1 Introduction

Wetlands make up 5% of the planet’s terrestrial surface (Mitsch and Gosselink 2007), provide many critical ecosystem services worldwide, and are hotspots of biological diversity. A wide variety of wildlife, including birds, amphibians, mammals, and invertebrates, rely on wetlands for all or portions of their lifecycles (Dudgeon et al. 2006; Junk et al. 2006). In addition, wetlands provide groundwater recharge, act as storage and filtration sites for runoff following precipitation events, and are critical components of nutrient cycles (Howard-Williams 1985, Mitsch and Gosselink 2007; Mitsch et al. 2013). Wetlands can also store greater amounts of carbon than upland sites (Lal 2008), making them important carbon sinks. Despite these ecosystem services and benefits to human societies, wetlands worldwide have undergone extensive degradation and destruction due to human activities, particularly from agricultural development (Dahl et al. 1991; Davidson 2014). While many remaining wetlands are still vulnerable to these ongoing pressures, accelerating climate change is increasingly a predominant threat to the ecological integrity and conservation of these ecosystems (Erwin 2009).

Shallow depressional inland wetlands are particularly vulnerable to climate change (Johnson et al. 2005; Brooks 2009; Burris and Skagen 2013). Depressional wetlands form in geographic low areas and are often isolated from other water sources like lakes and rivers (Gala and Young 2015); thus, they usually rely on direct precipitation or overland flow of precipitation runoff as major water sources (Tiner 2003; Gala and Young 2015). Although some accumulated water that fills these basins after precipitation can be returned to the water table providing groundwater recharge (Van der Kamp and Hayashi 1998; Gurdak and Roe 2010), most water is lost to the atmosphere through evaporation or is transpired by vegetation (Parkhurst et al. 1998; South et al. 1998). As a result, weather is a critical driver of wetland hydroperiod (i.e., timing and duration of flooding), particularly in regions where evapotranspiration exceeds precipitation. The length and frequency of weather-driven wet-dry cycles in these wetlands influence ecosystem processes like nutrient cycling (Busnardo et al. 1992; Lockaby et al. 1996; Battle and Golladay 2001), as well as plant and animal communities (Fay et al. 2016; Van der Valk and Mushet 2016; Euliss et al. 2004). Worldwide, climate change is altering precipitation and temperature patterns, fundamentally changing wetland hydrologic cycles and function (Erwin 2009).

Climate projections for the Southern Great Plains of North America predict warmer year-round temperatures and increasing variability of precipitation patterns throughout the year (Kloesel et al. 2018; IPCC 2021). Spring and summer are expected to become drier with fewer, more-intense precipitation events, while predictions for winter and fall indicate either limited changes in precipitation or a shift to wetter conditions. As a result, the numerous depressional wetlands in this region face an uncertain future under these unprecedented conditions. Depressional wetlands in the Southern Great Plains provide important surface water resources and critical stopover habitat for migratory waterfowl and shorebirds using the North American Central Flyway to travel to and from their northern breeding grounds (Bolen et al. 1989; Gersib et al 1992; Albanese and Davis 2015). As the region shifts toward warmer, drier conditions in the future, wetland availability may decline (Uden et al. 2015), reducing stopover habitat and connectivity for these species (Albanese and Haukos 2017). However, despite the likelihood of substantial effects of climate change in this region, few studies have evaluated the potential effects of changing weather patterns on wetland inundation dynamics, and most studies that have been conducted focused on the spring migration period (Uden et al. 2015). This leaves a substantial knowledge gap regarding the potential effects of climate change throughout the year. Understanding the effects of changing weather across seasons is critical it may change wetland hydroperiods and cause shifts in plant and animal communities and potentially limit the capacity of wetlands to adequately provide ecosystem services such as nutrient cycling and carbon storage (McKenna and Sala 2016; Owen et al. 2020). The lack of comprehensive information regarding wetland inundation patterns throughout the year and the lack of associated year-round projections of inundation patterns relative to future climate conditions limit the development and implementation of management and conservation strategies designed to maintain biodiversity and critical ecosystem services under changing climatic conditions.

