Global ionospheric sporadic E intensity prediction from GNSS RO using a novel stacking machine learning method incorporated with physical observations

Sporadic E (Es) layers, the irregularities of enhanced electron density commonly occurring in the ionospheric E region, are affected by the interactions between distinct atmospheric layers. Es intensity (EsI) is a crucial parameter to describe Es layer characteristics, while there still lacks the method for high-precision EsI prediction due to its complex spatiotemporal variation and physical driving mechanisms. We propose a novel stacking machine learning (SML) method for global EsI prediction, which combines the advantages from different ML models to obtain better performance than using a single ML model. Various Es-related physical observations, including vertical ion convergence, gravity wave potential energy, and solar and geomagnetic indices, are incorporated as the inputs of SML together with the EsI derived from global navigation satellite system (GNSS) radio occultation (RO) measurements. SML performs well in both long-term and short-term EsI prediction and characteristics reconstruction. SML-predicted EsI is in good agreement with GNSS RO-derived EsI, with the mean error (ME) of 0.032 TECU km-1 and root mean square error (RMSE) of 0.158 TECU km-1. Taking ionosonde observations as reference, SML has the RMSE of 1.064 MHz, which is reduced by 20.1 %–40.5% compared to existing prediction methods. The higher accuracy of our method than methods not incorporating physical observations illustrates the significance of considering multiple related physical factors when constructing the Es prediction model. The proposed method can be expected to provide valuable information for not only ionospheric irregularities monitoring and space weather forecasting but also the mechanisms of Es layer formation and atmospheric coupling.