2.1. Study Area
The 12,989-square-kilometer (km
2) study area consists of the drainage area upstream of the streamflow gaging stations with identifying (ID) numbers 10, 14, 18, 21, and 22, as depicted in
Figure 1a. It includes parts of the Current, Little Black, Eleven Point, Spring, and Strawberry River basins and contains 22 United States Geological Survey (USGS) streamflow gaging stations (
Figure 1a). Each of these five rivers are sub-basins of the Black River (
Figure 1). The Black River is entirely contained within the Salem Plateau Subdivision of the Ozark Plateau Physiographic Region, which is characterized as gently rolling topography with an abundance of karst features such as springs, sinkholes, and caves [
32]. Greer Spring on the Eleven Point River, Mammoth Spring on the Spring River, and Big Spring on the Current River are the three largest springs in the Ozark Plateaus whose geographic extent is depicted in
Figure 1b [
32].
The study area does not contain any major reservoirs, and the land area is predominantly forest and pastureland [
32]. For the Current River Basin, approximately 80.1, 16, and 0.1% of its area is classified as forest, grassland, and urban land, respectively. The Little Black River Basin is classified to be approximately 54.9, 26.1, 17.7, and 0.1% forest, grassland, cropland, and urban land, respectively [
33]. The Eleven Point River Basin is classified to be 65, 34, and 0.4% forest, grassland/cropland, and urban land, respectively [
34]. Forest, grassland/cropland, and urban land classifications cover approximately 48.3, 49.1, and 2.4% of the Spring River Basin, respectively. The Strawberry River Basin is reported to be approximately 66 and 31.9% forest and pastureland/cropland, respectively.
Thirty-meter resolution raster datasets representative of 2001 and 2021 from the National Land Cover Database (
https://www.mrlc.gov/data, accessed on 1 December 2023) were used to quantify land use changes from 2001 to 2021 for each of the five systems. Developed land increased by 0.3% within the Current River Basin and 0.2% for the remaining four basins. Forested/planted-cultivated land decreased by 0.3%/0.4%, 1.0%/0.3%, 1.6%/0.6%, 3.0%/0.8%, and 2.6%/0.8%, while shrubland/herbaceous land increased by 0.2%/0.1%, 1.4%/−0.4%, 0.8%/1.1%, 1.2%/2.4%, and 1.4%/1.8% within the Current, Little Black, Eleven Point, Spring, and Strawberry River Basins, respectively.
Adamski et al. [
32] summarized the climate of the Ozark Plateaus (
Figure 1b). It is characterized as temperate with its thunderstorm dominated severe weather season primarily occurring during the months from March to June. Wilkerson [
33] reported the months from April to June to be the wettest for the Current and Little Black River Basins. Miller and Wilkerson [
34] reported that March through May were the wettest months for the Eleven Point River Basin.
Figure 2 summarizes the monthly mean precipitation climatology for each of the five basins, computed using the gridded Parameter-elevation Relationships on Independent Slopes Model (PRISM) monthly climate dataset representative of the period 1981–2010 [
35]. The graphs in
Figure 2 depict a trimodal distribution for the monthly mean precipitation climatology across all five basins with two larger modes occurring in May and November and a smaller mode in July. The months of April, May, and November were consistently the three wettest individual months across all five basins. Except for the Little Black River Basin, January, February, and August were consistently the three driest months. For the Little Black River Basin, January, February, June, and August were the four driest months.
Adamski et al. [
32] indicated a general southeast directed increase for mean annual precipitation from minimum values in the north of the Ozark Plateaus to maximum values near its southern boundary. This trend is generally observed in
Figure 3a, which depicts the mean annual precipitation computed using the PRISM monthly climate dataset.
Figure 3b depicts the PRISM gridded mean annual precipitation dataset for a region surrounding the study area’s five river basins. The PRISM data-computed mean annual precipitation values for the Current, Little Black, Eleven Point, Spring, and Strawberry River basins were 1189, 1217, 1195, 1199, and 1218 millimeters (mm), respectively. This large degree of homogeneity in the mean annual precipitation climatology across the five systems is observed in
Figure 3.