To improve understanding and prediction of climate change effects on hydrologic cycles of depressional wetlands, we developed and implemented a model that used variables for weather conditions, wetland characteristics, and land use in surrounding watersheds to describe wetland inundation patterns over time in the Pleistocene Sand Dunes Ecoregion of central Oklahoma, USA. This ecoregion represents an important area for conservation due to its high density of wetlands and the large numbers of migratory birds that use the area as stopover habitat (Heitmeyer and Vohs 1984; Albanese and Davis 2015). We expand on previous studies by using remote sensing data (Landsat) collected throughout the year, which allows modeling of wetland inundation patterns at relatively fine temporal extents across all seasons. We then used this model, along with regionally downscaled climate projections, to predict wetland availability and inundation in the spring, summer, fall, and winter across a range of climate scenarios for the middle and end of the century (2036–2050 and 2084–2099, respectively). Describing wetland inundation in relation to weather, wetland, and land use variables provides wetland managers and conservation planners with crucial information to develop strategies to maintain wetlands and their critical ecosystem functions and services under a changing climate.

2 Methods

2.1 Study site

The study area included the eastern half of the Pleistocene Sand Dunes Ecoregion (Omernik 1987) adjacent to the Cimarron River in central Oklahoma, encompassing portions of four counties (Garfield, Kingfisher, Logan, and Major) within an approximately 1300 km2 area (Fig. 1). Climate is semi-arid to subtropical humid with annual evapotranspiration greatly exceeding average precipitation. Average yearly rainfall is approximately 83 cm in the eastern part of the study area and decreases to 75 cm in the west. Rainfall is highly variable between years, with yearly rainfall ranging from 53 cm (2012) to 145 cm (2007) during the period of our study (Oklahoma Mesonet 2015). On the leeward terraces of the Cimarron River, aeolian dune formation and migration have occurred during episodic periods of extreme aridity, with the last period of significant dune formation occurring approximately 800–900 years ago. In the valleys of the now stabilized dune fields, the area has abundant, temporary, interdunal depressional wetlands (Lepper and Scott 2005). The study area is estimated to be the largest aggregation of wetlands in central Oklahoma, with an average density of 4 wetlands/km2 (Dvorett et al. 2016). The majority of wetlands in the study area are small with more than 80% of them < 0.5 ha and approximately 6% > 2 ha. Inundation of these wetlands is primarily dependent on precipitation within their small watersheds. The predominant land use in the study area is agriculture, with the primary crops being rye (Secale cereale) and wheat (Triticum aestivum).

Fig. 1
figure 1

(A) Map of the study area in the Pleistocene Sand Dunes Ecoregion of Oklahoma, USA, and aerial image of a portion of the study area highlighting a location of high wetland density. Common migratory bird species that occur in wetlands of the Pleistocene Sand Dunes Ecoregion include (B) male and female Mallard (Anas platyrhincos) and (C) long-billed dowitcher (Limnodromus scolopaceus, front center) and Hudsonian godwit (Limosa haemastica, three individuals in back). Photo credit B. York

Full size image

2.2 Wetland inundation

Wetland inundation was determined using 57 classified Landsat 5 and Landsat 7 images from a single scene (Path 28 Row 35) from the winter of 1995 to the winter of 2011 (USGS 2015). Images were classified into water and upland pixels using a decision tree analysis on Landsat spectral bands 3, 4, and 5 that was specifically designed for the region, resulting in a high accuracy (kappa > 0.93). A detailed description of the decision tree classification and imagery validation is in Dvorett et al. (2016). Following classification, we applied the change detection tool in ENVI 5.2 (Exelis Visual Information Solutions, Boulder, CO) to identify wetlands that became inundated and those that remained dry between subsequent Landsat scenes. A wetland was deemed inundated when it was completely dry in the first image of a pair and contained at least one pixel (0.09 ha) of water in the second image. A wetland was deemed dry if it did not contain water in either the first or second scene. While each Landsat scene is collected every 16 days, paired images for the change detection were 48 to 96 days apart because of the availability of Landsat scenes with < 10% cloud cover and no popcorn clouds (isolated cumulus or stratocumulus clouds) (Collins et al. 2014).

