# A probabilistic gridded product for daily precipitation extremes over the United States

## Abstract

Gridded data products, for example interpolated daily measurements of precipitation from weather stations, are commonly used as a convenient substitute for direct observations because these products provide a spatially and temporally continuous and complete source of data. However, when the goal is to characterize climatological features of extreme precipitation over a spatial domain (e.g., a map of return values) at the native spatial scales of these phenomena, then gridded products may lead to incorrect conclusions because daily precipitation is a fractal field and hence any smoothing technique will dampen local extremes. To address this issue, we create a new “probabilistic” gridded product specifically designed to characterize the climatological properties of extreme precipitation by applying spatial statistical analysis to daily measurements of precipitation from the Global Historical Climatology Network over the contiguous United States. The essence of our method is to first estimate the climatology of extreme precipitation based on station data and then use a data-driven statistical approach to interpolate these estimates to a fine grid. We argue that our method yields an improved characterization of the climatology within a grid cell because the probabilistic behavior of extreme precipitation is much better behaved (i.e., smoother) than daily weather. Furthermore, the spatial smoothing innate to our approach significantly increases the signal-to-noise ratio in the estimated extreme statistics relative to an analysis without smoothing. Finally, by deriving a data-driven approach for translating extreme statistics to a spatially complete grid, the methodology outlined in this paper resolves the issue of how to properly compare station data with output from earth system models. We conclude the paper by comparing our probabilistic gridded product with a standard extreme value analysis of the Livneh gridded daily precipitation product. Our new data product is freely available on the Harvard Dataverse (https://bit.ly/2CXdnuj).

## Keywords

Extreme value analysis Precipitation Spatial statistics Nonparametric bootstrap Global Historical Climatology Network Gaussian processes Gridded daily precipitation## 1 Introduction

Gridded data products of precipitation are a popular substitute for daily measurements of rainfall from weather stations. The products are generated by interpolating the station measurements onto a uniform space-time grid to create a spatially and temporally complete, continuous, and homogeneous version of the raw data. For this reason, gridded products are often used to summarize the climatological properties of extreme precipitation, for example maps of return values, and then to evaluate the performance of climate models with respect to extremes. In traditional analysis of precipitation extremes using gridded products (see, e.g., Wehner 2013; Sylla et al. 2013, and many others), the extreme climatology is estimated separately for each grid cell using a univariate extreme value analysis. However, we assert that gridded daily precipitation products are problematic data sources for constructing these extreme climatologies because daily precipitation is well-known to exhibit fractal scaling (e.g., Lovejoy et al. 2008; Maskey et al. 2016, and numerous references therein), and therefore any spatial smoothing or averaging during the gridding process will diminish variability and extreme values. Additionally, a recent thread of research (King et al. 2013; Gervais et al. 2014; Timmermans et al. 2019) explicitly questions the appropriateness of using gridded products as a substitute for observed extremes.

As an alternative to using gridded data products, irregularly observed weather station measurements can be used to obtain spatially-complete summaries of extreme precipitation by utilizing the diverse statistics literature on spatial extreme value analysis. These statistical tools collectively analyze extremes over space for processes such as precipitation, specifically allowing one to estimate the distribution of extreme precipitation even for locations where no data is available. The methods broadly fall into one of four categories: max-stable processes (Haan 1984; Smith 1990; Schlather 2002; Kabluchko et al. 2009), which provide mathematically-founded models for characterizing spatial dependence among extremes; copula-based approaches (Hüsler and Reiss 1989; Demarta and McNeil 2005; Sang and Gelfand 2010; Fuentes et al. 2013; Krupskii et al. 2018), which construct a joint multivariate distribution for spatial extremes via careful modeling of transformed marginal (univariate) distributions; Bayesian methods (Reich and Shaby 2012; Shaby and Reich 2012; Morris et al. 2017), which use a hierarchical framework to build numerically tractable models based upon the mathematics of max-stable processes; and nonparametric mixture models (Gelfand and Kottas 2002; Wang et al. 2011; Kottas et al. 2012), which simultaneously analyze both the average and extreme values of precipitation. See Davison et al. (2012) for an excellent review of these approaches.

In this paper, we seek to characterize the climatology of extreme precipitation based on measurements from approximately five thousand Global Historical Climatology Network (GHCN) stations (see Sect. 2) over the contiguous United States (CONUS). Unfortunately, practically speaking, many of the aforementioned statistical methods are only suitable for a small number of weather stations. For example, Davison et al. (2012) use a data set with just 36 stations, Saunders et al. (2017) employ 173 stations to derive a max-stable process to model extreme precipitation in Australia, and Shaby and Reich (2012) analyze the largest data set considered to date (to our knowledge) comprised of approximately one thousand stations. Regardless of the number of weather stations of interest, a second limitation of the aforementioned methods is that they are only appropriate for homogeneous spatial domains. CONUS, on the other hand, is a highly heterogeneous spatial domain with variable topography interacting with a diverse set of physical phenomena that produce extreme precipitation. These phenomena include atmospheric rivers on the west coast, tropical cyclones on the east coast, and mesoscale convective systems in the Great Plains.

In light of these shortcomings in existing methodologies for spatial extremes, note that we do not need to simulate realistic fields of seasonal maxima, which is a complicating factor in the aforementioned max-stable and Bayesian approaches to spatial extreme value analysis (furthermore, note that we do not intend to model the full space-time distribution of daily precipitation, as in Serinaldi and Kilsby 2014). Instead, we simply wish to characterize the statistics of extreme precipitation over CONUS. In this case, one apparently viable option would be to employ latent variable or conditional independence models (Cooley et al. 2007; Craigmile and Guttorp 2013; Mannshardt et al. 2013), which provide flexible methodologies suitable for larger data sets and heterogeneous spatial domains. These methods assume that daily precipitation totals over space occur independently of one another and are conditional on latent processes that characterize spatial dependence in extremes. The assumption of independence invalidates the application of latent variable methods to modeling extremes of an atmospheric process such as precipitation because daily precipitation measurements at nearby stations are certainly not independent. This dependence is induced by the spatial coherence of storm systems so that if one weather station experiences an extreme on a specific day it is more likely that a nearby station will also experience an extreme on the same day. Failing to account for this so-called “storm dependence” means that the resulting maps of extreme statistics will be misleading since the true spatial signals will be commingled with noise.

