Regionalization of IDF curves for mainland China: a comparative evaluation of machine learning versus spatial interpolation techniques

Abstract. Regionalization of Intensity-Duration-Frequency (IDF) curves is essential for designing stormwater drainage systems, especially in regions without rainfall data of high temporal resolution. However, most studies have not thoroughly compared regionalization methods using sub-daily site observations versus gridded daily precipitation products. The potential of machine learning (ML) methods driven by daily gridded precipitation remains largely underexplored. This study addresses these gaps by regionalizing the IDF curves across mainland China for durations ranging between 1 and 72 hours and return periods ranging from 2 to 1,000 years. Five interpolation methods based on hourly observations from 2363 stations and five machine learning methods based on a gridded daily dataset were tested for accuracy. Both ML and traditional interpolation methods showed robust performances based on the Kling-Gupta Efficiency (KGE) performance measure. The most successful interpolation method was Kriging with External Drift using mean annual precipitation, with KGE > 0.96 for 1-hr-5-yr and 24-hr-5-yr storms and KGE > 0.84 for 1-hr-100-yr and 24-hr-100-yr storms, while Gradient Boosting was the best-performing ML model, with KGE > 0.94 for 1-hr-5-yr and 24-hr-5-yr storms and KGE > 0.87 for 1-hr-100-yr and 24-hr-100-yr storms. Notably, despite ML using daily data and interpolation using hourly data, the accuracy of ML gradually improved, eventually approaching or even surpassing the interpolation methods as duration and return period increased. Consequently, a regionalized dataset on IDF curves for mainland China with a spatial resolution of 0.1 degrees (and optionally 0.5 degrees) was generated using the optimal regionalization method.