2.3 Variable development

This study included 497 wetlands that ranged in size from 0.36 to 22.23 ha and had no indication of hydrological alteration due to excavating or ditching. Inundation at these wetlands was identified from the paired classified Landsat images described above. We included three weather variables that were summarized over multiple temporal extents (1 month, 2 months, and 3 months prior to each Landsat image used to determine wetland inundation status) including total precipitation (mm), maximum 24-h precipitation event, and average daily temperature. We used multiple temporal extents to account for the fact that different weather variables may affect wetland inundation dynamics over different characteristic scales. Additionally, we included the Palmer Drought Severity Index (PDSI) calculated for the 6 months prior to an image to evaluate the effects of prolonged drought conditions on inundation status. Weather variables were derived from PRISM 4-km resolution daily weather maps and were acquired from PRISM Climate Group (PRISM Climate Group, 2017). Historical PDSI data was downloaded from the National Oceanographic and Atmospheric Administration (NOAA) National Center for Environmental Information (NOAA 2015). To account for the role of landscape context in determining wetland hydrologic cycles, we included four variables describing land use and wetland characteristics, including watershed size, wetland size, percent agriculture in the watershed, and average soil drainage class in the watershed. We selected these variables as they have been shown to be important for influencing wetland hydrologic cycles in similar Great Plains wetland ecosystems (Cariveau et al. 2011; Bartuszevige et al. 2012). We delineated watersheds for each wetland using the depression evaluation tool in ARC MAP 10.1 (ESRI, Redlands, CA) and 1-m spatial resolution light detection and ranging (LiDAR) elevation dataset. LiDAR data were generated by the USDA Natural Resources Conservation Service (NRCS) in January 2012 with 18.5-cm vertical accuracy and 1.4-m point spacing (NRCS 2012). Wetland size was derived from wetland maps created by Dvorett et al. (2016). We calculated percentage agricultural land use in each watershed delineated for study wetlands using the 2011 National Land Cover Dataset (Homer et al. 2015). For each wetland, a soil drainage class was calculated as a weighted average in the surrounding watershed from the NRCS Soil Survey Geographic Database (SSURGO) (NRCS 2015). Drainage classes A through D were converted to 1 through 4, with 4 generally having the lowest hydrologic conductivity and highest water table (Soil Survey Division Staff 1993).

2.4 Data analysis

We used an Akaike information criteria (AICc) model selection approach to select the best model describing probability of wetland inundation (Burnham and Anderson 2002). We only used data from 1995 to 2006 to train the model, while the remaining dataset (2007–2011) was withheld for model evaluation. For all model sets described, we considered the model with the lowest AICc value as the top ranked model for that model set. In cases where more than one model was within 2 AICc units of the top ranking model, we considered these models to be competitive, and to have similar levels of support relative to the top model (Burnham and Anderson 2002). To avoid overfitting models by including spurious parameters and to simplify our final model, we used the model with the fewest variables among competitive models for subsequent analytical steps. We used binomial generalized linear mixed effects models (GLMMs), using inundation status of individual wetlands (inundated or not inundated) as a binary response variable. For all models, we used a crossed random effect with wetland ID and year of the study as separate random variables. This random effects structure was used to account for yearly variation in conditions, and for the fact that each wetland was observed multiple times throughout the study. We evaluated pairwise correlations among variables using Pearson’s correlation coefficients (r), and any combination of variables that were r >|0.6| were not included together in the same model. As a result, we did not use total rainfall and maximum rainfall variables in the same model as they were highly correlated (r > 0.72). Models were fitted using the R package glmmTMB (Magnusson et al. 2017), and model selection was performed using the package AICcmodavg (Mazerolle and Mazerolle 2017).

Prior to final model selection, we screened each of the three weather variables that were measured over multiple temporal scales (average daily temperature, total precipitation, and maximum rainfall) and selected the temporal scale that provided the best fit for the data. This was done by using AICc to compare models with two-way interactions between season and one of the three temporal scales for each weather variable. We used interactions with season for all of the weather variables to account for potentially changing relationships between weather and inundation across seasons. Seasons were defined using meteorological divisions as follows: March through May (spring), June through August (summer), September through November (fall), and December through February (winter). We then used a sequential model building process to develop our final model for wetland inundation. First, we considered all combinations of variables describing wetland characteristics and land use (wetland size, watershed size, percent crop with a watershed, and soil drainage class), and then used the variables from the top ranked model in subsequent models. After selecting the top model for wetland and land use characteristics, we considered a second model set containing all possible combinations of weather variables (total rainfall, maximum rainfall, average daily temperature, and PDSI) in addition to the selected wetland and land use variables. As with all previous steps, models were ranked with AICc. The final model was further evaluated for multicollinearity with variance inflation factors (VIFs), and the variable with the highest VIF value was removed until all variables, apart from interaction terms, had VIF values below 3 (Zuur et al. 2009).

