Climate Change Will Impact Surface Water Extents and Dynamics Across the Central United States

Abstract

Climate change is projected to impact river, lake, and wetland hydrology, with global implications for the condition and productivity of aquatic ecosystems. We integrated Sentinel‐1 and Sentinel‐2 based algorithms to track monthly surface water extent (2017–2021) for 32 sites across the central United States (U.S.). Median surface water extent was highly variable across sites, ranging from 3.9% to 45.1% of a site. To account for landscape‐based differences (e.g., water storage capacity, land use) in the response of surface water extents to meteorological conditions, individual statistical models were developed for each site. Future changes to climate were defined as the difference between 2006–2025 and 2061–2080 using MACA‐CMIP5 (MACAv2-METDATA) Global Circulation Models. Time series of climate change adjusted surface water extents were projected. Annually, 19 of the 32 sites under RCP4.5 and 22 of the 32 sites under RCP8.5 were projected to show an average decline in surface water extent, with drying most consistent across the southeast central, southwest central, and midwest central U.S. Projected declines under surface water dry conditions at these sites suggest greater impacts of drought events are likely in the future. Projected changes were seasonally variable, with the greatest decline in surface water extent expected in summer and fall seasons. In contrast, many north central sites showed a projected increase in surface water in most seasons, relative to the 2017–2021 period, likely attributable to projected increases in winter and spring precipitation exceeding increases in projected temperature.

Plain Language Summary

Climate change is expected to impact rivers, lakes, and wetlands. In this effort we used multiple satellites to track monthly surface water extent (2017–2021) for 32 sites across the central United States. The average amount of surface water was highly variable across sites. Individual statistical models, relating meteorological variables to surface water extent, were developed for each site. The models were then updated with climate change adjusted variables. Most sites were projected to show a decline in surface water extent, with drying most consistent across the southeast central, southwest central, and midwest central U.S. Projected declines under dry conditions at these sites suggest greater impacts of drought events are likely. Projected changes were seasonally variable, with the greatest decline in surface water extent expected in summer and fall seasons. In contrast, many north central sites showed a projected increase in surface water in most seasons, likely attributable to projected increases in winter and spring precipitation.

1. Introduction

Freshwater ecosystems, including wetlands, ponds, lakes, rivers, and reservoirs, provide vital ecosystem services including water quality improvement, flood abatement, erosion control, drinking water, recreation, and habitat for wildlife (Hansson et al., 2005; Rahmani et al., 2018; Russi et al., 2013). However, many freshwater ecosystems have been destroyed or degraded over time (Brooks et al., 2016; Fluet‐Chouinard et al., 2023), with most of the United States (U.S.) freshwater wetland losses between 1986 and 2009 occurring in the Great Plains of the central U.S. (Dahl, 2000, 2011). Wetlands occur in a transition zone between aquatic and terrestrial ecosystems and are therefore sensitive to minor changes in hydrology (Burkett & Kusler, 2007), which can shift wetland type and function (Dahl, 2011). Globally, substantial wetland loss is expected by 2100 under higher climate change emission scenarios (Xi et al., 2021). Conservation of freshwater ecosystem services will be supported by consideration of how climate change will alter the hydrology and functioning of these systems (Muhammad et al., 2018; Salimi et al., 2021; Zhang et al., 2021). Freshwater ecosystems respond to episodic, seasonal, and inter‐annual variability in water availability, as well as long‐term trends in climate and land use (McKenna et al., 2017; Sun et al., 2011; Watts et al., 2012). Therefore accurate projections of climate change effects include changes in regional patterns of surface water extent across a range of drought‐to‐deluge conditions.

Many studies that rely on remote sensing have documented the relationship between climate conditions and terrestrial inundation, or surface water, extent, where both terms refer to not only freshwater ecosystems such as wetlands, lakes, and rivers, but also the episodic storage of surface water in flood‐prone zones. Surface water extent typically depends on antecedent conditions (Haque et al., 2022) with changes positively correlated with precipitation and negatively correlated with temperature and evapotranspiration (Mishra & Cherkauer, 2011; Song et al., 2018; Tulbure & Broich, 2019; Xia et al., 2019). However, the response of surface water extents to climate inputs will also depend on topography, groundwater and soil properties as well as land cover (Fuentes et al., 2019), hydrological alterations (e.g., drained, intact, dammed) (Haque et al., 2022; Rajib et al., 2020), and floodplain dynamics (Heimhuber et al., 2017), so that changes in surface water extent can be non‐linear and lagged in response to changes in precipitation or water availability (Silio‐Calzada et al., 2017; Song et al., 2018; Vanderhoof et al., 2018). Site‐specific analyses account for geographic variability in the response of surface water to climate when characterizing the climate‐surface water relationship.

The most common approach for projecting impacts of climate change on surface water extent or alternatively on wetlands, alone, across the central U.S., has been to statistically relate spatial variability in climate and land use variables with wetland abundance or density, then update study period climate with a range of possible climate change scenarios (Garris et al., 2015; Sofaer et al., 2016; Zhang et al., 2021). Garris et al. (2015), for example, projected wetland expansion in Iowa and Northeastern Missouri, while projecting that Northern Minnesota and Michigan were at risk for wetland losses. Most of this research, however, has focused on the Prairie Pothole Region of the north‐central United States and west‐central Canada, characterized by poorly drained landscapes where relatively small increases in surface water level can produce substantial increases in surface water extent (Vanderhoof et al., 2017). Across this region, most global circulation model (GCM) scenarios project a strengthening of the existing precipitation gradient, in which the western and central Prairie Pothole Region show decreased annual precipitation, while the eastern Prairie Pothole Region, a region dominated by agriculture, shows a small increase in annual precipitation (Ballard et al., 2014; Johnson et al., 2010). However, increases in evapotranspiration associated with consistently higher temperatures may offset minor increases in precipitation, so that drying across the Prairie Pothole Region is projected with associated shrinking of wetland area and reduced habitat for waterfowl (Johnson & Poiani, 2016; Sofaer et al., 2016). Alternatively, McIntyre et al. (2019) projected increases in surface water across the western U.S. Prairie Pothole Region. Geographic variability in the directionality of impacts is a distinct possibility. Across the conterminous U.S., Zou et al. (2018) documented diverging trends over the 1984 to 2016 period in which surface water has decreased in water‐poor regions and increased in water‐rich regions. In contrast, Petrakis et al. (2022), documented non‐significant increases in surface water across much of the central U.S., using Moderate Resolution Imaging Spectroradiometer (MODIS) data from 2003 to 2019. Annual projections may also obscure seasonal changes, for instance, across the northern U.S. the fraction of precipitation falling as rain, instead of snowfall is likely to increase (Shook & Pomeroy, 2012), and may impact the hydroperiod of wetlands, which across the conterminous U.S. are driven by precipitation, spring snowmelt and evapotranspiration (Park et al., 2022).

While many efforts to model the impact of climate change on surface water extent in the U.S. have focused on the Prairie Pothole Region (e.g., Johnson & Poiani, 2016; Sofaer et al., 2016), interest in services provided by wetlands, lakes and rivers extends well beyond the Prairie Pothole Region. For instance, water is critical to support agriculture across the central U.S., which in 2018–2020 contributed 68% of the principal crops harvested across the entire U.S. (USDA, 2021). Past harvest failure events, however, have been attributed to both drought and flood conditions along the Mississippi and Missouri Rivers (Lall et al., 2018). Our aim was to fill some of the current data gaps about climate’s current and projected effects that are changing the dynamic behavior of surface water extent and persistence. Our research characterized contemporary (2017–2021) surface water extents for 32 sites across the central United States (U.S.) (Figure 1) using a multi‐sensor approach of Sentinel‐1 and Sentinel‐2 based algorithms (Vanderhoof et al., 2023) and projected how climate change scenarios could modify surface water extent temporally across a range of aridity conditions. Our research questions include:

Figure 1.

The 32 study area sites across the central United States with the National Wetland Inventory (NWI) wetland extent (USFWS, 2019) shown to indicate the variability in the amount, type, and arrangement of the surface water.

1. How variable are current surface water dynamics across the central U.S.?

2. How does the response of surface water to climate inputs differ across the central U.S.?

3. How will climate change impact average, seasonal and outlier surface water extents in different parts of the central U.S.?

Our work produces the first suite of climate change projections of surface water extent across the central U.S. that explicitly considers short‐term variability in how surface water responds to climate inputs and enables projections across a range of climate and land use conditions.

2. Methods

2.1. Study Areas and Time Period

A total of 32 sites, averaging 4,523 km² in size, were selected across the central U.S. (Figure 1). Sites were named using the two‐letter acronym of the U.S. state in which the site was primarily located within. Aridity, annual precipitation divided by annual evapotranspiration, shows a strong west‐east gradient from dry to wet across the study area, as does precipitation, ranging from 1,428 mm in the southeast (LA1) to 469 mm in the northwest (ND4) (Figure 2). The study area also shows a strong temperature gradient from north to south with the average minimum temperature ranging from −2.8 to 15°C for ND1 and TX4, respectively (PRISM Climate Group, 2022) (Table 1, Figure 2). The sites were selected to not only represent diverse temperature, precipitation, and aridity conditions, but to also vary in landcover as well as the amount and type of surface water storage (Table 1). To facilitate the interpretation of findings across the study area, the 32 sites were grouped into four climate groups using the Köppen‐Geiger climate classes (Beck et al., 2018). As half of the sites were located within the warm temperate, fully humid, hot summer class, this class was further bifurcated using annual, average aridity (Figure 2), so that sites were identified as part of the (a) north central (NoC), (b) midwest central (MWC), (c) southeast central (SEC), or (d) southwest central (SWC) group within the central U.S. (Table 1).

