Gaona et al. (2026) Spatial Downscaling of the CHIRPS Rainfall Product Using Machine Learning Methods: The Catamayo–Chira Transboundary Basin (Ecuador-Peru) Case

Identification

• Journal: Hydrology
• Year: 2026
• Date: 2026-03-10
• Authors: Jessica K. Gaona, Luis-Felipe Duque, Raul F. Vazquez, Candy L. Ocaña
• DOI: 10.3390/hydrology13030089

Research Groups

• Carrera de Ingeniería Ambiental, Universidad Nacional de Loja (UNL), Ecuador
• Grupo de Investigación en Hidroinformática, Universidad Nacional de Loja (UNL), Ecuador
• Laboratorio de Ecología Acuática (LEA), Facultad de Ciencias Químicas, Universidad de Cuenca, Ecuador
• Departamento de Ingeniería Civil, Facultad de Ingeniería, Universidad de Cuenca, Ecuador
• Instituto Binacional de Investigación para el Desarrollo Sostenible de la Cuenca del Marañón de Perú y Ecuador, Universidad Nacional de Jaén (UNJ), Peru

Short Summary

This study spatially downscaled the 5 km CHIRPS rainfall product to 1 km for the Catamayo–Chira Transboundary Basin (Ecuador-Peru) using various single-variable and multivariable machine learning methods, demonstrating significant improvement in precipitation estimates and successfully capturing "El Niño" event differences.

Objective

• To identify a data-learning method capable of generating acceptably precise precipitation datasets for the Catamayo–Chira Transboundary Basin by downscaling a Satellite Precipitation Product (SPP).

Study Configuration

• Spatial Scale: Catamayo–Chira Transboundary Basin (Ecuador-Peru), covering 17,199 km², with elevation ranging from 0 to 4000 m above sea level. Downscaling was performed from 5 km to 1 km resolution.
• Temporal Scale: January 2001 to December 2023 (23 years), analyzed at annual and mean monthly scales.

Methodology and Data

• Models used: Simple Linear Regression (LR), Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Machine (SVM) with linear (SVM-lin), radial basis function (SVM-rbf), polynomial (SVM-poly), and sigmoid (SVM-sig) kernels, and Artificial Neural Networks (ANN).
• Data sources:
  • Satellite/Reanalysis Precipitation: Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) product (0.05° / ~5 km spatial resolution, daily values aggregated to monthly/annual).
  • Observed Precipitation: Data from 10 gauged meteorological stations (INAMHI, Ecuador; SENAMHI, Peru) for the period 2001–2023.
  • Predictor Variables:
    • Normalized Difference Vegetation Index (NDVI): MOD13A3.061 product (MODIS, Terra satellite, 1 km spatial, 16-day temporal resolution).
    • Land Surface Temperature (LST): MOD11A2.061 product (MODIS, 1 km spatial, 8-day temporal resolution).
    • Altitude: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) (1 km spatial resolution).
    • Longitude (Long) and Latitude (Lat): Derived from the 1 km DEM.

Main Results

• Downscaling significantly improved precipitation estimates, with RMSE values for the 1 km downscaled product being substantially lower than those for the original 5 km CHIRPS dataset across the entire study period.
• For annual precipitation, the most effective single-variable methods were simple linear regression (LR) using Longitude (Long) (Mean Absolute Error (MAE) = 16.1 mm, Root Mean Square Error (RMSE) = 17.7 mm) and Latitude (Lat) (MAE = 18.4 mm, RMSE = 21.1 mm).
• Among multivariable non-linear methods for annual precipitation, Support Vector Machine with radial basis function kernel (SVM-rbf) and Artificial Neural Networks (ANN) performed best (RMSE < 56 mm, MAE < 41 mm).
• Surprisingly, single-variable linear methods (LR for Long and Lat) often outperformed more complex multivariable non-linear methods in terms of RMSE for annual precipitation prediction.
• For mean monthly precipitation, simple LR with Lat and Long predictors consistently showed the smallest errors across most months. SVM with a linear kernel (SVM-lin) was the best-performing non-linear multivariable method at the monthly scale (0.1 mm ≤ MAE ≤ 45.6 mm and 0.1 mm ≤ RMSE ≤ 61.4 mm).
• The performance of both linear and non-linear methods tended to be better during the drier months.
• Downscaled annual precipitation distributions successfully captured the spatial anomalies and differences between "El Niño" years (e.g., 2017) and normal years, showing expected patterns like increased coastal precipitation during "El Niño" events.

Contributions

• Generated the first 1 km downscaled CHIRPS precipitation maps for the Catamayo–Chira Transboundary Basin, addressing a critical data gap for water resource management in this binational region.
• Provided a comprehensive comparative evaluation of single-variable and multivariable machine learning downscaling methods under the complex local conditions of a Pacific–Andean system.
• Demonstrated that simple linear regression models using geographic coordinates (longitude and latitude) can achieve comparable or superior performance to more complex non-linear multivariable methods in regions where dominant geographic gradients strongly influence precipitation patterns.
• Developed a replicable modeling protocol for spatial downscaling of satellite precipitation products, applicable to other data-scarce basins.

Funding

• The Article Processing Charge (APC) was funded by Universidad Nacional de Jaén (Cajamarca, Perú).

Citation

@article{Gaona2026Spatial,
  author = {Gaona, Jessica K. and Duque, Luis-Felipe and Vazquez, Raul F. and Ocaña, Candy L.},
  title = {Spatial Downscaling of the CHIRPS Rainfall Product Using Machine Learning Methods: The Catamayo–Chira Transboundary Basin (Ecuador-Peru) Case},
  journal = {Hydrology},
  year = {2026},
  doi = {10.3390/hydrology13030089},
  url = {https://doi.org/10.3390/hydrology13030089}
}