% accuracy on variable brightness backgrounds using parabolic background approximation [45]. Support vector machines demonstrated strong performance with 95% detection rates for large filaments while effectively discriminating filaments from sunspots [4]. Hybrid systems combining image processing with neural network verification reduced false acceptance rates to 4% while maintaining operational viability [19].

Computational Efficiency: Machine learning approaches showed variable computational profiles. Neural network implementations were computationally efficient during inference once trained [25], while SVM classifiers offered moderate processing speeds suitable for near real-time applications [14]. The critical limitation remained the requirement for careful feature engineering and dataset preparation prior to model training [51].

Strengths and Limitations: Machine learning methods excel at feature discrimination and classification tasks, particularly in distinguishing filaments from other solar features and handling variable background conditions [19,25]. These approaches also demonstrated effectiveness in specialized tasks such as fragment merging, with neural networks achieving 92% true-positive rates in merging broken filament components [23]. However, they remain constrained by dependency on hand-crafted features, limited generalization beyond training data distributions, and sensitivity to image quality variations that differ from training conditions.

3.3. Deep Learning Architectures

Deep learning represents the current state of the art in solar filament detection, employing end-to-end learning to automatically extract relevant features from raw image data without extensive preprocessing or feature engineering.

Performance Characteristics: Deep learning models consistently achieve superior performance metrics across multiple evaluation dimensions. U-Net variants report average true positive rates of 0.9145 and Dice similarity coefficients of 0.8944 [26], while Mask R-CNN implementations achieve 96% precision without intensity normalization [44]. Recent hybrid architectures including attention mechanisms and transformer components further advance performance, with F1-scores exceeding 91% [41,42].

Computational Efficiency: The computational profile of deep learning approaches shows significant variation based on model architecture and implementation. Standard U-Net training requires substantial resources (49 min on GPU to 57 h on CPU) [26], while inference times can be optimized to sub-second performance (0.42–0.458 s per image) [44,52]. Lightweight architectures like Flat U-Net demonstrate that competitive performance can be maintained while significantly reducing computational requirements [43].

Practical Deployment Barriers: Despite these advances, the computational demands of deep learning models impose non-trivial constraints on both research and operational deployment. The gap between different architectures is substantial: training a Conditional GAN (CGAN) for filament segmentation can require approximately 80 h, whereas an optimized U-Net variant completes training in roughly 1.2 h on comparable hardware [47]. More resource-intensive models such as CondInst demand up to 30 h on a single NVIDIA RTX 2080 GPU [52], while hybrid CNN-Transformer architectures operating on moderate batch sizes (e.g., 4, with 12 Transformer layers and 512 × 512 input resolution [42]) can push memory requirements beyond the capacity of consumer-grade hardware. These costs are compounded by data requirements: state-of-the-art models typically require thousands of annotated samples for supervised training [47,52], and the manual annotation of filament boundaries in full-disk H$\alpha$ images remains a labor-intensive bottleneck. For operational deployment—particularly in resource-constrained environments such as onboard spacecraft processors or small observatory servers—model compression techniques including network quantization (reducing 32-bit weights to 8-bit integers), knowledge distillation, and structured pruning offer practical pathways to bridge the gap between research-grade accuracy and deployable efficiency [43]. The emergence of ultralightweight architectures explicitly designed for this purpose underscores the recognition that computational pragmatism must accompany accuracy-driven innovation for filament detection systems to achieve widespread operational adoption.

Strengths and Limitations: The primary advantages of deep learning include exceptional segmentation accuracy, robustness to noise and image quality variations, and elimination of manual feature engineering requirements [26,30]. These models also demonstrate superior generalization across different observatories and imaging conditions when properly trained [27]. However, significant challenges remain including substantial computational requirements for training, dependency on large annotated datasets, limited model interpretability, and sensitivity to domain shift between training and deployment conditions [41,42].

3.4. Cross-Paradigm Performance Analysis

The comparative analysis reveals distinct trade-offs between the three technical approaches. Traditional image processing methods offer the advantages of computational transparency and minimal data requirements but suffer from limited accuracy and robustness. Classical machine learning approaches provide improved feature discrimination and classification capability while maintaining reasonable computational demands, though they require careful feature engineering. Deep learning architectures deliver state-of-the-art performance and robustness but demand substantial computational resources and annotated datasets.

The optimal approach selection depends heavily on application requirements. For resource-constrained environments or applications requiring high interpretability, traditional methods or classical machine learning may be preferable. For maximum accuracy and robustness in well-resourced environments, deep learning approaches represent the current state-of-the-art. Future research directions increasingly focus on hybrid approaches that combine the strengths of multiple paradigms while mitigating their respective limitations.