Figure 2.

(a) Aridity, annual precipitation divided by annual evapotranspiration, (b) total annual precipitation, (c) percent change in annual precipitation (2006–2025 to 2061–2080) for the RCP4.5 scenario, and (d) RCP8.5 scenario, (e) the average maximum June, July, August (JJA) temperature, and (f) and the change in maximum temperature (JJA) for the RCP4.5 scenario and (g) RCP8.5 scenario. Panels (a, b, and e) show GRIDMET data (2006–2022; Abatzoglou, 2013), while remaining panels represent MACAv2 20‐model average (Abatzoglou & Brown, 2012).

Table 1.

Site (Ordered North to South Using State Abbreviations) Characteristics Including Land Cover (Homer et al., 2020) and Average Annual Total Precipitation (PPT), Average Annual Maximum (Tmax) and Average Annual Minimum (Tmin) Temperature Climate Normal (PRISM Climate Group, 2022)

Region	Site	Size (km2)	FP (%)	Wetland (%)	Emergent wetland (relative %)	Forest-shrub wetland (relative %)	Riverine (relative %)	Lakes—ponds (relative %)	Dominant land cover	PPT (mm)	Tmax (°C)	Tmin (°C)
North Central	ND1	3,969	2.1	7.7	70.0	0.3	11.4	18.2	Cultivated crops (80%)	502	8.9	−2.8
	ND4	3,994	4.4	14.1	57.9	0.2	0.4	41.4	Cultivated crops (58%)	469	10.3	−2.3
	ND2	4,011	22.7	4.6	67.0	5.4	13.6	14.0	Cultivated crops (78%)	529	10.4	−1.1
	ND3	4,017	9.9	24.3	46.5	0.3	0.6	52.5	Cultivated crops (50%)	508	10.0	−1.7
	MN1	4,046	21.3	31.7	17.9	47.7	0.7	33.7	Deciduous forest (27%), woody wetlands (23%)	659	10.2	−1.7
	MN2	4,047	31.2	48.3	12.2	54.3	0.7	32.8	Woody wetlands (34%)	697	10.2	−1.7
	MN3	4,091	23.0	32.7	17.1	31.3	0.7	50.9	Deciduous forest (33%)	682	10.5	−1.3
	MN4	4,091	23.1	41.7	11.7	72.9	0.9	14.4	Woody wetlands (34%), deciduous forest (32%)	712	10.7	−1.4
Midwest Central	SD2	4,400	9.0	12.0	60.6	2.6	1.9	34.9	Cultivated crops (76%)	614	12.1	0.4
	SD1	4,412	13.2	13.1	67.3	0.4	0.1	32.2	Cultivated crops (42%), grassland (32%)	594	11.8	0.3
	SD3	4,467	19.0	12.2	69.8	0.8	0.8	28.6	Cultivated crops (63%)	615	12.3	0.6
	SD4	4,513	11.7	10.3	77.6	1.4	0.3	20.7	Cultivated crops (63%)	641	13.3	1.3
	IA2	4,690	13.9	6.6	8.7	29.1	35.5	26.7	Cultivated crops (33%), deciduous forest (29%)	894	13.7	2.8
	IA1	4,717	10.7	3.2	19.1	54.2	15.6	11.1	Cultivated crops (79%)	920	14.0	2.6
	IA3	4,762	21.1	5.4	24.7	46.0	12.1	17.1	Cultivated crops (63%)	921	14.9	3.6
	IA4	4,766	15.0	6.3	8.6	37.3	44.8	9.3	Cultivated crops (65%)	916	15.0	3.9
Southwest Central	MO1	5,036	18.4	6.0	15.1	33.7	25.2	26.0	Pasture/hay (30%), cultivated crops (30%)	983	18.2	6.5
	OK2	4,533	19.2	9.4	1.1	12.4	8.5	78.0	Pasture/hay (49%), deciduous forest (28%)	1,113	21.3	8.8
	OK1	4,533	15.0	6.2	3.5	16.4	45.8	34.2	Deciduous forest (38%), grassland (37%)	1,042	21.7	9.2
	OK4	4,533	23.5	13.2	2.6	13.0	8.8	75.6	Deciduous forest (36%), pasture/ hay (29%)	1,167	22.3	10.2
	OK3	4,585	11.5	3.6	7.8	18.6	23.7	49.9	Grassland (55%)	819	22.5	9.5
	TX1	5,140	16.0	5.4	9.3	19.9	16.0	54.8	Grassland (37%), developed (23%)	985	24.2	11.9
	TX2	4,775	9.1	4.4	1.3	4.1	22.2	72.4	Shrub/scrub (50%), evergreen forest (26%)	761	25.5	12.3
	TX4	4,106	13.4	4.3	1.1	0.5	24.6	73.9	Shrub/scrub (66%)	672	28.4	15.0
Southeast Central	MO2	5,048	29.8	7.4	16.3	39.3	29.1	15.2	Cultivated crops (53%), pasture/ hay (22%)	1,052	18.0	6.3
	MO4	5,117	17.8	11.3	13.2	34.2	7.2	45.4	Pasture/hay (34%), deciduous forest (32%)	1,116	19.1	6.8
	MO3	5,127	21.3	8.1	12.2	57.0	11.2	19.7	Pasture/hay (43%), cultivated crops (27%)	1,109	19.2	7.1
	AR1	4,533	35.6	28.7	0.8	78.3	14.8	6.1	Cultivated crops (47%), woody wetlands (25%)	1,333	23.3	11.6
	AR2	4,551	33.9	11.9	10.1	67.9	9.6	12.3	Evergreen forest (42%), woody wetlands (34%)	1,403	23.9	10.7
	MS1	5,057	57.8	34.8	1.6	77.3	13.5	7.6	Cultivated crops (44.1%), woody wetlands (31.8%)	1,417	24.0	12.0
	LAI	4,540	11.8	5.1	1.5	37.5	17.4	43.7	Evergreen forest (51%)	1,428	24.3	11.3
	TX3	4,540	25.6	15.0	6.9	29.6	8.6	54.9	Evergreen forest (40%)	1,350	25.3	12.3

Note. Wetland (USFWS, 2019) and floodplain (Woznicki et al., 2019) are shown as the percentage of each site, while wetland type is the relative percentage. FP, floodplain.

The period (2017–2021) was selected to coincide with the production of the Sentinel‐1 and Sentinel‐2 surface water extent algorithms (Vanderhoof et al., 2023) and to represent current climate and land use conditions. To characterize the range of climate conditions within the 2017–2021 period, the Gridded Surface Meteorological (GRIDMET) 5‐day, 4 km resolution Palmer Drought Severity Index (PDSI; 1980–2021; Abatzoglou, 2013) was converted to a rank percentile. As PDSI may not directly reflect the range of surface water extents represented by the 5‐year period, we used the Joint Research Center (JRC) global surface water (GSW) monthly collection (v1.4, 30 m resolution, March 1984 to January 2022), derived from Landsat (Pekel et al., 2016). Total water was calculated within each site and month over the 38‐year record and converted to a rank percentile. Timesteps in which >5% of the water data were missing were excluded, but substantial, with 50.2% of the JRC GSW monthly observations excluded due to missing data (e.g., cloud, cloud shadow, snow).

The 5‐year period showed a large range of both climate and water conditions at most sites, relative to the past 42 and 38 years, respectively. Of the 32 sites, 11 sites showed a PDSI rank percentile range of >90% and 31 of 32 had a PDSI rank percentile range of >70% (Table 2). Within the 2017–2021 period, peak surface water was highly represented. The 5‐year period contained a maximum water percentile of >95% for all sites with 14 of the 32 sites containing the 100% maximum water extent. Similar to PDSI, of the 32 sites, 12 sites showed a water rank percentile range of >90% and 29 of 32 sites had range of >70% (Table 2).The median PDSI and water rank percentiles during the 2017–2021 period averaged 70.1% and 62.2%, respectively (Table 2), suggesting that both climate and water during this 5‐year period were wetter, on average, across the sites, compared to the 1980–2016 and 1984–2016 periods, respectively.

Table 2.

The Rank Percentiles for the 2017–2021 Period, Relative to the 1980–2021 Period Using the Palmer Drought Severity Index (PDSI) (Abatzoglou, 2013), and Relative to the 1984–2021 Period Using the Global Surface Water (GSW) Monthly Data Set (Pekel et al., 2016)

