A novel framework for gap-filling meteorological time series data

A novel ensemble gap‑filling framework is introduced, integrating two complementary components: (1) a bias‑correction protocol tailored for applying satellite-derived data from the ECMWF Reanalysis 5th Generation Land Component (ERA5-Land) in semi‑arid regions and (2) a machine learning (ML) – based gap‑filling system designed to reconstruct missing ground observations using ensemble predictive modelling. Four bias correction statistical methods applied were broadly comparable; however, Linear Scaling (LS) consistently demonstrated higher accuracy for most indices across stations. Through the assessment of seven ML algorithms, tree‑based and boosting models achieved higher accuracy at inland stations, while linear models performed competitively at both coastal and inland stations. This spatial heterogeneity underscores the importance of adaptive ensemble selection to ensure robust and geographically consistent predictions of climate variables. A voting ensemble ML model for each variable demonstrated strong predictive skill for temperature variables, with R² values consistently exceeding 0.9 across most stations. Moreover, sea level pressure predictions demonstrated exceptional accuracy across all sites, with R² consistently above 0.98. For dew point temperature, performance was lower at inland climates, such as Kharga and Dakhla stations, where R² values were 0.85 and 0.87, respectively. Wind speed prediction proved most challenging, with several stations recording R² values below 0.75 and Asyut showing the weakest performance (R² = 0.47). The proposed methodology effectively addressed gaps and inconsistencies in long‑term meteorological records, producing high‑quality, continuous time series of climate variables from 1990 to 2020 across 13 stations in Egypt, to support research in climate change impact assessment, hydrological modelling, and agricultural development planning.