Understanding low-latitude F-layer ionospheric electron density (Ne) under severe geomagnetic conditions is crucial for various GNSS applications. Existing ionospheric models utilizing machine learning (ML) have struggled to accurately capture the complex dynamics of Ne, particularly under extreme geomagnetic conditions. In this study, we propose the attention-based recurrent ResNet18 (ABRR-18) model to predict ionospheric Ne using radio occultation (RO) data obtained from multiple satellite missions between 2002 and 2023. The proposed model integrates ResNet18 and bidirectional-long short-term memory (Bi-LSTM) with a spatial attention mechanism (SAM). Besides, it incorporates various space weather indicators such as solar flux, sunspot number, disturbance storm time, and interplanetary magnetic field (IMF). Experimental results revealed that ABRR-18 outperformed other applied models, such as artificial neural network (ANN)-international reference ionosphere (IRI), ANN-TDD, least-squares boosting (LSBoost), Bi-LSTM, and AlexNet-Bi-LSTM-SAM, achieving a correlation of 0.9674 and a root-mean-square error (RMSE) of$1.0295 \times 10^{5}$ele/cm3. ABRR-18 showed superior performance under severe geomagnetic conditions and during high solar activity years over the IRI-2016 model. In addition, the ABRR-18 model outperforms the IRI-2016 and IRI-2020 models, with predictions closely aligning with incoherent scatter radar (ISR) observations, particularly during extreme conditions. Compared to the IRI model (IRI-2016 and IRI-2020), ABRR-18 demonstrated superior accuracy in characterizing global ionospheric spatial–temporal properties. This study underscores the potential of deep learning (DL) techniques in ionospheric modeling by exhibiting superior performance. The ABRR-18 model introduces an innovative approach, offering notable advancements in comprehending and predicting ionospheric Ne in challenging conditions.

Medium-scale traveling ionospheric disturbances (MSTIDs) are periodic wave-like structures in the ionosphere that can significantly alter the local ionospheric conditions leading to the performance degradation of the radio wave communication and satellite navigation. Due to their complex origins and evolving dynamics, traditional monitoring methods often fail to effectively extract the characteristic features of MSTIDs. Leveraging advances in deep learning, this study proposes the state-of-the-art MSTID intelligent recognition and propagation parameter inversion method based on the “You Only Look Once” (YOLO) series models. The method consists of three main stages: MSTID target detection, MSTID blob instance segmentation, and inter-frame matching of segmented blobs. A dataset comprising 3,422 annotated images for target detection and 236 images for instance segmentation was constructed using Differential Total Electron Content (DTEC) maps from Japan’s GEONET network under the different solar activity conditions. The method automatically extracts key propagation parameters from consecutive image pairs including velocity, azimuth, wavelength, period, and coverage area. Experimental results show that YOLO v9m achieves the highest detection accuracy (78.34%) for MSTID targets, while YOLO v8m-seg excels in instance segmentation (95.58%). All of the models satisfy the real-time processing requirements. Blob features including color, centroid, area, Hu moments and topology, are extracted to compute feature difference scores across consecutive frames for the optimal matches determined by the Hungarian algorithm. Ellipse fitting and pixel-to-geographic coordinate conversion are then employed to calculate propagation parameters. Comparative validation with the traditional three-station cross-spectral and keogram methods demonstrates good consistency, with errors in parameters such as propagation velocity and azimuth within 20%. This AI-based approach offers a promising solution for the advancing intelligent, accurate, and real-time ionospheric disturbance detection.

Accurate and timely forecasting of ionospheric irregularities is of great significance for the reliable and stable operation of global high-precision communication and navigation systems at low latitudes. In this study, we implement a hybrid ensemble model (HEM) that combines multiple machine learning models for forecasting the occurrence and intensity of ionospheric irregularities instead of considering ionospheric irregularity forecasting as classifications. Meanwhile, this model is trained with the GNSS-derived rate of total electron content index (ROTI) maps from approximately 147 ground-based global navigation satellite systems (GNSS) receivers in Brazil sector (35° S–5° N, 30° W–75° W). Meanwhile, a diverse set of input features, including interplanetary magnetic field (IMF) components, F2 layer critical frequency (foF2), peak height F2-layer (hmF2), F10.7, flow pressure, and SYM-H indices, are carefully selected during January 1, 2022 to 15 October 31, 2022. Regarding the relative importance of various input features, results demonstrate that the performance of the HEM model trained by the ROTI and hmF2 observations for predicting ionospheric irregularities is superior to that of other input features. Furthermore, the deviations of forecasting ionospheric irregularities from the HEM model occur mainly in the southern equatorial ionization anomaly (EIA) regions, and the accuracy of the HEM model with daily standard deviation (STD) and root mean square (rms) is less than 0.1 TECU/min. Hence, the HEM model is more stable and greater than other predicted models. Additionally, the HEM algorithm can forecast the ionospheric irregularity structures and intensity for 30 min over Brazilian territory. It is expected to improve the accuracy of short-term ionospheric irregularities forecasting at low latitudes.

