Daily maps of the Great Lakes surface temperatures, with a horizontal resolution of ∼ 1.3 km, were obtained from the Great Lakes Surface Environmental Analysis (GLSEA) produced at the NOAA Great Lakes Environmental Research Laboratory (Schwab et al., 1992). The lake surface temperatures are estimated using clear-sky infrared brightness temperatures from the Advanced Very High Resolution Radiometer on board multiple polar-orbiting satellites. If a surface retrieval is not possible on a given day due to cloud cover, then a smoothing algorithm is applied to the previous analysis to maintain complete coverage. Only satellite observations are used to produce the daily lake surface temperature analyses, which have been shown by Schwab et al. (1992) to have a small bias and root mean square error (RMSE; 1–1.5 ∘C) when compared to buoys. The daily GLSEA analyses were used to overwrite the simulated surface temperatures for Great Lakes grid points at 00:00 UTC on each day in the YNT_SST, YNT_SSN, and YNT_SSNG simulations using the WPS. Replacing the coarse-resolution (0.25∘) GFS-FNL surface temperatures (Fig. 2a) with the GLSEA analyses (Fig. 2b) led to warmer lake temperatures near the shoreline, especially along the northern parts of Lake Michigan where temperatures were > 2 K warmer, and cooler temperatures across the rest of the lake, when averaged over the 22 May–22 June 2017 time period (Fig. 2c). This spatial pattern indicates that the finer horizontal resolution of the GLSEA dataset allows it to capture warmer temperatures in shallower waters near the shoreline while also depicting the cooler mid-lake temperatures due to the cooler-than-normal weather conditions that prevailed across the region in May (NCEI, 2017).

GVF is the photosynthetically active fractional green vegetation cover within a grid cell, with higher values indicating more extensive actively transpiring vegetation. It is a key parameter in an LSM because vegetation representation is used to partition the incoming solar radiation into sensible, latent, and ground heat fluxes, where the latent heat flux is largely due to vegetation transpiration (e.g., Yin et al., 2016). Surface latent heat flux is sensitive to GVF because vegetation roots are able to access deeper soil moisture that would not otherwise be able to evaporate (Miller et al., 2006). For this study, we used daily global GVF derived using observations from the Visible Infrared Imaging Radiometer Suite (VIIRS; Vargas et al., 2015) instead of the default monthly climatology to constrain the evolution of vegetation in the YNT_GVF and YNT_SSNG simulations. The VIIRS GVF composite product is generated daily at 4 km resolution and available from the NOAA Comprehensive Large Array-data Stewardship System (CLASS). Ding and Zhu (2018) have shown that the VIIRS GVF product has smaller errors and bias than other satellite-derived GVF datasets because of reduced atmospheric influences, improved observing capabilities in high biomass regions, better representation of vegetation canopies, and reduced bidirectional reflection distribution function effects. The real-time daily GVF analyses were used to overwrite the default monthly climatological vegetation fraction data used by the WRF model at 00:00 UTC on each day. Using real-time, satellite-derived GVF instead of a monthly GVF climatology has been shown to improve the representation of the surface energy budget and subsequent model forecasts during the warm season (Case et al., 2014). In Fig. 2f, it is evident that use of the real-time GVF led to lower leaf area index (Fig. 2e; computed internally by the WRF model) across most of the domain compared to the climatological vegetation data (Fig. 2d), with the exception of some forested regions in the northern portion of the domain and bands of enhanced leaf area index surrounding metropolitan areas such as Chicago. The lower leaf area index in agricultural areas is consistent with delayed crop growth due to the cool spring weather, whereas the bands of higher leaf area index represent the impact of urban sprawl, since the climatological vegetation data shown in Fig. 2d were generated using satellite observations from the late 1980s and early 1990s (see Gutman et al., 1995).

