Spatial estimation of standardized precipitation index based drought severity across Türkiye using machine learning models
Abstract
Reliable spatial estimation of extreme drought intensity represented by the minimum SPI value is essential for water resources management under increasing climate variability. However, conventional assessments often rely on station-based temporal indices, providing limited representation of spatial heterogeneity and topographic controls. This study developed a machine learning framework to estimate the spatial distribution of extreme drought intensity represented by the minimum SPI value across Türkiye using long-term precipitation records from 102 meteorological stations. For each station and SPI accumulation period, extreme drought intensity was represented by the minimum (most negative) SPI value observed over the 1935–2024 record; this minimum SPI value was used as the machine-learning target, while latitude, longitude, and elevation were used as geographic predictors. Seven machine learning models were evaluated; among them, GBM yielded the highest accuracy (R ² = 0.991, RMSE = 0.007), followed by XGBoost and CatBoost. These results indicate that ensemble tree-based methods reproduce the observed spatial pattern of extreme drought intensity across the station network with high accuracy; establishing predictive skill at unmonitored locations would require evaluation under a station-grouped partition. Furthermore, feature importance and SHAP analyses showed that spatial predictors (longitude, latitude, and elevation) were substantially more influential than the time-scale variable, although their relative importance varied among the evaluated models. This framework offers a robust tool for interpretable and spatially explicit drought monitoring in topographically complex regions.
Related
- Development of a nonparametric spatial standardized precipitation–temperature index for regional drought characterization
- RainFlow: Flow Matching for Infrared-Only Quantitative Precipitation Estimation
- Spatial context importance in deep learning for global horizontal irradiance estimation from satellite imagery
- Composable multi-satellite precipitation estimation for evolving observing systems
- Dependence of the precipitation intensity distribution on spatial and temporal resolution in observations
BibTeX
@article{atasever2026spatial,
title = {{Spatial estimation of standardized precipitation index based drought severity across Türkiye using machine learning models}},
author = {Umit Haluk Atasever and Talha Taşkanat and Aykut Erol and Ümran Köylü and Hatice Citakoglu},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-73493-9},
url = {https://www.nature.com/articles/s41598-026-73493-9},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Spatial estimation of standardized precipitation index based drought severity across Türkiye using machine learning models
AU - Umit Haluk Atasever
AU - Talha Taşkanat
AU - Aykut Erol
AU - Ümran Köylü
AU - Hatice Citakoglu
PY - 2026
DA - 2026/10/01
JO - Scientific Reports
DO - 10.1038/s41598-026-73493-9
UR - https://www.nature.com/articles/s41598-026-73493-9
AB - Reliable spatial estimation of extreme drought intensity represented by the minimum SPI value is essential for water resources management under increasing climate variability. However, conventional assessments often rely on station-based temporal indices, providing limited representation of spatial heterogeneity and topographic controls. This study developed a machine learning framework to estimate the spatial distribution of extreme drought intensity represented by the minimum SPI value across Türkiye using long-term precipitation records from 102 meteorological stations. For each station and SPI accumulation period, extreme drought intensity was represented by the minimum (most negative) SPI value observed over the 1935–2024 record; this minimum SPI value was used as the machine-learning target, while latitude, longitude, and elevation were used as geographic predictors. Seven machine learning models were evaluated; among them, GBM yielded the highest accuracy (R ² = 0.991, RMSE = 0.007), followed by XGBoost and CatBoost. These results indicate that ensemble tree-based methods reproduce the observed spatial pattern of extreme drought intensity across the station network with high accuracy; establishing predictive skill at unmonitored locations would require evaluation under a station-grouped partition. Furthermore, feature importance and SHAP analyses showed that spatial predictors (longitude, latitude, and elevation) were substantially more influential than the time-scale variable, although their relative importance varied among the evaluated models. This framework offers a robust tool for interpretable and spatially explicit drought monitoring in topographically complex regions.
SN - 2045-2322
ER -