In order to create a viable alternative analytical technique, we first conduct an extreme value analysis to estimate the extreme statistics of precipitation for each station and then interpolate these statistics to a fine grid. In order to account for spatial coherence in these extreme statistics, we use second-order nonstationary spatial Gaussian process models that account for heterogeneities in the climatology of extreme precipitation and enable statistical inference at locations where we do not have observations of daily precipitation. By construction, this framework accounts for both first-order heterogeneities in the expected values due to topography and second-order heterogeneities in the co-variability. Then, we use a nonparametric, or “block”, bootstrap approach to characterize uncertainty in the extreme statistics of precipitation while accounting for storm dependence. Intuitively, our approach can be seen as commuting the order of operations relative to existing analytical methods that in general first apply spatial and temporal interpolation to station data onto a regular grid and then calculate extreme statistics from that gridded product. We demonstrate that our method yields a larger signal-to-noise ratio in estimates of extreme statistics relative to the ratios yielded by more traditional analytical methods that do not borrow strength spatially.

Focusing on daily measurements of precipitation from the Global Historical Climatology Network (GHCN) over the contiguous United States (CONUS), the result of our analysis is a probabilistic gridded product that describes the spatial climatological distribution of extreme precipitation. To illustrate why our methodology provides an improved characterization of extreme precipitation statistics, consider two nearby towns in a homogeneous region, e.g., the Great Plains. While these two towns might experience different storms for any given day or month, our general hypothesis is that these two towns will exhibit similar statistics of extreme precipitation because of their similar climatology. The same argument holds over the rest of CONUS, especially after correcting for the effects of elevation. As a result, it makes more sense to borrow strength over space and conduct smoothing on the statistics of extremes rather than daily weather itself. A traditional extreme value analysis using gridded daily products does not incorporate this important borrowing of strength over space.

Two threads of research should be mentioned in relation to the novel methodology described in this paper. The first is Diaconescu et al. (2015) and references therein, which explores the order of operations when regridding data products and climate model output for comparison. In this paper, they compare a “first-step” procedure (which first interpolates daily weather fields and then computes a derived index on the common grid) and a “last-step” procedure (which first computes the derived index and then interpolates to a common grid), finding that the last-step procedure yields much smaller errors from the regridding process (a result that is robust to the interpolation method). The intuition for our approach supports a preference for the last-step procedure and is indeed quite similar, except that (1) we propose re-gridding station data instead of preprocessed data products, and (2) we interpolate parameters from a statistical model instead of data summaries (e.g., the largest seasonal daily precipitation total): hence, our results immediately yield a variety of gridded climatological summaries (e.g., return levels or return probabilities for any return period) instead of requiring a separate interpolation scheme for each desired summary. Second, our methodology is related to regional frequency analysis (RFA) (Wallis et al. 2007; see also Norbiato et al. 2007; Yang et al. 2010) which, as with the work at hand, seeks to provide spatially-complete maps of the climatology of extreme precipitation based on the interpolation of pointwise summary statistics calculated for each of a set of weather stations. However, the methodologies are quite different: the cornerstone of RFA is that “data from sites within a homogeneous region can be pooled to improve the reliability of the magnitude-frequency estimates” (Wallis et al. 2007), and hence RFA requires the specification of such regions. This is a nontrivial task, particularly for a large, heterogeneous domain such as CONUS. Our approach, on the other hand, allows the data itself to determine the degree of local homogeneity in the climatology of extreme precipitation across a set of weather stations, and is hence more easily applied to a generic spatial domain.

Finally, we note that the methodology outlined in this paper resolves an outstanding problem for evaluation of Earth System Models (ESMs) with respect to extremes: namely, how to properly compare irregularly observed station data with output from ESMs. Comparison of ESMs against station data poses several major challenges. First, the volume of the atmosphere sampled by station data is typically orders of magnitude smaller than the volumes represented by a typical ESM grid cell. While ESM sub-grid parameterizations are designed to emulate the parent distributions of measurable quantities and to report averages across these distributions, stations can record samples from any part of the parent distribution of their measurable quantities, including the “tails”. Second, the typical integrations of ESMs forced just with external boundary conditions are designed to mimic the climatological statistics of typical weather conditions for each surface station, but not the precise deterministic time-series of the weather actually recorded by those stations. Thus, the question is how best to compare deterministic time series of point measurements against a statistical characterization of the mesoscale climate conditions consistent with those measurements. The methodology developed in this paper addresses this question by framing the model-data intercomparison in terms of the extreme statistics from the outset, and by developing a method for interpolating statistical properties between stations that produces mesoscale and synoptic-scale statistics directly comparable to those from an ESM.

The paper proceeds as follows: in Sect. 2 we introduce the GHCN data used to create our probabilistic gridded product, and in Sect. 3 we describe the extreme value analysis, spatial smoothing, and block bootstrap approach. In Sect. 4, we present the results of our analysis and give a summary of the probabilistic gridded product, and in Sect. 4.3 we conduct a comparative analysis using the Livneh gridded daily precipitation product. Section 5 concludes the paper.

## 2 Data

The data used for the following analysis consist of daily measurements of total precipitation (in millimeters) obtained from the Global Historical Climatology Network (GHCN; Menne et al. 2012a, b) over the contiguous United States (CONUS). The GHCN is quite extensive over CONUS, consisting of over twenty thousand weather stations with measurements dating back to the late nineteenth century, although of course the length and quality of individual records are highly variable. In addition to daily precipitation measurements, the database includes three quality control flags, providing for each day an indication of the data quality (“QFLAG”), source quality (“SFLAG”), and measurement quality (“MFLAG”). Quality control for the raw (nonmissing) daily data values involved the following criteria: (1) values were kept only if the QFLAG field was blank, meaning the measurement did not fail any quality assurance checks; (2) values were kept only if the SFLAG field was not equal to “S” (which implies that the measurement may differ significantly from “true” daily data); (3) any values with MFLAG equal to “T” were set to a measurement of 0 mm (“T” indicates a “trace of precipitation”). After processing the daily values based on the quality control flags, we selected the subset of stations that had a minimum of 66.7% nonmissing daily precipitation measurements over December 1, 1949 through November 30, 2017. This procedure resulted in a high-quality set of daily precipitation measurements for \(n = 5202\) stations (see Supplementary Figure D.1) covering 68 years.