A customized version of the Land Information System (LIS; Kumar et al., 2006) run at the Short-term Prediction Research and Transition Center (SPoRT) was used to generate high-resolution soil moisture and soil temperature analyses. Version 3.6 of the Noah LSM (Chen and Dudhia, 2001) was run on a 1 km resolution domain covering the central and eastern United States and nearby portions of southern Canada. Required inputs to run the Noah LSM were obtained from hourly analyses of surface pressure, 2 m temperature, 2 m specific humidity, 10 m wind speed, and downwelling shortwave and longwave radiation from the North American Land Data Assimilation System Phase 2 (NLDAS-2; Xia et al., 2012). No observations were assimilated during the LIS runs. Quantitative precipitation estimates (QPEs) were obtained from the Multi-Radar/Multi-Sensor System (MRMS) gauge-adjusted radar product (Zhang et al., 2016), the Global Data Assimilation System (GDAS; Wang et al., 2013), and NLDAS-2. A simple blending methodology was used to incorporate the multiple sources of QPEs because an evaluation of the real-time SPoRT-LIS product (Case 2016; Case and Zavodsky 2018; Blankenship et al., 2018) and preliminary LIS experiments during this study revealed that the NLDAS-2 and MRMS precipitation products have a dry bias across the region. To reduce this bias, the precipitation forcing used the average of the highest two values of the MRMS, GDAS, and NLDAS-2 QPE datasets. Inspection of the blended precipitation product showed that the precipitation bias was reduced, while preserving small-scale spatial details in the MRMS QPE product. Daily VIIRS GVF composites were also used to constrain vegetation during the offline LIS-Noah simulation.

Following an initial spin-up of LIS using NLDAS-2 forcing data from 2012–2016 to remove the memory of the prescribed initial conditions, the final analysis from this run was used to restart the simulation on 1 January 2012 using NLDAS-2 atmospheric forcing data, VIIRS GVF, and the merged QPE product. Soil moisture and soil temperature analyses from this LIS simulation were then used to replace the corresponding variables in the YNT_SOIL, YNT_SSN, and YNT_SSNG simulations at 00:00 UTC on each day from 12 May–22 June 2017 using the WPS. Comparison of the 0–10 cm soil temperatures from the GFS (Fig. 2g) and LIS (Fig. 2h), averaged over the 22 May–22 June 2017 period, shows that the topsoil temperatures are noticeably cooler in the LIS data across most of the region, except for the northern parts of Wisconsin and Michigan. The cooler temperatures are most prominent in suburban regions, where the largest increases in GVF also occurred (Fig. 2f). For 0–10 cm soil moisture, the LIS analyses are generally wetter across the domain (Fig. 2l), with the largest increases across the forested regions of Wisconsin and Michigan. Deeper soil layers exhibited similar differences between the GFS-FNL and LIS datasets (not shown).

To examine regional differences in model performance, Fig. 4 shows the 2 m temperature bias and RMSE computed separately for each station using hourly observations from 22 May–22 June 2017. For the AP-XM simulation, there is a north–south gradient in the RMSE, with the largest errors across northern Illinois and Indiana (Fig. 4a). Stations near Lake Michigan have the smallest RMSE due to its moderating influence on local weather conditions. Similar to the RMSE, the smallest biases occurred in the northern part of the domain and along the eastern shoreline; however, biases along the western shoreline are larger and of comparable magnitude to those at inland locations across Wisconsin and Illinois. Overall, the AP-XM simulation had an RMSE of 3.03 K and a bias of -0.14 K when averaged across all stations (Fig. 3a–b). Switching to the YNT parameterization suite greatly reduced the RMSE by 25.83 % across the entire domain (Fig. 3b); however, the bias increased to 0.55 K (Fig. 3a). The largest RMSE reductions (up to 45 %) occurred in rural areas of northern Illinois, with similar RMSEs found across the entire domain (Fig. 4b). The larger positive temperature bias in the YNT baseline simulation is primarily due to larger errors in Wisconsin and within densely populated urban areas along the western Lake Michigan shoreline from Chicago to Milwaukee (Fig. 4f). A mixed pattern of larger and smaller biases occurred elsewhere across the domain.

Inspection of the YNT sensitivity experiments shows that the smallest RMSEs occurred during the YNT_SOIL, YNT_SSN, and YNT_SSNG simulations, with the average RMSE reduced by 30.32 % to 32.5 % relative to the EPA baseline (Fig. 3b) and from 6.0 % to 9.0 % relative to the already greatly improved YNT baseline (Fig. 3c). On an individual basis, the high-resolution soil dataset (YNT_SOIL) had the largest positive impact at most stations (Fig. 4d), whereas slightly larger RMSEs were observed when using nudging (YNT_N2KM) (Fig. 4j). Comparison of the YNT_SSN and YNT_SSNG simulations (Fig. 4l, p) shows that inclusion of the VIIRS GVF dataset during the YNT_SSNG simulation led to slightly larger RMSE for stations near the lakeshore but similar or smaller errors for stations located further inland.

