Remotely sensed land skin temperature (LST) is increasingly being used to improve gridded interpolations of near-surface air temperature. The appeal of LST as a spatial predictor of air temperature rests in the fact that it is an observation available at spatial resolutions fine enough to capture topoclimatic and biophysical variations. However, it remains unclear if LST improves air temperature interpolations over what can already be obtained with simpler terrain-based predictor variables. Here, the relationship between LST and air temperature is evaluated across the conterminous United States (CONUS). It is found that there are significant differences in the ability of daytime and nighttime observations of LST to improve air temperature interpolations. Daytime LST mainly indicates finescale biophysical variation and is generally a poorer predictor of maximum air temperature than simple linear models based on elevation, longitude, and latitude. Moderate improvements to maximum air temperature interpolations are thus limited to specific mountainous areas in winter, to coastal areas, and to semiarid and arid regions where daytime LST likely captures variations in evaporative cooling and aridity. In contrast, nighttime LST represents important topoclimatic variation throughout the mountainous western CONUS and significantly improves nighttime minimum air temperature interpolations. In regions of more homogenous terrain, nighttime LST also captures biophysical patterns related to land cover. Both daytime and nighttime LST display large spatial and seasonal variability in their ability to improve air temperature interpolations beyond simpler approaches.
Near-surface air temperature is an important driver of various biotic and physical processes. Accurate spatial representations of air temperature are thus critical for distributed environmental modeling and spatial analyses in hydrology (e.g., Livneh et al. 2013), ecology (e.g., Hijmans et al. 2005; Dobrowski et al. 2009), epidemiology (e.g., Kloog et al. 2012), and agriculture (e.g., Kucharik and Serbin 2008). However, the generation of locally relevant gridded temperature fields at spatial resolutions less than 10 km remains a challenge. Long-term temperature observations are only found at limited single-station points (Willmott et al. 1991; Pielke et al. 2002) and larger-scale atmospheric outputs from global climate models and atmospheric reanalyses (e.g., Kalnay et al. 1996) fail to capture important local topoclimatic variations (Daly 2006). Subsequent inaccuracies in local temperature distributions can have significant impacts on environmental analyses, especially in regions of complex terrain (Minder et al. 2010).
Statistical interpolation of station-point observations is the most common approach for creating more locally relevant high-resolution gridded temperature fields (Thornton et al. 1997; Maurer et al. 2002; Hijmans et al. 2005; Daly et al. 2008; Livneh et al. 2013; Oyler et al. 2015). Temperature interpolation methods have included various forms of inverse distance weighting (e.g., Willmott et al. 1985), regression (e.g., Thornton et al. 1997; Daly et al. 2008), splines (e.g., Hijmans et al. 2005), and kriging (e.g., Kilibarda et al. 2014; Oyler et al. 2015). Elevation is frequently used as a spatial predictor variable because of its strong relationship with temperature (Daly 2006) and the global availability of digital elevation models (DEMs). Some interpolation approaches assume a static environmental lapse rate of −6.5°C km−1 (Maurer et al. 2002), whereas others use regression to estimate spatially and temporally varying linear lapse rates (Thornton et al. 1997; Daly et al. 2008).
It is well documented that local spatial patterns of temperature are a function not just of elevation but also of numerous other topoclimatic and biophysical factors including cold-air drainage potential, slope, aspect, water bodies, and land cover (Daly 2006; Dobrowski et al. 2009). As a result, interpolation methods that assume simple linear lapse rates may fail to capture key local variations in temperature (Daly et al. 2008; Minder et al. 2010; Oyler et al. 2015). For instance, cold dense air from higher elevations in complex terrain can drain into flat valley bottoms where it pools and leads to a thick inversion layer (e.g., Lundquist et al. 2008). This significantly complicates the interpolation of air temperature because of subsequent nonlinear variations with elevation. Because radiative cooling and relatively stable atmospheric conditions are required for temperature inversions to form, cold-air drainage most frequently affects spatial patterns of nighttime minimum temperatures, but strong inversions in winter can persist both day and night under stable synoptic conditions (Daly 2006). Techniques applied to account for cold-air drainage, inversions, and other topoclimatic factors include the calculation of lapse rates weighted toward a prediction point’s specific physiographic environment (Daly et al. 2008) and the use of DEM-based topographic indices as predictor variables (Holden et al. 2011). However, these techniques remain limited by the fact that topographic indices are only proxy variables, and DEM-based variables may not capture important time-varying land surface biophysical and topoclimatic variations (Pielke and Avissar 1990).