Region	Site	PDSI (min, %)	PDSI (max, %)	PDSI (median, %)	GSW monthly count	Water (min, %)	Water (max, %)	Water (median, %)
North Central	ND1	4.2	98	40.1	26	20	95.2	43.1
	ND4	1.2	95.9	49.4	24	27.4	91.3	65.8
	ND2	0.9	100	59.8	29	2.6	100	48.9
	ND3	0.8	99.7	37.1	26	27.5	94.9	67.9
	MN1	1	100	57.9	28	0.5	100	45.4
	MN2	9.3	100	73.8	19	1.2	99.3	32
	MN3	2.5	100	72.9	25	0.6	99.3	33.9
	MN4	4	100	89.2	25	0.5	98.9	43.4
Midwest Central	SD2	7.9	100	61.7	26	12	98.3	65.3
	SD1	6.7	100	62.2	27	27.4	98.4	74.6
	SD3	17.7	100	72.3	21	6	100	72
	SD4	7.1	100	67.3	29	25	100	70.9
	IA2	15.8	100	91.5	29	0.5	100	65.7
	IA1	21.1	99.5	77.7	34	27.7	98.5	69.1
	IA3	16.7	98.8	78.1	31	15.3	99	51.4
	IA4	25.8	100	83.4	30	2.3	100	66.5
Southwest Central	MO1	14.7	99.1	64.2	32	11.1	100	74.7
	OK2	21.1	100	70.1	31	4.2	99.1	75.1
	OK1	18.2	100	70.1	32	13.1	100	76.9
	OK4	14.4	100	87.1	32	17.7	98.7	67
	OK3	11.6	96.3	57.4	35	24.1	99.5	72.9
	TX1	19	98.3	72.8	31	25.6	100	80.9
	TX2	13.2	96.7	61.1	38	28.6	100	72.9
	TX4	21.2	94.9	56.7	29	4	50	28.6
Southeast Central	MO2	9.8	92.7	64.4	37	9.1	100	76.3
	MO4	22	98.8	69.4	32	10.4	98	66.6
	MO3	25.1	100	70.5	29	20.3	100	68
	AR1	19.6	100	89.8	35	13.3	100	70.8
	AR2	28.4	99.1	90.1	32	9.8	100	65.1
	MS1	27.5	100	85.2	31	13.9	100	68.6
	LA1	22.1	100	81.3	38	3	99.1	50.8
	TX3	24.8	100	79.3	34	21	92.2	60.3
	All	14.2	99	70.1	30	13.3	97.2	62.2

2.2. Surface Water Extent

2.2.1. Sentinel‐1 and Sentinel‐2 Surface Water Classification Algorithms

Surface water extent over time at each of the 32 sites was derived using previously published algorithms based on Sentinel‐1 (C‐band synthetic aperture radar in dual polarization mode, 20 × 22 m resolution) and Sentinel‐2 (optical sensor with l3 spectral bands, 10–20 m resolution) (Vanderhoof et al., 2023). These algorithms used gradient boosted classifiers to classify images at 20 m resolution into non‐water, Open Water (OW) (e.g., lakes, ponds, rivers) or Vegetated Water (VW) (e.g., emergent wetlands, forested/shrub wetlands, riparian corridors). Sentinel‐2 images with snow and ice were excluded from 16 of the 32 sites using site‐specific seasonal bounds of persistent snow cover (Vanderhoof et al., 2023) and masking out pixels mapped as snow cover by the MODIS snow cover product (Hall & Riggs, 2021). Additionally, large precipitation events can cause temporary, wide-spread soil saturation that induce commission error (Bazzi et al., 2019). VW, but not OW, was therefore masked where a same‐day precipitation event was classified as more than light rain (>5 mm as measured by daily GRIDMET; Endarwin et al., 2014). When previously validated with 7,200 validation points, selected randomly with a stratified disproportionate sampling scheme, derived from 36 high‐resolution (WorldView‐2 [1.8 m, Maxar, Westminster, CO], WorldView‐3 [1.2 m, Maxar, Westminster, CO], PlanetScope [3–5 m, Planet Lab, San Francisco, CO]) images, the Sentinel‐1 algorithm showed an omission and commission error of 3.1% and 0.9% for OW, and a 28.4% and 16.0% commission error for VW, respectively. The Sentinel‐2 algorithm showed an omission and commission error of 3.1% and 0.5% for OW, and a 10.7% and 7.9% commission error for VW, respectively. Additional processing details can be found in Vanderhoof et al. (2023). In this effort the Sentinel‐1 and Sentinel‐2 algorithms were applied to all Sentinel‐1 and Sentinel‐2 imagery (2017–2021), as described above, overlapping each of the 32 sites.

2.2.2. Derivation of Monthly Surface Water Extent

The classified Sentinel‐1 and classified Sentinel‐2 time series were consolidated at a monthly time step. Within each month, pixel values were assigned as the majority classification, (a) non‐water or (b) water, where water included both OW and VW (Figure 3). If observations of water and non‐water were equal, then, depending on the classified observations present, OW was selected over VW and non‐water, and non‐water was selected over VW (Figure 3). This choice was based on the higher‐class accuracy of OW, relative to VW (Vanderhoof et al., 2023).

Figure 3.

Flow chart of surface water time series generation, model development, and application of climate change projections.

An individual water mask was then derived for each site and applied across the time series. Each water mask defined the maximum surface water extent and was used to further limit commission error within the surface water time series. Pixels classified as water outside of the water mask were re‐classified as non‐water. To generate each water mask, the Sentinel‐1 OW, Sentinel‐1 VW, Sentinel‐2 OW, and Sentinel‐2 VW percentile rasters, representing the full time period (2017–2021; Vanderhoof et al., 2023) were manually reviewed for each study area (Figure 3). Percentile thresholds were selected, below which the frequency of erroneously classified water pixels visually exceeded the frequency of correctly classified water pixels (Table A1). Ancillary data were used to inform the threshold selection, including the National Wetland Inventory (NWI) data set (USFWS, 2019), the 2019 National Land Cover Database (NLCD; Homer et al., 2020), and aerial imagery (source: Esri, Maxar, Earthstar Geographics). The spatial extent where water pixels were retained was defined as the 100‐year floodplain (Woznicki et al., 2019), to account for short‐term flood events, as well as areas where the water percentile was greater than the selected threshold in any of the four 5‐year percentile rasters (Table A1). This site-specific approach enabled us to include the sub‐pixel VW class, while maximizing the accuracy of the surface water time series at each site.

Monthly surface water time‐steps where >5% of the data were missing were reviewed to check for anomalous declines in surface water area that could be attributable to the presence of no data available (NA) pixels, for example, where NA pixels occur over a lake or in a high wetland density area. Of the 1920 observations (32 sites × 12 months × 5 years), 26 observations had >5% NA and of those 16 of the observations were removed. An additional 12 observations were removed because of excessive omission or commission error. The number of monthly observations excluded totaled 1.5% and ≤3 observations per site.

The monthly surface water time series was validated using PlanetScope imagery (source: Planet Labs PBC, San Francisco, CA; Dove Classic, Dove R, and Super Dove; 3–5 m resolution). Within each site, two high‐resolution images were selected (total of 64 images) to represent different years and months within and across the sites (Table A2). Errors of omission and commission as well as Overall Accuracy (OA) were calculated both with and without the water masks applied to the monthly surface water composites. Relative Bias (RB) was also calculated as a percent, where values above zero indicate that the classification overestimated surface water and values below zero indicate an underestimation (Padilla et al., 2014). Additional details regarding the derivation of the surface water validation can be found in Appendix A.

2.3. Development of Surface Water‐Climate Models

The time series of surface water at each site was characterized using summary statistics (median, standard deviation, range, coefficient of variation, kurtosis, and skew), and consequently modeled as a function of meteorological variables using Scikit‐learn Python module’s random forest regression models (Pedregosa et al., 2011) (Figure 3). Random forest models use a bootstrapping approach to generate hundreds of regression trees and make no prior assumptions about cause‐and‐effect relationships or correlations among variables (Hastie et al., 2009). The daily GRIDMET data set (Abatzoglou, 2013) was reduced to a monthly time‐step and (a) precipitation, relative humidity ((b) minimum and (c) maximum), temperature ((d) minimum and (e) maximum), (f) vapor pressure deficit (VPD), (g) specific humidity (SPH), and (h) surface downward shortwave radiation were averaged across each site for each month in the time series. Averaging climate values across each site at each time step can obscure local scale climate variability, therefore the spatial meteorological variability within each timestep was compared to the temporal variability in meteorology over the time series for select climate variables (Figure A1). Temporal variability was much larger than the within site variability, suggesting that this approach was reasonable. In addition to monthly values, values for each variable, accumulated or averaged over the prior 2, 3, 6, 9, and 12 months (8 variables × 6 accumulation periods = 48 variables total), were tested for inclusion in the statistical models. Inclusion of variables beyond precipitation and temperature was intended to better represent climatic drivers of surface water dynamics (Song et al., 2018; Tulbure & Broich, 2019; Xia et al., 2019).

For each model we concurrently ran a variable and hyperparameter selection process, where the root mean square error (RMSE) of potential models were compared using a nested cross‐validation, KFold with six splits (Cawley & Talbot, 2010). For model hyperparameters, the number of trees was tested as 200, 300, or 500. Maximum tree depth and minimum samples per leaf were set to limit tree over‐fitting and were tested as 2 or 3, while the maximum samples per tree was set to 0.8. While random forest techniques are generally insensitive to multi-collinearity, the inclusion of highly correlated variables can make it more challenging to identify the most predictive variables and deflate or bias variable importance values (Gregorutti et al., 2016). Additionally, the inclusion of many variables can produce models that are difficult to interpret (Murphy et al., 2010). A stepwise forward selection routine was implemented where the set of potential predictors were sequentially tested. The predictor that contributed most to reducing the model’s RMSE was selected. During each step, the remaining predictors were removed if they had a correlation value of 0.8 or greater with any of the selected predictors. This process was iterated until the improvement in the model’s RMSE was <0.001 with any additional variables (Sherrouse & Hawbaker, 2023).

Small sample sizes within each of the models (n = 60) required consideration of the train‐test split (Vabalas et al., 2019). Two approaches were tested, (a) Leave‐one‐out cross validation (CV) and (b) a stratified train‐test split, where the stratification allowed the model to be trained on the maximum range of surface water extent, but 20% of the data was reserved for model testing. The preferred model was selected as the variable, hyperparameter and test/train split producing the lowest RMSE (Sillmann et al., 2013), and the R² and RMSE were reported for each selected model. Additionally, the R² and RMSE values using a KFold cross‐validation with 6 (or 10%) splits were also reported for each model. Variable importance was calculated with Python Scikit‐learn as the permutation importance, which reports the mean decrease in accuracy.

