Assessing present and future coastal moderation of extreme heat in the Eastern United States

Climate models suggest a rapid increase of extremely hot days in coming decades. Cool marine air currently ventilates extreme heat in populous coastal regions, diminishing its impacts, but how well climate models capture this effect is uncertain. Here we conduct a comprehensive observational analysis of coastal extreme-heat ventilation—its length scale, magnitude, and regional patterns—and evaluate two ensembles of downscaled global climate models along the eastern US coast. We find that coastal areas are 2 °C–4 °C cooler than ∼60 km inland, resulting in reductions near 50% in population exposure to temperatures above 35 °C. Large seasonal and inter-regional variations are closely linked with land-sea temperature contrasts. High-resolution models underestimate coastal cooling by 50%–75%, implying that substantial and spatiotemporally varying model bias correction is necessary to create accurate projections of coastal extreme heat, which is expected to rise considerably with anthropogenic forcing. Our results underline the importance of regionally- and observationally-based perspectives for assessing future extreme heat and its impacts, and for positioning effective heat-risk management for communities and jurisdictions that span coast-to-inland areas.


Introduction
Extreme heat can have multiple and severe impacts in temperate and subtropical regions (Miller et al 2008, Dunne et al 2013, Horton et al 2016, Mora et al 2017. Fine-scale processes, such as local ocean-atmosphere or land-atmosphere interactions, often play a crucial role in climate extremes (Diffenbaugh et al 2005, Lebassi et al 2009. Such interactions prevail in coastal areas, which frequently experience warm-season daytime cooling. Well-defined sea breezes occur where the coast-to-inland temperature difference is large, but coastal cooling may be observed in the absence of a sea breeze as well (Lebassi-Habtezion et al 2011, Meir et al 2013.
Several previous studies have noted the importance of sea breezes and coastal moderation of extreme temperatures for ameliorating heat and pollution in coastal areas of Southern California (Clemesha et al 2018) and New York (Melecio-Vázquez et al 2018). Regional analyses have found important regional heterogeneity in projected future changes in these effects (Zhao et al 2011). However, no studies have focused directly on surveying coastal moderation of extreme high temperatures, nor how that influences population exposure. In addition, complex interacting atmospheric and marine processes make regional generalizations and comparisons difficult. Therefore, establishing a comprehensive and region-specific observational basis for coastal extreme-heat moderation is crucial for better understanding its spatial patterns and evaluating large projected increases in extreme heat (Gao et al 2012, Zobel et al 2017. Recent work on US temperature extremes has used global Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. climate models (GCMs) or reanalysis datasets that are too spatially coarse to resolve coast-to-inland temperature gradients and, consequently, the coastal-cooling phenomenon (Thibeault and Seth 2014, Wuebbles et al 2015, Ashfaq et al 2016, Papalexiou et al 2018; studies that do use high-resolution products typically take a broader view in their analysis (Gao et al 2012, Ning et al 2015, Zobel et al 2018. For those projections that rely on absolute temperature thresholds, bias corrections are necessary and often implemented, but there has been little analysis of how such corrections alter the picture of coastal exposure to extreme heat.
As heat extremes rapidly increase (Horton et al 2016), coastal moderation of extreme heat is paramount to understand and quantify, particularly its potential changes in a warming climate. This coastal moderation affects the heat exposure of about 50 million people in the eastern US alone, and nonlinear increases in health and economic impacts at the hottest temperatures make even small reductions meaningful (Wu et al 2014, Coffel et al 2018, Coffel et al 2019. Accurate assessment of the spatial footprint of future heat extremes is essential to local-and regionalscale efforts to manage heat exposure and its risks, as it enables financial, educational, medical, and other resources to be allocated according to need. For jurisdictions which include both coastal and inland communities, ongoing management of heat risks will necessarily need to be informed by the present magnitude of coastal cooling, as well as its future changes.
Here we conduct a systematic analysis of the extent to which historical heat extremes are moderated by marine influences in the eastern US, focusing on the 60 km wide coast-to-inland swath along the Atlantic and Gulf coasts. Using station and gridded observations, we analyze regional patterns of coastal cooling over the recent historical record to position our consideration of how projections of future heat extremes must be adjusted, due to the above-described model challenges, to reflect their true spatial distribution in coastal and near-coastal zones.

