Damiani et al. (2025) Spatially generalizable bias correction of satellite solar radiation for regional climate assessment—a case study in Japan

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2025
• Date: 2025-11-14
• Authors: Alessandro Damiani, Noriko N. Ishizaki, Tomomi Watanabe, Yuta Tamaki, Raúl R. Cordero, Sarah Féron, Hitoshi Irie
• DOI: 10.1016/j.jag.2025.104947

Research Groups

• National Institute for Environmental Studies, Tsukuba, Japan
• Universidad de Santiago de Chile, Santiago, Chile
• University of Groningen, Leeuwarden, the Netherlands
• Chiba University, Chiba, Japan

Short Summary

This study develops a physics-informed eXtreme Gradient Boosting (XGBoost) model to bias-correct Himawari satellite surface solar radiation (SSR) estimates over Japan, achieving significant improvements in accuracy and spatial consistency, particularly over snow-covered and complex terrain, and uses the corrected data to evaluate a regional reanalysis.

Objective

• To evaluate JAXA’s Himawari satellite surface solar radiation (SSR) against ground-based observations, then combine sparse ground measurements with high-resolution meteorological data to train an eXtreme Gradient Boosting (XGBoost) model that captures and spatiotemporally extrapolates systematic bias between satellite SSR and ground truth data, and finally use the bias-corrected satellite SSR data to assess a novel high-resolution regional reanalysis.

Study Configuration

• Spatial Scale: Central Japan, with data at various resolutions (1 km × 1 km, 0.025° × 0.025°, 5 km × 5 km) unified to 0.025° × 0.025° for analysis.
• Temporal Scale: 2016–2022 for model training (daily accumulated SSR), 2016–2020 for regional reanalysis evaluation.

Methodology and Data

• Models used:
  • eXtreme Gradient Boosting (XGBoost) for bias correction.
  • Pstar3 radiative transfer scheme (Himawari SSR algorithm).
  • Empirical snow model (AMGSDS-NARO snow depth).
  • Nonhydrostatic regional model coupled with a local ensemble transform Kalman filter (RRJ-Conv).
• Data sources:
  • Satellite: JAXA Himawari-8/9 (SSR, transmittance, principal components of SSR), MODIS (white sky albedo).
  • Observation (Ground-based): Japan Meteorological Agency (JMA) pyranometer observations (21 stations), Chiba University station pyranometer observations.
  • Reanalysis/Gridded: Agro-Meteorological Grid Square Data System (AMGSDS-NARO) for mean/max/min temperature, wind speed, precipitation, snow depth, and SSR; Regional Reanalysis for Japan with Assimilating Conventional Observations (RRJ-Conv) for SSR.
  • Topographic: GEBCO_2023 elevation data.
  • Ancillary: JMA global chemical transport model Meteorological Research Institute Chemistry Climate Model v2.1 (ozone data).

Main Results

• Himawari SSR estimates showed a systematic positive bias of approximately 10% throughout most of the year, shifting to a negative bias (exceeding –20%) during winter over snow-covered surfaces (snow depth > 40 cm).
• The XGBoost bias correction model explained 50–60% of the variance, with a Mean Absolute Error (MAE) of 0.6–0.7 MJ/m²/day (4–5%) and a Root Mean Square Error (RMSE) of 0.9 MJ/m²/day (6–7%).
• The correction nearly eliminated the Mean Bias Error (MBE) (from 1.2 MJ/m²/day to approximately 0.0 MJ/m²/day) and reduced RMSE by 40–50% (from 1.85 MJ/m²/day to approximately 0.9 MJ/m²/day).
• Model predictions were primarily driven by Himawari-derived irradiance, minimum temperature, and snow depth, with physics-informed impacts (e.g., lower minimum temperatures and higher snow depth leading to negative bias correction).
• Spatial transferability analysis showed that MAE between corrected Himawari and AMGSDS-NARO SSR decreased uniformly across distances from JMA stations, with the MAE–distance slope remaining statistically similar, indicating improved accuracy without increasing the distance penalty.
• The RRJ-Conv regional reanalysis reproduced expected SSR spatial patterns and overall magnitudes well but tended to slightly overestimate SSR under cloudy skies (5–10% higher, especially in northern regions and complex terrain).
• Under clear-sky, high-altitude, high-albedo conditions, RRJ-Conv data were valuable for refining bias-corrected satellite SSR where the machine learning model struggled to extrapolate, reducing localized low-bias patches.

Contributions

• Development and application of a physics-informed, interpretable eXtreme Gradient Boosting (XGBoost) model for spatially generalizable bias correction of satellite-derived Surface Solar Radiation (SSR).
• Demonstrated significant improvement in Himawari SSR accuracy over Japan, particularly in snow-affected and complex terrain, by nearly eliminating mean bias and reducing RMSE by approximately 50%.
• First study to demonstrate domain-wide bias correction of Himawari SSR using an interpretable ensemble learning framework.
• Comprehensive evaluation of a novel high-resolution regional reanalysis (RRJ-Conv) using the bias-corrected satellite data, highlighting its strengths and limitations.
• Assessment of spatial transferability of the correction method using an independent ground-based gridded SSR product, identifying topography, land cover, and climate regimes as key drivers of spatial discrepancies.
• Proposed a synergistic approach integrating satellite, ground-based, and reanalysis data to enhance SSR estimation, especially in extreme conditions where machine learning models might struggle.

Funding

• Climate Change Adaptation Research Program of the National Institute for Environmental Studies.
• Japan Science and Technology Agency (grant no. JPMJPF2013).
• Japan Society for the Promotion of Science (KAKENHI grant no. 24K04409).

Citation

@article{Damiani2025Spatially,
  author = {Damiani, Alessandro and Ishizaki, Noriko N. and Watanabe, Tomomi and Tamaki, Yuta and Cordero, Raúl R. and Féron, Sarah and Irie, Hitoshi},
  title = {Spatially generalizable bias correction of satellite solar radiation for regional climate assessment—a case study in Japan},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2025},
  doi = {10.1016/j.jag.2025.104947},
  url = {https://doi.org/10.1016/j.jag.2025.104947}
}