Global Ionospheric F-layer Electron Density Prediction Based on Multiple Radio Occultation Data using Attention-Based Deep Learning Model

Understanding low-latitude F-layer ionosphericelectron density (Ne) under severe geomagnetic conditionsis crucial for various GNSS applications. Existingionospheric models utilizing machine learning (ML) havestruggled to accurately capture the complex dynamics ofNe, particularly under extreme geomagnetic conditions. Inthis study, we propose the Attention-Based RecurrentResNet18 (ABRR-18) model to predict ionospheric Neusing radio occultation data obtained from multiplesatellite missions between 2002 and 2023. The proposedmodel integrates ResNet18 and Bi-LSTM with a spatialattention mechanism. Besides, it incorporates variousspace weather indicators such as solar flux, sunspotnumber, disturbance storm time, and interplanetarymagnetic field. Experimental results revealed thatABRR-18 outperformed other applied models, such asANN-IRI, ANN-TDD, LSBoost, Bi-LSTM, andAlexNet-Bi-LSTM-SAM, achieving a correlation of 0.9674and a root mean square error of(Formula presented). ABRR-18 showed superiorperformance under severe geomagnetic conditions andduring high solar activity years over the IRI-2016 model.Additionally, the ABRR-18 model outperforms theIRI-2016 and IRI-2020 models, with predictions closelyaligning with incoherent scatter radar observations,particularly during extreme conditions. Compared to theinternational reference ionosphere model (IRI-2016 and IRI-2020), ABRR-18 demonstrated superior accuracy incharacterizing global ionospheric spatial-temporalproperties. This study underscores the potential of DLtechniques in ionospheric modeling by exhibiting superiorperformance. The ABRR-18 model introduces aninnovative approach, offering notable advancements incomprehending and predicting ionospheric Ne inchallenging conditions.