A 10-km version of the NCEP Eta Model has been run over a roughly 1000 km x 1000 km domain centered over the upper Midwest for 20 cases where heavy warm season rainfall occurred from mesoscale convective systems to investigate the response of the precipitation forecasts to improvements in the depiction of mesoscale features at initialization time. Modifications to the initial conditions included (i) use of a cold pool initialization scheme, (ii) inclusion of mesonetwork surface observations using the model’s own vertical diffusion formulation to allow the surface data to be assimilated into a deeper layer through a simulated initialization period, and (iii) addition of water vapor at points covered by radar echo to ensure relative humidities greater than 80%. All of these modifications were implemented in runs using both the operational Betts–Miller–Janjic (BMJ) and Kain– Fritsch (KF) convective parameterizations. In addition, simulations were also run with a doubling of the convective time step, alternation of the two convective schemes within one run, and exclusion of a convective scheme in another run. For all 20 cases, 14 variants in the model initilization/moist physics were used, creating a high grid resolution (10-km grid spacing) ensemble. Although techniques (i) and (ii) both resulted in initial surface fields agreeing better with available observations, average skill scores for precipitation forecasts did not change appreciably when (i) was used, with (ii) resulting in a modest improvement in equitable threat score (ETS), with an increase in the bias that already exceeded 1.0 for most precipitation thresholds in the BMJ runs. Skill scores among the cases varied widely; no single adjustment consistently improved the scores. Interestingly, the simplest modification, the addition of water vapor in relatively dry atmospheric regions at points where radar echo was present, had the greatest positive impact on ETSs for most precipitation thresholds. Although the impacts were greatest in the first 6 h of the forecasts, some improvements occurred through the full 24-h integration period. Variations among the runs for a given case were far greater when different convective schemes were used than when initialization modifications were made, further supporting other recent research suggesting that high grid resolution short-range ensembles may benefit from the use of a variety of models or physical parameterizations.