The bias pattern for the YNT simulations is more complex. Overall, the bias was largest (0.67 K) in the YNT_N2KM simulation, with the smallest biases occurring in the YNT_GVF (-0.03 K) and YNT_SSN (-0.09 K) simulations (Fig. 3a). Switching from the AP-XM to YNT baseline configurations led to larger biases across most of the domain, especially along the southwestern shoreline of Lake Michigan (Fig. 4e–f). The high-resolution SST dataset had a minimal impact on the biases (Fig. 4g), whereas they were smaller in the YNT_SOIL (Fig. 4h) and YNT_GVF (Fig. 4m) simulations relative to the YNT baseline. The use of these two land datasets, however, led to much larger negative biases along the eastern shoreline of Lake Michigan. When 2 km analysis nudging was used (YNT_N2KM), larger positive biases occurred from Chicago to Milwaukee, with smaller biases along the eastern shoreline (Fig. 4n). The increased RMSE and bias near the western shoreline compared to locations further inland during the YNT_N2KM simulation suggests that the modified nudging routine (applied to heights above 2 km instead of above the PBL) may not work well for areas near Lake Michigan due to the moderating influence of the lake on the PBL. Because the PBL tends to be more stable and shallower for locations over and near Lake Michigan due to the cooler surface temperatures, this means that confining analysis nudging to above 2 km limits its ability to constrain the evolution of the lower troposphere during the YNT_N2KM simulation. This behavior could also be due to deficiencies in the YNT configuration over complex urban–lake transition zones.

For the 2 m water vapor mixing ratio, switching to the YNT physics suite led to a reduction of nearly 15 % in the station-averaged RMSE during the YNT simulation relative to the AP-XM baseline (Fig. 3e), with additional incremental reductions occurring in all sensitivity experiments, except for YNT_N2KM (Fig. 3f). The lower RMSE in all of the YNT simulations is primarily due to the large reduction in bias (Fig. 3d). Whereas the AP-XM configuration had a large moist bias (0.60 g kg-1), the YNT bias was much smaller and also became negative (-0.20 g kg-1). The bias was further reduced during most of the sensitivity experiments, with only a slight increase during the YNT_SSNG simulation. Overall, the YNT_SSN simulation had the smallest RMSE and a bias close to zero when averaged across all of the stations.

Looking more closely at the individual stations (Fig. 5), it is evident that most of them have a positive (e.g., moist) bias when the AP-XM configuration is used (Fig. 5e). The largest biases are located in the southern portion of the domain, especially for stations near the lakeshore. In contrast, about two-thirds of the stations exhibit a negative bias during the YNT simulation (Fig. 5f). The spatial pattern of the biases is similar during all of the YNT sensitivity experiments; however, their magnitudes are generally smaller, which is consistent with the overall statistics (Fig. 3d). For RMSE, the largest errors in the AP-XM simulation occur primarily along the southern end of Lake Michigan, with generally smaller errors in the northern half of the domain (Fig. 5a). The RMSE during the YNT simulation is smaller at most locations, especially along the shoreline, though a few stations near the western shoreline have larger errors (Fig. 5b). The use of the SOIL and GVF datasets reduced the errors at these nearshore locations (Fig. 5d, i), with the smallest errors at most stations occurring during the combination experiments (YNT_SSN and YNT_SSNG). As was the case with 2 m temperature, the most accurate 2 m water vapor analyses were obtained during the YNT_SSN simulation.

Compared to the temperature and water vapor fields, changes to the 10 m wind speed statistics were much more modest during the YNT simulations. Switching from the AP-XM configuration to the YNT configuration led to a 3.26 % reduction in the RMSE but a larger bias that also changed sign from negative to positive (Fig. 3g). For the YNT experiments, the average RMSE was slightly smaller during the YNT_SOIL and YNT_N2KM simulations (-1.21 % and -1.78 %, respectively) but slightly larger (0.95 %) during the YNT_SST simulation compared to the YNT baseline (Fig. 3i). The use of the GVF surface dataset led to a 7.64 % increase in the RMSE during the YNT_GVF simulation, primarily due to a larger wind speed bias. Overall, the most accurate wind speed analyses were achieved during the YNT_SSN simulation, with an RMSE reduction of 6.47 % across all stations.