2.4. Climate Change Data and Projections

Climate change projections were derived from the MACA‐CMIP5 (MACAv2‐METDATA) ensemble, accessed through the Northwest Knowledge Network (Abatzoglou & Brown, 2012). This data set applies the Multivariate Adaptive Constructed Analogs statistical method to downscale a set of 20 CMIP5 Global Climate Models (GCMs) from their native resolution to 4 km resolution. The downscaling process used daily GRIDMET data (4 km resolution, 1979–2012) to train and bias correct the GCMs. This enabled the down‐scaled climate data to better reflect daily patterns of near‐surface meteorology (Abatzoglou & Brown, 2012). Simulated changes in GCMs future experiments, under Representative Concentration Pathways (RCP) emission scenarios, were defined by their total radiative forcing in the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (IPCC, 2013). We included RCP emission scenarios 4.5 (stabilization at 4.5 W/m² by 2100 without overshoot) and 8.5 (rising radiative forcing leading to 8.5 W/m² by 2100) (van Vuuren et al., 2011).

In addition to using multiple emission scenarios, differences between individual GCMs can also be used to bound the uncertainty of climate change projections. A subset of the individual GCMs were selected by reviewing the GCMs precipitation and temperature change projections (RCP8.5, 2070–2099) for the central U.S. Projected change in annual precipitation was graphed as a function of change in June‐August maximum temperature for all 20 GCMs. In addition to the 20‐model mean, five individual models were selected from the precipitation-temperature scatter graph that bounded the full range of GCM climate change projections. Models selected included MIROC5, MIROC‐ESM, MRI‐CGCM3, GFDL‐ESM2M, and IPSL‐CM5A‐MR. For each GCM included in the analysis, the monthly MACAv2‐METDATA variables selected for inclusion in each site’s random forest model were processed including: precipitation, relative humidity (minimum and maximum), temperature (minimum and maximum), VPD, SPH, and surface downward shortwave radiation.

Historical data within the MACAv2‐METDATA is defined as 1950–2005, so to reflect the surface water date range (2017–2021), the change in meteorological variable values was calculated as the change between current, defined as 2006–2025 and future, defined as 2061–2080. Projected changes to each climate variable were calculated as a monthly (January, February, March, etc.) delta for temperature, and percent change for non-temperature variables (McIntyre et al., 2019; Miller et al., 2021), between the current (2006–2025) and future (2061–2080) conditions for each variable and site. So that the projected percent change to January precipitation, for instance, was calculated as the percent change between the averaged January 2061–2080 and the averaged January 2006–2025 values. The projected per month meteorological changes were then applied to the monthly, observed GRIDMET data (2017–2021). Variables accumulated over multiple months were re‐calculated using the climate‐change adjusted monthly GRIDMET data. A total of 12 sets ([1 20‐model mean + 5 individual models] × 2 emission scenarios = 12 climate change projections) of climate‐change adjusted model variables were used in each existing random forest model to predict an updated time series of surface water extents, producing a total of 12 projections of climate change adjusted monthly surface water time series for each site (Figure 3).

Model predicted values can be expected to be biased toward the observed mean, with the degree of bias depending on the model’s explanatory power (Sillmann et al., 2017). While quantile mapping is a common approach for this type of bias correction (Trentini et al., 2023), the approach assumes stationarity in the error correction values. Scaled distribution mapping does not assume stationarity, while also accounting for the likelihood of individual events (Switanek et al., 2017). Therefore, before comparing projections to the observed surface water extents, scaled distribution mapping was applied, with gamma distribution, to the climate change projected surface water extents using the python package bias‐correction (Kumar, 2021). Projecting changes to the tails of a distribution is particularly challenging (Sillmann et al., 2017), and common bias correction methods often show poor performance within the tails of Cumulative Distribution Functions (Trentini et al., 2023). Therefore, in evaluating projected climate change impacts to the wettest and driest conditions, identified per site as the observations with the most (top 15%) and least (bottom 15%) observed surface water extent, climate change projections were compared to the random forest predicted values (Garris et al., 2015; Hanlon et al., 2013). When summarizing seasonal results, seasons were defined as spring (March, April, May), summer (June, July, August), fall (September, October, November), and winter (December, January, February). In addition, projections of percent change to surface water extent for the 20‐model mean were correlated, using Spearman correlation with a Bonferonni correction applied, with the projected change to climate variables and surface water characteristics across the 32 sites.

3. Results

3.1. Accuracy of Surface Water Time Series

Across the 32 sites, errors of omission and commission for monthly surface water extent averaged 1.9% and 6.5%, respectively, with an OA of 97.2% (Table 3). By climate group, the NoC and SEC sites had lower OA, 96.5% and 96.6%, respectively, as well as higher commission error, 8.7% and 6.9%, respectively, relative to the MWC and SWC sites (Table 3). The SWC sites had the greatest OA of 98.1% and the lowest errors of omission and commission at 0.9% and 5.0%, respectively. Omission error was 3% or less in all four groups (Table 3). The use of a water mask improved accuracy across all four groups, and on average reduced commission errors by 3.9% and RB from 9.9% to 4.4%, while minimally impacting omission error. In addition to the validation, the median surface water extent for each site over the 5‐year period was highly correlated (R = 0.93, p < 0.01) with the percentage of each site mapped as wetland by the NWI, indicating the surface water extent was consistent with other sources of wetland data (Figure A2).

Table 3.

Errors of Omission (OE) and Commission (CE) for Water, Overall Accuracy (OA) of Both Water and Non-Water, and the Relative Bias (RB) Indicating the Magnitude That Water Is Over- or Under-Estimated for 2017–2021

Region	OE (%)	CE (%)	OA (%)	RB (%)
Water mask applied
All Sites	1.9	6.5	97.2	4.4
North Central	2.0	8.7	96.5	6.8
Midwest Central	2.0	5.3	97.7	3.3
Southwest Central	0.9	5.0	98.1	4.1
Southeast Central	3.0	6.9	96.6	4.0
No mask applied
All Sites	1.6	10.4	95.8	9.9
North Central	1.1	13.8	94.8	12.8
Midwest Central	2.0	10.9	95.7	9.0
Southwest Central	0.7	8.0	97.1	7.3
Southeast Central	2.9	9.7	95.5	6.9

3.2. Surface Water Dynamics Under Current Climate

The median surface water extent as a percent of the site area was highly variable across the 32 sites, ranging from 3.9% in OK3 to 45.1% in MN2. By group, the NoC sites had the most amount of water on average, 26.8% but sites in South Dakota, Arkansas and Mississippi also displayed high amounts of surface water (Table 4). While the NoC sites had, on average, a higher standard deviation and range in surface water, the SEC sites showed the highest average coefficient of variation (Table 4). At a site‐scale, temporal variation was substantial at sites with large floodplains, such as MO2 and ND2 (Figure 4), but expansion and contraction of wetlands and lakes was also evident at sites such as OK4 and TX1 (Figure 4). Sites with floodplains had the highest coefficient of variation, for example, ND2, which contains the Red River floodplain, and MS1, which contains the Mississippi River and Yazoo River backwater. In contrast, temporal variability was lowest at sites where surface waters were concentrated in large lakes (e.g., MN3, OK4, SD1, OK2) (Table 4).

Table 4.

iSurface Water Dynamics Including Median, Standard Deviation (Std Dev), Range, and Distribution Metrics for Surface Water Extent Over the Time Series at Each Site

Region	Site	Median (% of area)	Std dev (km2)	Range (km2)	Coefficient of variation	Kurtosis	Skew
North Central		26.8	2.34	10.23	10.4	−0.1	0.4
	ND1	7.9	1.24	5.12	15.7	0.3	0.8
	ND4	17.5	1.26	6.45	7.2	−0.1	0
	ND2	11.9	2.72	11.64	22.8	−0.1	0.7
	ND3	26.9	1.76	7.71	6.5	−0.7	0
	MN1	29.1	1.76	7.30	6	−0.6	0.5
	MN2	45.1	3.95	18.30	8.8	1.1	0.9
	MN3	39.4	2.06	8.16	5.2	−0.8	0
	MN4	36.6	3.97	17.13	10.8	−0.2	0.2
Midwest Central		13.6	1.25	6.00	11.1	0.4	0.6
	SD2	16.9	2.06	8.87	12.2	−0.7	0.3
	SD1	26.1	1.41	7.53	5.4	1.1	0.7
	SD3	24.2	1.80	9.61	7.5	1	0.3
	SD4	14.6	1.36	6.92	9.3	1.2	0.9
	IA2	9.1	0.86	3.62	9.4	0.1	0.7
	IA1	4.7	0.98	4.40	20.8	0.4	0.9
	IA3	4.4	0.59	2.74	13.5	0.1	0.6
	IA4	8.7	0.93	4.33	10.7	0	0.5
Southwest Central		7.8	0.73	3.50	10.2	1.8	1
	MO1	5.7	1.11	4.99	19.5	0	0.8
	OK2	8.0	0.43	1.80	5.3	−0.2	0.7
	OK1	5.2	0.41	2.63	8	11.1	2.7
	OK4	15.9	0.92	4.82	5.8	1.8	1
	OK3	3.9	0.47	2.64	12.2	2.6	1.1
	TX1	9.8	0.70	3.17	7.1	−0.4	0.4
	TX2	4.5	0.28	1.29	6.3	−0.1	0.5
	TX4	9.0	1.53	6.68	17	−0.3	0.6
Southeast Central		13.7	2.02	8.63	13.8	0.3	0.8
	MO2	8.2	1.62	8.95	19.8	4	1.6
	MO4	12.2	0.92	3.49	7.5	−0.8	0
	MO3	8.0	1.17	4.88	14.6	−0.2	0.7
	AR1	22.2	2.59	12.04	11.7	−0.2	0.2
	AR2	11.2	1.71	6.96	15.3	−0.2	0.8
	MS1	26.9	6.13	25.11	22.8	0.3	0.8
	LA1	5.5	0.44	1.95	8	0.3	1
	TX3	15.1	1.57	5.63	10.4	−0.5	0.9

Note. Bolded values indicate averages for each region.