Observational data
We use historical daily-maximum temperature [T max ] data for 1981-2015 from the 4 km resolution parameter regression on independent slopes model (PRISM) (Daly et al 2008). PRISM takes station data as input and processes it using terrain-and coast-aware interpolations to produce a best-estimate gridded product (Daly et al 2003). In the eastern US PRISM employs a coastal-advection model that assumes a grid point's coastal influence is a function of distance from the coast, with bays and inlets treated as transition zones and terrain effects assumed negligible. Stations with similar coastal influence are weighted more heavily when computing grid point variables. Gridded 4 km resolution data is sufficient for capturing coastal cooling as observed from weather stations (Novak andColle 2006, Lebassi-Habtezion et al 2011), a validation we also perform using the Global Surface Hourly and Global Historical Climatology Network-Daily datasets (Menne et al 2012). We conduct temperature-gradient and population-exposure analyses (see sections 2.4, 2.5) for seven regions (table 1, figure 1).

Model data
We employ data from two ensembles of daily-resolution downscaled GCMs (table S1 is available online at stacks.iop.org/ERL/14/114002/mmedia). From a dataset produced by Zobel et al (2017Zobel et al ( , 2018, we use an ensemble of five GCM model variants dynamically downscaled with the Weather Research and Forecasting (WRF) model to 0.1°(∼11 km) resolution (hereafter referred to as the WRF ensemble). These data are for 1995-2004 (historical) and 2085-2094 (future, RCP8.5). From the localized constructed analogs (LOCA) project (Pierce et al 2014(Pierce et al , 2015, we use a statistically-downscaled (∼6 km resolution) ensemble of historical runs  and future projections for the high-emissions RCP8.5 scenario (2075-2099) for 14 GCMs (Meinshausen et al 2011). These 14 models are selected to span much of the range of the full CMIP5 suite while maintaining a variance that enables greater comparability with the WRF ensemble; Table 1. (Columns 1 and 2) The number of coastal grid points and regional hot days resulting from the PRISM analysis. The total number of points comprising each regional distribution is thus the product of these two columns. (Column 3) Summary of the means of the coastalcooling intensity calculation discussed in the text. Intensity ranges span the 5th-95th percentiles of the distribution, making the cooling significant based on a two-tailed t-test.

Region
Coastal however, other model choices may produce slightly different results.

Defining coastal and inland areas
To ensure accurate comparison among gridded products, we match grid points based on their region and distance from the model-defined coastline, rather than by their absolute geographical location. We focus our analysis on coastline sections where coastal weather stations face the open ocean and that do not have bays or estuaries larger than 50 km in width, as these can introduce complex weather patterns that could confound the coastal cooling effects we seek to identify (Novak and Colle 2006). Terrain-complexity concerns also motivate our focus on the eastern US, where terrain variations are small within 100 km of the coast. For each section of coastline we define an 'inland' area located 60 km away, perpendicular to the local coastline direction. This 60 km distance is far enough inland to be beyond the reach of daytime coastal effects (Finkele 1998, Hu andXue 2016), and small enough that differences due to synoptic-scale weather conditions are minimized.

Defining hot days and coastal cooling
We focus on characterizing coast-to-inland temperature differences rather than attempting to attribute such differences to driving processes, such as sea breezes, clouds, or precipitation. We define hot days as the top decile of daily T max in the warm season (defined as May-September), based on a daily grid point climatology temporally smoothed with a Gaussian filter-a common method for avoiding spurious dayto-day variations (Freychet et al 2018). For each region, 'regional hot days' are then defined as those for which >50% of regional inland grid points are experiencing a hot day, following Smith et al (2013).
We calculate the 'coast-to-inland temperature difference' on a regional hot day as the difference between daily T max along the coast (averaged over each set of three adjacent coastal grid points within that region) and daily T max 60 km inland (averaged over each set of three corresponding inland grid points). We define the magnitude of coastal cooling for each regional hot day as being proportional to this coast-toinland temperature difference. We choose to quantify the coastal-cooling magnitude in this way because temperature does not always steadily increase moving inland (figure S1). Instead, the temperature gradient can be nonmonotonic. To this end, we define 'coastalcooling intensity' as equal to 75% of the coast-toinland temperature difference (figure S1). Selecting a percentage higher (lower) than 75% results in a larger (smaller) value of coastal-cooling intensity, but does not affect the regions relative to one another according to our sensitivity analysis (figure S2). Averaging over many grid points and hot days (table 1) allows for statistically robust conclusions.