Spatially, there is a latitudinal gradient in wind speed errors during the AP-XM simulation. The largest RMSEs are located across the southern part of the domain (Fig. 6a), with mostly negative wind speed biases (up to 2 m s-1) in the same region transitioning to a mix of negative and positive biases in northern Wisconsin and Michigan (Fig. 6e). The RMSE and bias were much smaller for stations around the southern shoreline of Lake Michigan during the YNT simulation; however, slightly larger RMSEs are present across inland locations in the northern part of the domain (Fig. 6b). A similar spatial pattern of changes relative to the AP-XM baseline occurred during the YNT sensitivity experiments, though the errors are generally larger during the YNT_GVF simulation (Fig. 6i, m) and smaller during the YNT_SOIL (Fig. 6d, h) and YNT_N2KM (Fig. 6j, n) simulations. The poor performance of the YNT_GVF and YNT_SSNG simulations is primarily due to larger errors across inland areas of Wisconsin, where there are large positive wind speed biases (Fig. 6m, p), with similar errors elsewhere in the domain.

Figure 7 shows the diurnal evolution of RMSE and bias for 2 m temperature, 2 m water vapor mixing ratio, and 10 m wind speed at hourly intervals starting at 19:00 local standard time (LST). The time series were computed by averaging over data from all stations on the 1.3 km domain. Overall, it is apparent that the AP-XM simulation contains very different diurnal error patterns than the YNT simulations. For example, the 2 m temperature bias exhibits a prominent diurnal cycle (Fig. 7b) characterized by large positive and warm (negative and cool) biases during the night (day), resulting in an overall damping of the diurnal temperature cycle. The warm biases exceed 2.0 K during most of the night (22:00–03:00 LST), and the cold biases are < -2 K for several hours during the daytime (09:00–13:00 LST). These results indicate that the small temperature bias in the summary statistics for the AP-XM simulation (Fig. 3a) is misleading because it obscures the presence of substantial biases of opposite signs during the day and night. The RMSE is also much larger during the AP-XM simulation (Fig. 7a), with local maxima of 3.5 K at 11:00 and 23:00 LST, respectively, corresponding to peaks in the temperature biases. Switching to the YNT baseline greatly reduces the temperature RMSE, and the bias time series is no longer characterized by the highly amplified diurnal pattern seen in the AP-XM simulation. Examination of the YNT sensitivity experiments shows similar error patterns to the YNT baseline. The largest differences occur at night when the use of the GVF and SOIL datasets leads to smaller biases. In contrast, confining the analysis nudging to above 2 km a.g.l. (above ground level; YNT_N2KM) slightly increases the RMSE and bias during the nighttime relative to the YNT baseline.

For water vapor, the AP-XM simulation again exhibits a larger bias and RMSE than the other simulations (Fig. 7c, d). It has a large moist bias that ranges from 0.2 g kg-1 shortly after sunrise to 0.9 g kg-1 near 19:00 LST, before decreasing to a relatively stable bias of 0.6 g kg-1 during the night. The RMSE is smaller in the YNT baseline simulation, with a dry bias evident for all but the evening hours (19:00–22:00 LST). As is the case for temperature, the RMSE is smallest during the late-night hours and then steadily increases during the day before reaching its maximum in the evening. All of the YNT sensitivity experiments have similar RMSE and bias patterns to the YNT baseline, with the smallest (largest) spread between simulations occurring during the nighttime (daytime) hours, possibly due to differences in the PBL depth and surface energy balance (see Fig. 8). Comparison of the 10 m wind speed time series reveals that the AP-XM simulation has the smallest bias (∼ 0.15 m s-1) during the night but that the wind speeds are weaker than observed during the daytime, with the largest biases (-0.95 m s-1) occurring at noon (Fig. 7f). This diurnal pattern in the AP-XM simulation, characterized by winds that are too strong (weak) during the night (day), stands in contrast to the mostly positive biases in the YNT simulations. The biases are tightly clustered in all of the YNT experiments during the nighttime hours (22:00–07:00 LST), with the exception of the two simulations employing the GVF dataset (YNT_GVF and YNT_SSNG) that are characterized by persistently larger positive biases. These two simulations also have the largest RMSE (Fig. 7e). Further research is necessary to determine why the incorporation of the high-resolution GVF dataset leads to larger surface wind speed errors.

Near-surface atmospheric conditions can be strongly impacted by the partitioning of net surface radiation into sensible, latent, and ground heat fluxes (Santanello et al., 2018). To examine this more closely, Fig. 8 shows a time series depicting the average diurnal evolution of the PBL height, net surface radiation, and sensible, latent, and ground heat fluxes during 22 May–22 June 2017 computed using data from stations on the 1.3 km domain to maintain consistency with earlier results. Because in situ flux and PBL height observations are not available across the entire domain, the aim is not to examine the accuracy of the simulated surface energy fluxes and PBL height but rather to use these variables to help explain differences in the near-surface temperature, water vapor, and wind speed errors in the model simulations. All of the variables were obtained directly from the WRF output files. The net surface radiation is defined as the sum of the upward and downward shortwave and longwave radiation fluxes at the surface.

