A Cross-Validated Regional POT-EGPD Framework for Sub-Daily IDF Curves from Adjusted Satellite Precipitation in Mediterranean Climates

Sub-daily intensity duration frequency (IDF) curves underpin flood risk management and infrastructure design, yet they remain unavailable or highly uncertain in many Mediterranean and semi-arid regions due to sparse high-resolution gauge networks. Meanwhile, satellite precipitation products provide spatially continuous coverage but can exhibit systematic biases in magnitude, frequency, and extremes, particularly at short durations. This study examines whether machine-learning-adjusted, high-resolution satellite precipitation can support reliable sub-daily IDF estimation across an entire region, using Historical Palestine (Israel and the West Bank) as a climatically heterogeneous Mediterranean case study.

We propose a regional extreme-value framework that links bias-corrected satellite precipitation to sub-daily Intensity Duration Frequency (IDF) curves through a Peak-Over-Threshold approach with an Extended Generalized Pareto Distribution (POT-EGPD) parameterized using L-moments. Regionalization is performed using Gaussian Mixture Models (GMM), trained on calibration gauges only, with predictors combining extreme-shape information (L-moment ratios) and physiographic metadata (elevation, latitude, longitude, and climatic class). To evaluate generalizability, we implement station-based cross-validation, ensuring regional representativeness during splitting while preventing leakage.

We compare three methodological variants designed to isolate the effects of frequency and magnitude biases in satellite extremes: (i) a baseline regional POT model (U1) using calibration-gauge thresholds, tail parameters, and exceedance rates; (ii) a frequency-adjusted variant (U2) that corrects satellite exceedance rates using calibration-derived regional scaling; and (iii) an annual-maxima benchmark using regional GEV modelling (AM-GEV) with satellite index scaling. Satellite inputs include raw and machine-learning-adjusted products (e.g., IMERG raw versus IMERG adjusted via LightGBM).

Preliminary results indicate that machine-learning adjustment improves the consistency of satellite-based extreme DDF behaviour relative to gauges, and that explicitly correcting exceedance frequency further stabilizes tail behaviour across regions. The framework is intended for scalable regional IDF production in environments where fine-resolution gauges are scarce, supporting design-rainfall estimation and climate-resilient planning.