To improve upon DEM-based predictor variables, there has been a recent focus on the use of remotely sensed land skin temperature (LST) as a spatial predictor of air temperature (Vogt et al. 1997; Kawashima et al. 2000; Jones et al. 2004; Florio et al. 2004; Mostovoy et al. 2006; Vancutsem et al. 2010; Fu et al. 2011; Hengl et al. 2012; Lin et al. 2012; Kloog et al. 2012; Benali et al. 2012; Kilibarda et al. 2014; Oyler et al. 2015). Unlike surface temperature at shelter height, which is normally defined as air temperature at a height of 1.5–2.0 m, remotely sensed LST is the radiometric temperature of the ground or canopy surface (Jin and Dickinson 2010). LST spatial patterns are sensitive to land surface properties such as land cover, albedo, soil moisture, and surface roughness and their interaction with atmospheric conditions (Mostovoy et al. 2006; Jin and Dickinson 2010). LST represents a different physical parameter than air temperature, but the two variables are physically related and can be strongly correlated (Jin and Dickinson 2010).
One of the main approaches for incorporating LST into air temperature interpolations is to use linear regression or other similar statistical methods to model air temperature as a function of LST (Vogt et al. 1997; Kawashima et al. 2000; Jang et al. 2004; Florio et al. 2004; Fu et al. 2011; Hengl et al. 2012; Kloog et al. 2012; Benali et al. 2012; Kilibarda et al. 2014; Oyler et al. 2015; Zeng et al. 2015). To build the statistical models, remotely sensed daily or climatological LST observations are typically matched with in situ station-based air temperature observations. Variants of this approach range from single global linear regression models (Benali et al. 2012) to more complex methods such as mixed regression with daily varying slopes (Kloog et al. 2012), local moving-window regression kriging (Oyler et al. 2015), and spatiotemporal regression kriging (Kilibarda et al. 2014).
As a statistical predictor of air temperature, LST has two main limitations: 1) the period of record of LST datasets is often too short for long-term air temperature interpolations (Oyler et al. 2015); and 2) LST is only available under clear-sky conditions, which may result in a large number of missing values (Crosson et al. 2012). Additional methods must be used to account for missing values (Neteler 2010; Hengl et al. 2012; Crosson et al. 2012; Kloog et al. 2012; Kilibarda et al. 2014) and incorporate LST into long-term temperature interpolations (Oyler et al. 2015). Although LST has shown potential to improve air temperature interpolations in specific cases (Oyler et al. 2015), it remains unclear if LST provides improved predictive power beyond simpler DEM-based statistical models (Florio et al. 2004; Lin et al. 2012). In other words, is LST redundant to elevation and other DEM-based predictor variables, or does it capture additional topoclimatic and land surface biophysical spatial variations crucial for air temperature interpolation?
The objective of this analysis was to determine the ability of LST to improve spatial interpolations of climatological air temperature patterns across the conterminous United States (CONUS; Fig. 1) beyond that which can be obtained with basic X, Y, Z positional predictors (XYZ; longitude, latitude, and elevation). Spatial interpolations of climatological air temperature patterns are one of the most widely used types of climate data products (Daly et al. 2008) and form the foundation of spatiotemporal data products based on climatologically aided interpolation (Willmott and Robeson 1995; Daly et al. 2008; Vose et al. 2014; Oyler et al. 2015). We specifically focused on seasonal variability in the relationship between monthly climatological means of LST and air temperature and whether the relationship is strong but redundant with XYZ, unique, or even poor. For seasons and areas where the relationship was found to be unique, our second objective was to determine if the additional spatial variation captured by LST was mainly topoclimatic (e.g., cold-air drainage patterns, coastal influences) or biophysical (e.g., land cover, moisture status, albedo). Because of previously noted dissimilarities in the predictive power of nighttime and daytime LST (Vancutsem et al. 2010; Fu et al. 2011; Oyler et al. 2015), we also examined differences in the relationship between daytime LST and maximum air temperature (Tmax), and that of nighttime LST and minimum air temperature (Tmin).
2. Materials and methods
a. Input data
1) Remotely sensed LST
For observations of daytime and nighttime LST, we used the CONUS 30-arc-s-resolution LST climatology of Oyler et al. (2015). The Oyler et al. LST climatology is based on the Moderate Resolution Imaging Spectrometer (MODIS) Aqua satellite MYD11A2 8-day, 1-km product (Wan and Dozier 1996; Wan 2008) and contains 10-yr (2003–12) climatological LST means for each month. As compared with the ~1030/2230 LT overpass times of the MODIS Terra satellite, Aqua’s overpass times of ~0130/1330 LT more closely correspond to the diurnal timing of Tmax and Tmin in the CONUS (Crosson et al. 2012). The original MYD11A2 LST product uses a split-window algorithm to account for atmospheric attenuation of the thermal infrared signal and land-cover-specific emissivities to estimate surface emissivity (Wan and Dozier 1996; Snyder et al. 1998). To calculate monthly means, the Oyler et al. LST climatology uses the eight MYD11A2 8-day observations centered on a respective month where each 8-day MYD11A2 observation is the average of nonmissing daily clear-sky LST observations.