Inspection of Fig. 8 reveals large differences between the AP-XM and YNT simulations. The PBL is ∼ 50–150 m deeper in the AP-XM simulation during the nighttime but then becomes much shallower than the YNT simulations from mid-morning through the afternoon (10:00–16:00 LST), with the daytime maximum in PBL height occurring ∼ 2 h later (Fig. 8a). The AP-XM simulation is also characterized by a smoother and less amplified diurnal evolution. For the YNT simulations, the PBL heights are tightly clustered during the night (21:00–07:00 LST) but begin to diverge during the morning and reach their largest differences during the afternoon. In particular, simulations employing the high-resolution soil moisture analyses (YNT_SOIL, YNT_SSN, and YNT_SSNG) have average PBL heights that are ∼ 100 m lower than the other YNT simulations. These three simulations also have slightly lower sensible heat flux (Fig. 8c) and higher latent heat flux during the afternoon (Fig. 8d), which is consistent with the wetter and cooler topsoil layer in the SPoRT LIS analyses (Fig. 2g–l) and cooler 2 m temperatures (Figs. 3a, 7b). Using the SST and GVF datasets and confining analysis nudging to above 2 km had minimal impact on the PBL heights in the YNT_SST, YNT_GVF, and YNT_N2KM simulations; however, sensible and latent heat fluxes are slightly smaller during the afternoon in the YNT_GVF simulation.

Comparison of the AP-XM and YNT simulations also reveals large differences in the surface energy flux time series. For example, the AP-XM simulation has much smaller sensible heat flux during the daytime (Fig. 8c), and the latent heat flux remains relatively large during the night (Fig. 8d). Though the AP-XM and YNT simulations produce similar magnitudes of latent heat flux during the day, the afternoon maximum is delayed by 2 h in the AP-XM simulation. The combination of a shallower PBL during the day (Fig. 8a) and higher latent heat flux at night likely contributes to the persistently large moist bias in the 10 m water vapor mixing ratio (Figs. 3d, 7d) during the AP-XM simulation. Another noteworthy feature of the AP-XM simulation is that the ground heat flux remains negative at all times. This unphysical behavior stands in sharp contrast to the more realistic evolution during the YNT simulations, where the positive (negative) ground heat flux during the night (day) indicates that heat is being transferred from (toward) the ground toward (from) the atmosphere due to cooler (warmer) surface temperatures. These results indicate that the poor performance of the AP-XM simulation on the 1.3 km domain, when assessed using near-surface moisture, temperature, and wind observations, is likely due to the presence of vastly different and sometimes unphysical surface energy fluxes.

The lower accuracy of the AP-XM simulation could be due to limitations in the parameterization schemes when used at higher spatial resolution. This possibility is supported by Fig. 9, which shows the evolution of the PBL height and surface fluxes on the 12 km domain computed using simulated data from all stations on the 1.3 km domain. Differences between the AP-XM and YNT simulations are much smaller both in timing and magnitude on the 12 km domain. For example, the time series for PBL height, sensible heat flux, and latent heat flux are very similar for all of the simulations. Though the ground heat flux time series for the AP-XM simulation continues to be an outlier at this resolution, it now has the correct diurnal cycle with positive (negative) values during the night (day). The improved simulation of surface fluxes on the 12 km domain likely contributes to the more accurate temperature and wind speed analyses in the AP-XM simulation at that resolution (Fig. 3a–b and g–h). The presence of persistently higher latent heat flux (Fig. 9d) leads to a positive moisture bias in the AP-XM simulation (Fig. 3d–e); however, the bias is smaller on the 12 km domain than it was on the 1.3 km domain. Inspection of the surface energy fluxes and PBL height on the 4 km domain revealed larger differences between the AP-XM and YNT simulations (not shown) but not as large as those on the 1.3 km domain. Though it is not the focus of this research, differences in the PBL height between the AP-XM and YNT simulations could be due to differences in the vertical mixing strength and entrainment flux in the AMC2 and YSU PBL schemes (e.g., Hu et al., 2010). Together, these results show that the AP-XM simulation performs well at 12 km resolution but that its accuracy decreases with increasing model resolution.