Populations and avoided exposure
We estimate coastal populations using the 1 km resolution Gridded Population of the World dataset Heavy colored lines mark sections of coastline incorporated into the PRISM regional averages, as described in section 2.3. (b)-(h) Differences between inland and coastal daily maximum temperature in PRISM on regional hot days (left, top 10%) and all other days (right, bottom 90%). Green squares indicate the median of differences between the three inland-coastal station pairs for each region.
(CIESIN 2016); our estimate of 50 million people within 60 km of the Atlantic and Gulf coasts aligns well with previous reports (Wilson and Fischetti 2010). To assess the benefits of coastal cooling, we estimate the avoided exposure to extreme heat due to coastal cooling. We calculate avoided exposure by assuming that T max at 60 km inland represents a counterfactual case for the coast-that is, what coastal temperatures would have been if not for coastal-cooling effectsand take its difference from the actual coastal temperature. We multiply this difference by the grid point population and define it as the avoided populationweighted extreme-heat exposure.
To assess the effect of model bias on projected extreme-heat exposure in coastal regions, we calculate two cases of population exposure to daily maximum temperatures >35°C. The first case is a best estimate where we bias-correct the WRF and LOCA projections with PRISM observations by adding the modelderived 21st-century changes to the PRISM climatology. In the second case, the model projections are not bias-corrected. For both cases, we decompose the future changes in population exposure into two components: the contribution from changes in mean temperatures, and the contribution from changes in the coast-to-inland temperature gradient. We do this by generating an additional coast-to-inland temperature profile with the historical gradient but the future mean temperature. The mean-change contribution is obtained by comparing the historical temperature profile with this artificial one, while the gradientchange contribution is estimated by comparing the artificial profile with the future profile. In all cases, we assume that populations are fixed in size and spatial distribution.

Characterization of observations
High-resolution, spatially complete PRISM data reveals that coastal daily-maximum temperatures are suppressed by several degrees Celsius relative to nearby inland areas (figure 1), in agreement with previous studies over more limited domains in the eastern US (Meir et al 2013, Melecio-Vázquez et al 2018. These coastal-cooling intensities differ considerably across regions-median values range from around 2°C for the Southeast US to near 3.5°C for Texas and 4°C for Northern New England (figure 1). Within each region, large temporal variability in coastal-cooling intensity exists; however, each region's interquartile range is only 1°C-2°C, allowing for the differences between the largest and smallest coastal-cooling-intensity regions to be statistically significant at the 95% level.
A positive land-sea temperature contrast exists in all regions during summer, and in both seasonal and regional terms, it is closely associated with the coastalcooling intensity (figure 2). This finding is consistent with the land-sea contrast being a known driver of ephemeral sea breezes (Lebassi-Habtezion et al 2011, Sequera et al 2015. The coastal-cooling effects we identify are consistently present on nearly all hot days (figure 1), however, underscoring their importance in reducing extreme-heat exposure.
Comparing the distributions of PRISM and GHCN-station-derived coastal-cooling intensities using quantile-quantile plots, we find PRISM exhibits biases for days with coastal cooling greater than 10°C and also for coastal warming ( figure S3). These appear to be related to PRISM's data interpolation, manifested particularly in a poor representation of days with fast-changing synoptic conditions and large coast-inland temperature differences. (We consider differences between PRISM and stations in more detail in the Discussion.) However, PRISM biases are less than about 1°C-1.5°C for coastal-cooling values in the 0°C-5°C range, which comprise the majority of hot days-ensuring the validity of our subsequent analyses and conclusions for these days.
Together, these results emphasize three important aspects of observed extreme heat in coastal areas. Firstly, for the majority of days, PRISM gridded climatology captures the cooling magnitudes from moretargeted station data, suggesting that PRISM is an appropriate observational basis against which to evaluate downscaled models ( figure 1); secondly, the coastal-cooling phenomenon is present on nearly all warm-season days and across regions (figure 1); and thirdly, the magnitude of observed warm-season coastal cooling is highly regionalized, in part because it is closely associated with the magnitude of regional land-sea contrasts (figures 1, 2).

