Information on extreme precipitation is essential to managing weather-related risks and designing hydraulic structures. Research attention to frequency analyses based on remotely sensed precipitation datasets, such as those obtained from weather radars and satellites, has been rapidly increasing owing to their potential to provide information for ungauged regions worldwide. Together with the ability to measure the areal scale directly, these analyses promise to overcome the sampling limitations of traditional methods based on rain gauges. This focused review of the literature depicts the state of the art after a decade of efforts, and identifies the crucial gaps in knowledge and methodology that currently hinder the quantitative use of remotely sensed datasets in water resources system design and operation. It concludes by highlighting a set of research directions promising immediate impact with regard to the separation of the sources of uncertainty currently affecting applications based on remotely sensed datasets, the development of statistical methods that can cope with the peculiar characteristics of these datasets, and the improvement of validation methods. Important gains in knowledge are expected from the explicit inclusion of the areal dimension in the analyses and from the fine-scale investigation of extreme precipitation climatology.
The lack of knowledge on precipitation frequency over ungauged areas introduces a significant source of uncertainty in relevant engineering designs and risk estimation procedures. Radar-based observations offer precipitation information over ungauged areas and thus have gained increasing attention as a potential solution to this problem. However, due to their relative short data records and inherent uncertainty sources, their ability to provide accurate estimates on the frequency of precipitation extremes requires evaluation. This study involves the evaluation of at-site precipitation frequency estimates from NEXRAD Stage IV radar precipitation dataset. We derive precipitation annual maxima series from the 16yrs record (2002-2017) of NEXRAD and we compare against 539 long-term (50yrs) hourly gauge records. In addition, Intensity-Duration-Frequency (IDF) curves are estimated from both radar and gauge dataset and compared. IDF estimation is based on fitting the Generalize Extreme Value distribution to annual precipitation maxima. Evaluation is carried out over the contiguous United States and results are grouped and presented for five dominant climate classes and for a range of return period and precipitation durations. NEXRAD was shown to overestimate intensities at shorter durations (1- and 3-hr) and low quantiles, while it tends to underestimate higher quantiles at longer durations (24hr). In addition, evaluation of the IDF curves estimated from NEXRAD revealed a distinct geographic dependence with certain regions exhibiting a tendency to overestimation (e.g. east of the Rocky Mountains) or underestimation (Midwest). Overall, this analysis suggests that, while significant discrepancies may exist, there are several cases where NEXRAD provide estimates within the uncertainty bounds of the reference rain gauge dataset. The climate/geographic region and the temporal duration are important aspects to consider. Findings provided in this work on these aspects will hopefully serve as a general guideline for those interested in using NEXRAD estimates for further research or applications on precipitation extremes.
This study expands the Metastatistical Extreme Value (MEV) framework to sub-daily rainfall frequency analysis and compares it to extreme value theory methods in presence of short records and measurement errors. Ordinary events are identified based on the temporal autocorrelation of hourly data and modeled with a Weibull distribution. MEV is compared to extreme value theory methods in the estimation of long return period quantiles from actual data (160 rain gauges with at least 60-year record in the contiguous United States) and on synthetic data perturbed with measurement errors typical of remote sensing rainfall estimation. MEV tends to underestimate the 100-year return period quantiles of hourly rainfall when 5–20 years of actual data are used, but presents diminished uncertainty. When a good model of the ordinary events and adequate number of events per year are available, MEV is able to provide information on the 100-year return period quantiles from 10–20, or even 5 years of data with significantly reduced uncertainty (\textless30% uncertainty for 5-year records). MEV estimates of 100-year return period quantiles from short records are much less sensitive than extreme value theory methods to additive/multiplicative errors, presence of cap values in the estimates, and missing of extreme values. Results from this study strongly support the use of MEV for rainfall frequency analyses based on remotely sensed datasets.