To establish some notation, define \(Z_{tk}({\mathbf{s}}_{i})\) to be the precipitation measurement, in mm, for day \(k = 1, \ldots ,m_{t}\) in a fixed 3-month season (e.g., December, January, and February or DJF) of year \(t = 1950, \ldots , 2017\) at station \({\mathbf{s}}_{i} \in {\mathcal {S}} = \{\mathbf{s}_{1}, \ldots , {\mathbf{s}}_{n}\} \subset G\), where \(m_{t}\) is the number of daily observations in a fixed season of year *t*, \({\mathcal {S}}\) denotes the \(n = 5202\) stations shown in Supplementary Figure D.1, and *G* denotes the contiguous United States. Note that the year represents a “season year” such that, for example, the 1950 DJF is defined as December, 1949 to February, 1950.

## 3 Statistical methods

Recall from Sect. 1 that the essence of our method is to first obtain estimates of the climatological features of extreme precipitation based on station data (Sect. 3.1) and then use a spatial statistical approach to infer the true underlying climatology over a fine grid (Sect. 3.2), accounting for uncertainty and storm dependence using a nonparametric block bootstrap approach (Sect. 3.3). We now provide specific details on each step of this analysis.

### 3.1 Stage 1: extreme value analysis for individual stations

*t*as \(Y_{t} ({\mathbf{s}})\), that is, \(Y_{t} ({\mathbf{s}}) = \max _{k} \{ Z_{{tk}} ({\mathbf{s}}):k = 1, \ldots ,m_{t} \}\). Coles (2001) (Theorem 3.1.1, page 48) shows that (for large \(m_{t}\)) the cumulative distribution function (CDF) of \(Y_{t} ({\mathbf{s}})\) is a member of the GEV family

The GEV framework is commonly referred to as the “block maxima” approach to extreme value analysis. The point process or “peaks over threshold” (POT) approach is often preferred to the block maxima approach because, as in the threshold excess model, estimates of the climatological coefficients are obtained from all extreme values (i.e., those that exceed a high threshold) instead of the single maximum over a block of time. But, as discussed in Coles (2001) Section 4.3.1, the main challenge is identifying a threshold for what is considered “extreme”: too small a threshold violates the mathematical (asymptotic) basis of the POT approach, leading to biased coefficient estimates, while too large a threshold results in a very small number of exceedances, yielding large variance. Also, in practice, when conducting station-specific extreme value analysis, the POT approach resulted in a larger number of numerical optimization errors. Finally, in this case where we wish to characterize trends in extremes over time, it is not clear that a temporally-constant threshold is appropriate. Therefore, we opted to use the GEV framework for this analysis. However, as a sensitivity analysis we also conducted the full analysis using the POT approach and found that results were similar (see Supplementary Figure D.9 for a comparison).

*climatological coefficients*for location \({\mathbf{s}}\), as these values describe the climatological distribution of extreme precipitation in each year. Note that the trend model (2) averages over both inter-annual variability (e.g., the El Ni no/Southern Oscillation or ENSO) and lower frequency modes of variability (e.g., the Pacific Decadal Oscillation or PDO), such that we only attempt to characterize temporally smooth trends in extremes. Some authors (e.g., Cooley et al. 2007; Risser and Wehner 2017) have also permitted the scale parameter to contain covariates allowing the width of the GEV distribution to co-vary. Uncertainty in the magnitude of the shape parameter is generally quite large, negating any benefits of allowing it to co-vary. In this paper, however, we consider only (2) because in a statistical sense it performed as well as more sophisticated trend models for individual stations.

*r*-year return value (sometimes called the return level). The

*r*-year return value, denoted \(\phi _{t}^{{(r)}} ({\mathbf{s}})\), is defined as the seasonal maximum daily precipitation total (i.e., \(Y_{t} ({\mathbf{s}}) = \max _{k} \{ Z_{{tk}} ({\mathbf{s}})\}\)) that is expected to be exceeded on average once every

*r*years. In other words, \(\phi _{t}^{{(r)}} ({\mathbf{s}})\) is an estimate of the \(1 - \frac{1}{r}\) quantile of the distribution of seasonal maximum daily precipitation in year

*t*at station \({\mathbf{s}}\), i.e., \(P(Y_{t} ({\mathbf{s}})> \phi _{t}^{{(r)}} ({\mathbf{s}})) = 1/r\). Because of how the year-specific distribution (1) has been defined, \(\phi _{t}^{{(r)}} ({\mathbf{s}})\) can be written in closed form in terms of the climatological coefficients:

*x*in year

*t*, denoted \(\rho _{t}^{{(x)}} ({\mathbf{s}})\), which indicates that there is a one in \(\rho _{t}^{{(x)}} ({\mathbf{s}})\) chance that an event at least as large as

*x*will occur in year

*t*at location \({\mathbf{s}}\). In other words, \(\rho _{t}^{{(x)}} ({\mathbf{s}})\) is the inverse probability of the seasonal maximum daily precipitation total in year

*t*exceeding

*x*, i.e.,\(P(Y_{t} ({\mathbf{s}})> x) = 1/\rho _{t}^{{(x)}} ({\mathbf{s}})\). The return period is available in closed form by inverting (4):

In summary, this first stage of our method involves applying the GEV analysis defined by (2) and (3) to the observed seasonal maxima, independently for each station. For this step, we use the climextRemes software package (Paciorek 2016), which is an R/Python package for conducting extreme value analysis with climate data. We use climextRemes functionality to obtain maximum likelihood estimates of the climatological coefficients, denoted \(\{ \hat{\mu }_{0} ({\mathbf{s}}),\hat{\mu }_{1} ({\mathbf{s}}),\hat{\sigma }({\mathbf{s}}),\hat{\xi }({\mathbf{s}})\}\) for \({\mathbf{s}} \in {\mathcal{S}}\), which are estimates of the true climatological coefficients \(\{ \mu _{0} ({\mathbf{s}}),\mu _{1} ({\mathbf{s}}),\sigma ({\mathbf{s}}),\xi ({\mathbf{s}})\}\).

### 3.2 Stage 2: spatial statistical modeling for the climatological coefficients

In order to account for dependence across the climatological coefficients over the spatial domain *G*, we use second-order nonstationary spatial Gaussian process models for each of the spatially-varying climatological coefficients in (2). Gaussian processes (GPs) are an extremely popular tool in statistical modeling, and are broadly applied in spatial and environmental statistics as well as machine learning and emulation of complex mathematical models. GPs are a form of “nonparametric” or nonlinear regression, as they characterize a nonlinear relationship between a set of inputs and an output: in our application, the inputs are geographic coordinates and the output (or response) is one of the estimated climatological coefficients \(\{ \mu _{0} ,\mu _{1} ,\log \sigma ,\xi \}\). Intuitively, GPs interpolate or smooth the output variable to infer a nonlinear relationship with the inputs.