A comparative overview of the key performance characteristics across the three technical paradigms is provided in

Table 2. Performance Comparison Across Detection Paradigms.

Characteristic	Traditional Image Processing	Classical Machine Learning	Deep Learning
Typical Accuracy	72–96%	82–95%	90–96%
Processing Speed	1–60 s/image	1–3 s/image	0.4–3 s/image
Training Requirements	None	Moderate	Substantial
Data Dependency	Low	Medium	High
Robustness to Noise	Low-Medium	Medium	High
Interpretability	High	Medium	Low (Improving via Attention maps/XAI)
Adaptability	Low	Medium	High

.

3.5. Critical Perspectives: Paradigm Conflicts and Epistemic Gaps

Beyond the quantitative improvements in detection accuracy, the historical trajectory of solar filament detection reveals fundamental epistemological tensions that warrant critical examination. The field currently faces a “success paradox,” where metrics improve while physical understanding stagnates.

3.5.1. The Semantic Gap: Pixel Precision vs. Physical Reality

A fundamental disconnection persists between data-driven optimization and physical insight. Deep learning models, while achieving high pixel-level segmentation precision (high IoU), essentially perform sophisticated texture matching without an understanding of the underlying magnetohydrodynamic (MHD) physics. This creates a semantic gap: a model may successfully segment a filament based on H$\alpha$ absorption features but fail to recognize physically impossible topologies, such as disjointed segments that violate magnetic connectivity constraints. Current algorithms treat filaments as “dark shapes” rather than magnetic flux ropes suspended by Lorentz forces, leading to detections that are visually plausible but physically inconsistent for downstream modeling.

3.5.2. Metric Misalignment: Geometric Accuracy vs. Prognostic Value

The field suffers from “inward-looking” evaluation metrics. The pervasive reliance on generic computer vision metrics (e.g., F1-score, Pixel Accuracy) creates an evaluation system disconnected from scientific utility. In operational forecasting, not all filaments are equal; a small, active-region filament with high chirality is significantly more critical for eruption prediction than a large, quiescent filament. However, standard metrics weight all pixels equally. A model could achieve a 98% F1-score by perfectly segmenting stable quiescent filaments while missing the subtle, pre-eruptive signatures of an active region filament—rendering the model statistically superior but operationally hazardous.

3.5.3. Fragmentation in Benchmarking and Reproducibility

A significant impediment to progress is the lack of standardized evaluation protocols. Performance metrics are highly dependent on specific datasets, which vary wildly in preprocessing, resolution, and annotation standards across studies. This heterogeneity makes cross-study comparisons largely unreliable and hinders meaningful progress assessment.

• The Role of Standardization:Addressing this fragmentation is critical. The introduction of standardized datasets, such as MAGFILO v1.0 [53], represents a pivotal step forward. By providing a consistent benchmark for fair algorithm comparison, MAGFILO v1.0 addresses the critical gap in reproducibility, allowing for rigorous assessment of different architectures on a level playing field. However, until such benchmarks are universally adopted, the community risks a “Tower of Babel” scenario where results are incomparable across different research groups.

3.5.4. The “Instrumental Overfitting” Crisis

Closely related to fragmentation is the issue of instrumental overfitting. Models trained predominantly on limited datasets from major observatories (e.g., strictly BBSO or MLSO) inevitably learn not only filament characteristics but also instrument-specific artifacts and annotator biases. This observational bias manifests as performance degradation when applied to new instruments or extreme solar conditions. The models have not learned “what a filament is,” but rather “what a filament looks like through a specific telescope,” revealing the fundamental fragility of current approaches in the face of domain shifts.

3.5.5. Technological Monoculture and the “Black Box” Barrier

Finally, the overwhelming focus on deep learning architectures, particularly minor variations in U-Net, threatens methodological diversity. While effective, this “black-box” nature complicates validation in operational settings where trust is paramount. Forecasters require not just a binary detection mask, but a confidence interval and an explainable rationale—capabilities that are largely absent in current end-to-end segmentation networks, creating a barrier to their deployment in real-time space weather pipelines.

4. Application-Oriented Challenges and Advances

The transition of solar filament detection from research laboratories to operational environments has revealed several critical challenges that must be addressed for reliable real-world deployment. Simultaneously, significant advances have been made in overcoming these barriers, moving toward more robust and applicable detection systems.

4.1. Cross-Instrument Generalization

A fundamental challenge in operational solar monitoring involves the deployment of detection algorithms across diverse observational platforms with varying instrumental characteristics, resolutions, and imaging conditions.