We assessed model fit using the conditional and marginal R2 values (Nakagawa and Schielzeth 2013). These metrics allowed us to determine how much of the variance in inundation status was explained by the fixed effects of our model (marginal R2) and the entire model including random effects (conditional R2). To estimate a probability cutoff for inundation status, we calculated the receiver operating characteristic (ROC) curve for the final model and chose the value that maximized the true positive rate while minimizing the false positive rate. This resulted in a probability cutoff of 0.45. From the ROC curve for the final model, we calculated the area under the curve (AUC) which provides an estimate of a model’s predictive accuracy with values close to 1 indicating high predictive ability and values close to 0.5 indicating the model preforms no better than would be expected by chance. To further evaluate the final model’s ability to predict wetland inundation, we estimated the probability of inundation for the evaluation dataset (2007–2011). We used the probability cutoff determined from the training data to classify each wetland as inundated or not. We then calculated a new AUC for the testing data, and confusion matrices for the overall model and each of the four seasons. For each confusion matrix, we report the overall accuracy, model sensitivity (estimate of the model’s ability to identify true positives), and model specificity (estimate of the model’s ability to identify true negatives).

2.5 Projections of wetland inundation

To predict effects of climate change on wetlands, we used downscaled general circulation models (GCMs) from three Representative Concentration Pathways (RCPs), with each RCP representing a different emissions scenario for greenhouse gases. We used the RCP 2.6 scenario as a best-case scenario, as this scenario represents the situation where global greenhouse gas emissions peak around 2020 and decline through the century. We used RCP 8.5 as a worst-case scenario as this path predicts carbon emissions will remain unchanged through the twenty-first century. RCP 4.5 was used as an intermediate case as this scenario predicts that emissions will peak midcentury and decline through 2100. From each RCP, we downloaded weather forecasts for two 15-year periods, one in the middle of the century (2036–2050) and one at the end of the century (2086–2099) from 10 randomly selected GCMs. We used 10 GCMs for each emissions scenario so that we could capture the range of variability of predicted wetland inundation across climate models within an RCP scenario and between RCP scenarios (Sofaer et al. 2016).

Using the final wetland inundation model from the preceding analysis, we predicted wetland inundation status for each of the 497 wetlands for each year in the mid and end of century periods for each GCM. To standardize our estimates across years, we estimated inundation for the midpoint of each of the seasons resulting in four predictions for each wetland per year. To compare predicted wetland numbers to current conditions, we first calculated the average number of inundated wetlands per season for each GCM in each period (midcentury, end of century). We then calculated the proportion of years each wetland was predicted to be inundated under each GCM. These summary statistics were then compared to the same statistics calculated for the 15-year period of the study (1995–2011).

3 Results

Probability of wetland inundation was influenced by variables that describe both the characteristics of individual wetlands and weather conditions within each season (Supplemental Information Tables 1a, b, c). During the model selection step for individual wetland characteristics, four models were considered competitive (< 2 AICc units from the top model; Supplemental Information 1c). To avoid overfitting and simplify the final model, we used the wetland characteristics model with the fewest variables for our final model selection process. The top ranked model resulting from this process contained the main effects for wetland size, percent cropland within a watershed, PDSI, and the interactions between season and average daily temperature (averaged over 3 months) and between season and total precipitation (summed over 2 months) (Table 1 and Supplemental Information Table 1c). No other models were considered competitive in the final model set (Table 1). The final model showed relatively good fit for the data (marginal R2 = 0.61; conditional R2 = 0.72). The AUC statistics for the training dataset (1995–2006) and evaluation dataset (2007–2011) were 0.93 and 0.82, respectively. The final model showed relatively good predictive ability for the testing dataset, correctly predicting wetland inundation across all seasons 80% of the time (Supplemental Information Table 2). Model accuracy was somewhat variable among seasons, with the model being most accurate for fall and winter (model accuracy 83% for fall and 97% for winter), and least accurate for summer (model accuracy = 53%; Supplemental Information Table 2) while model accuracy was intermediate for the spring period (model accuracy = 77%; Supplemental Information Table 2).