*C*is often assumed to be second-order stationary, meaning that the covariance is fully defined by the separation vector \({\mathbf{s}} - {\mathbf{s^{\prime}}}\). However, for the application at hand, this is a highly restrictive assumption, because the spatial covariance for environmental processes likely varies across the domain. Therefore, we use a

*nonstationary*covariance function, which allows the covariance to depend on location and can account for second-order heterogeneities over the spatial domain of interest. A GP that uses a nonstationary covariance function is often referred to simply as a nonstationary GP. For further details on this subject, we refer the interested reader to Risser (2016).

*d*-variate Gaussian distribution with mean vector \({\mathbf{a}}\) and covariance matrix \({\mathbf{B}}\). In (7), \(\varvec{\delta }_{\theta} = \big (\delta _{\theta} ({\mathbf{s}}_{1}), \ldots , \delta _{\theta} ({\mathbf{s}}_{n}) \big )\) and the

*i*,

*j*element of \(\mathbf{\Sigma }_\theta\) is \(C_{\theta } ({\mathbf{s}}_{i} ,{\mathbf{s}}_{j} )\).

To complete the model specification in (6), we must specify the form of \(\delta _{\theta} ({\mathbf{s}})\) as well as the covariance function \(C_{\theta}\); for full details, see Appendix 2. In short, the mean function \(\delta _{\theta} ({\mathbf{s}})\) will be linear in an elevation-based covariate with a spatially-varying effect, and the covariance function for \(C_{\theta}\) is second-order nonstationary such that \(C_{\theta } ({\mathbf{s}},{\mathbf{s^{\prime}}}) \ne C_{\theta } ({\mathbf{s}} - {\mathbf{s^{\prime}}})\). Again, intuitively, using a nonstationary covariance function allows the second order properties of \(u_{\theta } (\cdot)\) to vary over space, including both variance and the magnitude and direction of the spatial extent of the smoothing. A statistical treatment of (7) uses a likelihood-based approach (e.g., maximum likelihood estimation) to estimate the statistical parameters associated with the mean and covariance.

### 3.3 Nonparametric bootstrap for uncertainty quantification

Finally, we specify a systematic framework for combining the GEV likelihood in (3) and the Gaussian process priors for the coefficient fields. However, note that we have not yet specified how to interrelate the likelihood (3) across stations, and these relationships are a critical component of a statistical model for daily precipitation extremes over space. As mentioned in Sect. 1, the simplest approach is the conditional independence or latent variable approach (see, e.g., Cooley et al. 2007; Craigmile and Guttorp 2013; Mannshardt et al. 2013), which assumes independence across stations in the likelihood and relies on the Gaussian process priors to capture dependence in extremes. However, while this approach is somewhat feasible for large data sets and heterogeneous spatial domains, the fact that it does not account for storm dependence makes it theoretically incorrect. And, as mentioned in Sect. 1, existing approaches for more appropriately modeling extremes over space are insufficient for daily precipitation measurements from a large network like GHCN over CONUS.

*kriging predictor*in a traditional geostatistical framework. Here, \(\widehat{\mathbf{\Sigma }}_\theta \big [ \widehat{\mathbf{\Sigma }}_\theta + \widehat{\mathbf{D}}_{\varvec{\theta }}\big ]^{-1}\) is the matrix version of [signal]/[signal + noise], so we can see that best estimate \({\widetilde{{\varvec{\theta }}}}\) is the sum of the spatial mean (\(\widehat{\varvec{\delta }}_\theta\)) and a spatial residual term (\({\widehat{{\varvec{\theta }}}} - \widehat{\varvec{\delta }}_\theta\)) that is re-scaled based on the relative magnitude of the signal and noise. In fact, this relationship can be generalized to a generic location \({\mathbf{s^{\prime}}} \in G\) for which we do not have observations of daily precipitation (e.g., on a fine grid). Similar to (10), our best estimate of the true climatological coefficient at \({\mathbf{s^{\prime}}}\) is

*i*. Then, for \(b=1, \ldots , B\), the bootstrap sample is obtained by drawing \(T = 68\) years from \(\{ 1950, \ldots , 2017 \}\), denoted \(\{a_{1}^*, a^*_{2}, \ldots ,a^*_{T} \}\), so that the

*b*th bootstrap sample for station

*i*is

*r*-return value estimate in year

*t*, is calculated as

## 4 Results

### 4.1 Spatial statistics and uncertainty quantification

In addition to the increased signal-to-noise ratio, the use of spatial statistics in our novel approach yields insight into the behavior of extreme precipitation. As described in Sect. 3.2, nonstationary Gaussian process models allow the spatial length-scale (i.e., magnitude and direction of spatial dependence) to vary across the spatial domain. As an illustration, we show results for the location intercept coefficient \(\mu _{0} ({\mathbf{s}})\) in DJF, which characterizes the center of the extreme value distribution (constant over time) for each spatial location and hence drives the magnitude of the return values. Our best estimates of \(\mu _{0} ({\mathbf{s}})\) in DJF are shown in Fig. 2a, where we show statistically-gridded estimates (using Eq. 11) as it is easier to visualize the spatial distribution with a “filled-in” map. Figure 2 also shows the elevation-based spatial mean (\(\hat{\delta }_{\theta } ({\mathbf{s^{\prime}}})\) in Eq. (11) in Fig. 2b, as well as the spatially-dependent residual (\(\widehat{\mathbf{c}}^\top _{{\varvec{\theta }}, \theta ({\mathbf{s}}^{\prime})} \big [ \widehat{\mathbf{\Sigma }}_{\theta} + \widehat{\mathbf{D}}_{\varvec{\theta }}\big ]^{-1}\big ( {\widehat{{\varvec{\theta }}}} - \widehat{\varvec{\delta }}_{\theta} \big )\) in Eq. (11) in Fig. 2c. (Corresponding plots for MAM, JJA, and SON are shown in the Supplementary Figures D.4, D.5, and D.6.) The spatial mean in Fig. 2b represents the estimated first-order properties of \(\mu _{0} ({\mathbf{s}})\), and characterizes a linear relationship between elevation and the center of the extreme value distribution. The use of a spatial Gaussian process yields the spatially-dependent residual term, which characterizes additional nonlinear relationships over space in \(\mu _{0} ({\mathbf{s}})\) that are not captured by the elevation covariate.