Challenges: Different solar observatories employ distinct imaging systems with unique characteristics in terms of spatial resolution, spectral bandwidth, temporal cadence, and noise properties. This variability poses significant challenges for detection algorithms trained on data from a single source. Studies by Yuan et al. [20] and Schuh et al. [54] highlighted how performance degradation occurs when methods developed for one instrument are applied to data from another observatory without retraining or adaptation.

Recent Advances: Several approaches have demonstrated improved cross-instrument generalization. Multi-observatory training strategies, where models are trained on combined datasets from multiple sources, have shown promising results. Diercke et al. [27] achieved 92% segmentation accuracy across ChroTel, GONG, and KSO datasets through semi-supervised learning. Similarly, Yuan et al. [20] demonstrated robustness across four different observatories (BBSO, KSO, OACT, YNAO) with accuracy exceeding 96% for filament number detection.

Domain adaptation techniques and data augmentation strategies have also proven effective. Methods incorporating synthetic data generation and style transfer between instruments help bridge the domain gap, improving model generalization without requiring extensive labeled data from each target instrument.

4.2. Real-Time Processing and Computational Efficiency

The operational requirements of space weather forecasting demand detection algorithms capable of processing solar images in near real-time, presenting significant computational challenges, particularly for resource-intensive deep learning models.

Challenges: Early detection methods faced substantial computational bottlenecks. Region growing and morphological operations required significant processing times, with some implementations taking up to one minute per image [2,15]. While classical machine learning approaches improved speeds to 1–3 s per image [19], these remained insufficient for high-cadence monitoring applications.

Recent Advances: Substantial progress has been made in computational optimization. Lightweight deep learning architectures such as Flat U-Net [43] maintain competitive performance (precision ∼0.93) while significantly reducing computational requirements. Optimized implementations of modern architectures achieve inference times under 0.5 s per image [44,52], enabling true real-time processing.

Parallel processing approaches have demonstrated remarkable efficiency gains. Andrijauskas and Gradvohl [49] achieved 326× speed-up for filter-based diffusion methods using hybrid OpenMP-MPI parallelization, supporting high-throughput processing of large solar images. These advances facilitate the integration of detection algorithms into operational solar observatories and space weather monitoring systems.

4.3. Integration with Space Weather Forecasting

The ultimate application of solar filament detection lies in improving space weather prediction capabilities, requiring seamless integration with forecasting pipelines and operational decision-making systems.

Challenges: Significant gaps remain between detection outputs and actionable space weather forecasts. Current systems often operate in isolation, with limited integration between filament detection algorithms and coronal mass ejection (CME) prediction models. The accurate identification of filament disappearance events and their correlation with eruptive phenomena remains challenging, with existing methods showing limited accuracy and reliance on empirical thresholds [17,18].

Current State of Eruption and CME Prediction: Contemporary CME prediction methods can be broadly classified into empirical, statistical, and physics-based approaches. Empirical models rely on observed correlations between filament parameters—such as length, chirality, and magnetic flux—and eruption probability, but are inherently limited by the quality and completeness of input filament catalogs derived from automated detection pipelines [5]. Statistical and machine learning-based methods, including logistic regression and neural network classifiers, have shown promise in identifying pre-eruptive configurations; however, their predictive skill is critically dependent on the accuracy and consistency of upstream filament feature extraction [17,21]. Physics-based approaches, such as MHD simulations of flux rope instability and loss of equilibrium [55], offer the most physically grounded eruption forecasts but require precise filament magnetic field measurements that current automated detection pipelines do not routinely provide. Across all three paradigms, a fundamental bottleneck persists: the reliability of any CME prediction is ultimately bounded by the fidelity of the filament detection and characterization that feeds into it. Improvements in automated detection—particularly in robust physical parameter extraction (tilt angle, magnetic configuration, and kinematic signatures)—can therefore directly enhance the input quality for all classes of prediction models, narrowing the critical gap between raw observations and actionable space weather forecasts.

Recent Advances: Progress has been made in developing end-to-end frameworks that link detection with forecasting. Automated tracking algorithms now enable continuous monitoring of filament evolution with ∼90% accuracy [56], providing crucial temporal context for eruption prediction. Machine learning-based disappearance detection systems have improved the reliability of identifying pre-eruptive signatures [17,18].

The extraction of physically meaningful parameters represents another significant advancement. Methods for quantifying filament area, length, orientation, and kinematic properties provide valuable inputs for empirical and physics-based eruption models [5,57]. These parameters enable statistical analyses across solar cycles, enhancing our understanding of filament behavior and its relationship to space weather events.

4.4. Handling Data Quality Variability

Operational environments must contend with substantial variations in image quality due to atmospheric conditions, instrumental artifacts, and solar activity levels, presenting persistent challenges for reliable detection.