Figure 4.

Examples of the variability in surface water extent for ND2, MO2, OK4, and TX1, in a wet month (a, e, i, m), and a dry month (b, f, j, n), and corresponding examples of Copernicus Sentinel‐2 imagery from these month and years (c, d, g, h, k, l, o, p). Panel (a) shows flooding along the Red River and (e) shows flooding along the Missouri River.

Kurtosis and skew help characterize the distribution of the surface water values over the time series for each site. A skew of 0 represents a normal distribution, while a positive skew means the outliers are further to the right (or wetter) and the mean is closer to the left (or drier). Surface water extents were normally distributed at four sites (ND4, ND3, MO4, MN3), whereas distributions at the remaining 28 sites were positively skewed (Table 4). OK1, MO2, and OK3, and OK4 had the largest positive skew, meaning surface water extent tended to be relatively stable with infrequent increases in surface water extent. Of the 32 sites, MO4 and MN3, had the lowest kurtosis values, while 16 sites had negative kurtosis values, meaning that the distribution was flatter than a normal distribution, or timesteps with surface water conditions that were much wetter or much drier than the median condition were relatively common. Like skew, OK1, MO2, OK3 and OK4 showed the highest kurtosis values, meaning the wettest and driest conditions were less common and deviated less from the median surface water extent (Table 4).

3.3. Surface Water—Climate Statistical Models

Temporal variability in surface water extent at each site was modeled as a function of meteorological variables. The random forest models explanatory power ranged from R² = 0.29 at TX4 to R² = 0.80 at ND3 (Table 5). Open (lakes, ponds, rivers) and vegetated (wetlands, riparian corridors) water was summed and modeled over time at all sites except OK1 and OK2, where only OW was modeled (Table 5), as the random forest model explanatory power was minimal at these two sites using the combined open and VW. The number of explanatory variables selected by each model ranged from 2 to 6 variables and averaged 4 variables. Most models selected a relative humidity (minimum or maximum; 24 of 32 models) or precipitation variable (23 of 32 models) (Figure A3). Minimum temperature (12 months; 7 of 32 models), minimum relative humidity (9 months; 6 of 32 models), and precipitation at 2, 6 and 12 months (each selected by 6 of 32 models) were the variables most frequently selected (Figure A3).

Table 5.

Model Parameters and Performance for Each Site

Region	Site	Response variable	Correlation threshold	Min leaf samples	Trees (count)	Train-test split	Model R2	Model RMSE	R2: K-fold CV (splits = 6)	RMSE: K-fold
North Central	ND1	OW + VW	0.8	2	200	Stratified	0.57	60.58	0.43	93.32
	ND4	OW + VW	0.8	2	300	LOOCV	0.48	90.80	0.42	95.42
	ND2	OW + VW	0.8	2	300	Stratified	0.47	157.33	0.32	222.69
	ND3	OW + VW	0.8	3	500	Stratified	0.80	66.12	0.55	116.97
	MN1	OW + VW	0.8	3	500	Stratified	0.41	68.58	0.33	90.12
	MN2	OW + VW	0.8	2	300	LOOCV	0.41	123.64	0.45	118.73
	MN3	OW + VW	0.8	2	300	Stratified	0.40	72.11	0.20	103.94
	MN4	OW + VW	0.8	3	500	Stratified	0.31	63.53	0.15	83.65
Midwest Central	SD2	OW + VW	0.8	2	200	LOOCV	0.42	155.35	0.41	156.33
	SD1	OW + VW	0.8	2	200	LOOCV	0.34	113.81	0.42	107.02
	SD3	OW + VW	0.8	2	200	LOOCV	0.54	120.56	0.54	121.74
	SD4	OW + VW	0.8	3	300	Stratified	0.73	53.51	0.61	83.96
	IA2	OW + VW	0.9	2	200	LOOCV	0.60	53.66	0.58	55.35
	IA1	OW + VW	0.8	2	500	LOOCV	0.48	69.96	0.46	71.06
	IA3	OW + VW	0.8	2	300	LOOCV	0.36	46.92	0.38	46.32
	IA4	OW + VW	0.9	3	500	Stratified	0.37	56.92	0.28	77.93
Southwest	MO1	OW + VW	0.8	2	200	LOOCV	0.46	66.82	0.51	70.83
Central	OK2	OW	0.8	2	500	Stratified	0.43	27.37	0.26	36.31
	OK1	OW	0.8	2	500	Stratified	0.67	15.99	0.27	34.98
	OK4	OW + VW	0.8	2	300	LOOCV	0.33	74.46	0.34	73.77
	OK3	OW + VW	0.8	2	300	Stratified	0.47	26.61	0.32	38.73
	TX1	OW + VW	0.8	2	200	Stratified	0.56	37.13	0.33	56.49
	TX2	OW + VW	0.8	2	300	Stratified	0.49	16.07	0.45	20.94
	TX4	OW + VW	0.8	3	300	LOOCV	0.29	127.74	0.21	134.92
Southeast Central	MO2	OW + VW	0.8	2	300	Stratified	0.55	93.98	0.39	135.96
	MO4	OW + VW	0.8	2	300	Stratified	0.62	104.34	0.48	147.23
	MO3	OW + VW	0.95	2	500	LOOCV	0.48	282.44	0.47	284.46
	AR1	OW + VW	0.8	2	200	LOOCV	0.45	190.17	0.40	198.45
	AR2	OW + VW	0.8	2	500	LOOCV	0.65	162.15	0.68	155.01
	MS1	OW + VW	0.8	2	500	LOOCV	0.53	414.60	0.51	426.07
	LA1	OW + VW	0.8	2	200	LOOCV	0.39	306.48	0.38	309.87
	TX3	OW + VW	0.8	2	200	LOOCV	0.54	105.18	0.59	99.55

Note. Maximum tree depth of 3 was identified in all hyperparameter selections. OW, open water; VW, vegetated water; LOOCV, leave one out cross validation; RMSE, root mean square error; CV, cross validation.

The selected variables typically represented a range of lag periods. A single‐scene or 2‐month lag variable was included in 26 of the 32 models. Similarly, 29 of the 32 models included a 9 or 12‐month lag variable (Figure A3). A geographic pattern in the lag period was also observed. Across the climate groups, the NoC and MWC, which experience persistent snow cover, averaged longer lag periods (8.5 and 7.8 months, respectively) for the variable showing the greatest importance within the model, relative to the SEC and SWC (4.0 and 4.1 months, respectively), which lack persistent snow cover, with the patterns driven by long lag periods for the North and South Dakota sites (11.3 and 9.0 months, respectively), and short lag periods for the Arkansas‐Louisiana‐Mississippi and Texas sites (2.3 and 2.8 months, respectively) (Figure A3).

3.4. Projected Changes in Climate and Surface Water Extent

By 2061–2080, temperatures and VPD are projected to increase, with associated reductions in relative humidity, compared to the 2006–2025 period. These changes are projected under all climate scenarios at all sites, with greater and lesser increases in temperature and VPD following a north to south gradient, respectively (Figures 2 and A4). Annually for the 20‐model MACAv2‐METDATA mean (RCP8.5), precipitation is expected to increase in the northern half of the study area but decrease in the southern half (Figures 2 and A4). Seasonally, precipitation is projected to increase in the spring months for most sites (24 of 32 sites) but decrease in the summer months for all sites. Projected MACAv2‐METDATA precipitation changes in the fall and winter months are more variable but generally conditions are expected to be wetter in the northern portion and drier in the southern portion of the study area (Figure A4).

Seasonal and annual projected increases or decreases in surface water refer to how projected changes in meteorology, between 2061–2080 and 2006–2025, will update observed GRIDMET data in the random forest models, and change estimates of surface water extent, relative to current surface water extents (2017–2021). Annually, 6 of 8 MWC sites under RCP4.5, were projected to show an average decline in surface water, relative to the 2017–2021 period, increasing to 7 of 8 MWC sites under RCP8.5. Additionally, 6 of 8 sites in both the SWC and SEC were projected to have less surface water under RCP8.5 (Figure 5). In contrast, 5 of 8 NoC sites were projected to have more surface water under both scenarios (Figures 5 and A5). Across all sites, nine were projected to show an average annual decline of more than 4% in surface water extent under RCP8.5 (maximum of − 16.2% at IA1), while six sites were projected to show an average annual increase in surface water extent of >5% (maximum of 11.4% at MO1) (Figure A5). Examples of the climate change projections over the entire time series for sites with projected increases, decreases, or no net change are shown in Figure 6. We note that, IA1, OK3, and MO1, all of which were projecting a greater degree of change also contained some of the smallest amounts of surface water, relative to the other sites considered (Table 4), therefore projected percent changes were converted to projected change in surface water extent in Figure 6b. Projection variability was greater between GCMs (median range = 9.8% between GCMs for annual percent change) than between emission scenarios (median range = 1.4% between RCP4.5 and RCP8.5 for 20‐model average) (Figure A5).

Figure 5.