2) Remotely sensed biophysical factors
To determine the biophysical properties captured by LST, we also used remotely sensed observations of three related land surface biophysical factors: the normalized difference vegetation index (NDVI), land cover, and snow cover. In this analysis, NDVI and snow cover vary from month to month while land cover is a static variable. For observations of NDVI, we used the MODIS Terra MOD13A3 monthly composite 1-km NDVI product (Huete et al. 2002). As a proxy for vegetation density, higher NDVI values generally correspond to cooler daytime LST values because of evaporative cooling (Nemani and Running 1989; Sandholt et al. 2002) and the lower thermal storage capacity of leaves relative to the ground surface (Oke 1988; Prihodko and Goward 1997). We expected NDVI to be more related to LST spatial patterns during the growing season. For land cover, we used the 30-m National Land Cover Database (NLCD) 2011 product (Jin et al. 2013). NLCD 2011 is based on a decision-tree classification of 2011 Landsat imagery and contains 16 land-cover classes. Land cover has a direct influence on LST because of land-cover-based variations in surface roughness, thermal properties, latent and sensible heat fluxes, and albedo (Oke 1988). To quantify snow cover, we used the MODIS Terra MOD10A2 8-day 500-m snow-cover product, which is based on the normalized difference snow index (Hall et al. 2006). Snow mainly influences LST by increasing the albedo of the land surface and decreasing shortwave radiation absorption (Van De Kerchove et al. 2013). We expected snow cover to influence LST spatial patterns during late fall, winter, and early spring, especially in regions of complex terrain. For NDVI and snow cover, we calculated 10-yr 2003–12 climatological means to correspond with the LST climatology. In the case of NDVI, we calculated the 10-yr mean NDVI value for each month. For snow cover, we calculated the 10-yr mean number of days with snow cover within each month. Last, we resampled the resulting NDVI, land-cover, and snow-cover grids to match the 30-arc-s grid of the Oyler et al. LST climatology.
3) Weather stations
We used weather station Tmin and Tmax observations (Fig. 1a) primarily from the daily Global Historical Climatology Network (GHCN-D; Menne et al. 2012). To supplement GHCN-D with additional observations in the more remote and topographically complex areas of the western United States (Fig. 1b), we also incorporated stations from the Snowpack Telemetry network (SNOTEL) and from the Remote Automatic Weather Stations network (RAWS). We quality assured and homogenized all Tmin and Tmax observations using the quality assurance procedures of Durre et al. (2010) and the homogenization procedures of Menne and Williams (2009). For inclusion in the analysis, we required a station to have at least 5 yr of nonmissing daily observations in each month over the 10-yr 2003–12 time period of the LST climatology. This resulted in a total of 8018 input stations (Fig. 1a). For each station, we aggregated daily observations to monthly and calculated 10-yr monthly climatological means for Tmin and Tmax to correspond with those of the LST, NDVI, and snow-cover observations. To improve estimation of the 10-yr means for stations with missing values, we applied the Oyler et al. (2015) missing value infilling procedure. The Oyler et al. procedure uses neighboring stations and atmospheric reanalysis fields (Kalnay et al. 1996) to infill missing daily observations at an individual station.
For each station point, we extracted values for XYZ position, land cover, and the 10-yr monthly climatological means of LST, NDVI, and snow cover. We used each station’s provided longitude, latitude, and elevation for XYZ position. For continuous gridded variables, we extracted values to a station point using a bilinear interpolation of the four nearest 30-arc-s grid cells. For the categorical gridded land-cover variable, we extracted the value of the nearest grid cell.
b. Moving-window analyses
We used a spatial moving-window approach for our analyses of LST and air temperature. This allowed us to examine if the nature of the relationship between LST and air temperature is relatively static or heterogeneous across different regions of the CONUS. We performed the moving-window analyses on a regular 1°CONUS grid (Fig. 1d). Each local moving-window analysis centered at a specific grid point was based on LST and air temperature observations from the 100 closest stations to the grid point. We performed three main analyses: a classification of the relationship between LST and air temperature [section 2b(1)], a performance assessment of air temperature predictive models based on LST [section 2b(2)], and an analysis of the spatial variability of LST relative to air temperature [section 2b(3)]. We performed each analysis separately for daytime LST and Tmax, and nighttime LST and Tmin. We summarized results with respect to the nine CONUS climate regions (Fig. 1c) defined by the National Centers for Environmental Information (Karl and Koss 1984).