Adamski et al. [
32] summarized mean monthly temperatures in the Ozark Plateaus to range from 30 to 38 degrees Fahrenheit (°F) during January, generally its coolest month, and from 78 to 82 °F in July, typically the warmest month. The mean monthly mean and minimum temperature values presented in
Figure 4 for the Current, Little Black, Eleven Point, Spring, and Strawberry River basins, computed using the gridded PRISM dataset, support their mean temperature climatology summary [
32] that January and July are the coolest and warmest months, with computed values of 0.4, 1.0, 1.0, 1.6, 2.3 and 25.2, 26.0, 25.5, 25.9, and 26.4 degrees Celsius (°C), respectively. The mean monthly mean temperature values presented in
Figure 4 are above 0 °C across all months for all five watershed systems. However, the PRISM mean monthly minimum temperature values presented in
Figure 4 are below 0 °C for all five river basins during the months of January, February, and December. Across the Current, Little Black, Eleven Point, Spring, and Strawberry River basins, March was consistently the fourth coolest month, with mean March minimum temperature values of 1.0, 1.9, 2.2, 1.6, and 2.7 °C, respectively.
Figure 5 and
Figure 6 and
Table 1 present the spatial distribution and summary statistics of elevations and basin slopes throughout the Current, Little Black, Eleven Point, Spring, and Strawberry River Basins. The scale of the raster digital elevation model data presented in
Figure 5 is one arc second, and its source is the U.S. Geological Survey 3D elevation program. The basin slopes presented in
Figure 6 were computed using the digital elevation model data shown in
Figure 5. For the Current, Little Black, Eleven Point, Spring, and Strawberry River Basins, the computed maximum basin reliefs were 389, 177, 374, 304, and 219 meters (m), respectively. The hypsometric curves shown in
Figure 7a present a fair degree of similarity across the five watershed systems that is not easily apparent upon examination of
Figure 5 and
Table 1. By contrast, the plots presented in
Figure 7b of basin specific land surface slopes reinforce the information presented in
Figure 6 and
Table 1.
Adamski et al. [
32] summarized the streams of the Black River Basin as fast flowing with minimum and maximum monthly streamflows within the Ozark Plateaus generally occurring between July and October and between March and May, respectively. Wilkerson [
33] reported flood frequencies from Alexander and Wilson [
36] for several USGS streamflow gaging stations located within the Current River Basin, including for the stations with IDs 1, 3, 8, 9, and 10, as depicted in
Figure 1a. They listed values of 27,300, 50,700, 68,700, 93,500, 113,000, and 185,000 cubic feet per second for 2-, 5-, 10-, 25-, 50-, and 100-year return periods for the station with ID number 10, as shown in
Figure 1a. Miller and Wilkerson [
34] also listed flood frequency values from Alexander and Wilson [
36] for two USGS streamflow gaging stations located within the Eleven Point River Basin, viz., the stations with ID numbers 16 and 17 as depicted in
Figure 1a. Southard and Veilleux [
37] computed and reported flood frequency values for 14 of the 22 USGS streamflow gaging sites shown in
Figure 1a, i.e., stations with ID numbers 1, 3, 6–8, 9, 10, and 15–21, as depicted in
Figure 1a.
2.4. Methods
In this study, we applied the EVT-based MSP modelling approach introduced by Asadi et al. [
22]. The modelling analysis involved two parts, one to model the spatially variable marginal distributions and another to properly account for the spatial dependence among the observed flood data [
6,
7,
8,
22]. Asadi et al. [
22] comprehensively outlined their MSP based modelling approach, and herein we only highlight a few of its essential features. We encourage the interested reader to refer to their work for the full details [
22]. The following section, which discusses fitting the marginal distributions, includes a description of aspects that were unique to this study.
2.4.1. Marginal Fitting
In univariate EVT, it can be shown that a distribution is max-stable if and only if it is the generalized extreme value (GEV) distribution [
39]. Mathematical nondegenerate limit law expressions of max-stability exist in the multivariate and spatial process settings [
10]. In either case, univariate EVT results guarantee that the marginal distributions of an MSP are max-stable GEV distributions, possibly with GEV model parameters that may vary spatially. Ribatet [
10,
11], Ribatet et al. [
9], Davison et al. [
12], and Cooley et al. [
13] provided thorough summaries of MSPs and MSP based modelling.
Asadi et al. [
22] presented a threshold exceedance Poisson point process independence likelihood for marginal fitting:
where
and
denote the GEV shape, scale, and location parameters at fixed streamflow gaging site locations on the river network,
, the number of years of observations at location
, and the empirical
-quantile,
, of the data
for location
, respectively, and wherein
. With
the parameters
and
equal those in the GEV distribution for annual maxima.