Table 1 Model coefficients and 95% confidence intervals for the top ranked model describing inundation for depressional wetlands in the Pleistocene Sand Dunes Ecoregion of Oklahoma, USA, in 1995–2006
Full size table

Both total precipitation and daily average temperature were good predictors of probability of inundation across all four seasons. However, each of these variables was associated with wetland inundation at a different temporal extent; total rainfall provided the best model fit when measured over 2 months, whereas average daily temperature provided the best fit at the 3-month extent (Table 1, Supplemental Table 1c). In all seasons, probability of inundation increased with greater rainfall totals (Table 1, Fig. 2a). Changes in total precipitation had the greatest effect on probability of inundation in spring and winter. For example, a 10-mm increase in precipitation during these seasons resulted in a 4% increase in probability of inundation (Fig. 2a), while a similar increase in precipitation in summer resulted in a 2% increase in probability of inundation. Average daily temperature over the previous 3 months had a negative effect on probability of inundation in all seasons (Table 1, Fig. 2b). Like total precipitation, average daily temperature had the greatest effect in spring; a 1 °C increase in average temperature during this season resulted in a 4.7% decrease in inundation probability, while the same increase in average temperature in summer caused a 3% decrease in inundation probability (Fig. 2b). Drought over the previous 6 months (PDSI) was negatively related to probability of wetland inundation (β = 0.22; 95% CI = 0.17, 0.27) (Fig. 2c).

Fig. 2
figure 2

Relationship between probability of inundation and (A) total precipitation (measured over previous 2 months), (B) average daily temperature (measured over previous 3 months), (C) Palmer Drought Severity Index (PDSI), (D) wetland size, and (E) percent cropland within a watershed, for depressional wetlands in the Pleistocene Sand Dunes Ecoregion of Oklahoma, USA, in 1995–2006. Gray bands indicate 95% confidence intervals (confidence intervals omitted from panels A and B for clarity)

Full size image

Smaller wetlands and wetlands in areas with lower proportions of row crops in the watershed were less likely to be inundated than larger wetlands and those in watersheds dominated by row crops. For every hectare increase in wetland size, inundation probability increased by approximately 1.7% when holding all other variables at their mean (β = 0.34; 95% CI = 0.25, 0.42; Fig. 2d). Wetlands in watersheds dominated by row crops were more likely to be inundated than wetlands surrounded by grassland (β = 0.48, 95% CI = 0.20, 0.77), but the marginal effect for percent cropland was relatively small at 0.89% difference in probability of inundation between watersheds entirely covered in row crops versus grasslands (Fig. 2e).

3.1 Climate projections

GCMs across all RCPs showed a consistent pattern of greater average temperatures across all seasons compared to current conditions, with midcentury temperature predictions for different RCPs overlapping considerably (Supplemental Figure S1). There was greater uncertainty regarding changes in precipitation, but GCMs generally predicted less precipitation in spring and summer, with a trend toward greater precipitation in fall and winter. Predictions for precipitation were similar between midcentury and end of century, but GCMs, especially those in RCP 8.5, showed a trend toward greater increases in temperature by the end of century. As a result, all three emissions scenarios indicate a general trend toward there being fewer inundated wetlands in spring and summer as soon as the middle of the century (2036–2050) (Figs. 3 and 4). Using the average from all ten GCMs in each emissions scenario, the average number of inundated wetlands in the spring was 39%, 52%, and 42% lower in 2036–2050 compared to 1995–2011 for RCP 2.6, 4.6, and 8.6, respectively. On average, individual wetlands were inundated in spring in 46% (7 years) of years from 1995 to 2011, but projections indicate average springtime inundation frequency could decline to as little as 26% (4 of 15 years), 20% (3 of 15 years), and 26% (4 of 15 years) of years for RCP 2.6, 4.6, and 8.6 respectively by midcentury (Fig. 5, Supplemental Information Figure S2, S3). The GCMs predicted an average 77%, 73%, and 79% decrease of inundated wetlands in summer for 2036–2050 for RCP 2.6, 4.6, and 8.6 respectively (Fig. 3). Frequency of inundation was much lower in summer compared to other seasons under current conditions with individual wetlands containing water an average 2 out of 15 years (13%). By midcentury, wetlands are projected to be dry during the summer in most years (Fig. 4). Estimates for the end of the century (2084–2099) showed considerable overlap with midcentury predictions (Fig. 4, Supplemental Information Figures S2 and S3). The effect of climate change on wetland inundation in winter and fall is less clear, as GCMs in all three emissions scenarios predicted both higher and lower numbers of inundated wetlands compared to current conditions (Figs. 4 and 5).