Challenges: Traditional image processing methods are particularly vulnerable to image quality degradation. Noise, limb darkening, atmospheric distortions, and instrumental defects significantly impact detection reliability [5,28]. Early neural networks also exhibited sensitivity to image artifacts and conditions differing from their training data [50,58].

Recent Advances: Deep learning approaches have demonstrated enhanced robustness to image quality variations. Attention mechanisms, as implemented in models like Attention U²-Net [41], enable the network to focus on salient filament features while suppressing irrelevant background noise and artifacts. Multi-scale feature extraction in architectures such as hybrid CNN-Transformer models [42] provides resilience to resolution variations and partial occlusions.

Advanced preprocessing techniques have also evolved substantially. Adaptive normalization methods, learned enhancement operations, and noise-invariant training strategies improve performance under challenging observational conditions without requiring manual parameter tuning.

4.5. Operational Deployment Considerations

The transition from research prototypes to operational systems introduces additional practical considerations including reliability, maintainability, and continuous adaptation requirements.

Challenges: Operational deployment faces hurdles related to model maintenance, continuous performance monitoring, and adaptation to evolving solar conditions across solar cycles. The “black-box” nature of deep learning models complicates validation and trust in operational settings [41,42,59]. Additionally, the computational infrastructure requirements for training and deploying advanced models can be prohibitive for some operational environments.

Recent Advances: Developments in explainable AI and model interpretability are addressing trust and validation concerns. Attention visualization and feature importance analysis provide insights into model decision processes, facilitating operational oversight and validation [41].

Modular system design approaches enable incremental updates and component-level validation. The emergence of standardized interfaces and containerized deployment options simplifies integration with existing observatory infrastructure and forecasting systems.

A synthesized overview of these application-oriented challenges and the corresponding technical advances is presented in

Table 3. Operational Challenges and Technical Solutions.

Challenge Area	Key Challenges	Recent Advances and Solutions
Cross-Instrument Generalization	Performance degradation across different instruments Domain shift between training and deployment data Limited labeled data for new instruments	Multi-observatory training strategies [27] Domain adaptation and transfer learning Data augmentation and synthetic data generation
Real-Time Processing	Computational intensity of advanced algorithms Memory and storage constraints Latency requirements for operational use	Lightweight architectures (e.g., Flat U-Net) [43] Parallel processing and hardware acceleration [49] Model optimization and pruning techniques
Space Weather Integration	Disconnect between detection and forecasting systems Limited accuracy in disappearance event detection Inadequate physical parameter extraction	End-to-end detection and tracking frameworks [18] Machine learning-based eruption prediction [17] Advanced parameter extraction for forecasting models [5]
Data Quality Variability	Sensitivity to noise and artifacts Performance degradation under poor conditions Manual parameter tuning requirements	Attention mechanisms for robust feature selection [41] Multi-scale and noise-invariant architectures [42] Adaptive preprocessing and normalization
Operational Deployment	Model interpretability and validation challenges Maintenance and adaptation requirements Computational infrastructure demands	Explainable AI and visualization techniques [41] Modular system design and containerization Cloud-based and distributed computing solutions

.

4.6. Summary and Outlook

The application-oriented challenges in solar filament detection are being systematically addressed through targeted technological advances. Cross-instrument generalization benefits from multi-source training and domain adaptation techniques. Computational efficiency continues to improve through architectural optimization and parallel processing. Integration with space weather forecasting is advancing through end-to-end frameworks and improved physical parameter extraction.

Despite significant progress, important challenges remain. The development of standardized evaluation protocols across diverse operational scenarios would facilitate more systematic comparison of method performance. Increased focus on model interpretability and physical consistency will enhance trust in operational deployments. Finally, closer collaboration between algorithm developers and space weather forecasters will ensure that detection systems evolve to meet the practical requirements of operational forecasting environments.

5. Future Research Directions

Based on the comprehensive analysis of current limitations and emerging trends in solar filament detection, several critical research directions emerge that promise to advance the field toward more robust, efficient, and physically meaningful detection systems. These directions address fundamental challenges while leveraging new technological opportunities.

5.1. Standardized Datasets and Benchmarking

The lack of standardized, large-scale annotated datasets represents a fundamental limitation in current solar filament detection research. Future efforts should prioritize the development of community-wide benchmark datasets with consistent labeling protocols across multiple solar cycles and observational platforms [60].

Key Initiatives:

Establishing curated datasets that encompass diverse solar activity conditions, multiple instruments (H$\alpha$, EUV, magnetograms), and comprehensive filament annotations would significantly enhance model generalization and enable fair comparison of different approaches. These datasets should include challenging cases such as limb filaments, faint structures, and eruption precursors to drive algorithmic improvements. Community-driven annotation efforts and open data sharing initiatives, similar to those pioneered by Mouradian [3] but scaled for modern requirements, will be essential for creating these resources.

