Saghiry et al. (2025) Towards More Reliable Gridded Precipitation Estimates: Gauge-Based Multi-Scale Evaluation and Machine Learning Bias Correction

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-12-24
• Authors: Soumia Saghiry, Dalila Loudyi, Oussama Laassilia, Arfan Arshad, Latifa Dhaouadi, Riaz Ali, Meryem El Alaoui
• DOI: 10.1007/s41748-025-01000-7

Research Groups

• Process Engineering and Environment Laboratory (LGPE), Faculty of Sciences and Techniques of Mohammedia, Hassan II University of Casablanca, Casablanca, Morocco
• Laboratory of Engineering Sciences for Energy (LabSIPE), National School of Applied Sciences, Chouaib Doukkali University, El Jadida, Morocco
• Department of Biosystems and Agricultural Engineering, Oklahoma State University, Stillwater, OK, USA
• Regional Center for Research in Oasis Agriculture, Cartage University, Ariana, Tunisia
• National Key Laboratory of Green and Long-Life Road Engineering in Extreme Environment, Shenzhen University, Shenzhen, China
• Advanced Systems Engineering Laboratory, National School of Applied Sciences, Ibn Tofaïl University, Kénitra, Morocco

Short Summary

This study evaluates four high-resolution gridded precipitation products against 12 rain gauges in Morocco's Moulouya Basin across multiple spatio-temporal scales and applies machine learning (ML) for bias correction. It finds that GPM IMERG-Final (GPM-F) performs best uncorrected, and ML models (Random Forest for daily, eXtreme Gradient Boosting for monthly/seasonal) significantly enhance its accuracy, providing reliable precipitation inputs for hydrological applications in data-scarce regions.

Objective

• To evaluate and compare the performance of four high-resolution gridded precipitation products (PERSIANN-CDR, CHIRPS, ERA5-Land, and GPM IMERG-Final) against ground observations from 12 rain gauges in the Moulouya Basin, Morocco, to identify the most reliable dataset.
• To enhance the accuracy of the best-performing precipitation product (GPM IMERG-Final) by correcting for bias using three machine learning models: Artificial Neural Network (ANN), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost).

Study Configuration

• Spatial Scale: Moulouya Basin, Morocco (approximately 55,000 km²). Evaluation conducted at two spatial scales: pixel scale (point-to-pixel comparison at 12 rain gauge locations) and basin scale (basin-wide average).
• Temporal Scale: Common overlapping period from 01 June 2000 to 30 December 2023. Evaluation and bias correction performed at daily, monthly, and seasonal time scales. Annual scale was used for evaluation but excluded from bias correction due to limited data record length.

Methodology and Data

• Models used:
  • Gridded Precipitation Products: PERSIANN-CDR (v1r1), CHIRPS v2.0, GPM IMERG Final V06 (GPM-F), ERA5 Land (CDS Release 2019).
  • Machine Learning Models for Bias Correction: Artificial Neural Network (ANN), Random Forest (RF), eXtreme Gradient Boosting (XGBoost).
• Data sources:
  • Ground Observations: Daily precipitation data from 12 rain gauge stations in the Moulouya Basin, provided by the Moulouya Hydraulic Basin Agency.
  • Satellite/Reanalysis Data:
    • CHIRPS v2.0: High-resolution (0.05°) product combining satellite infrared (IR) data, climate model outputs, and ground observations.
    • GPM IMERG Final V06: Multi-satellite product (60°N-S latitude band) integrating GPM Core Observatory observations and calibrated with Global Precipitation Climatology Centre (GPCC) ground data.
    • PERSIANN-CDR (v1r1): Long-term (0.25°) product combining satellite IR data with an Artificial Neural Network (ANN) model, calibrated with GPCP.
    • ERA5 Land: High-resolution (31 km) reanalysis product focusing on land-surface variables, derived from ERA5.
  • Predictors for ML Models: Cyclic time descriptors, GPM, ERA5, and CHIRPS precipitation products, geographical predictors (elevation, longitude, latitude), and meteorological parameters extracted from ERA5 (minimum and maximum temperature, wind speed, atmospheric pressure).

Main Results

• Gridded Precipitation Product Performance (Uncorrected):
  • GPM-F consistently showed superior performance, particularly at the daily scale (median Pearson correlation coefficient (CC) ≈ 0.48), making it the optimal choice for high-temporal-resolution applications.
  • ERA5 Land excelled at monthly and seasonal scales (median CC ≈ 0.77 seasonally, median coefficient of determination (R²) ≈ 0.65 seasonally), demonstrating strong pattern representation but with notable biases.
  • CHIRPS exhibited the lowest relative bias (Rbias) and normalized root mean square error (NRMSE) at longer temporal scales, making it suitable for applications requiring accurate quantitative precipitation totals.
  • PERSIANN-CDR consistently underperformed across all metrics and scales in the Moulouya Basin, showing the lowest correlations and highest errors.
  • Product performance was elevation-dependent, generally better at mid-elevations (600–910 m) and degrading at higher elevations, with PERSIANN showing significant overestimation at lower elevations.
• Machine Learning Bias Correction of GPM-F:
  • ML bias correction substantially improved GPM-F estimates across all evaluation metrics (CC, R², Rbias, RMSE, NRMSE, Nash–Sutcliffe Efficiency (NSE)).
  • At the daily scale, Random Forest (RF) was optimal, increasing median CC to approximately 0.8, raising R² from 0.24 to 0.65, and reducing root mean square error (RMSE) by 61.8%.
  • At the monthly scale, eXtreme Gradient Boosting (XGBoost) slightly outperformed RF, showing more consistent improvements and superior bias reduction.
  • At the seasonal scale, XGBoost was optimal, achieving CC up to 0.96, R² above 0.85, RMSE reduced by 80.4%, and consistently high NSE values (> 0.85).
  • Artificial Neural Network (ANN) provided consistent but generally more moderate improvements compared to RF and XGBoost.
• Validation and Feature Importance:
  • 5-fold cross-validation and Leave-One-Station-Out (LOSO) validation confirmed the robustness and generalization ability of RF (daily) and XGBoost (monthly/seasonal) models.
  • Feature importance analysis revealed that RF primarily relies on direct precipitation inputs (61.19%), while XGBoost distributes importance more evenly across precipitation, topographic/spatial (22.09%), and atmospheric dynamical predictors (19.71%), explaining its better performance at seasonal scales.

Contributions

• First comprehensive multi-scale evaluation of four high-resolution gridded precipitation products and application of a comparative multi-scale machine learning bias correction framework in the Moulouya Basin, a data-scarce region with complex topography.
• Provides specific recommendations for optimal gridded precipitation products and ML bias correction methods tailored to different temporal resolutions (daily, monthly, seasonal) and physiographic contexts (elevation gradients).
• Demonstrates the significant potential of ML-based bias correction to transform raw satellite precipitation estimates into highly reliable inputs for hydrological modeling, flood forecasting, and water resource management in regions with limited ground observations.
• Highlights the critical importance of considering elevation-dependent performance and the interplay between satellite retrieval algorithms and topographic influences when selecting and correcting precipitation products.

Funding

This research received no external funding.

Citation

@article{Saghiry2025Towards,
  author = {Saghiry, Soumia and Loudyi, Dalila and Laassilia, Oussama and Arshad, Arfan and Dhaouadi, Latifa and Ali, Riaz and Alaoui, Meryem El},
  title = {Towards More Reliable Gridded Precipitation Estimates: Gauge-Based Multi-Scale Evaluation and Machine Learning Bias Correction},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-01000-7},
  url = {https://doi.org/10.1007/s41748-025-01000-7}
}