Focusing on the spatially-dependent residual, first note that the residuals are smooth over the southeast United States and upper Great Plains but highly heterogeneous across the Rocky Mountains and along the west coast. This is not surprising, but note that these differences can be explicitly characterized by the nonstationary Gaussian process, which estimates this length-scale directly from the data—and, these differences persist even after accounting for elevation directly via the spatial mean in Fig. 2b. (Note that elevation artifacts can still be seen in Fig. 2c, indicating some inadequacy in our characterization of the elevation-based mean.) To visualize these differences, consider Fig. 3, which shows a heuristic representation of the spatially-varying length scale (both magnitude and direction) via a set of ellipses, which are estimated locally across CONUS at the mixture component locations. Intuitively, a long/skinny ellipse (e.g., coastal Washington state in DJF) means that the length scale of spatial dependence is much larger in one direction than in the orthogonal direction; a circular ellipse (e.g., western Illinois in DJF) means that the length-scale is roughly the same in all directions. Similarly, a large ellipse corresponds to long-range spatial dependence while a small ellipse corresponds to shorter-range spatial dependence. Note that the Gaussian process estimates translate to the variations in heterogeneity noted in the gridded spatial residuals: in DJF, the Rocky mountain range and west coast generally has a shorter length-scale (i.e., highly heterogeneous) while the southeast US and upper Great Plains generally have longer length-scales (i.e., highly smooth or homogeneous). There are other interesting features to these ellipse maps, e.g., coastal effects (both east and west), as well as the southwest-to-northeast orientation of the ellipses across the southeast US (particularly in SON). Also, note the seasonal differences in the spatial length-scale, particularly DJF vs. JJA, which could be due to either the climatology or the different storm types leading to extreme precipitation in winter and summer.

As mentioned in Sect. 3.2, we include an orographic correction in the nonstationary Gaussian process mean for each GEV coefficient. Intuitively, this orographic correction involves estimating a linear relationship between elevation and each GEV coefficient empirically from the data, while allowing the magnitude and sign of this relationship to vary smoothly across CONUS. While the mean is linear in elevation, note that this does not completely specify the relationship between extreme precipitation and elevation because of the spatial smoothing induced by the Gaussian process (this is illustrated in Fig. 2). Figure 4 shows the data-driven, spatially-smoothed estimate of the relationship between extreme precipitation and elevation, across each season. Dark red areas indicate a strong positive relationship relationship between elevation and extreme precipitation (i.e., higher elevation corresponds to stronger storms), while blue areas indicate a negative relationship. Across all of the seasons, the relationship between elevation and extreme precipitation is generally positive, with the largest effects showing up in the western United States. Of course, elevation itself is rather homogeneous over much of the central United States. However, in the western United States (where elevation plays an important role in the climatology of precipitation), there are important seasonal differences in the relative elevation/precipitation relationship: for example, especially in California, the effect is larger for DJF than JJA. This is not surprising, as there is a strong wet/dry seasonal cycle in the western United States, but it is encouraging that the nonstationary GP was able to infer this directly from the data.

Of course, many gridded daily products also include an orographic correction, most notably the Parameter-elevation Relationships on Independent Slopes Model (PRISM; see Daly et al. 2008 and the references therein). While our orographic correction is applied to the climatological coefficients rather than daily precipitation itself, several notes of comparison should be made. First of all, like PRISM, our approach involves a linear relationship between the climate variable and elevation (compare Eq. 2 in Daly et al. 2008 with Eq. 14 in the Appendix). However, PRISM proceeds to incorporate a distance weighting scheme (see Eq. 3 in Daly et al. 2008) that is fixed *a priori* and spatially constant (albeit elevation-dependent). Our approach with nonstationary Gaussian processes, on the other hand, estimates the appropriate length scale for distance weighting directly from the data, using a non-constant weighting scheme over space. Furthermore, the Gaussian process prediction (via Eq. 11) implicitly accounts for the over-representation issue in distance-weighted averaging (such as in PRISM) while also explicitly removing noise in the data due measurement error. Finally, it is worth restating that the most important difference between our approach and PRISM is in *what* is actually being smoothed; whereas PRISM smooths precipitation itself, our method smooths GEV distribution parameters.

### 4.2 Probabilistic gridded product

Returning to the main goal of the analysis, we now present our new probabilistic gridded product. The product consists of spatially complete maps (on the \(0.25^\circ\) grid) of the four climatological coefficients in (2): a best estimate map calculated from the full data, as well as 250 smoothed maps calculated from the bootstrap data sets. With these gridded data in hand, one can create summary statistics for the entire spatial domain as well as bootstrap sampling distributions for confidence intervals or standard errors. Furthermore, these calculations are possible for any individual year in 1950–2017 (note, however, that we have not explicitly accounted for year-to-year variability due to, e.g., ENSO). The most common summaries would likely be return values (4) or return periods (5), but one could just as easily calculate risk ratios for comparing one time point with another (see, e.g., Risser and Wehner 2017) or even trends over time.

As an illustration, Fig. 5 shows spatially complete maps of the 20-year return value for each season, and the bootstrap standard errors are shown in Fig. 6. These maps are unique in that they are high resolution, and they produce return-value estimates consistent with what would be measured at a point (rather than over an area). As a result, the return value estimates from our method are substantially higher than those estimated from gridded datasets. For example, Figure A4b in Volosciuk et al. (2015), shows 20-year DJF return values estimated from the gridded Climate Prediction Center daily precipitation product (the equivalent of the DJF panel in our Fig. 5), and the highest daily return values shown are only about 125 mm, in comparison to values approaching 200 mm estimated using our method. We explore this comparison more deeply in Sect. 4.3.

The influence of topography, which is explicitly included in the spatial model (see Sect. 4.1) is clearly evident throughout Fig. 5, as the Sierra Nevada topographic variability associated with the Basin and Range province is clearly evident: high topography areas generally have larger return values. Notably, however, the influence of topography in the western U.S. is substantially damped during the summer convective season, with the Sierra Nevada and Cascade ranges hardly evident. This variation in orographic influence, which emerges naturally from the data itself, is consistent with what one might physically expect: orographic influence in the western U.S. is most prevalent during the winter season, when storm systems have a strong zonal flow that is roughly perpendicular to the dominant orography, and it is weaker during the season when precipitation is primarily convective. Note that Fig. 3 shows that the spatial dependence of DJF extreme precipitation in the western U.S. is more meridional than zonal and appears to be aligned with the dominant orientation of topography in the vicinity. This would indicate that though winter storms impinging on the western U.S. have a predominantly zonal flow, extremes tend to spatially co-occur along meridionally-oriented (and somewhat perpendicular) mountain ranges that intercept the flow. Intriguingly, however, orographic modulation of extremes is also absent in the southwestern U.S., which is the center of action during the monsoon season when large moisture fluxes from the south bring summer precipitation that one might expect to also cause orographically modulated precipitation. It is plausible that this is because the direction of the incoming moisture flux is roughly parallel, rather than perpendicular, to the dominant direction of ranges in the Basin and Range province, which may weaken the orographic effect.