Evaluation Frameworks: Developing standardized evaluation metrics and protocols is equally crucial. Current studies employ diverse performance measures, making direct comparisons difficult. Future work should establish comprehensive evaluation frameworks that assess not only segmentation accuracy but also computational efficiency, robustness to image quality variations, and generalization across instruments. Regular benchmarking challenges, similar to those in computer vision, could accelerate progress by focusing community efforts on common goals.

5.2. Towards Trustworthy and Edge-Deployable AI Systems

As deep learning architectures grow in complexity, the gap between computational demands and operational constraints widens. Future research must pivot from purely accuracy-driven metrics to developing systems that are both resource-efficient for spaceborne deployment and scientifically trustworthy through physical grounding.

On-board Intelligence and Edge Computing: The current paradigm of “downlink-then-process” is bottlenecked by limited telemetry bandwidth in deep-space missions (e.g., Solar Orbiter). Future detection systems must prioritize Edge AI, enabling real-time filament detection directly on spacecraft.

• Technological Pathway: Research should move beyond simple architecture scaling (e.g., Flat U-Net [43]) to advanced model compression techniques such as network quantization (reducing precision from 32-bit float to 8-bit integer) and knowledge distillation (training lightweight “student” models to mimic heavy “teacher” networks). This would facilitate “smart telemetry,” where spacecraft downlink only metadata or event-triggered image cutouts of active filaments, significantly optimizing data budgets.

Physics-Informed Neural Networks (PINNs): To overcome the “black-box” limitation, the next generation of models must incorporate inductive biases derived from solar physics. Instead of learning purely from pixel statistics, Physics-Informed Neural Networks (PINNs) can embed physical laws directly into the training process.

• Critical Insight: Future architectures could incorporate a “physics loss” function (${\mathcal{L}}_{phy}$) alongside standard segmentation loss. For instance, imposing constraints based on magnetic connectivity or enforcing topological continuity can penalize the model for predicting physically impossible fragmented structures. This ensures that the detected filaments are not just visually similar to the ground truth but are consistent with underlying MHD principles.

Explainability and Uncertainty Quantification: Trust is the prerequisite for operational adoption. To demystify the decision-making process of deep models, a dual approach is required:

• Visual Explainability (XAI): Techniques such as attention mechanisms [41,61], provide critical insights by visualizing which image regions drive the model’s decisions. Future work should advance from simple attention maps to sophisticated feature attribution methods (e.g., Integrated Gradients) that can verify whether the model is focusing on physical filament structures or irrelevant background artifacts.
• Uncertainty Quantification (UQ): Beyond knowing where the model looks, operators need to know how confident it is. Integrating Bayesian Neural Networks or Monte Carlo Dropout allows systems to generate probability heatmaps. By distinguishing between aleatoric uncertainty (data noise) and epistemic uncertainty (model ignorance), these systems provide forecasters with actionable risk assessments—flagging low-confidence detections for human review while automating high-confidence cases.

5.3. Multi-Modal Data Integration

Current detection methods predominantly focus on single-wavelength observations, particularly H$\alpha$ images. Future research should exploit the complementary information available across multiple observational modalities.

Multi-Wavelength Fusion: Developing frameworks that integrate H$\alpha$, EUV, magnetogram, and spectroscopic data will improve filament discrimination and characterization. As noted by Scholl and Habbal [62], multi-wavelength approaches can better distinguish filaments from other solar features and provide more comprehensive physical parameter extraction. Deep learning architectures specifically designed for multi-modal data fusion, potentially incorporating cross-attention mechanisms and shared representation learning, represent a promising research direction.

Magnetic Context Integration: Incorporating magnetic field information represents a particularly valuable extension. Filament formation and stability are intimately connected to the underlying magnetic field structure. Future detection systems should integrate vector magnetogram data to improve eruption prediction and enable more physically grounded tracking of filament evolution.

5.4. Advanced Tracking and Eruption Prediction

While current systems demonstrate competent detection capabilities, significant advances are needed in temporal analysis and predictive modeling.

Disappearance and Eruption Detection: Fully automated, accurate detection of filament disappearance and eruption events remains challenging. Future research should develop machine learning approaches that integrate temporal evolution patterns, kinematic signatures, and magnetic context to improve eruption prediction reliability. As highlighted by Atoum [17] and Zheng et al. [18], these capabilities are crucial for space weather forecasting but currently suffer from limited accuracy and reliance on empirical thresholds.

