Jeong et al. (2026) Enhancing spatial specification of runoff through a performance-informed multivariate weighting framework

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2026
• Date: 2026-01-22
• Authors: Minyeob Jeong, Jongho Kim
• DOI: 10.1007/s00477-026-03174-6

Research Groups

• School of Civil and Environmental Engineering, University of Ulsan, Ulsan, South Korea

Short Summary

This study proposes a performance-informed multivariate weighting framework to enhance runoff spatial specification in data-sparse or ungauged sub-basins. By integrating multiple weighted ancillary variables derived from linear regression and artificial intelligence models, the framework significantly outperforms traditional single-variable methods, improving accuracy and capturing spatial heterogeneity.

Objective

• To identify the most significant candidate ancillary variables influencing runoff estimation.
• To determine the relative importance (weights) of each variable based on its relevance to runoff.
• To apply these weights to a modified pycnophylactic-dasymetric mapping framework.
• To improve the spatial specification of runoff beyond existing interpolation approaches.

Study Configuration

• Spatial Scale: The Han River Basin, South Korea (approximately 34,428 km²), comprising 24 middle basins and 237 standard sub-basins. Data were processed at a GLDAS grid resolution of 0.25° (approximately 25 km) and downscaled to finer target grids of 0.125° (one-fourth the area). Validation was performed using 34 GLDAS cells.
• Temporal Scale: Daily time series data from January 1, 2000, through September 30, 2023. The full period was used for validation, while AI models used 80% (2000-2020) for training and 20% (2021-2023) for validation. Specific low-flow (5 September 2022), high-flow (31 August 2002), and extreme flood (9 August 2022) events were analyzed.

Methodology and Data

• Models used:
  • Linear Regression (LR) for variable weighting.
  • Long Short-Term Memory (LSTM) (an Artificial Intelligence model) for variable weighting, interpreted using SHapley Additive exPlanations (SHAP).
  • Modified Dasymetric Mapping (DM) for spatial specification.
  • Modified Pycnophylactic-Dasymetric Mapping (PPDM) for spatial specification.
• Data sources:
  • Time-variant variables (Antecedent Precipitation Index (API), Precipitation (P), Air Temperature (AT), Evapotranspiration (ET), Humidity (HU), Soil Temperature (ST), Soil Moisture (SM), Wind Speed (WS)) and runoff (Z) were obtained from GLDAS-Noah version 2.1.
  • Time-invariant variables (Elevation (EL), Potential Runoff Coefficient (PRC), Surface Manning’s Roughness (MN), Slope (S)) were compiled from various geospatial data sources, including digital elevation models and soil/land cover databases.
  • Watershed boundaries for the Han River Basin were constructed using shapefiles from Korea’s Water Management Information System (WAMIS).

Main Results

• The proposed multivariate framework significantly improved runoff specification accuracy (higher R², lower SI, lower PBIAS) compared to traditional methods relying on a single ancillary variable (e.g., elevation).
• Key principal ancillary variables consistently identified across methods and conditions include soil moisture (SM), antecedent precipitation index (API), humidity (HU), and rainfall (P), with soil temperature (ST) as a secondary factor.
• Variable importance is context-dependent: SM, API, and HU are most influential during dry conditions, while P, SM, API, and HU are dominant during wet conditions.
• The improvement in specification accuracy was more pronounced during high-flow events, where multivariable AI-based models more faithfully reproduced GLDAS runoff distributions, capturing peak magnitudes and multi-modal structures, unlike traditional methods which tended to oversmooth.
• While AI models showed superior predictive performance (R² ~0.8–0.9) in runoff prediction tasks compared to LR (R² ~0.5–0.6), both achieved similar accuracy in the spatial specification task, indicating that predictive power does not directly translate to better downscaling when total runoff volume is conserved.
• Spatially varying weights (grid-dependent, DE approach) did not yield significantly higher accuracy than spatially uniform weights (grid-independent, IN approach) within the relatively homogeneous Han River Basin, suggesting that uniform weighting can be a practical option in such environments.

Contributions

• Introduces a novel performance-informed multivariate weighting framework that integrates multiple hydrologically relevant ancillary variables with dynamically derived weights into dasymetric and pycnophylactic mapping.
• Demonstrates a substantial improvement in runoff spatial specification accuracy over conventional single-variable, ad-hoc approaches, particularly in capturing spatial heterogeneity and addressing the Modifiable Areal Unit Problem (MAUP).
• Provides a robust methodology for enhancing runoff estimation in ungauged or data-sparse basins, bridging the resolution gap between coarse global datasets and fine-scale local water management needs.
• Highlights the critical role of condition-specific variable selection (e.g., dry vs. wet periods) for accurate hydrological modeling and spatial specification.

Funding

• National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2022-NR070280).
• Korea Environment Industry & Technology Institute (KEITI) through Water Management Program for Drought Program, funded by Korea Ministry of Environment (MOE) (2022003610003).

Citation

@article{Jeong2026Enhancing,
  author = {Jeong, Minyeob and Kim, Jongho},
  title = {Enhancing spatial specification of runoff through a performance-informed multivariate weighting framework},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2026},
  doi = {10.1007/s00477-026-03174-6},
  url = {https://doi.org/10.1007/s00477-026-03174-6}
}