Projected changes in surface water extent for (a) annual, (b) wettest 15% of conditions, and (c) driest 15% of conditions for RCP4.5 and 8.5. Sites are organized north (left) to south (right). Error bars indicate the range of random forest model predictions within each scenario introduced by using multiple Global Circulation Models (Abatzoglou & Brown, 2012).

Figure 6.

(a) Observed and projected change in surface water extent in ND1 (net increase projected), (b) projected change to median surface water extent at RCP8.5 using the 20‐model average, as well as (c) observed and projected surface water extent in MN2 (minimal net change projected), (d) TX4 (net decrease projected), and (e) MS1 (net decrease projected). The time series panels show how the 2017–2021 surface water extents are projected to change over the time series, using climate change adjusted climate data instead of observed climate data.

Analyses using annual average surface water extent masks temporal variability in the direction and magnitude of projected changes. Seasonally across the central U.S., for instance, the greatest declines in surface water extent were projected for the summer (19 and 21 sites under RCP4.5 and 8.5, respectively) and fall seasons (20 and 24 sites under RCP4.5 and 8.5, respectively), likely attributable to an increase in VPD and temperature with a general decrease in summer precipitation, although the use of lagged variables in all random forest models complicates this generalization. By climate classes, the SEC showed the greatest declines in summer, followed by the MWC. In contrast, the greatest increases in surface water extent were projected for the spring season (15 and 16 of 32 sites under RCP4.5 and 8.5, respectively, Figures A6 and 7), followed by the winter season (14 and 13 of 32 sites under RCP4.5 and 8.5, respectively, Figure A3), consistent with projected increases in spring and winter precipitation for many sites, but again the interpretation is complicated by the importance of variables lagged over 9–12 months. Many sites (22 and 23 sites under RCP4.5 and RCP8.5, respectively) are expected to show an amplification of seasonal surface water dynamics, with this dynamic most prevalent in the NoC and SEC climate classes. Additionally, differences attributable to emission scenario were greater within a season, than annually, ranging from a median RCP4.5 versus RCP8.5 difference of 3.0% in the spring for the 20‐model average to a median RCP4.5 versus RCP8.5 difference of 1.7% in fall (Figure A6).

Figure 7.

Projected changes in surface water extent for (a) spring, (b) summer, (c) fall, and (d) winter for RCP4.5 and 8.5. Sites are organized north (left) to south (right). Error bars indicate the range of random forest model predictions within each scenario, introduced by using multiple Global Circulation Models (Abatzoglou & Brown, 2012).

For the driest 15% of observations at each site, further declines in surface water extent were projected (24 sites under RCP4.5, 23 sites under RCP8.5) with the largest projected declines expected in the SEC and SWC. Alternatively, increases in dry condition surface water extent, or less severe drought impacts, were projected for MO1, MO3, and four of the eight NoC sites (Figures 5 and A6). Under the wettest 15% of observations, which represent historical peak surface water extents (Table 2) under RCP4.5, 19 of the sites are projected to show increases in surface water extent, suggesting flood conditions may be exacerbated at many sites (Figure 5), but the projected direction of impact for the wettest 15% was less consistent by climate group, compared to the seasonal and dry period findings. Projected increases in surface water extent under the wettest conditions decreased to 13 of the 32 sites under RCP8.5, likely due to increasing temperatures and VPD moderating surface water extent under wetter conditions. Sites with major floodplains showed mixed directionality, with some sites projecting increased surface water extent (e.g., AR1, MS1), while other sites projected decreased surface water extent (e.g., ND2, AR2) (Figure A6).

Site‐based variability in the projected direction and magnitude of changes to surface water extent were correlated with surface water metrics and MACAv2‐METDATA climate projections. Across sites, the direction and magnitude of projected changes to surface water extent tended to be primarily explained by projected changes to climate, instead of total surface water extent or surface water variability. For instance, increased winter precipitation, using RCP8.5, was positively correlated with more surface water extent annually and more surface water in dry conditions (Table 6). Projected changes to annual and dry condition surface water extents were also found to be positively correlated with the annual change in minimum temperature, which may be indicative of a latitudinal relationship. Additionally, projected increases in summer VPD were correlated with projected declines in dry condition surface water extent (Table 6).

Table 6.

Across Site Correlations Between Percent Change to Annual, Dry, Wet, and Seasonal Surface Water Extent Using the 20-Model Means, and Surface Water Metrics As Well As Projected Climate Change Derived by the 20-Model Means

	Annual	Dry	Wet	Spring	Summer	Fall	Winter	Annual	Dry	Wet	Spring	Summer	Fall	Winter
Surface water metrics	Model average (RCP4.5)							Model average (RCP8.5)
Median	0.13	0.12	−0.04	0.08	0.13	0.19	0.03	0.04	0.1	0.08	0.03	0.03	0.3	0.06
Standard Deviation	0.03	0.05	−0.15	0.04	−0.06	−0.01	0.22	0.1	0.15	−0.12	0.06	−0.07	0.11	0.3
Range	0	0.03	−0.18	0.05	−0.06	−0.05	0.17	0.05	0.12	−0.16	0.05	−0.09	0.05	0.26
Range/Median	−0.18	−0.18	−0.11	−0.11	−0.31	−0.29	0.25	−0.04	−0.05	−0.3	−0.1	−0.2	−0.37	0.29
Std. Dev/Median	−0.15	−0.17	−0.1	−0.17	−0.33	−0.22	0.33	0.02	−0.01	−0.27	−0.09	−0.19	−0.31	0.36
Kurtosis	−0.2	0	−0.14	−0.05	−0.15	−0.1	−0.25	−0.21	−0.12	−0.09	−0.07	−0.12	−0.29	−0.21
Skew	−0.24	−0.02	−0.2	−0.14	−0.25	0.01	−0.17	−0.23	−0.06	−0.17	−0.11	−0.1	−0.28	−0.19
Projected change
Temp (min) (annual)	0.2	0.49	−0.22	0.03	0.35	0.18	0.01	0.39	0.53	−0.11	0.16	0.33	0.28	0.13
Precipitation (spring)	0.23	0.33	−0.03	0.03	0.2	0.15	0.15	0.28	0.3	−0.02	0.2	0.12	0.03	0.14
Precipitation (summer)	0	−0.31	0.14	−0.05	−0.4	0.15	0.27	0.07	0.09	0.15	0.18	0.17	−0.08	−0.22
Precipitation (fall)	−0.05	−0.2	0.02	−0.24	−0.27	0.04	0.32	0.12	0.27	−0.21	−0.04	0.07	0.12	0.02
Precipitation (winter)	−0.12	0.16	−0.37	−0.32	0.04	−0.04	−0.03	0.4	0.47	−0.1	0.14	0.23	0.23	0.25
Precipitation (annual)	0.04	−0.07	−0.03	−0.17	−0.24	0.13	0.31	0.15	0.23	−0.04	0.06	0.03	0.09	0.03
VPD (spring)	0.13	0.41	−0.17	−0.12	0.32	0.19	−0.02	0.03	0.1	0.04	−0.11	0.08	0.19	−0.03
VPD (summer)	−0.1	−0.05	−0.03	0.03	−0.18	−0.04	−0.1	−0.12	−0.41	0.1	−0.01	−0.4	−0.05	0.13
VPD (fall)	−0.03	0.36	−0.33	−0.11	0.22	−0.04	−0.18	0.14	−0.05	0.29	0.26	0.03	−0.07	0.11
VPD (winter)	0.07	0.4	−0.26	−0.08	0.27	0.15	−0.07	0.14	0.44	−0.14	−0.02	0.27	0.25	0.01
VPD (annual)	0.04	0.32	−0.22	−0.12	0.14	0.09	−0.05	0.15	0.29	0.01	−0.02	0.11	0.3	0.09

Note. Gray cells indicate significance at p < 0.01 after Bonferroni correction.

4. Discussion

4.1. Summary and Contextualized Findings

Surface water dynamics play a critical role across the central U.S. The Prairie Pothole Region, for example, provides habitat for 50%–80% of North American ducks each spring and summer (Batt et al., 1989), and the Mississippi River basin, which extends across the study area, drains 40% of the conterminous U.S. and produces most of the total harvested cropland in the country (USDA, 2021). Changes in surface water extent will therefore have considerable impacts on the ecology, productivity, and land use across the central U.S. Using time series of Sentinel‐1 and Sentinel‐2 derived surface water extent, we documented high between site variability in both the amount of surface water extent as well as how dynamic the surface water was, consistent with the findings of others (e.g., Petrakis et al., 2022; Zou et al., 2018). In modeling surface water at each site, antecedent meteorological conditions were found to be important, as previously documented by others (e.g., Song et al., 2018; Vanderhoof et al., 2018). Further, consistent with prior studies, surface water tended to be positively correlated with precipitation and negatively correlated with temperature (e.g., Tulbure & Broich, 2019; Xia et al., 2019). The site‐specific statistical models and projections enabled us to account for the influence of current land use and surface storage capacity on surface water dynamics (Vanderhoof et al., 2018).

