Review of “A bootstrap method to estimate the influence of rainfall spatial uncertainty in hydrological simulations”

Daniel Klotz2017 · DOI

Abstract

The present study proposes a bootstrap-based quantification of spatial rainfall uncertainty for rainfall-runoff modeling. The main body of the work uses the aforementioned bootstrap method to select between different numbers of stations and different groups of stations. The influence of the number and grouping is then quantified via the Nash-Sutcliffe Efficiency (NSE), and an exponential fit is used to derive a heuristic for the estimation of optimal densities of rainfall measurement stations. The analysis is based on three manually selected rainfall events from 2001, 2002, and 2006.

Introduction

Rainfall is notoriously "patchy," exhibiting significant spatial variability even over small distances. For hydrologists, this spatial uncertainty is a primary challenge in rainfall-runoff modeling. If a localized downpour is missed by a sparse rain gauge network, the resulting flood predictions will be significantly underestimated. While it is generally accepted that more gauges lead to better results, there is a point of diminishing returns where the cost of maintaining additional stations outweighs the marginal improvement in model accuracy. This study seeks to quantify this trade-off, providing a transferable method to estimate the remaining spatial uncertainty based on available gauge density.

Methodology

The authors employ a Bootstrap Resampling approach to "stress test" the hydrological model and evaluate its sensitivity to gauge placement. The methodology follows several key steps:

  1. Resampling: The researchers repeatedly sample different numbers ($N$) and combinations ($K$) of stations from an existing dense network.
  2. Spatial Interpolation: For each configuration, basin-wide rainfall is estimated using Thiessen Polygons.
  3. Hydrological Modeling: The DYRIM model is used to simulate runoff for each resampled station set.
  4. Optimization: Model parameters are calibrated using a Genetic Algorithm (GA).
  5. Heuristic Derivation: By plotting the Nash-Sutcliffe Efficiency (NSE) against station density and applying an exponential fit, the authors identify the "elbow" of the curve—the point where adding more stations ceases to provide significant performance gains.

Datasets and Study Area

The research focuses on the Quingjian Basin in China, with the Longxi Basin used for comparative scaling. The analysis is built upon three specific rainfall events from 2001, 2002, and 2006. A notable complexity in the dataset is the mixed temporal resolution, with some stations reporting every 2 hours and others every 6 minutes, requiring careful handling during the modeling process.

Key Insights and Results

  • Quantifying Diminishing Returns: The use of an exponential fit provides a mathematical framework to determine optimal gauge density, moving beyond simple qualitative assessments.
  • Spatial Sensitivity: The bootstrap method reveals that model performance is highly sensitive not just to the number of gauges, but to their specific spatial distribution.
  • Event-Driven Uncertainty: Results suggest that spatial uncertainty is dynamic, changing based on storm magnitude and the basin's runoff coefficient.
  • Predictive Heuristic: The study successfully derived a "prediction curve" that relates measurement density to the expected NSE of a simulation.

Critical Review and Future Directions

Peer review by D. Klotz highlights several critical areas for future improvement:

  • Computational Cost: Running thousands of simulations ($N \times K$) is computationally expensive. Future work should address parallelization and time-cost efficiency.
  • Metric Selection: While NSE is standard, the reviewer suggests that Mean Squared Error (MSE) might be a more appropriate objective criterion for this study, as NSE can sometimes obscure magnitude errors.
  • Methodological Updates: The use of Thiessen Polygons and basic Genetic Algorithms is noted as somewhat dated. Modern techniques like Kriging for interpolation or Shuffled Complex Evolution for optimization might yield more robust results.
  • Risk of Overfitting: Because the model was calibrated for each event individually, there is a risk of overfitting. Cross-validation—testing parameters from one event on another—is a necessary step to prove the model's generalizability.

Conclusion

This study provides a valuable framework for hydrologists to manage the inherent uncertainty of spatial rainfall data. By establishing a mathematical relationship between gauge density and simulation accuracy, it offers a practical tool for optimizing measurement networks and improving the reliability of flood predictions.

Summary updated

Abstract

The present study proposes a bootstrap based quantification of the spatial rainfall uncertainty for rainfall-runoff modelling.The main body of the work uses aforementioned bootstrap method is used to select between different numbers of stations and different groups of stations.The influence of the number and grouping is then quantified via the Nash Sutcliffe Efficiency and exponential fit is used to derive a heuristic for the estimation of the optimal densities of rainfall measurement stations.The whole analysis is based on three manually selected rainfall events from 2001, 2002 and 2006, C1
BibTeX
@misc{klotz2017review,
  author = {Daniel Klotz},
  title = {{Review of “A bootstrap method to estimate the influence of rainfall spatial uncertainty in hydrological simulations”}},
  year = {2017},
  doi = {10.5194/hess-2017-273-rc2},
}
Figure