1) Classification of relationship between LST and air temperature
The objective of our first moving-window analysis was to classify the ability of LST to improve spatial interpolations of climatological air temperature. For each moving-window grid point and month, we classified the spatial relationship of LST and air temperature as either poor, redundant with XYZ, or unique. We considered the relationship between LST and air temperature as poor if the absolute value of the correlation (|r|) between the two variables was less than an optimized correlation threshold (|r|-threshold) or statistically insignificant (p > 0.05). We considered LST to be redundant if |r| was ≥ |r|-threshold and statistically significant, but the absolute value of the partial correlation (|pr|) (Kim 2012) between LST and air temperature when controlling for XYZ was < |r|-threshold or statistically insignificant. The partial correlation can be interpreted as the correlation between LST and air temperature once the effect of XYZ is removed from both variables via linear regression. If LST and air temperature have a high |r| but a low, insignificant |pr|, it implies that LST and air temperature are correlated simply because they are both linearly related to XYZ. In other words, LST provides little additional information that is not already captured by XYZ. In contrast, a high |pr| suggests that LST can account for spatial variability in air temperature that is not linearly associated with XYZ and thus provide unique explanatory power beyond XYZ. We classified LST as unique if the |r| and |pr| between LST and air temperature were both ≥ |r|-threshold and statistically significant. For both |r| and |pr| we set |r|-threshold to 0.45, a value that optimized the correspondence of poor, redundant, and unique classifications with associated improvements or nonimprovements in air temperature model performance when LST was used as a statistical predictor [see section 2b(2)].
For each gridcell window and month, we also examined the main factors driving LST’s redundancy or uniqueness. For a redundant classification, we subsequently modeled LST as a function of XYZ (longitude, latitude, elevation) using linear regression. We then calculated the relative importance of each variable in the regression and we considered the variable with the highest relative importance as the main factor with which LST was redundant. We quantified relative importance using the Lindeman, Miranda, Gold method (LMG; Lindeman et al. 1980) of the relaimpo package (Grömping 2006) within the R environment for statistical computing (R Core Team 2014). The LMG relative importance measure reflects both a predictor’s overall direct influence on the dependent variable and how much of that influence is unique in relation to other correlated predictors (Grömping 2006). Similarly, for a unique classification, we modeled XYZ-detrended LST as a function of NDVI, land cover, snow cover, and XYZ-detrended air temperature. If air temperature had the highest relative importance, we assumed the relationship between LST and temperature to be more a reflection of topoclimatic variation. On the other hand, if NDVI, land cover, or snow cover had the highest relative importance, we assumed the relationship between LST and air temperature to be more representative of biophysical variation. In reality, LST is the result of complex interactions between both meteorological and biophysical land surface properties, and it is difficult to fully isolate the contributions of different factors in controlling LST spatial patterns (Prihodko and Goward 1997; Jin and Dickinson 2010). Nonetheless, this classification (topoclimatic vs biophysical) provides a simple quantitative framework for understanding the dominant reasons why LST might improve air temperature interpolations in specific seasons or regions.
2) Performance assessment of air temperature models
We examined if any unique spatial information captured by LST actually translated into more accurate predictions of Tmin and Tmax by conducting separate performance assessments of linear regression models at each moving-window grid point. The classification of the LST and air temperature relationship described in section 2b(1) provides guidance of where, when, and why LST will likely improve air temperature interpolations, but it does not directly quantify if LST will result in significantly more accurate interpolations. We started with a base linear regression model of air temperature as a function of XYZ. We then calculated the percent change in mean absolute error (MAE) for two additional models: 1) air temperature as a function of just LST and 2) air temperature as function of XYZ and LST. Assessing changes in MAE for both models is important for deciding whether LST contains enough redundant and unique information to completely replace XYZ, or if LST should only be used in combination with XYZ. For the performance assessments, we used a standard 10-fold cross validation. We partitioned the 100 stations associated with each moving window into 10 random subsamples of size 10. We then conducted the following procedure for each subsample: 1) hold out subsample as a validation set, 2) fit the regression model using the remaining 90 stations as the training set, and 3) validate model using the subsample validation set. The final MAE for a model at a specific moving-window location was then the average taken over the 10 cross-validation subsamples.
We focused the cross-validation analysis on linear regression models because linear regression is one of the main approaches for incorporating LST into air temperature interpolations (Vogt et al. 1997; Kawashima et al. 2000; Florio et al. 2004; Fu et al. 2011; Hengl et al. 2012; Kloog et al. 2012; Benali et al. 2012; Kilibarda et al. 2014; Oyler et al. 2015; Zeng et al. 2015). However, we acknowledge that other statistical methods have been used to predict air temperature from LST. Most notable is the temperature–vegetation index (TVX), which makes use of the negative relationship between NDVI and daytime LST, and the expectation that the LST of a dense canopy will be similar to air temperature (Nemani and Running 1989; Goward et al. 1994; Prihodko and Goward 1997; Mildrexler et al. 2011). We did not analyze a TVX approach because the assumption of a strong negative relationship between NDVI and LST is likely not applicable for nighttime LST (Zakšek and Schroedter-Homscheidt 2009). Sharp topoclimatic gradients and land-cover variations can also confound the derivation of air temperature from the NDVI and daytime LST relationship, making TVX difficult to calibrate and apply across wide regions of varying terrain (Nemani et al. 1993; Czajkowski et al. 1997).