Fig. 3
figure 3

Estimated number of inundated wetlands under three different greenhouse gas emissions scenarios in the (A) spring, (B) summer, (C) fall, and (D) winter for depressional wetlands in the Pleistocene Sand Dunes Ecoregion of Oklahoma, USA, for the middle of the century (2036–2050), and the end of the century (2084–2099). The three emissions scenarios are represented by RCP 2.6 (low emissions scenario), RCP 4.6 (intermediate emissions scenario), and RCP 8.6 (high emissions scenario). Triangles represent 10 randomly selected general circulation models (GCMs) from each emissions scenario. Dashed horizontal line represents the average number of inundated depressional wetlands in each season for the period of this study (1995–2011)

Full size image
Fig. 4
figure 4

Projections of percentage of years that individual wetlands are inundated during the period of the study (current conditions, 1996–2011) and during a 15-year period at the middle of the century (2036–2050), in each season for the Pleistocene Sand Dunes Ecoregion of Oklahoma, USA, for three greenhouse gas emissions scenarios (RCP 2.6, 4.6, 8.6). Future inundation for each RCP is estimated as the average of 10 general circulation models

Full size image
Fig. 5
figure 5

Percent change in years inundated during a 15-year period during the middle of the century (2036–2050) compared to current conditions, in each season for the Pleistocene Sand Dunes Ecoregion of Central Oklahoma, USA, for three greenhouse gas emissions scenarios (RCP 2.6, 4.6, 8.6). Future inundation for each RCP estimated as the average of 10 general circulation models

Full size image

4 Discussion

Substantial loss and degradation of wetlands has occurred worldwide, including in North America, during the last century (Dahl et al. 1991; Davidson 2014), and altered weather patterns resulting from climate change are increasingly a predominant threat to remaining wetlands (Erwin 2009). In this study, precipitation and temperature were the primary drivers of inundation patterns for depressional wetlands in the Pleistocene Sand Dunes Ecoregion of Oklahoma, and variables related to land use and landscape context had relatively minor effects. Similar to previous work (Uden et al. (2015), we found that the rising temperatures and more-sporadic precipitation expected in this region with climate change will reduce numbers of inundated wetlands in spring and summer, which are key seasonal periods when humans, plants, and wildlife depend on them for crucial resources and services. However, we also found that projected effects of climate change on wetland inundation in fall and winter are more variable, emphasizing the uncertainty of year-round effects of climate change in this region. The climate of the Southern Great Plains is highly variable within and between years (Clark et al. 2002; Zambreski et al. 2018), and wetlands in this region are characterized by highly variable water regimes with natural cycles of high inundation probability in spring and winter, and occasional dry periods, particularly in summer. Species that rely on wetlands in the region have evolved under this variable regime to have adaptions for coping with high inter- and intra-annual variability. However, our projections show that climate change will likely increase the frequency and severity of the drier extremes of wetland inundation cycles, potentially having important implications for ecosystem function and biodiversity.

Total precipitation over the previous 2 months was a better predictor of wetland inundation than any other precipitation variable, including maximum 24-h rainfall events over the previous 1, 2, and 3 months. This dependence of inundation on precipitation also was shown in previous studies in playa wetlands in the Southern Great Plains (Uden et al. 2015; Cariveau et al. 2011; Johnson et al. 2011; Bartuszevige et al. 2012). Direct precipitation and runoff from the surrounding landscape are the primary water sources for depressional wetlands, and higher rainfall totals in the preceding 2 months resulted in greater probabilities of inundation in all seasons. The amount of rainfall in individual maximum precipitation events likely does not adequately characterize the total amount of water a wetland receives in a season, because even though large storms are important for filling wetlands (Cariveau et al. 2011; Karim et al. 2012), smaller storms also contribute to maintaining water levels (Gerla 1992). However, because precipitation in this and many other regions is expected to become concentrated in fewer, more-extreme events (Prein et al. 2017), the occurrence of intense precipitation events may eventually play a larger role in influencing timing and duration of wetland inundation. Increased frequency of high-intensity storms may also increase sedimentation rates in wetlands, negatively impacting wetland function and reducing wetland hydroperiod (Burris and Skagen 2013; Tsai et al. 2007). As suggested by our results showing that higher temperatures also reduce inundation probability, effects of changing precipitation regimes may be exacerbated by the rising temperatures and increased frequency and severity of droughts that are projected for the Southern Great Plains (Bukovsky et al. 2017; Mullens and McPherson 2019). Specifically, more-flashy precipitation events, along with warmer temperatures and prolonged intervening dry periods, will lead to higher evapotranspiration rates that will likely cause even shorter hydroperiods in remaining inundated wetlands.

