A Guideline to Evaluate the Performance of Rainfall Disaggregation Methods in Representing Extreme Events

Abstract Submission: High-resolution rainfall records are crucial for hydrologic modeling and the characterization of extreme events used in hydrologic design and analysis. However, these high-resolution records are scarce and not easily accessible, with daily precipitation data being more common. This issue is valid both for historical records and future climate change projections, as the daily resolution is the finest widely available time step of General Circulation Models (GCMs) outputs. Therefore, methods to disaggregate daily records into sub-daily (e.g., hourly) rainfall series become essential. Some of these methods include Poisson point processes, Markov process models, and non-parametric methods like k-nearest neighbors (k-NN). Although these approaches have shown to serve their purpose, there are no defined criteria for selecting methodologies by their capability of reproducing extreme precipitation. This study provides a guideline for evaluating the ability of daily-to-hourly disaggregation methods to reproduce extremes. Two main objectives are set: (1) to replicate historical Intensity-Duration-Frequency (IDF) curves for various return periods, evaluated graphically and using accuracy-and-precision metrics; (2) to ensure the disaggregated series reproduce extreme events. To evaluate the proposal, three methods are applied across five case studies from different climate regions: (1) Socolofsky stochastic daily-to-hourly disaggregation method based on historical rainfall event characteristics; (2) k-NN disaggregation method; and (3) daily-to-hourly disaggregation based on local Huff’s curves. The Socolofsky and k-NN methods use leave-one-year-out with calibration cross-validation scheme. The Huff’s curves-based method, however, employs only cross-validation.

Learning Objectives/Expected Outcome (Optional) :