Our probabilistic gridded product is freely available via a public repository on the Harvard Dataverse (Risser et al. 2019; https://bit.ly/2CXdnuj). The data are packaged together into network common data form (netCDF) files. Three separate files are available, with data provided separately for each season: (1) best estimates of the GEV coefficients (e.g., Fig. 2a) as well as bootstrap standard errors, (2) 10-, 20-, 50-, and 100-year return values (with bootstrap standard errors) for 1955, 1965, 1975, 1985, 1995, 2005, and 2015, and (3) smoothed bootstrap estimates (250 total) for each GEV coefficient (these are used to calculate the bootstrap standard errors in the first two files). We also provide code that can be used to calculate return values, return periods, or any function thereof for any year in 1950–2017, as well as the corresponding bootstrap sampling distribution (which can be used to quantify uncertainty via standard errors or confidence intervals).

### 4.3 Comparison to the Livneh data product

## 5 Conclusions

In this paper, we have developed novel statistical methodology for conducting a spatial analysis of extreme precipitation for a large network of weather stations over a heterogeneous domain. Using Gaussian processes, our approach uses data-driven smoothing to borrow strength over space, which yields increased confidence in our subsequent estimates of the climatology of extreme precipitation. The result of our analysis is a new “probabilistic” gridded product specifically designed for characterizing extreme precipitation, which will be made freely available in a public data repository.

Compared to traditional gridded products, our results yield important differences in estimates of return values. Furthermore, our methodology is able to produce spatially-complete, high resolution maps of return values based on irregularly observed station data, and we are therefore able to characterize extreme statistics of precipitation at small spatial scales—especially relative to the large, spatially-averaged summaries provided in, for example, Kunkel et al. (2013) or Easterling et al. (2017). As a result, our new gridded product provides important insight into the underlying physics of precipitation over CONUS (both in the spatial length-scale differences and further exploration of the extreme statistics themselves), and the return value maps are of value with respect to impacts of extreme precipitation, resource management, and infrastructure design.

An important extension of the results presented in this paper involves characterizing high-resolution observed trends in extreme precipitation. In future work we plan to explore the presence and statistical significance of any trends over time in the statistics of extreme precipitation. Such an exploration naturally requires incorporation of more sophisticated trend models where the effects of interannual or decadal variability (e.g., the El Ni no/Southern Oscillation or the Pacific Decadal Oscillation) are explicitly accounted for. However, introducing additional time-varying covariate information into a statistical model for the extreme statistics of precipitation necessitates robust methodology for selecting the variables that are relevant for explaining year-to-year variability.

In conclusion, we note that the correct interpretation and usage of gridded products closely depends upon their construction. In the methodology presented here, the long period return values within a grid box are interpreted as representative of any point within that grid box. A simple interpretation is that for a given return period, any point within the grid box has the same return value. On the other hand, modeled daily precipitation is the total precipitation within a grid box integrated at all points within the box as constrained by the conservation properties of the model. Hence, when compared to return values from a climate model, observed gridded return values obtained by our method should be considered an upper bound. In other words, if during an extreme storm, simulated precipitation is such that every point within a grid box is simultaneously precipitating at a very high rate then we would expect the modeled return values to agree with our observational estimates. This situation could indeed occur if the model’s grid cells are small enough and if storm properties are relatively spatially uniform on this scale. Hence, we might expect that our estimates of mid-latitude winter storm extreme precipitation would agree with the very high resolutions of cloud system resolving models, but not for convective summer storms, which would always be simulated lower than our estimates because of these scale considerations.

Alternatively, one might expect that return values based on gridded daily observed precipitation products, as in Wehner (2013), might be more directly comparable to modeled precipitation if the gridding process is conservative. However, because the station density is low compared to model grid boxes, the probability of missing an extreme rainfall event with the grid box is high. Hence, return values calculated from a daily gridded precipitation product provide a lower bound on what climate models should be expected to produce. As a model evaluation strategy, we can utilize these two differing ways of estimating observed long period precipitation return values as model performance metrics by considering them as upper and lower bounds on expected model targets. Typically, models at CMIP5-class horizontal resolutions produce seasonal return values that are lower than observational estimate, because simulated gradients of temperature and moisture are weaker than observed. As resolution is increased to \(\sim 25\,{\text{km}}\), simulated mid-latitude winter extreme precipitation typically compares more favorably with this lower bound, but simulated summer extreme precipitation can be too high, even compared to the upper bounds, due to deficiencies in cumulus convection parameterizations (Wehner et al. 2014). It would be reasonable to expect that as higher resolution multi-decadal simulations become available that simulated mid-latitude winter extreme precipitation will be increased but should be less than the lower bound. Improvements in simulated mid-latitude summer extreme precipitation await convection-permitting resolutions.

## Notes

### Acknowledgements

The authors would like to thank members of the 2018 Statistical and Applied Mathematical Sciences Institute (SAMSI) working group on extreme precipitation (specifically Daniel Cooley, Brook Russell, Richard Smith, Ken Kunkel, and Whitney Huang) as well as Christina Patricola and Benjamin Timmermans for helpful discussions throughout the development of this work. Furthermore, we would like to thank the reviewers of this manuscript for helpful and constructive feedback that improved the quality of the paper. The data supporting this article are based on publicly available measurements from the National Centers for Environmental Information (at ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/daily/). This research was supported by the Director, Office of Science, Office of Biological and Environmental Research of the U.S. Department of Energy under Contract no. DE-AC02-05CH11231 and used resources of the National Energy Research Scientific Computing Center (NERSC), also supported by the Office of Science of the U.S. Department of Energy, under Contract no. DE-AC02-05CH11231. This document was prepared as an account of work sponsored by the United States Government. While this document is believed to contain correct information, neither the United States Government nor any agency thereof, nor the Regents of the University of California, nor any of their employees, makes any warranty, express or implied, or assumes any legal responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by its trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof, or the Regents of the University of California. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof or the Regents of the University of California.