Evaluation of downscaled GCMs
Compared to PRISM, the LOCA and WRF downscaled products ubiquitously underestimate observed coastal cooling: their typical mean cooling is 0.5°C-2°C, at least a factor of two (and up to a factor of 10) smaller than observed (figure 3). These model biases in the coast-to-inland temperature gradient are large, and consist of two primary types (figures S4, S5): a mean temperature bias at each grid point and a temperaturegradient bias concerning the difference between coastal and inland grid points (figure 3). Over the historical period, LOCA's mean coast-to-inland temperature gradient is no more than about 1°C (figure 3), suggesting that statistical downscaling does not sufficiently correct for the inability of the coarser parent model to represent the fine-scale coastal processes governing this gradient. Additionally, the LOCA methodology depends on using land-based stations to form analogs (Pierce et al 2014). As these stations are located on land, and few are near the coast, the LOCA reference points are tied to locations with little marine influence. This separation between the majority of LOCA stations and the coast makes the downscaling procedure less likely to capture the coastal-cooling effects we explore here despite its improved resolution.
Close to the coast, the dynamically-downscaled WRF ensemble performs somewhat better, reproducing the majority of the modest coastal-cooling effect for the Gulf Coast regions (figures 3(e)-(g)), but missing the large cooling magnitudes in other regions (figures 3(a)-(d)). The WRF ensemble is more skillful than LOCA for the Texas coast-to-inland temperature gradient, perhaps a function of WRF representing coastal atmospheric processes that are absent from the coarser global models that are the basis for the LOCA downscaling. Both products remain at a resolution where their ability to represent fine-scale coastal cooling is likely muted relative to observations (see Discussion).
The two ensembles point to a greater future increase in extreme heat at inland locations relative to the coast, amounting to a modest (∼0.5°C) strengthening of the coastal-cooling effect across all regions (figure 3). Even with such coastal-cooling increases, only the WRF future projections approach the magnitude of coastal cooling present in the historical observations, and only for the southernmost regions.

Population exposure
To understand how coastal cooling (and its model representation) influences future extreme-heat impacts assessments, we consider how models and observations vary in their estimated human exposure to extreme heat. Observationally, we find that, compared to the counterfactual case where temperatures in the coastal swath are identical to those 60 km inland, coastal cooling reduces present-day eastern-US population-weighted exposure to temperatures above 35°C by more than half for locations within 20 km of the coast (figure S6). The greatest reductions occur closest to the coast and for the highest temperatures-for example, 40°C almost never occurs anywhere along the immediate coast. Over the entire 60 km coastal swath and all regions considered, observed annual exposure to 35°C heat is about 200 million persondays, or 4 d per person, versus 7 d per person in the counterfactual case, consistent with other findings (Jones et al 2015).
If we bias-correct the LOCA and WRF ensembles with PRISM observations to provide a best estimate of future changes in coastal extreme heat, population exposure to 35°C extreme heat nearly triples by 2100 under a high-emissions scenario, to ∼650 million person-days ( figure 4(a)). The models also project an intensification of the coast-to-inland temperature gradient (figure 3), which could serve to partially counter the increased population exposure projected from mean temperature changes. As such, we decompose the change in population exposure to extreme heat into two components: one due to the mean warming across the 60 km coastal swath, and one due to the change in the temperature gradient within that swath. We find that the effects of mean warming dominate, with the model-projected increase in coastal cooling having a comparatively small effect ( figure 4(a)).
Due to their mean bias and relatively weak temperature gradients, models effectively systematically understate the magnitude of coastal cooling, thereby Figure 2. Correlation between mean land-sea contrast on regional hot days (ordinate) and mean corresponding coastal cooling (abscissa). Colors represent regions, for which the seasonal evolution of monthly averages from May to September is indicated by the arrows. As noted in the text, the correlation across all regions and months is 0.67, or 0.88 excluding Texas. overestimating coastal extreme heat (figure 3). We next investigate the characteristics of the bias correction that it is necessary to implement in order for models' future coastal extreme-heat risk estimates to be accurate. We recalculate population extreme-heat exposure, without bias-correcting the LOCA and WRF ensembles. We decompose the effects of the two different sources of model bias-the mean change (T bias) and the gradient bias (∇ bias)-on estimated population exposure ( figure 4(b)). This decomposition makes clear that, in both ensembles, a weak coastto-inland temperature gradient causes up to a 15% overestimate of late-21st-century population exposure to extreme heat within 16 km of the coast, though these effects become smaller further inland (solid bars in figure 4(b)). Mean-temperature biases (hatched bars) vary from moderately positive to strongly negative, the latter being about 50% of the projected changes. Members within each ensemble generally agree well, as indicated by the cross-model standard deviations.