3) Spatial variability of LST versus air temperature
In our final moving-window analysis, we examined the spatial variability of LST relative to air temperature. Examining differences in spatial variability between the two variables is important for understanding the potential of LST-based statistical models to impose or overemphasize spatial patterns that might not accurately reflect spatial variability in air temperature even in cases where LST is a strong statistical predictor. To quantify spatial variability for each moving-window grid point and month, we calculated the spatial standard deviation of LST (σLST) and air temperature (σair). We divided σLST by σair to produce a normalized standard deviation measure, σLST/σair. A σLST/σair value > 1.0 indicates that LST is more spatially variable than air temperature, whereas a value < 1.0 indicates that LST is less spatially variable than air temperature. We compared σLST/σair values for daytime LST and Tmax and those for nighttime LST and Tmin across the CONUS to see if there were any consistent differences between the two pairs of variables.
Because LST cannot be observed when there is cloud cover and has subsequent potential to bias air temperature spatial patterns toward clear-sky conditions, we also examined if there was a relationship between σLST/σair and monthly climatological percent cloud cover. To obtain estimates of percent cloud cover, we used 0.5°-resolution hourly total cloud-cover data from the Climate Forecast System Reanalysis (CFSR; Saha et al. 2010). We first extracted 2003–12 hourly total cloud cover to each station point using a bilinear interpolation of the four nearest 0.5° CFSR cells. We then aggregated the hourly values to 2003–12 monthly climatological averages. Finally, to estimate 2003–12 climatological percent cloud cover for a specific moving-window grid point and month, we took a simple average of values from the 100 stations associated with the moving-window grid point.
a. Moving-window classification
1) Poor classification
A poor relationship between LST and air temperature was most notable for daytime LST and Tmax in late spring, summer, and early fall across parts of the Ohio Valley, South, and Southeast (Figs. 2 and 3). This region of poor correlation accounted for nearly 30% of the CONUS land area from April through September (Fig. 4b). In contrast, no more than 6% of the CONUS had a poor relationship between nighttime LST and Tmin across all months (Fig. 4a). A poor relationship between nighttime LST and Tmin was limited to isolated areas in eastern Texas and the southern and central Appalachians (Fig. 3).
2) Redundant classification
Daytime and nighttime LST displayed significant differences in redundancy with XYZ (Figs. 2–4). Nighttime LST was redundant with latitude in many eastern climate regions in winter and early spring (Fig. 3). This region of redundancy represented 36%–46% of the CONUS land area from January through March (Fig. 4a). In contrast, daytime LST was redundant in many areas of the CONUS across all months (Figs. 2 and 4b). The portion of the CONUS in which daytime LST was redundant ranged from 49% in June to 84% in November (Fig. 4b). In winter, across the eastern climate regions, daytime LST was redundant with latitude (Fig. 2a). In areas of more complex terrain within the western climate regions and along the Appalachian Mountains in the eastern CONUS, daytime LST was most redundant with elevation (Fig. 2).
3) Unique classification
Daytime and nighttime LST were distinctly different with respect to both their overall ability to capture unique variation beyond XYZ and the main factors driving their uniqueness (Figs. 2–4). Nighttime LST uniqueness was more spatially and temporally consistent in western climate regions (Fig. 3) with the portion of the entire CONUS classified as unique ranging from 51% in March to nearly 92% in August (Fig. 4a). In contrast, for daytime LST, less than 25% of the CONUS was classified as unique across all months (Fig. 4b). The uniqueness of nighttime LST appeared to be mainly driven by its ability to capture important variations in topoclimate and land cover (Fig. 3). Nighttime LST topoclimatic uniqueness was dominant in the Florida Peninsula and the topographically complex western climate regions across all months, around the Great Lakes from summer through winter, and along the Appalachian Mountains up through the Northeast in summer and fall (Fig. 3). Nighttime LST uniqueness was more related to land cover in the South, northern plains, Ohio Valley, and Southeast in the spring, summer, and fall (Fig. 3). Unlike nighttime LST, daytime LST uniqueness was dominated more by biophysical factors as opposed to topoclimatic variation (Fig. 2). In the western climate regions, snow cover controlled daytime LST uniqueness during winter, whereas in summer NDVI was a more dominant factor (Fig. 2).
b. Moving-window performance assessment
1) Air temperature models with only LST
Across most months and climate regions, linear regression models that used daytime LST as a sole predictor of Tmax performed significantly worse than those based on XYZ (Figs. 5a and 6a). The overall CONUS MAE for the XYZ models was 0.67°C, while the overall MAE for the LST models was 0.99°C, a 47% increase in error. The most notable exception to this was centered on the northern Rockies and plains in winter (Fig. 5a). In parts of Wyoming, the MAE for daytime LST models of winter Tmax was 25%–33% less than the MAE for models based on XYZ (Fig. 5a). The MAE of daytime LST-based models was also 10%–19% lower than the MAE of XYZ models along the Pacific coast in summer (Fig. 5a).