Our projections of surface water dynamics under RCP4.5 and RCP8.5 can be compared to previous efforts across the central U.S. For example, projected increases in surface waters in many NoC sites were consistent with wetting trends documented in precipitation, streamflow, and wetland pond counts (McKenna et al., 2017; Niemuth et al., 2014; Zou et al., 2018) and projections of increased waterbody extent across the Dakotas (McIntyre et al., 2019). Alternatively, others have argued that wetlands in western and central Prairie Pothole Region, where most wetland conservation is concentrated, show drying (Ballard et al., 2014; Johnson et al., 2010), and that shrinking of wetland area across the region should be expected concurrent with decreases in water availability (Cook et al., 2014; Johnson & Poiani, 2016; Sofaer et al., 2016). These prior studies may help explain why the South Dakota and most western North Dakota sites showed projected declines in water extent in our analysis. While increased evaporative demand drove projected declines in water extent in the summer and fall months for many sites, at an annual scale, we found that the NoC sites were projected to show the greatest increase in annual precipitation, as well as high variable importance for variables with a longer lag period, which largely outweighed the increase in temperatures in these northern sites. Sites further south, such as the MWC group, showed more minor increases in precipitation, and the highest increases in temperature, which may further explain the projected declines in surface water at these sites. Our findings in Iowa, however, were counter to predictions by others of minor wetland expansion (Ballard et al., 2014; Garris et al., 2015; Johnson et al., 2010). In the SEC and SWC, our results were more consistent with existing literature. Our projections of declines in surface water extent across most of the SWC and SEC sites were consistent with previously documented drying trends across the U.S. Southwest (Ryberg & Chanat, 2022; Zou et al., 2018) and declines in high pulse precipitation events in the U.S. Southeast (Ryberg & Chanat, 2022).

In addition to projecting shifts in annual average surface water extent, our approach also documented projected shifts by season as well as for the wettest and driest conditions. Many sites showed a potential amplification of seasonal patterns in which spring, which is typically wet, may become wetter, while dry end of summers may become drier, with potential for corresponding amplification of surface water dynamics. However, changes to seasonal patterns of surface water extent will depend on the relative influence of short‐term (1–2 months) versus longer lag periods (9–12 months) on surface water extent, which in turn, was found to depend on latitude and the persistence of snow cover. Additionally, while many sites showed projected declines in dry condition surface water extents, our comparison between the JRC GSW 2017–2021 and 1984–2016 periods indicated that the surface water response to very dry conditions was under‐represented, relative to the historical range of conditions. This introduces potential uncertainty for projecting future impacts of drought on surface water extent. Further, we note that since projecting shifts in the tails of a distribution is particularly challenging, recent extremes may not necessarily be indicative of future extremes (Sillmann et al., 2017).

It is also important to note that our approach does not sufficiently consider how future changes in irrigation, tile drainage, flood management or sub‐surface groundwater storage, will compensate for or exacerbate climate-induced changes in surface water dynamics, which are likely to respond to seasonal amplification of surface water dynamics, for example. It will be the combined changes, induced by climate, land use, and groundwater, that will determine the future availability of surface water to adequately provide ecosystem services (Condon et al., 2020; Gaines et al., 2022; McKenna et al., 2019; Uden et al., 2015). In addition, the unique approach proposed in this analysis to project site‐specific responses to climate change requires discussion of challenges and sources of uncertainty within the surface water time series, statistical modeling, and climate change projections.

4.2. Surface Water Accuracy

Projecting impacts on a surface water time series requires an accurate and complete surface water data set. Using a multi‐sensor approach and integrating the Sentinel‐1 and Sentinel‐2 algorithms at a monthly time step improved the accuracy of the surface water classifications, relative to the single algorithm accuracy reported in Vanderhoof et al. (2023). This improvement largely occurred by allowing erroneous outlier classifications of water, to be disregarded when the classifications were composited at a monthly time step. Although the classification of subpixel water, such as VW, typically shows lower accuracy relative to OW (Goffi et al., 2020; Konapala et al., 2021; Vanderhoof et al., 2023), the inclusion of VW is critical to fully capture surface water dynamics (Jones, 2019), particularly in systems where much of the water is in forested wetlands or in vegetated riparian corridors. To further compensate for the lower accuracy of VW, we found that developing site specific water masks was an effective approach to reduce commission error outside of the 100‐year floodplain, with minimal increases in omission error. The monthly timestep enabled the Sentinel‐1 algorithm classification to gap‐fill when cloud or cloud shadows prohibited a Sentinel‐2 algorithm classification. This approach provided a substantial advantage relative to relying on Landsat alone. For the 2017–2021 period, for example, only 49.8% of the JRC GSW Landsat monthly data set contained complete data (i.e., >95% data coverage) compared to 98.6% by the Sentinel‐1 and Sentinel‐2 composites (Table 2). This difference suggests that a similar analysis relying on Landsat data would require twice as many years just to represent a similar amount of data and would still under‐represent snow‐cover months. However, a monthly timestep can potentially result in large episodic precipitation events or flood events being missed or under‐mapped. Figure 4 shows examples of flood events being mapped at the monthly timestep along the Red River and Missouri River, indicating that the approach still allows seasonal flood events to be documented and included in the analysis.

Reliance on Sentinel‐1 and Sentinel‐2 data sets to derive surface water extents limited the time series length over which the statistical models were developed. This could have contributed to statistical model uncertainty, attributable to both relatively small sample sizes, but also uncertainty around representing the full historical range of climate and water conditions. Contextualizing the 5‐year period using PDSI and the GSW data sets enabled us to demonstrate that the period represented a wide range of both climate and water conditions, relative to the past ~40 years (Table 2), suggesting that substantial seasonal and interannual variability in surface water conditions were included in the model development. Relying on a recent 5‐year period did allow the statistical models to be developed using contemporary land use and climate conditions. And as many waterbodies, including rivers (<30 m wide), streams, small wetlands, will be omitted by even moderate resolution data sets (Lechner et al., 2009; Vanderhoof & Lane, 2019), we prioritized using a 20 m resolution data set to maximize the number of waterbodies and wetlands included in the analysis.

4.3. Modeling Surface Water as a Function of Climate

The strength of the random forest models influences our confidence in the surface water projections. The random forest model performance metrics obtained in the analysis were moderate (median model R² = 0.48, K‐fold CV R² = 0.42). Although climate is a primary driver of variability in surface water extent over time (Xi et al., 2021), modeling variability in inundation extent as a function of meteorological variables, alone, typically produces a relatively weak explanatory power. Our random forest model performances were comparable or exceeded previous efforts to explain temporal variability in surface water extent using meteorological variables (e.g., Mishra & Cherkauer, 2011; Song et al., 2018; Tulbure & Broich, 2019; Xia et al., 2019). Surface water—climate models have previously been improved by including variables beyond temperature and precipitation, such as evapotranspiration or precipitation minus evapotranspiration (Sofaer et al., 2016; Song et al., 2018; Zhang et al., 2021). Similarly, in our analyses, 17 of the 32 models selected VPD, relative humidity or SPH as the variable showing the greatest model importance, compared to 13 that selected temperature or solar radiation, and only two that selected precipitation as the variable showing the greatest importance (Figure A3).

We also considered a range of antecedent conditions for the climate variables. Most efforts to explain temporal variability in surface water extent have used a single climate interval, whether that is daily (e.g., Heimhuber et al., 2017), monthly (e.g., Song et al., 2018), seasonal (e.g., Gaines et al., 2022; Tulbure & Broich, 2019), or annual (e.g., Xia et al., 2019), instead of testing climate variables averaged or accumulated over multiple intervals. Globally, lags of 1–2 months typically occur between peak rainfall and peak soil moisture (Papa et al., 2010), but in the Prairie Pothole Region and prairie plains, antecedent conditions drive the response of wetlands to additional precipitation (Haque et al., 2022), so longer periods of accumulation, such as 9 months, or even multiple years have been found to outperform shorter intervals in explaining patterns of surface water over time (Vanderhoof et al., 2018). The longer lags selected at more NoC sites in our analysis may also be explained by the lag between snowfall and snowmelt in the northern sites. Persistent snow cover occurs at the North and South Dakota, Minnesota, and Iowa sites. MACAv2‐METDATA has been recommended for use in snow‐dominated regions (Alder & Hostetler, 2019), but it is unclear if future changes in the precipitation type (e.g., rain, snow, rain on snow) and timing of thaw, which will influence spring surface water extent (Davitt et al., 2019; Mishra & Cherkauer, 2011), are adequately being accounted for.

In addition to temporal lags in meteorological data, issues related to spatial variability and resolution are also important. For example, averaging the meteorological variables per site at each timestep when modeling surface water may have limited potential uncertainty introduced from downscaling climate change projections, but in turn, introduced uncertainty by obscuring local scale variability in meteorology. While the relatively flat topography within the central U.S. sites meant that temporal variability in meteorology tended to dwarf within site variability in meteorology (Figure A1), this issue could limit our ability to apply this approach in regions with abrupt changes in elevation.

4.4. Modeling Approach

The approach taken to model future change is also important to consider. Methods in other efforts that project climate change impacts by modeling spatial variability in wetland density (e.g., Garris et al., 2015; Sofaer et al., 2016; Zhang et al., 2021) can be rapidly applied across large landscapes but produce projections of shifts in wetland averages only. Providing projected changes across seasons as well as in abnormally wet or dry conditions may be of equal or greater interest, for instance, to plan for future flood and drought impacts and upgrade deteriorating water infrastructure (Lall et al., 2018; Merz et al., 2020; Reiter et al., 2018). Additionally, modeling wetland density assumes that spatial variability in wetland storage responds similarly to climate as temporal variability in wetland storage. However, the capacity for surface water to contract and expand will depend on site topography, storage capacity and infiltration (Kuppel et al., 2015; Shaw et al., 2012; Vanderhoof et al., 2018). Further, the rate of surface water movement into or out of an area will be influenced by site specific hydrological alterations such as tile drainage, ditching, irrigation, and dams (Haque et al., 2022; McCauley et al., 2015; Rajib et al., 2020). These factors may help explain why adjacent sites at times showed opposite directionality in the projected impacts of climate on surface water. To produce wall‐to‐wall projections the site‐specific statistical approach will need to be scaled. Developing pixel‐specific models using Landsat time series has been used to track changes in land cover over time (Brown et al., 2020) and detect gradual and abrupt forest disturbances (Ye et al., 2021). Per‐pixel models may be less logical for multi‐pixel features, like waterbodies, but alternatively, area‐specific models could be developed for a series of moving windows, an approach that has been used in the disaggregation of remotely sensed land surface temperature (Gao et al., 2017), but where window size is optimized based on the size distribution of water features.