Trend surfaces were defined to support prediction throughout the entire river network, . Trend surfaces spatially model the location, , scale, , and shape, , parameters of the known GEV marginal distributions as a function of location s. For example, linear trend surfaces are of the form , , , where and are the parameters and covariates of the linear trend surface for , , and , respectively. Factors that are assumed or known to influence extreme flood hydrology in a drainage basin, for example, climatological, morphometric, and physiographic data, were candidates to be included as covariates.
It is important to model the spatial variation of the marginal parameters by carefully “building relevant trend surfaces including any relevant covariable” [
10]. Poor characterization of
,
, and
complicates estimation of the dependence parameters [
10,
40]. In this study, linear trend surfaces for the known GEV marginal parameters were developed by leveraging the theory of spatial extremes [
9,
10] and recent advances for fitting general linear models [
15,
16,
17,
31].
The elastic net penalty [
41] was applied to regression models to facilitate model selection from among the set of potential covariate models using the trend surface fitting methodology introduced by Love et al. [
31]. The elastic-net penalty is a convex combination of the penalties of ridge [
42,
43] and lasso [
44] regression, and the resulting estimates are able to retain properties of both approaches. Given observations
, an
matrix of covariates
, and an assumed linear model:
the elastic-net minimizes:
where
is non-negative and tuned to weight the penalty term;
controls the penalty term to vary from ridge to lasso regression at
and
, respectively; and
is the weight assigned to the ith observation [
9]. Ridge regression results in solutions that include all the predictors, whereas application of lasso regression yields sparse, much more easily interpretable solutions [
45]. The elastic-net penalty is a convex combination of these two penalties. As the parameter that weights the relative contributions of the L
1 and L
2 penalties increases from 0 to 1, the number of non-zero estimated coefficients increases from 0 to the sparsity of the lasso [
15].
Automatic variable selection was a primary aim for the marginal fitting analysis; therefore, we weighted the L
1 penalty more heavily so that the elastic-net performed much like lasso regression while retaining ridge regression’s capacity to collectively shrink the coefficients for any highly correlated covariables [
15,
46]. To select the tuning parameter, cross validation (CV) was employed with each elastic-net model fit. Each elastic-net model was fit using the R software (version 4.2.1) package ‘glmnet’ [
15]. For each model, the pseudo responses were made up of the three GEV univariate parameter estimates at each location, and a set of spatially varying covariates were used as covariates in the models. Independent elastic net-model fits were performed for
and
and guided subsequent spatial GEV model fitting and selection. We note that we set
, as in EVT; it is common to consider the GEV shape parameter in this manner [
18,
47], especially over homogeneous regions.
Figure 8 is a schematic diagram depicting the main elements of the marginal fitting method.
2.4.2. Dependence Model
Asadi et al. [
22] introduced an MSP dependence model for extreme flood data. Their modelling approach aimed to account for both the river distance between hydrologically connected gaging stations and for stations that are not connected but share common meteorological events. The former is simply the distance along the river, whereas the latter is termed the hydrologic distance and is defined to be the Euclidean distance between the weighted (e.g., using precipitation climatology or elevation) centroids of their upstream drainage areas. The overall distance metric that combines both river distance and hydrologic distance is defined as:
for any
where
,
,
,
,
,
,
, and
represent a weight that is assigned to the dependence term for flow-connected gaging stations (
), a weight assigned to the dependence term for gaging sites that are not flow-connected (
), weights that account for the proportions of extreme flood discharge values coming from each branch of the river network, the river distance between sites
, the distance beyond which inter-site correlation is essentially zero, a rotation and dilation matrix to account for geometric anisotropy, the hydrological location of a gaging location on the river network, and a variogram shape parameter, respectively. Understanding the desire to model any location in
, observed or not, Asadi et al. [
22] suggested the use of elevation as a surrogate for precipitation and that values for
be estimated by integrating elevations for the area upstream of each gaging station. Similarly, they suggested the hydrologic location be defined as the center of mass of the precipitation climatology (or elevation, as a replacement for precipitation) for each gaging site’s contributing drainage area [
22]. In order to account for potential anisotropy, the rotation and dilation matrix
is given by:
The parameters , , , , and were estimated via the fitting of the MSP. Large values of corresponded to weak dependence, whereas small values correspond to strong dependence.
The dependence measure
was constructed in the following manner. We defined
if
and
were flow-connected and
otherwise [
22]. The weights
reflect the number of bifurcations that occur in the river network between the two locations. We used the “linear with sill” covariance function given by
[
22]. For additional background on this and other covariances on river networks, refer to Ver Hoef et al. [
48] and Ver Hoef and Peterson [
49].