In contrast to daytime LST, nighttime LST models of Tmin had overall better performance than XYZ models, but with significant regional variability (Figs. 5c and 6c). The overall MAE for nighttime LST models of Tmin (0.94°C) was 14% less than the overall MAE for models based on XYZ (1.11°C). However, consistent and significant improvements in MAE were mainly confined to the western climate regions and the Florida Peninsula (Figs. 5c and 6c). In the winter and spring, a majority of eastern climate regions displayed increases in MAE (Figs. 5c and 6c).
2) Air temperature models with LST and XYZ
Larger decreases in cross-validated MAE were observed for Tmin than for Tmax when LST was added as a predictor to models already using XYZ (Figs. 5b,d and 6b,d). Overall Tmin MAE decreased from 1.11° to 0.87°C, a 21% decrease, while overall Tmax MAE decreased from 0.67 to 0.63, a decrease of 6%. Spatial patterns of MAE reductions for both Tmin (Fig. 5d) and Tmax (Fig. 5b) aligned with the uniqueness patterns of nighttime (Fig. 3) and daytime LST (Fig. 2). Significant decreases in Tmin MAE were consistent across all climate regions in summer (Figs. 5d and 6d). Large percent decreases of 33% in summer Tmin MAE were evident in the Southwest and West climate regions (Fig. 6d). However, in winter significant decreases in Tmin MAE were mainly confined to the western CONUS and the Florida Peninsula (Fig. 5d) because of nighttime LST’s redundancy with latitude (Fig. 3a). On average, winter Tmin MAE decreased by 24% in western climate regions and 9% in eastern climate regions (Fig. 6d). In comparison with Tmin, significant decreases in Tmax MAE were more isolated and, except for areas around the Great Lakes, were generally absent from the eastern CONUS (Figs. 5b and 6b). Similar to the daytime LST-only models, a notable decrease in Tmax MAE occurred in winter in the northern Rockies and plains where January Tmax MAE decreased by 17% (Figs. 5b and 6b). Significant percent decreases of 7%–30% in Tmax MAE were also found during summer in areas within the western climate regions (Figs. 5b).
c. Moving-window spatial variability
Nighttime LST and Tmin displayed similar spatial variability, whereas daytime LST was much more spatially variable than Tmax, especially during summer months (Fig. 7). Across all months and climate regions, average σLST/σair was 1.00 for nighttime LST and 1.79 for daytime LST. An annual cycle in daytime LST σLST/σair was discernable across all climate regions, but was most distinctive in climate regions with less complex terrain (Fig. 7b). In the Southeast, daytime LST σLST/σair reached a high of 2.82 in August and a low of 1.08 in December. In comparison with daytime LST, nighttime LST σLST/σair seasonality was relatively flat (Fig. 7a). However, nighttime LST σLST/σair had a statistically significant relationship with cloud cover (r = +0.53; p < 0.01; Fig. 8a) while the relationship of daytime LST σLST/σair and cloud cover was weak and borderline significant (r = −0.18; p = 0.07; Fig. 8b). Nighttime LST σLST/σair tended to values > 1.0 in seasons and climate regions with more cloud cover and to values < 1.0 in seasons and climate regions with less cloud cover (Fig. 8a). In September in the West climate region, average nighttime LST σLST/σair and percent cloud cover both reached overall low values of 0.76 and 19%, respectively. Average daytime LST σLST/σair was >1.0 across all months (Figs. 7b and 8b).
a. Daytime LST
The moving-window analyses suggest that daytime LST is generally a more redundant and poorer statistical predictor of air temperature than nighttime LST. This contrast can likely be attributed to the type and scale of the physical processes that operate on daytime versus nighttime LST. Daytime LST is strongly influenced both by overall exposure to incoming solar radiation and how biophysical properties (e.g., land cover, albedo, moisture, roughness) control the resulting temperature response of the surface (Roth et al. 1989; Nichol 2005; Mildrexler et al. 2011). Under the same atmospheric conditions, the LST of a shaded surface will be much less than an unshaded surface (Wan and Dozier 1996). Temporally, on a cloudy day, air temperature and LST will be more coupled, yet significantly different under the clear-sky conditions that are required for remotely sensed LST to be obtained (Jin et al. 1997; Gallo et al. 2011). On a clear and calm summer day, a dry unvegetated surface can be nearly 30°C greater than 2-m air temperature at midday (Vehrencamp 1953; Oke 1988), whereas the canopy LST of a dense forest will normally be closer to ambient conditions (Prihodko and Goward 1997; Mildrexler et al. 2011). Because of the strong control exerted by underlying land surface properties, daytime LST has a high degree of microscale variability (Nichol 2005). In contrast, Tmax is measured 1.5–2 m above the surface in a ventilated box shielded from direct radiation and is therefore less coupled to underlying daytime land surface properties. As a result, daytime LST is much more spatially variable than air temperature, especially during the higher solar radiation loads of late spring and summer (Fig. 7b). Unless there is a larger-scale climatic gradient (e.g., elevation) that is strong enough to emerge beyond daytime LST’s microscale variability, air temperature will likely not be well correlated with daytime LST. For instance, in the Southeast, where there was poor correlation between daytime LST and Tmax in the summer (Fig. 2c), the daytime LST spatial standard deviation calculated for station locations in August was 182% greater than the standard deviation of station-observed Tmax (σLST/σair = 2.82).