Long-term Evolution Modeling: Beyond immediate eruption prediction, understanding filament lifecycle evolution across multiple solar rotations represents an important research direction. Developing models that can track filament development, rotation, and interaction with other solar features would provide valuable insights for both fundamental solar physics and operational forecasting.

5.5. Generalized and Adaptive Detection

Ensuring reliable performance across diverse solar activity conditions and throughout the solar cycle remains a significant challenge.

Adaptation to Solar Cycle Variations: Future detection systems must maintain performance across different phases of the solar cycle, from minimum to maximum activity. Research should focus on developing adaptive algorithms that can automatically adjust to changing solar conditions without manual intervention. Continual learning approaches and domain adaptation techniques specifically designed for solar cycle variations represent promising directions.

Handling Extreme Conditions: Developing robustness to challenging observational conditions, including high noise levels, atmospheric distortions, and instrumental artifacts, requires continued attention. Future work should explore noise-invariant training strategies, self-supervised learning approaches that leverage unlabeled data from difficult conditions, and robust preprocessing techniques that require minimal parameter tuning.

5.6. Integration with Operational Forecasting

Translating detection advances into operational space weather forecasting capabilities represents the ultimate application goal.

End-to-End Forecasting Systems: Future research should develop integrated frameworks that connect filament detection directly with space weather prediction models. These systems should automatically extract physically meaningful parameters (orientation, magnetic configuration, kinematic properties) and feed them into eruption probability models and CME propagation forecasts. The work by Zheng et al. [18] represents an initial step in this direction, but more sophisticated integration is needed.

Operational Deployment Considerations: Research should address practical deployment challenges including computational infrastructure, real-time processing constraints, and system reliability requirements. Developing containerized, scalable detection systems that can be easily integrated with existing observatory infrastructure and forecasting pipelines will facilitate operational adoption.

5.7. Physical Parameter Extraction

Beyond binary segmentation, transforming detection outputs into quantitative physical parameters represents a critical step connecting pixel-level results to scientific analysis and operational forecasting.

Current Limitations: Most existing detection systems output only segmentation masks, leaving the extraction of physically meaningful quantities—such as filament tilt angles, chirality, spine length, area, magnetic flux, and line-of-sight velocity fields—to separate, often manual post-processing steps. This decoupling between detection and physical characterization hinders end-to-end automation and limits the immediate utility of detection results in physics-based eruption models [5,18].

Promising Directions: Future detection frameworks should evolve from pure segmentation tasks to multi-task learning systems that simultaneously produce detection masks alongside quantitative physical parameters. Architectures incorporating physics-aware decoder heads—specialized branches for regression or classification of physical quantities—can transform detection from a pattern recognition problem into a quantitative measurement tool.

• Integrated Parameter Extraction: Multi-task networks can jointly predict filament masks, spine coordinates, chirality classification, and kinematic properties in a single forward pass. This approach not only improves computational efficiency but also enables the network to learn shared representations that mutually benefit segmentation and parameter estimation tasks.
• Metrics for Physical Consistency: Establishing evaluation protocols beyond standard segmentation metrics—such as tilt angle error relative to manual measurements, chirality agreement rates, and velocity field consistency—will incentivize the development of scientifically meaningful detection systems.
• Large-Scale Statistical Applications: Automated parameter extraction pipelines enable comprehensive statistical studies of filament properties across solar cycles, facilitating investigations into hemispheric asymmetry patterns, poleward migration trends, and the identification of pre-eruptive signatures critical for space weather forecasting [5,57].

5.8. Historical Data Utilization and Long-Term Studies

The vast archives of historical solar data remain underutilized but offer unique opportunities for long-term studies.

Historical Data Digitization: Specialized pipelines for digitizing and analyzing historical records, such as the hand-drawn suncharts studied by Priyadarshi et al. [57], would extend our understanding of filament behavior across multiple solar cycles. Machine learning approaches tailored to historical data characteristics (variable quality, different observational techniques) could unlock century-long datasets for comprehensive solar cycle studies.

Cross-Century Analysis: Integrating contemporary detections with historical records would enable unprecedented studies of solar cycle evolution and long-term trends in filament behavior. These analyses could reveal connections between filament properties and broader solar magnetic cycle patterns, potentially improving our understanding of solar dynamo processes.

5.9. Cross-Disciplinary Opportunities: The Era of Foundation Models

The convergence of solar physics with general artificial intelligence offers transformative potential, particularly with the emergence of large-scale Foundation Models (FMs).

From Specialized Networks to Foundation Models: The field is witnessing a paradigm shift from training task-specific models (e.g., U-Net trained from scratch) to adapting generalist Foundation Models.