Our model projected fewer inundated wetlands in the Pleistocene Sand Dunes ecoregion, particularly in the spring and summer when average numbers of inundated wetlands could decline by as much as 42% and 79%, respectively, by the middle of the century under the most-extreme emissions scenario. Even under relatively optimistic conditions represented by RCP 2.6, models still predicted fewer wetlands will be inundated in spring and summer compared to current conditions. This is likely because, even if dramatic action is taken to reduce carbon emissions in the next decade, many regions will still experience continued increases in temperature due to buildup of greenhouse gases in the atmosphere from past emissions (Randers and Goluke 2020; IPCC 2021). In contrast to spring and summer, potential effects of climate change on wetland inundation in fall and winter were more ambiguous. Predictions based on individual general circulation models across all three RCPs showed a range of potential outcomes including both more and fewer inundated wetlands compared to current conditions. This uncertainty could arise due to a number of factors. Although there is high confidence in projections that spring and summer will become drier, predictions for winter and fall precipitation are less certain, with models predicting either no change (Easterling et al. 2017) or slightly higher precipitation (Guo et al. 2019). Any increases in fall and winter precipitation could counteract effects of increases in temperature, which could lead to unchanged or even increasing wetland inundation. Additionally, most plants are entering dormancy or are already dormant during fall and winter, and the combination of lower plant transpiration, reduced solar angles, and cooler average temperatures compared to summer, may result in a comparatively small effect of increasing temperatures on the wetland water budget during these seasons (Parkhurst et al. 1998; Anda et al. 2015; Yan et al. 2020).

Land use and wetland characteristics had a comparatively small effect on wetland inundation status compared to temperature and precipitation. Wetland size was the only wetland-related variable that influenced inundation status, with larger basins having higher probability of holding water. In contrast, percent cropland and average soil drainage class within wetland watersheds had little or no effect on inundation, a finding in contrast to previous studies showing that greater total cover of crops and higher amounts of permeable soils increase probability of wetland inundation (Cariveau et al. 2011; Bartuszevige et al. 2012; Uden et al. 2015; Starr and McIntyre 2020). The limited role of wetland and land use-related variables compared to other studies may be associated with the relatively small watershed areas and low slopes surrounding wetlands in our study region, both of which can influence the amount of water available to flow into a wetland (Cariveau et al. 2011; Bartuszevige et al. 2012). Differences in crop type and timing of harvest may also influence the total amount of runoff from agricultural fields (Carroll et al. 1997), which in turn influences subsequent inundation. The dominant crops in our study area were wheat and rye, which can maintain some vegetative cover throughout most of the year. Other studies that found a greater influence of crops on wetland inundation occurred in regions where the dominant crop type was corn (Zea mays) (Uden et al. 2015), which leaves little residual vegetation after harvest. Additionally, traditional farming practices commonly used in our study area, such as terracing, may limit or redirect flow of water away from wetlands during precipitation events. Our results, which differ from past studies, emphasize that consideration of landscape context is critical for evaluating factors that influence hydrologic cycles. Developing models that specifically account for regional topographic and land use patterns is critical for making effective management and adaptation plans.

One challenge associated with broad-scale studies that examine temporal changes in frequency and extent of water inundation is the limited temporal resolution of many remote sensing products. For example, Landsat data have a minimum temporal resolution of 16 days between images, but as much as 90 days could pass between usable images due to the frequent occurrence of cloud cover in our study site. As a result, our analysis likely underestimates the proportion of total wetlands inundated during at least a portion of each season because smaller ephemeral wetlands may undergo several undetected wet-dry cycles in the time period between successive images. A study of playa wetlands in central Texas found playas can have hydroperiods as short as 17 days under dry conditions, with the mean hydroperiod across the region varying considerably based on rainfall and methods used to predict hydroperiod (Collins et al. 2014). While use of remote sensing techniques allows continuous monitoring of large spatial areas, complimentary ground-based studies will also be needed to better monitor changes in water levels over short periods of time. Such information that accounts for both occurrence of water in wetland basins and water persistence on the landscape will be essential for better predicting effects of climate change on wetland ecosystems. Additionally, many previous efforts to model influences of weather on inundation have relied on annual wetland surveys (Johnson et al. 2011; Uden et al. 2015; Bartuszevige et al. 2012); although these studies have made important contributions to understanding climate effects on wetlands, this “snapshot” approach limits our ability to assess how short-term weather variation affects wetland inundation throughout the year. Despite the variable temporal resolution of the Landsat data used in this study, our methods improve our understanding of year-round water dynamics in Southern Great Plains wetlands.