The inconsistent ability of daytime LST to improve Tmax interpolations can also likely be attributed to XYZ already accounting for most of the spatial variability in Tmax. Under well-mixed atmospheric conditions, elevation and Tmax will have a strong linear correlation (Daly 2006). In contrast, the linear relationship between elevation and Tmin tends to be not as strong because of the inversions and cold-air drainage patterns associated with stable and stratified nighttime atmospheric conditions (Daly 2006). Subsequently, relative to Tmin, there is generally less potential for Tmax interpolations to be improved beyond what can be obtained with just XYZ models (Oyler et al. 2015).
Although daytime LST tends to be a redundant and poorer statistical predictor of Tmax relative to XYZ, it is worth investigating the limited environmental settings in which the addition of daytime LST to XYZ did improve model performance. Moderate model improvements were most apparent in the semiarid and arid mountainous regions of the western CONUS (Fig. 5b). In these regions, the unique portion of daytime LST associated with snow cover in winter and NDVI in summer (Fig. 2) suggests that daytime LST can capture patterns of albedo, aridity, and evaporative cooling that have an influence on Tmax independent of elevation. Additionally, even given the more biophysical nature of daytime LST, it appears to be able to capture topoclimatic phenomena unique from XYZ under regionally specific conditions. For instance, the significant improvement in Tmax MAE in winter in the northern Rockies and plains (Fig. 5b) is partially the result of daytime LST capturing a Tmax inversion pattern. Daytime LST maintains a linear relationship with Tmax (Fig. 9a), whereas the relationship between elevation and LST is not well represented by a linear model (Fig. 9b). Improvements to Tmax MAE around the Great Lakes and along the California Pacific coast (Fig. 5b) are likely the result of daytime LST capturing coastal influences. Nonetheless, the isolated improvement provided by daytime LST in these regions is seasonally specific and disappears in seasons where the relationship between elevation and Tmax becomes more linear (Figs. 9c,d) or coastal influences are diminished (Fig. 5b).
b. Nighttime LST
The overall greater model improvement afforded by nighttime LST can likely be attributed to the stronger coupling between nighttime LST and Tmin. Unlike daytime LST, without direct solar radiation, nighttime LST spatial variability is more influenced by local and mesoscale atmospheric processes important for air temperature (Nichol 2005). Moreover, because the atmospheric boundary layer typically displays increased stability and reduced advection at night, near-surface air temperature is more tied to the biophysical properties of the underlying land surface (Voogt and Oke 2003; Nichol 2005; Pielke et al. 2007). In direct contrast to daytime LST and Tmax, nighttime LST and Tmin thus maintain similar spatial variability throughout the annual seasonal cycle (Fig. 7a).
The stronger coupling between nighttime LST and Tmin results in improved model performance across regions of varying climate and topography but appears to be most beneficial in areas of complex terrain (Fig. 5d). Because of the topoclimatic uniqueness of nighttime LST across the mountains of the semiarid and arid western CONUS (Fig. 3), nighttime LST is likely capturing local inversion and cold-air drainage patterns that cannot be represented with a linear XYZ model (Daly et al. 2008; Lundquist et al. 2008; Holden et al. 2011). For example, the Big Hole Valley in southwestern Montana (45.60°N, 113.51°W) is a broad, enclosed high-elevation valley that frequently experiences cold-air drainage (Fig. 10a). An XYZ linear model of August Tmin for the Big Hole region fails to capture the valley inversion (Fig. 10b) and produces a Tmin value +5.97°C warmer that what is observed in the valley (Wisdom, Montana; 45.62°N, 113.45°W, 1847 m). In contrast, a linear model that uses both XYZ and nighttime LST produces a more definitive inversion pattern with a Tmin value +0.68°C warmer than observed (Fig. 10c).