• Zero-Shot Segmentation: Models such as the Segment Anything Model (SAM) have demonstrated remarkable zero-shot generalization capabilities. Future research should investigate the applicability of SAM-based architectures for solar filament detection. By utilizing prompt engineering (e.g., providing box prompts or point prompts of filaments) rather than full retraining, researchers could bypass the bottleneck of massive annotated datasets.
• Vision-Language Integration: The rise of Multi-modal Large Language Models (MLLMs) offers a novel pathway for semantic understanding. Integrating visual encoders with language models could enable systems that not only segment filaments but also generate natural language descriptions of their morphology (“sigmoid,” “barb-bearing”) or query databases using semantic prompts (“Show me all active region filaments with high chirality”), effectively bridging the gap between pixel data and scientific query.

Domain Adaptation and “Solar-Foundations”: While terrestrial foundation models are powerful, they lack domain-specific knowledge of plasma physics and magnetic topologies. A critical cross-disciplinary opportunity lies in developing Solar Foundation Models. By pre-training large-scale transformers on petabytes of unlabeled multi-wavelength solar data (via self-supervised learning strategies like Masked Image Modeling), the community can create a robust backbone specifically for the sun. This “Solar-Encoder” could then be efficiently fine-tuned for various downstream tasks, including filament detection, CME tracking, and flare prediction, with minimal labeled examples.

Stellar Astronomy Extension: The methodologies developed for solar filament detection hold significant potential for stellar astronomy. By adapting these techniques to stellar observational data, researchers can extend filament studies to other stars, enabling comparative analysis of stellar activity and magnetic phenomena. This “Sun-as-a-star” approach represents a promising frontier, connecting solar physics with stellar astrophysics through shared detection and analysis frameworks.

The future of solar filament detection lies in developing intelligent, adaptive systems that combine physical understanding with data-driven approaches. By addressing these research priorities, the field can move toward more reliable, efficient, and physically meaningful detection capabilities that significantly enhance both scientific understanding and operational space weather forecasting.

A comprehensive summary of the future research directions outlined above, including their key objectives and relative priority, is provided in

Table 4. Enhanced Future Research Directions for Solar Filament Detection.

Research Area	Key Objectives & Critical Insights	Priority
Standardized Data & Benchmarks	Critical Insight: Address dataset bias and evaluation metric limitations; shift from single-instrument optimization to cross-domain generalization Develop multi-instrument, multi-cycle, multi-wavelength benchmark datasets spanning diverse solar conditions Establish physics-utility evaluation protocols that prioritize scientifically critical cases (eruption precursors, active region filaments) Create challenge datasets specifically targeting generalization across observatories and solar cycles Implement standardized annotation protocols to minimize subjective biases in ground truth data	High
Trustworthy & Physics-Informed AI	Critical Insight: Bridge data-driven optimization with physical understanding; develop systems that are both resource-efficient for spaceborne deployment and scientifically trustworthy through physical grounding Develop physics-constrained neural networks incorporating MHD principles and magnetic connectivity constraints as physics loss functions Advance model compression (network quantization, knowledge distillation) for on-board edge deployment and smart telemetry in deep-space missions Integrate explainable AI (XAI) and uncertainty quantification (Bayesian methods, Monte Carlo Dropout) to provide confidence intervals alongside detection masks Implement feature attribution techniques to verify that models focus on physical filament structures rather than instrumental artifacts	High
Generalized & Adaptive Detection	Critical Insight: Move beyond one-size-fits-all approaches; enable robust performance across diverse observational conditions Develop semi-supervised and self-supervised learning frameworks to address annotation scarcity Design context-aware models that adapt complexity to specific task requirements and data characteristics Create lightweight architectures for resource-constrained environments and real-time applications Build systems that maintain performance across different solar cycles and activity levels	High
Multi-Modal Data Fusion	Critical Insight: Leverage complementary physical information across wavelengths; enhance detection reliability and physical understanding Integrate H$\alpha$, EUV, magnetogram, and spectroscopic data using cross-modal attention mechanisms Incorporate magnetic field topology and evolution context for improved eruption prediction Develop unified frameworks that simultaneously analyze multiple observational modalities Enable comprehensive filament characterization through synergistic multi-wavelength analysis	Medium
Advanced Tracking & Eruption Prediction	Critical Insight: Move beyond static detection to temporal understanding; enable reliable identification of pre-eruptive signatures and lifecycle evolution Improve disappearance and eruption detection by incorporating temporal evolution patterns, kinematic signatures, and magnetic context Develop long-term evolution models that track filament development, rotation, and interaction across multiple solar rotations Integrate machine learning approaches to replace empirical thresholds in eruption prediction with data-driven decision criteria	High
Operational Forecasting Integration	Critical Insight: Translate detection advances into actionable space weather forecasts; connect research outputs directly to operational decision-making pipelines Develop end-to-end frameworks that connect filament detection directly with CME propagation and geomagnetic storm prediction models Create containerized, scalable detection systems optimized for deployment in operational space weather monitoring pipelines Establish real-time processing capabilities that balance detection accuracy with the latency constraints of operational forecasting environments	High
Physical Parameter Extraction	Critical Insight: Transform detection from segmentation task to quantitative physical analysis; enable direct scientific insights Develop methods that simultaneously output detection masks and quantitative physical parameters (tilt angles, velocity fields, magnetic properties) Create evaluation metrics that assess parameter extraction accuracy and physical consistency Build integrated pipelines that seamlessly connect detection with subsequent physical analysis Enable statistical studies of filament properties across solar cycles through automated parameter quantification	High
Historical Data Analysis	Critical Insight: Extend temporal coverage for long-term studies; leverage century-scale datasets for solar cycle analysis Develop specialized machine learning pipelines for historical data digitization and analysis Create techniques adapted to hand-drawn synoptic charts and early photographic plates Conduct cross-century filament behavior studies to understand long-term solar activity patterns Leverage century-long datasets to analyze filament evolution across multiple solar cycles	Medium
Foundation Models & Cross-Disciplinary Applications	Critical Insight: Leverage large-scale pre-trained models and multi-modal learning to overcome data scarcity; extend solar-validated methodologies to stellar observations Investigate zero-shot segmentation via foundation models (e.g., SAM) with prompt engineering to bypass the annotation bottleneck Explore vision-language integration (MLLMs) for semantic filament description and natural-language database querying Develop Solar Foundation Models through self-supervised pre-training on petabyte-scale multi-wavelength solar archives Extend detection methodologies to stellar astronomy for comparative studies of magnetic activity across different stellar types	Medium

