Filling ionospheric monthly medians missing data: a machine learning approach

Abstract

Ionospheric databases, like most time series of experimental parameters, contain missing data owing to either natural or technical reasons, limiting the statistical methods that can be used to analyse these series. In this work, two machine learning (ML) techniques, gradient boosting regression (GBR) and random forest regression (RFR), are applied to estimate missing values in the hourly monthly median foF2 datasets from eight ionospheric stations, covering the period 1975–2022. Their performance is compared with that of traditional linear and multiple regressions using comparative statistical indices, which indicate that ML techniques outperform conventional methods, reaching an explained variance of 97%, and with absolute errors between 40 and 60% smaller, on average, than linear and multiple regressions. These findings highlight the efficiency of ML in reconstructing ionospheric datasets, enhancing the reliability and continuity of long-term ionospheric observations.