Discussion and conclusions
The consistency of the coastal cooling that we observe across most warm-season days, in all regions (figure 1), leads us to propose that sea breezes are only part of a larger set of phenomena that determine the magnitude and inland extent of coastal cooling. Morning fog, midday clouds, a weak and ill-defined sea breeze, or a shallow coastal inversion layer are some of the alternative possibilities for propagation mechanisms, although investigating their contributions will require significant future research to determine how they may vary by region, season, or time of day. We find that latitudinal differences in mean cooling intensity are well-correlated with regional and seasonal land-sea temperature contrasts (r=0.67) (figure 2); as the land-sea contrast decreases over the summer, so does the associated cooling effect. This correlation rises to 0.88 when Texas is excluded, indicating that it exhibits a markedly different behavior, which we propose to explain by noting that the Texas coast experiences strong low-level onshore flow due to the North Atlantic subtropical high during summer (Liang et al 2006, Hu and Xue 2016). Prior results have also described deep inland infiltration of marine air under persistent onshore-flow conditions (Arritt 1993, Gilliam et al 2004, Misra et al 2011. As some of the only high-resolution multi-decadal simulations spanning the entire eastern US, the LOCA and WRF ensembles have a spatial comprehensiveness that enables estimation of the degree to which downscaled products overstate coastal extreme heat in both current and future climates. Our analysis suggests that this overstatement is considerable (several degrees Celsius), particularly for the large observed temperature gradients within 10-20 km of the coast in New England (figure 3), meaning that heat-exposure analyses using high-resolution downscaled ensembles nonetheless require observationally based bias correction to avoid major biases in heat exposure. For coarser-scale model output, bias correction is even more important; our results contextualize and evaluate this correction, which is implemented as a common remedy for their insufficient representation of coastal temperature gradients.
While a significant part of the downscaled products' bias is likely attributable to their coarser spatial resolution relative to PRISM, several pieces of evidence suggest the importance of additional factors. Firstly, the LOCA and WRF ensembles (6 km and ∼11 km resolution, respectively) are more similar to each other than to 4 km PRISM; secondly, their biases vary substantially between regions; and thirdly, the WRF ensemble generally exhibits less bias than LOCA despite its coarser spatial resolution. Such large and pervasive model biases are observed in other US coastal areas (Lebassi et al 2009, Wang andKotamarthi 2014), though often varying considerably by region.
Therefore, an essential question that this study raises concerns the origin of the model biases that we have identified for the coast-to-inland temperature gradients. The accuracy of these gradients is crucial for assessing the magnitude and location of present and future extreme heat exposure, and thus for climate risk management (Kunreuther et al 2013). Statistically downscaled products, such as the LOCA ensemble, are sensitive to GCM physics and weather-station boundary conditions, whereas dynamically downscaled products, such as the WRF ensemble, are sensitive to high-resolution model physics and GCM boundary conditions. Our investigation suggests that dynamical downscaling may be better suited to capture coastal cooling, given the importance of sub-grid-scale, subdaily dynamics. Nevertheless, statistical downscaling that incorporates sufficient marine observations could also strongly aid in reducing the persistent biases of dynamical models, whatever their resolution. Quantifying the biases associated with each strategy, and their variation according to location, season, and synopticscale meteorological pattern, is critical for making further progress in extreme-heat projections, especially because the choice of downscaling method is the single largest source of uncertainty for high-resolution climate projections (Li et al 2012, Xie et al 2015, Zhang and Soden 2019. This likely would encompass studying the parameterizations of sub-grid-scale processes in each GCM and in regional models such as WRF. However, the modest intermodel spread in coastal cooling (figures S7, S8) suggests that selecting parent models on the basis of skill in representing important physical processes-the strategy proposed by Maraun et al (2017)-presents only a partial solution, as it is the downscaling method itself that generates a significant portion of the uncertainty. What is most evident from our analysis is that careful, regionalized bias correction is essential prior to evaluating coastal moderation of extreme heat. Such analysis greatly improves confidence in model projections, and is consistent with the practice of retaining the signal associated with model products (Hall 2014).