In areas of homogenous topography where the benefit of Tmin was not as great, the more land-cover-based uniqueness of nighttime LST suggests that nighttime LST can nonetheless represent important biophysical variation. For instance, a model using both XYZ and nighttime LST clearly displays a summer urban heat island in the city of Indianapolis, Indiana (39.77°N, 86.15°W), while a XYZ model does not (Fig. 11). The significant decrease in Tmin MAE within the Florida Peninsula (Fig. 5d) also suggests that LST can capture coastal influences that are nonlinear with XYZ.
Despite the benefits of nighttime LST across all seasons (Figs. 5d and 6d), the positive relationship between cloud cover and nighttime LST σLST/σair (Fig. 8a) indicates that there is potential for nighttime LST to overemphasize certain spatial patterns. Under cloudy and unstable atmospheric conditions, Tmin spatial patterns that are a function of inversions, cold-air drainage, and underlying land surface properties are typically less variable and distinctive. As suggested by the higher nighttime LST σLST/σair values in cloudier regions and seasons, nighttime LST likely cannot capture this decrease in Tmin spatial variability. In contrast, the nighttime LST σLST/σair values < 1.0 during less cloudy seasons in the arid and semiarid western climate regions are likely an issue of scale; the 1-km LST resolution smooths out clear-sky spatial variability inherent in the Tmin station-point observations. Even with the positive relationship between cloud cover and nighttime LST σLST/σair (Fig. 8a), nighttime LST and Tmin still maintain much more similar spatial variability across all regions and seasons than daytime LST and Tmax (Fig. 7).
On the basis of moving-window analyses of the relationship between climatological LST and air temperature, we make the following conclusions: 1) Daytime LST has a strong biophysical influence. As a result, it is more spatially variable than Tmax (Fig. 7b) and a generally poorer statistical predictor of Tmax than XYZ positional variables (Figs. 5a and 6a). Nonetheless, for specific settings daytime LST does appear to be able to capture summer contrasts in aridity and evaporative cooling that influence air temperature in mountainous semiarid and arid regions (Fig. 2). In addition, in regions and seasons where the linear relationship between XYZ and Tmax breaks down (e.g., summer along the Pacific coast, winter in the northern Rockies; Fig. 9b), daytime LST can maintain a linear relationship with Tmax (Fig. 9a) and improve model performance (Fig. 5b). 2) Nighttime LST appears to have a large topoclimatic component (Fig. 3), especially in mountainous semiarid and arid regions where cold-air drainage and inversion patterns are common (Daly et al. 2010; Lundquist et al. 2008; Holden et al. 2011). As such, it has a strong and consistent ability to improve Tmin interpolations in these areas (Fig. 5d) despite a potential to overemphasize clear-sky spatial patterns in more cloudy seasons and regions (Fig. 8a). Within regions of more homogenous terrain, nighttime LST also appears to be able to capture important Tmin variations related to land cover (Fig. 4). This is most evident during the more stable atmospheric conditions of summer (Figs. 5d and 11) when thermal and radiative land surface properties are likely more significant drivers of Tmin spatial patterns and nighttime LST is not dominated by the large-scale latitudinal gradients associated with winter (Fig. 3).
In the context of these conclusions, we note several caveats and limitations. We only focused on the spatial relationship between 10-yr monthly climatological values of LST and air temperature. Relationships between LST and air temperature could vary for different time scales. For instance, if cloud cover is minimal, daily LST might be less redundant with latitudinal gradients in winter because of the more variable position of fronts and air masses on a daily time step. Additionally, the XYZ model used in this analysis provided an important baseline for which to compare the LST-based air temperature models, but it is not indicative of the performance of more advanced DEM-based air temperature interpolation models (e.g., PRISM; Daly et al. 2008). Nonetheless, the analysis provides a good indication of the general strengths and weaknesses of LST throughout the CONUS.
Despite the recent widespread use of LST in statistical models of air temperature, this analysis shows that basic XYZ positional variables can significantly outperform an individual LST predictor (Figs. 5 and 6). Therefore, when developing spatial LST-based models of air temperature, spatially aware interpolation methods (e.g., inverse distance weighting, kriging) and simpler predictor variables should be evaluated in combination with LST. In the end, regional strengths and inadequacies of LST need to be carefully assessed to make sure LST actually provides more accurate spatial representations of air temperature.
This study was based on work supported by the National Science Foundation under EPSCoR Grant EPS-1101342 and the U.S. Geological Survey North Central Climate Science Center Grant G-0734-2. Additional support was provided by a NASA Applied Wildland Fire Applications award (Agreement NNH11ZDA001N-FIRES). Support for SZD was provided by NSF (DEB; 1145985). Any opinions, findings, and conclusions or recommendations expressed in this article are those of the authors and do not necessarily reflect the views of the National Science Foundation.
Current affiliation: Earth and Environmental Systems Institute, The Pennsylvania State University, University Park, Pennsylvania.