.

6. Overall Synthesis and Conclusions

This comprehensive review has systematically traced the remarkable evolution of solar filament detection methods, charting a clear trajectory from labor-intensive manual approaches to increasingly sophisticated intelligent systems. This journey from manual to intelligent detection represents not merely a technological shift, but a fundamental transformation in how we observe, analyze, and understand solar phenomena.

Historical Progression and Current State: The foundational era of traditional image processing established the basic principles of automated detection, with techniques such as thresholding, region growing, and morphological operations providing the first steps toward automation. While demonstrating feasibility, these methods were constrained by sensitivity to image quality variations, requirement for manual parameter tuning, and limited robustness across diverse observational conditions. The integration of classical machine learning marked a significant paradigm shift, introducing learned feature discrimination and classification capabilities through artificial neural networks and support vector machines. The current deep learning revolution represents the most transformative development, with architectures such as U-Net variants and Mask R-CNN achieving unprecedented segmentation accuracy through end-to-end learning while demonstrating superior robustness.

Throughout this evolutionary progression, consistent patterns emerge: detection accuracy has improved from moderate levels (72–85%) to consistently high performance (90%+), computational efficiency has advanced from minutes to sub-second inference times, and validation scope has expanded from single-observatory testing to comprehensive multi-instrument evaluations.

Persistent Challenges: Despite these advances, significant challenges persist in translating technological sophistication into reliable operational systems. Key limitations include the scarcity of standardized annotated datasets, substantial computational demands of advanced models, limited cross-instrument generalization, and the “black-box” nature of deep learning. The integration of detection outputs with space weather forecasting models remains incomplete, constraining the practical impact on operational forecasting capabilities. The field also faces fundamental issues including the disconnection between data-driven optimization and physical insight, inward-looking evaluation metrics, and the risk of technical path monoculture.

Future Outlook: Looking forward, the field stands at a critical juncture where technological sophistication must be matched by practical applicability. Priority research directions include developing standardized benchmarks and datasets, creating efficient and interpretable AI systems, advancing multi-modal data integration, improving eruption prediction capabilities, and addressing operational deployment constraints. The ultimate goal remains the development of intelligent, adaptive detection systems that combine physical understanding with data-driven approaches to provide reliable, timely information for space weather forecasting.

The evolution from manual to intelligent detection reflects broader trends in solar physics and computational science, illustrating how advances in machine learning and artificial intelligence are transforming our ability to extract knowledge from complex observational data. This progression from human expertise to algorithmic efficiency to learned intelligence represents not just a technical achievement, but a fundamental expansion of our observational capabilities. The continued convergence of solar physics with advanced computational techniques points toward a future where intelligent detection systems operate as seamless extensions of scientific inquiry, transforming vast data streams into actionable knowledge for both research and operational applications in space weather prediction.