5. Conclusion

In conclusion, we tracked and projected climate change adjusted time series of surface water extents for 32 sites across the central U.S., providing novel data for much of the MWC, SWC and SEC sites, where few surface water‐related climate efforts have focused. Surface water extents were highly variable across the 32 sites, with some sites dominated by a few large lakes, others containing large rivers with substantial floodplains, and others dominated by a high density of wetlands and lakes. Statistical model strength also showed variability, as did the variables selected and the average period of accumulation for the selected variables. Across the central U.S. we demonstrated a novel approach of (a) using a multi‐sensor approach to track surface water dynamics over time, (b) accounting for site specific responses of surface water to meteorological conditions, and (c) projecting changes to not only the annual average but also seasonally, as well as for dry and wet outlier conditions. Results suggest latitudinal patterns, where the surface water extents were projected to increase at many NoC sites but decrease across most MWC, SWC and SEC sites. Projected seasonal changes suggest an amplification of seasonal patterns at many sites, with wetter springs but drier summers and falls. And while the impacts of droughts on surface water extent are expected to be exacerbated at most sites, impacts of flood conditions are projected to decrease under RCP8.5, relative to RCP4.5, with greater temperatures and VPD. Focusing on scaling the approach as well as incorporating projected changes in land use could provide information useful for water management planning. Projected changes in surface water extents can inform future estimates of regional and national surface water storage, which can guide conservation and mitigation efforts (Johnson & Poiani, 2016) as well as prepare for future flood and drought events (Sillmann et al., 2017).

Key Points:

• Surface water across the central U.S. responds to episodic, seasonal, and interannual variability in water availability

• Most sites showed projected declines in surface water extents under RCP4.5 and 8.5, peaking in summer‐fall and increasing drought impacts

• Select sites, concentrated in the north central U.S., projected increases in surface water, associated with greater precipitation

Appendix A: Surface Water Classification Validation

PlanetScope (source: Planet Labs PBC, San Francisco, CA) validation images were acquired between the tenth and the twentieth of a given month, to best represent monthly surface water conditions, and were further limited to snow‐free conditions with <20% cloud cover to maximize the accuracy of visually attributing validation data. Validation points (n = 200) were randomly generated per image (total = 12,800 points). Visual cues from Planet imagery were used to manually interpret each point as water or non‐water. The NWI wetland type and water regime attributes, monthly precipitation (Abatzoglou, 2013), and aerial imagery (source: Esri, Maxar, Earthstar Geographics) were used to support point attribution. Surface water is often a minority cover class, so a stratified disproportionate sampling scheme was employed, where water was required to represent at least 20% of the points, respectively, when feasible, to increase validation precision (Strahler et al., 2006). Points were moved as needed to the nearest water feature to reach the minimum of 20% water pixels per image. Validation data were retained at the original (3–5 m) spatial resolution so that a validation point could represent a mixed pixel in the classified image (20 m).

Table A1.

Thresholds Selected From 5-Year Surface Water Sentinel-1 (S1) and Sentinel-2 (S2) Algorithm Percentiles to Account for Variable Accuracy Between Sites, Sensors, and Classes (Open Water [OW]) Compared to Vegetated Water (VW)

Site	S1 OW threshold (%)	S1 VW threshold (%)	S2 OW threshold (%)	S2 VW threshold (%)
ND1	10	25	10	30
ND4	12	25	10	30
ND2	12	35	10	35
ND3	8	30	10	25
MN1	20	~	12	20
MN2	12	30	10	25
MN3	10	25	10	25
MN4	10	25	10	25
SD2	10	40	5	30
SD1	30	20	5	20
SD3	15	30	10	20
SD4	15	~	5	20
IA2	12	20	10	20
IA1	5	25	10	23
IA3	25	40	5	35
IA4	5	28	10	34
MO1	10	28	5	15
OK2	5	40	10	40
OK1	10	20	10	~
OK4	10	30	10	20
OK3	5	30	5	20
TX1	10	15	10	~
TX2	10	20	5	~
TX4	5	30	5	20
MO2	10	35	14	30
MO4	10	~	10	35
MO3	10	~	10	50
AR1	10	40	10	30
AR2	5	30	5	30
MS1	5	~	15	30
LA1	5	10	5	~
TX3	5	30	5	30

Note. The symbol: ~ indicates that this output was excluded from the water mask.

Table A2.

Images (Source: PlanetLabs PBC, San Francisco, CA) Included in the Validation oftheMonthly Surface Water Composites; Sites Organized North to South

Site	Date	Satellite	Site	Date	Satellite
ND1	10/13/2017	PS dove classic	MO1	9/18/2018	PS dove classic
ND1	8/14/2021	PS dove classic	MO1	4/15/2020	PS dove classic
ND4	6/15/2018	PS dove classic	MO2	6/15/2017	PS dove classic
ND4	4/17/2021	PS dove classic	MO2	10/14/2019	PS dove R
ND2	7/13/2018	PS dove classic	MO4	2/14/2019	PS dove classic
ND2	4/16/2019	PS dove R	MO4	8/14/2020	PS dove classic
ND3	7/14/2017	PS dove classic	MO3	7/13/2018	PS dove classic
ND3	10/13/2020	PS dove R	MO3	11/18/2021	PS dove R
MN1	4/13/2017	PS dove classic	OK2	11/16/2017	PS dove classic
MN1	6/13/2020	PS dove R	OK2	6/17/2020	PS dove R
MN2	8/16/2019	PS dove classic	OK1	7/18/2019	PS dove classic
MN2	5/17/2021	PS dove R	OK1	12/13/2021	PS super dove
MN3	10/17/2018	PS dove classic	OK4	1/16/2018	PS dove classic
MN3	7/18/2020	PS dove R	OK4	9/16/2019	PS dove classic
MN4	9/12/2017	PS dove classic	OK3	3/15/2018	PS dove classic
MN4	5/15/2019	PS dove classic	OK3	8/14/2020	PS dove classic
SD2	7/17/2017	PS dove classic	TX1	2/15/2017	PS dove classic
SD2	11/13/2020	PS dove classic	TX1	7/13/2021	PS dove classic
SD1	5/15/2019	PS dove R	TX3	6/12/2019	PS dove classic
SD1	9/14/2021	PS dove classic	TX3	1/14/2021	PS dove classic
SD3	8/13/2018	PS dove classic	TX2	12/15/2018	PS dove classic
SD3	4/17/2021	PS dove classic	TX2	5/14/2020	PS dove R
SD4	10/16/2017	PS dove classic	TX4	8/14/2017	PS dove classic
SD4	6/16/2020	PS dove R	TX4	11/15/2019	PS dove classic
IA2	7/16/2018	PS dove classic	AR1	12/12/2017	PS dove classic
IA2	10/17/2021	PS dove classic	AR1	4/12/2021	PS super dove
IA1	5/14/2017	PS dove classic	AR2	8/16/2019	PS dove R
IA1	9/16/2019	PS dove classic	AR2	10/14/2020	PS dove classic
IA4	3/16/2019	PS dove classic	MS1	3/14/2018	PS dove classic
IA4	8/15/2021	PS super dove	MS1	10/17/2021	PS dove R
IA3	6/16/2017	PS dove classic	LA1	3/14/2018	PS dove classic
IA3	11/16/2020	PS super dove	LA1	10/17/2021	PS dove R

Figure A1.

A comparison of the relative magnitude of temporal versus within site or spatial variability in (a) precipitation, (b) vapor pressure deficit, and (c) surface downward shortwave radiation (Abatzoglou, 2013). The range of values occurring within a site was calculated for each monthly time step. The median value of the within site ranges was then used to characterize the within site spatial variability in meteorology (blue bar). This was compared to the temporal variability (green bar), calculated as the maximum monthly site average minus the minimum monthly site average over the January 2017 to December 2021 period.

Figure A2.

Correlation between the percentage of each site classified as wetland or deepwater habitat in the National Wetland Inventory (USFWS, 2019) and the median Sentinel‐1 (S1) and Sentinel‐2 (S2) surface water extent for 2017–2021.

Figure A3.

Permutation variable importance values for the variables selected for each site’s model. Importance values are color coded from low value (yellow) to high value (maroon). The number of months each variable was accumulated over is shown in parentheses. Row colors (white, tan and gray) indicate variable type. VPD, vapor pressure deficit; Tmin and Tmax, minimum and maximum temperature; SPH, specific humidity; RHmin and RHmax, minimum and maximum relative humidity; SR, solar radiation.

Figure A4.

Projected percent change in precipitation, vapor pressure deficit (VPD) and projected offset of maximum temperature (Tmax) using the 20‐model mean and RCP4.5 and RCP8.5 emission scenarios (Abatzoglou & Brown, 2012; IPCC, 2013). Red indicates warmer, drier, blue indicates wetter, cooler.

Figure A5.

Projected percent change to annual, average surface water extent by site, scenario, and model for the RCP4.5 and RCP8.5 emission scenarios.

Figure A6.

Projected percent change in surface water (negative and positive values indicate less and more water, respectively) for dry and wet conditions as well as seasonally using the 20‐model mean for the RCP4.5 and RCP8.5 emission scenarios.

Data Availability Statement

The surface water data produced for this analysis are published and available (Vanderhoof et al., 2024).