## Supplementary material

## References

- Barnard J, McCulloch R, Meng XL (2000) Modeling covariance matrices in terms of standard deviations and correlations, with application to shrinkage. Stat Sin 10:1281–1311Google Scholar
- Coles S (2001) An introduction to statistical modeling of extreme values. Springer, London, United KingdomCrossRefGoogle Scholar
- Cooley D, Nychka D, Naveau P (2007) Bayesian spatial modeling of extreme precipitation return levels. J Am Stat Assoc 102(479):824–840. https://doi.org/10.1198/016214506000000780 CrossRefGoogle Scholar
- Craigmile P, Guttorp P (2013) Can a regional climate model reproduce observed extreme temperatures? Statistica 73(1):103–122. https://doi.org/10.6092/issn.1973-2201/3988 Google Scholar
- Daly C, Halbleib M, Smith JI, Gibson WP, Doggett MK, Taylor GH, Curtis J, Pasteris PP (2008) Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. Int J Climatol 28(15):2031–2064CrossRefGoogle Scholar
- Daniels MJ, Kass RE (2001) Shrinkage estimators for covariance matrices. Biometrics 57(4):1173–1184CrossRefGoogle Scholar
- Davison AC, Padoan SA, Ribatet M (2012) Statistical modeling of spatial extremes. Stat Sci 27(2):161–186. https://doi.org/10.1214/11-STS376 CrossRefGoogle Scholar
- Demarta S, McNeil AJ (2005) The \(t\) copula and related copulas. Int Stat Rev 73:111–129CrossRefGoogle Scholar
- Diaconescu EP, Gachon P, Laprise R (2015) On the remapping procedure of daily precipitation statistics and indices used in regional climate model evaluation. J Hydrometeorol 16(6):2301–2310CrossRefGoogle Scholar
- Easterling DR, Arnold J, Knutson T, Kunkel K, LeGrande A, Leung LR, Vose R, Waliser D, Wehner M (2017) Precipitation change in the United States. In: Wuebbles DJ, Fahey DW, Hibbard KA, Dokken DJ, Stewart BC, Maycock TK (eds) Climate science special report: fourth national climate assessment, vol I. U.S. Global Change Research Program, Washington, DC, USA, pp 207–230. https://doi.org/10.7930/J0H993CC Google Scholar
- Fuentes M, Henry J, Reich B (2013) Nonparametric spatial models for extremes: application to extreme temperature data. Extremes 16(1):75–101. https://doi.org/10.1007/s10687-012-0154-1 CrossRefGoogle Scholar
- Gelfand AE, Kottas A (2002) A computational approach for full nonparametric Bayesian inference under Dirichlet process mixture models. J Comput Graph Stat 11(2):289–305. https://doi.org/10.1198/106186002760180518 CrossRefGoogle Scholar
- Gervais M, Tremblay LB, Gyakum JR, Atallah E (2014) Representing extremes in a daily gridded precipitation analysis over the united states: impacts of station density, resolution, and gridding methods. J Clim 27(14):5201–5218CrossRefGoogle Scholar
- Haan LD (1984) A spectral representation for max-stable processes. Ann Probab 12(4):1194–1204. https://doi.org/10.1214/aop/1176993148 CrossRefGoogle Scholar
- Hannart A, Naveau P (2014) Estimating high dimensional covariance matrices: a new look at the Gaussian conjugate framework. J Multivar Anal 131:149–162CrossRefGoogle Scholar
- Holland DM, De OV, Cox LH, Smith RL (2000) Estimation of regional trends in sulfur dioxide over the eastern United States. Environmetrics 11(4):373–393CrossRefGoogle Scholar
- Hüsler J, Reiss RD (1989) Maxima of normal random vectors: between independence and complete dependence. Stat Probab Lett 7:283–286CrossRefGoogle Scholar
- Kabluchko Z, Schlather M, de Haan L (2009) Stationary max-stable fields associated to negative definite functions. Ann Probab 37(5):2042–2065. https://doi.org/10.1214/09-AOP455 CrossRefGoogle Scholar
- King AD, Alexander LV, Donat MG (2013) The efficacy of using gridded data to examine extreme rainfall characteristics: a case study for Australia. Int J Climatol 33(10):2376–2387CrossRefGoogle Scholar
- Kottas A, Wang Z, Rodríguez A (2012) Spatial modeling for risk assessment of extreme values from environmental time series: a Bayesian nonparametric approach. Environmetrics 23(8):649–662. https://doi.org/10.1002/env.2177 CrossRefGoogle Scholar
- Krupskii P, Huser R, Genton MG (2018) Factor copula models for replicated spatial data. J Am Stat Assoc 113(521):467–479CrossRefGoogle Scholar
- Kunkel KE, Karl TR, Brooks H, Kossin J, Lawrimore JH, Arndt D, Bosart L, Changnon D, Cutter SL, Doesken N (2013) Monitoring and understanding trends in extreme storms: state of knowledge. Bull Am Meteorol Soc 94(4):499–514CrossRefGoogle Scholar
- Livneh B, Rosenberg EA, Lin C, Nijssen B, Mishra V, Andreadis KM, Maurer EP, Lettenmaier DP (2014) A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States: update and extensions. J Clim 27(1):478CrossRefGoogle Scholar
- Lovejoy S, Schertzer D, Allaire V (2008) The remarkable wide range spatial scaling of TRMM precipitation. Atmos Res 90(1):10–32. https://doi.org/10.1016/j.atmosres.2008.02.016. http://linkinghub.elsevier.com/retrieve/pii/S0169809508000562
- Mannshardt E, Craigmile PF, Tingley MP (2013) Statistical modeling of extreme value behavior in North American tree-ring density series. Clim Change 117(4):843–858. https://doi.org/10.1007/s10584-012-0575-5 CrossRefGoogle Scholar
- Maskey ML, Puente CE, Sivakumar B, Cortis A (2016) Encoding daily rainfall records via adaptations of the fractal-multifractal method. Stoch Environ Res Risk Assess 30(7):1917–1931. https://doi.org/10.1007/s00477-015-1201-7 CrossRefGoogle Scholar
- Menne MJ, Durre I, Korzeniewski B, McNeal S, Thomas K, Yin X, Anthony S, Ray R, Vose RS, EGleason B, Houston TG (2012a) Global Historical Climatology Network—Daily (GHCN-Daily), version 3, 10.7289/V5D21VHZ, nOAA National Climatic Data Center (accessed 19 March 2018)Google Scholar
- Menne MJ, Durre I, Vose RS, Gleason BE, Houston TG (2012b) An overview of the global historical climatology network-daily database. J Atmos Ocean Technol 29(7):897–910CrossRefGoogle Scholar
