Past studies have found that LST has a considerable effect on the development of lake-effect snowstorms [Hjelmfelt, 1990; Kristovich and Laird, 1998]. For instance, numerical experiments conducted over the Great Salt Lake [Onton and Steenburgh, 2001] indicated that a 2°C increase (decrease) in LST results in 32% more (24% less) lake-effect precipitation. Here, we first examined LST averaged over the individual lakes of the Great Lakes to delineate the differences between the two sets of LST boundary conditions (prior to the simulations). The grids of the individual lakes were determined by land use types provided by the United States Geological Survey (USGS). Figure 2 depicts the monthly mean differences between NARR LST and MODIS LST from December 2003 to February 2008. In general, over these winter months, NARR LST tends to be warmer than MODIS LST by about 3°C, with an exception in February 2004 when NARR LST is cooler than MODIS LST over Lake Erie (Figure 2d). The maximum difference between these two LST data sets reached as high as 8.6°C in December 2005 over Lake Erie (Figure 2d).
Figure 2. The differences in the monthly LST between MODIS and NARR (NARR minus MODIS; °C) for the Great Lakes during the winters (December, January, February) from 2003 to 2008. The grids of each lake were determined by land use type data from the USGS. (a) Lake Superior, (b) Lake Michigan, (c) Lake Huron, (d) Lake Erie, and (e) Lake Ontario.
Download figure to PowerPoint
 During winter, LSTs are normally (but not necessarily) warmer than the air temperature near the surface. An intuitive expectation is that a higher LST creates greater instability when the overlying air mass is colder, which may lead to higher precipitation. To examine this, monthly precipitation simulated from MLST_Ice and NLST_Ice, as well as observations, is shown in Figure 3 for the 5-year average. In terms of the geographical distribution of precipitation, the two simulations agree with each other, but their precipitation magnitudes are noticeably different (Black hatchings are significant at 90% confidence level; Figures 3j, 3k, and 3l). For example, in December over central Lake Superior, the precipitation difference reached a maximum of 42 mm/month (Figure 3j). Meanwhile, when compared to observations, both simulations show a systematic bias of over-predicted precipitation, such as on the southern shores of Lake Superior (Figures 3a, 3b, 3d, 3e, 3g, and 3h). However, around Lakes Huron, Erie, and Ontario, the MLST_Ice experiment produces less precipitation than NLST_Ice, making it closer to the observations. In eastern Lake Ontario, the simulated precipitation is distinctly greater than the observations in both experiments. From the terrain map of WRF (not shown), the elevation of Lake Ontario is 75 m while the nearest mountain to the east of the lake is 500 m. Therefore, the overestimated precipitation in eastern Lake Ontario may also be attributed to orographic lifting that is too strong, as was documented in Niziol et al. . Such an orographic effect has also been found over Lake Michigan [Hjelmfelt, 1992]. Regarding the spatial pattern of the precipitation differences, Figure 4 shows the spatial correlation between the simulations and the observations. The correlation indicates that while both simulations capture the major pattern of observed precipitation, MLST_Ice appears to perform slightly yet consistently better in terms of spatial agreement.
Figure 3. Monthly precipitation (mm/month) averaged over the period of 2003–2008. Observations: (a) December, (b) January, and (c) February; MLST_Ice: (d) December, (e) January, and (f) February; NLST_Ice: (g) December, (h) January, and (i) February; NLST_Ice minus MLST_Ice: (j) December, (k) January, and (l) February. Black hatchings indicate the 90% confidence level (CL) based upon Student's t-test.
Download figure to PowerPoint
 To examine quantitatively whether MLST_Ice improves the precipitation simulation as compared to NLST_Ice, we evaluated their mean error (ME) and RMSE:
where N is the number of grid points for evaluation, and P and O represent the model results and observations, respectively. To facilitate comparisons, we interpolated model grids to those of the observed precipitation. Both ME and RMSE (Figure 5) were computed for (a) the entire domain and (b) the lake-effect areas combined, calculated grid by grid between simulations and observations with monthly data. Here, ME indicates bias in the monthly magnitude of precipitation (perfect = 0), while RMSE quantifies bias in the precipitation variation (perfect = 0). Although ME fluctuates month to month, the mean values (red dots) suggest that MLST_Ice produces consistently less error in the precipitation magnitude than NLST_Ice, especially over the lake-effect domains (Figure 5b). There, the maximum difference between the two experiments occurs in December, with a mean value of 6.2 mm/month in MLST_Ice and 16.4 mm/month in NLST_Ice. Likewise, RMSE suggests that the precipitation of MLST_Ice is more realistic than that of NLST_Ice, as is reflected in their mean values (Figures 5c and 5d); this result corresponds to the lower LST as shown in Figure 2. These experiments are discussed further in section 3.3.
Figure 5. The box and whisker plots of the ME and RMSE of the simulated monthly precipitation (mm/month) for the winters (December, January, February) for the period of 2003–2008. (a and c) For the entire domain. (b and d) For the lake-effect domains. The thick horizontal line in each box, the top and bottom edges of the box, and the upper and lower whiskers represent the ME (Figures 5a and 5b) and RMSE (Figures 5c and 5d) for each of the five years. The upper (lower) whisker shows the maximum (minimum) value, and the red point represents the mean value.
Download figure to PowerPoint