Machine Learning‐Based Correction of Reanalysis Surface Radiation Fluxes Using Ground‐Measured Data in the Southern Brazilian Pampa Biome
International Journal of Climatology · 15.11.2026 · 46 (13)
DOI: 10.1002/joc.70528 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Climatology, online erschienen am 05.08.2026. Verfasst von Olusola Samuel Ojo und 5 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Universidade Federal de Santa Maria.
Abstract (Originalsprache)
Surface radiation fluxes are key drivers of land–atmosphere exchanges and climate variability. This study assesses the performance of ERA‐5 and MERRA‐2 surface radiation products using hourly observations from 2014–2023 at three stations (Santa Maria, Aceguá, and Pedras Altas) in the Southern Brazilian Pampa, considering incoming shortwave radiation (S↓), reflected shortwave radiation (S↑), atmospheric longwave radiation (L↓), surface longwave radiation (L↑), and net radiation ( Q *). The equivalent datasets for the three stations were also obtained from the European Centre for Medium‐Range Weather Forecasts Reanalysis 5th Generation (ERA‐5) and Modern‐Era Retrospective Analysis for Research and Applications, Version 2 (MERRA‐2). Statistical comparative evaluation was assessed using correlation and error‐based metrics. Analyses showed that MERRA‐2 overestimated S↓ with a peak bias of 35.92 W m −2 (20.07%) at Pedras Altas, whereas ERA‐5 showed lower biases (< 10%). Daily analyses revealed consistent U‐shaped error structures, with ERA‐5 producing improved performance (e.g., S↓ bias = −6.61 W m −2 (3.2%) at Aceguá; 4.01 W m −2 (1.9%) at Santa Maria; and 11.27 W m −2 (5.8%)). Monthly results demonstrated that ERA‐5 also outperformed MERRA‐2 for all radiation components, with minimum S↓ RMSE of 5.94 W m −2 (3.1%) and bias of 4.05 W m −2 (2.1%) at Santa Maria. Violin plots further indicated narrower bias distributions for ERA‐5, particularly for S↑ and Q *. A bivariate regression model was developed for gap‐filling using reanalysis predictors and refined using Random Forest (RF), K‐Nearest Neighbours (KNN), and Multilayer Perceptron (MLP) models. The regression achieved high predictive skill ( R 2 = 0.986 for S↓ and 0.983 for Q *), while RF provided the best performance ( R 2 = 0.9995 for S↓ and 0.9928 for Q * at Santa Maria), demonstrating strong potential for radiation flux reconstruction in subtropical grassland environments.
Förderung
- Universidade Federal de Santa Maria
Zitieren
APA
Ojo, O. S., Stefanello, M. B., Mergen, A., Pinheiro, M. E. O., Souza, V. D. A., & Roberti, D. R. (2026). Machine Learning‐Based Correction of Reanalysis Surface Radiation Fluxes Using Ground‐Measured Data in the Southern Brazilian Pampa Biome. International Journal of Climatology, 46(13). https://doi.org/10.1002/joc.70528
Vancouver
Ojo OS, Stefanello MB, Mergen A, Pinheiro MEO, Souza VDA, Roberti DR. Machine Learning‐Based Correction of Reanalysis Surface Radiation Fluxes Using Ground‐Measured Data in the Southern Brazilian Pampa Biome. International Journal of Climatology. 2026;46(13). doi:10.1002/joc.70528
BibTeX
@article{Ojo2026,
author = {Ojo, Olusola Samuel and Stefanello, Michel Baptistella and Mergen, Alecsander and Pinheiro, Maria Eduarda Oliveira and Souza, Vanessa de Arruda and Roberti, Débora Regina},
title = {Machine Learning‐Based Correction of Reanalysis Surface Radiation Fluxes Using Ground‐Measured Data in the Southern Brazilian Pampa Biome},
journal = {International Journal of Climatology},
year = {2026},
volume = {46},
number = {13},
doi = {10.1002/joc.70528}
}
EUVIMED