Abstract
The reliability of gridded precipitation and temperature products as alternatives to ground observations remains a matter of debate, particularly in data-sparse regions with complex terrain, such as Afghanistan. Bias correction is often essential; however, selecting between simple and advanced methods is challenging. We evaluated 13 precipitation products spanning satellite, reanalysis, merged, and gauge-based datasets (CHIRPS, CMORPH, CPC, ERA5, GLDAS, GMCP, GPCP, GPM calibrated/uncalibrated, GSMaP calibrated/uncalibrated, MERRA-2, MSWEP, and PERSIANN) and 6 temperature products (CPC, ERA5, GLDAS, MERRA-2, MODIS, and NCEP) against 82 observation stations covering Afghanistan’s diverse topography and climate zones. The temperature products showed consistently high accuracy without the need for bias correction, with CPC and GLDAS exhibiting the best performance. Among the precipitation datasets, the merged satellite–gauge–reanalysis products GMCP and MSWEP exhibited the best performance, followed by the calibrated satellite product GPM; however, bias correction was required for all precipitation products. Subsequently, simple, moderate, and complex statistical and machine learning (ML) models were applied to correct for precipitation bias. A representative subset was first used to analyse ML hyperparameter behaviour and to test combinations of meteorological features and elevation; the optimal feature set was used to optimize ML models across the full dataset for each product. The Kling–Gupta Efficiency, Root Mean Square Error, Percentage Error in Volume, and the coefficient of determination were used as the main evaluation metrics. Statistically bias-corrected precipitation data were suitable for monthly applications, with deviations of approximately 10–20% of total water volume compared with observed data. However, for daily applications and gap-filling of station observations, ML methods enriched with additional meteorological features yielded improved accuracy, reducing volume error by approximately 5–10%. Overall, the models demonstrated better performance at the monthly scale than at the daily scale. Therefore, the choice of bias-correction method heavily depends on the intended application.
| Original language | English |
|---|---|
| Journal | International Journal of Remote Sensing |
| Volume | 47 |
| Issue number | 1 |
| Pages (from-to) | 337-375 |
| ISSN | 0143-1161 |
| DOIs | |
| Publication status | Published - 2026 |
Bibliographical note
Publisher Copyright:© 2025 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
Keywords
- Hindu Kush Himalaya
- Machine learning
- statistical bias correction
Cite this
- APA
- Standard
- Harvard
- Vancouver
- Author
- BIBTEX
- RIS