- Morris SA, Reich BJ, Thibaud E, Cooley D (2017) A space-time skew-\(t\) model for threshold exceedances. Biometrics 73(3):749–758CrossRefGoogle Scholar
- Norbiato D, Borga M, Sangati M, Zanon F (2007) Regional frequency analysis of extreme precipitation in the eastern Italian Alps and the August 29, 2003 flash flood. J Hydrol 345(3–4):149–166CrossRefGoogle Scholar
- Paciorek C (2016) climextRemes: Tools for analyzing climate extremes. R package version 0.1.2. https://CRAN.R-project.org/package=climextRemes. Accessed 15 Oct 2016
- Paciorek CJ, Schervish MJ (2006) Spatial modelling using a new class of nonstationary covariance functions. Environmetrics 17(5):483–506CrossRefGoogle Scholar
- Reich BJ, Shaby BA (2012) A hierarchical max-stable spatial model for extreme precipitation. Ann Appl Stat 6(4):1430–1451. https://doi.org/10.1214/12-AOAS591 CrossRefGoogle Scholar
- Risser MD (2016) Nonstationary spatial modeling, with emphasis on process convolution and covariate-driven approaches. arXiv preprint arXiv:161002447
- Risser MD, Calder CA (2015) Regression-based covariance functions for nonstationary spatial modeling. Environmetrics 26(4):284–297CrossRefGoogle Scholar
- Risser M, Calder C (2017) Local likelihood estimation for covariance functions with spatially-varying parameters: the convoSPAT package for R. J Stat Softw 81(14):1–32. https://doi.org/10.18637/jss.v081.i14. https://www.jstatsoft.org/v081/i14
- Risser MD, Wehner MF (2017) Attributable human-induced changes in the likelihood and magnitude of the observed extreme precipitation during Hurricane Harvey. Geophys Res Lett. https://doi.org/10.1002/2017GL075888 Google Scholar
- Risser MD, Paciorek CJ, Wehner MF, O’Brien TA, Collins WD (2019) A probabilistic gridded product for daily precipitation extremes over the United States. https://doi.org/10.7910/DVN/LULNUQ
- Russell BT, Cooley DS, Porter WC, Heald CL (2016) Modeling the spatial behavior of the meteorological drivers’ effects on extreme ozone. Environmetrics 27(6):334–344CrossRefGoogle Scholar
- Sang H, Gelfand AE (2010) Continuous spatial process models for spatial extreme values. J Agric Biol Environ Stat 15(1):49–65. https://doi.org/10.1007/s13253-009-0010-1 CrossRefGoogle Scholar
- Saunders K, Stephenson AG, Taylor PG, Karoly D (2017) The spatial distribution of rainfall extremes and the influence of El Niño Southern Oscillation. Weather Clim Extremes 18:17–28CrossRefGoogle Scholar
- Schäfer J, Strimmer K (2005) A shrinkage approach to large-scale covariance matrix estimation and implications for functional genomics. Stat Appl Genet Mol Biol. https://doi.org/10.2202/1544-6115.1175 Google Scholar
- Schlather M (2002) Models for stationary max-stable random fields. Extremes 5(1):33–44. https://doi.org/10.1023/A:1020977924878 CrossRefGoogle Scholar
- Serinaldi F, Kilsby CG (2014) Simulating daily rainfall fields over large areas for collective risk estimation. J Hydrol 512:285–302CrossRefGoogle Scholar
- Shaby BA, Reich BJ (2012) Bayesian spatial extreme value analysis to assess the changing risk of concurrent high temperatures across large portions of european cropland. Environmetrics 23(8):638–648. https://doi.org/10.1002/env.2178 CrossRefGoogle Scholar
- Smith RL (1990) Max-stable processes and spatial extremes. Unpublished manuscriptGoogle Scholar
- Stein ML (1999) Interpolation of spatial data: some theory for kriging. Springer Science & Business Media, New YorkCrossRefGoogle Scholar
- Sylla M, Giorgi F, Coppola E, Mariotti L (2013) Uncertainties in daily rainfall over Africa: assessment of gridded observation products and evaluation of a regional climate model simulation. Int J Climatol 33(7):1805–1817CrossRefGoogle Scholar
- Timmermans B, Wehner M, Cooley D, O’Brien T, Krishnan H (2019) An evaluation of the consistency of extremes in gridded precipitation data sets. Clim Dyn. https://doi.org/10.1007/s00382-018-4537-0 Google Scholar
- Tye MR, Cooley D (2015) A spatial model to examine rainfall extremes in Colorado’s Front Range. J Hydrol 530:15–23CrossRefGoogle Scholar
- Volosciuk C, Maraun D, Semenov VA, Park W (2015) Extreme precipitation in an atmosphere general circulation model: impact of horizontal and vertical model resolutions. J Clim 28(3):1184–1205. https://doi.org/10.1175/JCLI-D-14-00337.1 CrossRefGoogle Scholar
- Wallis J, Schaefer M, Barker B, Taylor G (2007) Regional precipitation-frequency analysis and spatial mapping for 24-hour and 2-hour durations for Washington State. Hydrol Earth Syst Sci Discuss 11(1):415–442CrossRefGoogle Scholar
- Wang Z, Rodríguez A, Kottas A (2011) A nonparametric mixture modeling framework for extreme value analysis. Technical Report UCSC-SOE-11-26, University of California, Santa Cruz. http://www.soe.ucsc.edu/research/technical-reports/ucsc-soe-11-26. Accessed 1 Nov 2016
- Wehner MF (2013) Very extreme seasonal precipitation in the NARCCAP ensemble: model performance and projections. Clim Dyn 40(1–2):59–80CrossRefGoogle Scholar
- Wehner MF, Reed KA, Li F, Bacmeister J, Chen CT, Paciorek C, Gleckler PJ, Sperber KR, Collins WD, Gettelman A (2014) The effect of horizontal resolution on simulation quality in the Community Atmospheric Model, CAM5.1. J Adv Model Earth Syst 6(4):980–997CrossRefGoogle Scholar
- Westra S, Alexander LV, Zwiers FW (2013) Global increasing trends in annual maximum daily precipitation. J Clim 26(11):3904–3918CrossRefGoogle Scholar
- Yang T, Shao Q, Hao ZC, Chen X, Zhang Z, Xu CY, Sun L (2010) Regional frequency analysis and spatio-temporal pattern characterization of rainfall extremes in the Pearl River Basin, China. J Hydrol 380(3–4):386–405CrossRefGoogle Scholar

## Copyright information

**OpenAccess**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.