GNSS (Global Navigation Satellite System) observation insufficiency limits the development of the voxel-based computerized ionospheric tomography (CIT) technique. Electron densities of voxels without observation cannot be accurately estimated by the commonly used algebraic reconstruction techniques. In this study, we proposed a compressed sensing technique (CST) based on dictionary learning for ionospheric tomography. Specifically, the K-SVD (singular value decomposition) algorithm was used for dictionary learning based on training sets that are derived from the NeQuick model, hereafter referred to as the CST_NeQuick algorithm. K-SVD uses the orthogonal matching pursuit (OMP) for sparse coding and the SVD approach for dictionary updating. Both simulations and real experiments demonstrated the feasibility and superiority of the CST algorithm when compared to the widely used multiplicative algebraic reconstruction technique (MART). It was found that the CST_NeQuick algorithm’s tomographic performances were mostly superior to those of the MART algorithm in comparison with the independent slant total electron content (STEC) references. Another CST-based tomographic experiment was performed by using the MART-based solutions for dictionary learning, hereafter referred to as the CST_MART algorithm. It showed that the CST_MART algorithm can reduce the average root mean square (RMS) of the CIT-derived STEC by 37.3 % and 20.2 %, respectively, when compared to the MART and CST_NeQuick algorithms. Besides, the CST_MART algorithm’s electron density profiles also showed more agreement with the electron density profiles that were derived from radio occultation data.

Ionospheric monitoring and modeling have been difficult for a long time over the data-void or data-sparse oceans. As an emerging remote sensing technique, GNSS reflectometry (GNSS-R) has presented great potential in ionosphere sounding over these regions. However, the conventional approach to generate delay-Doppler-map (DDM) involved in the GNSS-R total electron content (TEC) retrieval process ignores the effects of tropospheric delay and the topside ionospheric delay above the GNSS-R receiver. This would cause certain errors in retrieved TEC results. In this contribution, an improved method to estimate ionospheric TEC over oceans using the GNSS-R technique is proposed, which considers the influence of the tropospheric delays and the topside ionospheric delays above the spaceborne GNSS-R receiver. To achieve the best matching between measured and simulated DDM, this paper employs the least squares (LS) fitting method for elastic matching. Additionally, the assessment was performed in May 2015 and 2017 at different solar activities, by comparing the ionospheric TEC derived from our proposed method with that from two ionospheric empirical models (NeQuick2 and IRI-2016), the Global Ionospheric Maps (GIMs) final products, as well as the measured GNSS TEC. The results show good consistency between these models. Meanwhile, when considering the topside ionospheric delay and tropospheric delay in DDM, the TEC accuracy has significantly improved. Especially, the improvements of root mean square error (RMS) can reach 5.3% and 23.5% during high and low solar activities, respectively, versus GNSS TEC. It is expected to benefit the application of GNSS-R in ionospheric modeling and application over the ocean area.

In this work, a three-step algorithm for the tomographic reconstruction of ionospheric electron density (IED) is proposed. In the new algorithm, the Taylor series expansion of the multiplicative algebraic reconstruction technique is first introduced to perform an adaptive adjustment for the relaxation parameter vector. Second, the horizontal and vertical constraints are imposed on the tomography system aiming to reduce the dependence on the initial iteration values for those voxels without any rays traversing them. Third, the uniformed voxel size is replaced with the variable voxel size by adopting the unequal grid interval in the altitudinal direction. The feasibility and superiority of the three-step algorithm are validated by devising a numerical simulation scheme. Finally, the new algorithm is successfully used to obtain the 3-D IED distribution under magnetically quiet conditions. The comparisons of IED profiles and the statistics of the reconstruction error of different algorithms further validate the superiority of the three-step algorithm.