Our predictions assume that the limited number of variables included in the final model are the primary drivers of wetland inundation in the Pleistocene Sand Dunes Ecoregion. However, wetland hydrologic cycles are regulated by complex interactions between biotic and abiotic environment (Mitsch and Gosselink 2007, McKenna et al. 2018), and our models are likely an oversimplification of the dynamic processes that influence wetland inundation. As a result, while our models may accurately predict wetland inundation under current conditions, the relationships we documented may not remain valid under future climate change if our model does not accurately reflect all processes that will affect inundation or if the nature of these relationships change. A number of process-based models have been developed that better incorporate the complex interactions influencing wetland hydrology, and these models may provide greater projection accuracy by incorporating this greater hydrological nuance (Johnson et al. 2005, Stocker et al. 2014, McKenna et al. 2021). However, due to their complexity, such models are not always generalizable to different types of wetland systems, and their results can be challenging to interpret, which complicates inferences about how specific weather changes will influence wetlands. Research that incorporates both approaches may allow for improved predictions under future conditions and a fuller understanding of how a changing climate will affect wetland hydrology.

Changes in inundation patterns have many effects on wetlands and species that depend on them. Recent studies show that drier conditions and reduced flooding may shift vegetation toward more upland-dominated communities or alter the persistence of emergent plants (van der Valk and Mushet 2016; Owen et al. 2020). Further, many amphibian and invertebrate species require wetlands that remain inundated for prolonged periods in order to complete specific life cycle stages (Pechmann et al. 1989; Snodgrass et al. 2000; Tarr et al. 2005). As the region becomes hotter and drier, fewer wetlands that meet these criteria of prolonged hydroperiods will be available. Wetlands of the Southern Great Plains represent critical wintering and stopover habitat for migratory waterfowl (e.g., ducks and geese) and shorebirds (species in the family Charadriidae, including sandpipers and plovers) during fall and spring migration (Albanese and Davis 2015; Reese and Skagen 2017). The reduced wetland inundation predicted in this study and by others in the Central Flyway (Larson 1995, Uden et al. 2015, Sofaer et al. 2016, Mckenna et al. 2021), especially for spring, may hamper migration by reducing migratory stopover habitat for these species (Albanese and Haukos 2017), negatively impacting their ability to successfully reach northern breeding grounds. Further, the loss of stopover habitat and food resources associated with wetlands, such as invertebrates and seed-producing plants, can negatively impact survival and reproductive success for individuals that do reach their breeding areas (Aharon-Rotman et al. 2016). Finally, because reduced inundation and changing wet-dry cycles can alter nutrient and chemical cycles in wetlands (Busnardo et al. 1992; Lockaby et al. 1996; Battle and Golladay, 2001), climate change may threaten their capacity to serve as carbon sinks (McKenna and Sala 2016).

Despite their relatively small surface area globally, wetlands provide a wide variety of essential services and functions to humans (Mitsch and Gosselink 2007), and they are also critical for maintaining biodiversity (Dudgeon et al. 2006; Junk et al. 2006). We show that shifts of climate regimes toward hotter, drier conditions will likely result in fewer inundated wetlands in the Southern Great Plains of North America, but similar changes to wetlands are likely worldwide, with associated far-reaching impacts to ecosystem functioning and services. Current climate predictions show that future changes in weather patterns are unavoidable, thus proactive management measures aimed at protecting remaining wetlands is critical, as these ecosystems and the services they provide will become increasingly threatened in the future. This is especially important given that our models predict that wetlands in the Southern Great Plains will likely experience reduced inundation in spring and summer regardless of the amount of future greenhouse gas emission and resultant climate warming. While altering the path of climate change may be difficult, reducing the influence of non-climate stressors, such as conversion of wetlands to other land uses, may be an effective strategy to adapting to future conditions and making these wetland ecosystems more resilient to changing climates.