Ahmed et al. (2026) Enhanced spatial precipitation maps by integrating XGBoost machine learning, terrain indices, and optimal interpolation

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-04-10
• Authors: Peshawa Bakhtyar Salih Ahmed, Nawbahar Faraj Mustafa, Marius Băban
• DOI: 10.1007/s00704-026-06177-z

Research Groups

• Water Resources Department, College of Engineering, University of Sulaimani, Sulaymaniyah, Iraq
• Basic Science Department, College of Dentistry, University of Sulaimani, Sulaymaniyah, Iraq

Short Summary

This study developed a novel integrated framework combining geostatistical interpolation, terrain optimization, and XGBoost machine learning to enhance spatial precipitation estimation in topographically complex, data-scarce regions. The framework achieved superior predictive accuracy (R² = 0.87, RMSE = 70.9 mm) using a multivariate model that incorporated temperature, Topographic Ruggedness Index (TRI), and an optimized Vector Ruggedness Measure (VRM-153).

Objective

• To optimize baseline precipitation surfaces by systematically evaluating Ordinary Kriging (OK), Kernel Interpolation with Barriers (KIB), and Empirical Bayesian Kriging (EBK) interpolation methods, along with determining the optimal Digital Elevation Model (DEM) resolution through resampling techniques, validated using Leave-One-Out Cross-Validation (LOOCV) metrics.
• To derive and optimize terrain parameters, specifically Topographic Ruggedness Index (TRI) and Vector Ruggedness Measure (VRM), from the optimal DEM resolution by establishing ideal window sizes through spike analysis and statistical evaluation to enhance predictive capacity.
• To develop a robust XGBoost-based machine learning model for spatial precipitation prediction that systematically incorporates predictor variables (temperature, precipitation, and optimized terrain indices) in progressively complex configurations (individual, two-parameter, and three-parameter combinations) to identify the most effective predictive feature set.
• To reconstruct and validate high-resolution precipitation maps using the optimal predictive configuration, employing comprehensive statistical evaluation metrics and visual inspection to determine the most accurate and transferable precipitation prediction model for the study region.

Study Configuration

• Spatial Scale: Kurdistan Region of Iraq (KRI), approximately 45,000 square kilometres, with elevations ranging from 168 metres to 3601 metres above mean sea level.
• Temporal Scale: Average 30-year precipitation and temperature records (1995–2025).

Methodology and Data

• Models used:
  • Geostatistical Interpolation: Ordinary Kriging (OK), Kernel Interpolation with Barriers (KIB), Empirical Bayesian Kriging (EBK).
  • Machine Learning: XGBoost (Extreme Gradient Boosting).
• Data sources:
  • Meteorological data: Average 30-year precipitation and temperature records from 26 meteorological stations provided by the General Directorate of Meteorological and Seismology Ministry of Transport and Communications.
  • Topographic data: SRTM one-arc-second Digital Elevation Models (DEMs) (approximately 30 metres resolution) obtained from the United States Geological Survey (USGS) Earth Explorer platform.
  • Derived data: Topographic Ruggedness Index (TRI) and Vector Ruggedness Measure (VRM).

Main Results

• The optimal resampling technique for DEM was cubic convolution, and the optimal resolution was 100 metres. This resolution preserved 62.8% of the original slope correlation while achieving a 92% reduction in computational processing time and an 88% reduction in storage requirements compared to the original DEM.
• Empirical Bayesian Kriging (EBK) was identified as the most dependable interpolation method, yielding the lowest error and bias (R² = 0.80, RMSE = 82.49 mm, MAE = 67.82 mm, PBIAS = 0.81%), outperforming Ordinary Kriging (RMSE = 87.06 mm) and Kernel Interpolation with Barriers (RMSE = 94.11 mm).
• An optimal Vector Ruggedness Measure (VRM) window size of 153 pixels was established, which reduced solo prediction error by 27.6% (RMSE improved from 156.61 mm to 113.32 mm) and increased R² from 0.35 to 0.66 compared to a 3-pixel window.
• The XGBoost-based multivariate model incorporating temperature, TRI, and VRM-153 achieved the highest predictive accuracy (R² = 0.87, RMSE = 70.9 mm, MAE = 51.36 mm, PBIAS = 0.67%), substantially outperforming univariate approaches (e.g., temperature alone: R² = 0.72, RMSE = 103.15 mm).
• Kernel density estimation confirmed optimal distributional correspondence for the three-parameter configuration (temperature, TRI, VRM-153), indicating the model successfully captured the underlying variability and central tendency of precipitation.

Contributions

• Introduces and validates a novel, integrated methodological framework that synergistically combines advanced geostatistical interpolation with machine-learning-driven feature selection for enhanced spatial precipitation estimation in topographically complex, data-scarce regions.
• Provides a systematic optimization process for resampling techniques, DEM resolution, and geostatistical interpolation methods, identifying cubic convolution, 100 metres resolution, and Empirical Bayesian Kriging as optimal for the study area.
• Optimizes terrain indices by determining an ideal window size for Vector Ruggedness Measure (153 pixels), significantly improving its predictive capacity for precipitation.
• Demonstrates the superior performance of a multivariate XGBoost model (integrating temperature, TRI, and VRM) over univariate and bivariate approaches for high-resolution precipitation mapping.
• Establishes a robust, transferable, and scalable framework for high-resolution precipitation mapping using globally accessible data (SRTM DEM) and temperature, addressing a critical research gap in regions with limited observational networks.
• Offers immediate practical implications for improved water resource management, agricultural planning, and climate adaptation strategies in semi-arid mountainous environments.

Funding

Not specified in the paper.

Citation

@article{Ahmed2026Enhanced,
  author = {Ahmed, Peshawa Bakhtyar Salih and Mustafa, Nawbahar Faraj and Băban, Marius},
  title = {Enhanced spatial precipitation maps by integrating XGBoost machine learning, terrain indices, and optimal interpolation},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06177-z},
  url = {https://doi.org/10.1007/s00704-026-06177-z}
}