It should not be overlooked that PRISM also has certain biases and uncertainties relative to station data, reported here as well as in the dataset description (Daly et al 2008). These biases derive from the incomplete spatial coverage of the station data which PRISM ingests, causing it to miss some coastal microclimates and weather systems, as well as possibly affecting the accuracy of its regional-scale coastal-proximity interpolation coefficients. Additional refinements in gridded observational products and interpolation methodologies, focusing on carefully evaluating coastlines and their environs, would aid in reducing these issues.
We find that historical coastal cooling reduces instances of daily maxima above 35°C by an average of nearly 60% within 30 km of the coast, and by 35% with 60 km of the coast, infiltrating far enough inland to affect major portions of the metropolitan areas of Houston, Tampa, Miami, New York City, and Boston. These reductions of 1°C-4°C are critical for human health, as mortality rises nonlinearly for daily-maximum temperatures above 35°C (Gosling et al 2007, Wu et al 2014. Applying temperature-mortality relationships from Petkova et al (2014) for New York City (assuming equivalent dose responses for all regions) yields a rough estimate that observed eastern-US coastal cooling reduces mortality by ∼20%, amounting to around 1000 fewer deaths per year for the historical total annual exposure of 200 million persondays. This calculation omits additional economic savings (such as reduced need for air conditioning). Such coastal-cooling-based reductions in population exposure are much larger than those obtained from downscaled simulations like those we evaluate here, suggesting an important shortcoming of current climate projections along coasts. Our geophysical results are likely invariant to changes in coastlines due to sealevel rise; however, these exposure numbers would almost certainly vary based on future spatial population redistributions, which we do not consider.
The projected intensification in future coastal cooling of ∼0.5°C (figure 3) gives the coastal-cooling effect a continuing importance in mitigating population exposure to extreme heat in a world that is rapidly warming (Jones et al 2018). It is also consistent with strong correlations between coastal cooling and landsea temperature contrast (figure 2) and with the expected circulation response to increases in warm-season land-sea temperature contrast (Dong et al 2009, Joshi et al 2008. However, such changes derive from a combination of offsetting atmospheric and oceanic responses (Kamae et al 2014), meaning that connecting them directly to coastal cooling would require targeted modeling experiments in follow-up work. Highresolution coupled atmosphere-ocean modeling, focusing on coastal atmospheric boundary layers, would be essential to verify and better understand regional differences in present and future coastal cooling. However, the dominant factor for extreme-heat projections along the coast, as elsewhere, is the 3°C-6°C average eastern-US summer warming by the late 21st century under RCP8.5 (figure 4(a)) (Lynch et al 2016, Vose et al 2017. Intermodel agreement about the sign and relative magnitude of these future changes, despite considerable differences over the historical period, is likely due to the downscaling method's preservation of the forced response to global-mean warming (Hall 2014).
We find here that future extreme heat will vary widely over distances that are too small for state-ofthe-art global models to properly simulate, even when downscaled. Understanding this variation is valuable in evaluating adaptation strategies and allocating resources for mitigation of impacts, particularly in polities that include both coastal and inland areas. Local fine-scale processes must therefore be considered carefully in order to ensure an accurate assessment of the present and future risks posed by extreme heat along the coastline of the eastern US.

Acknowledgments
Part of this work was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.

Data availability
The data that support the findings of this study are openly available: PRISM