Microearthquake waveforms and spectra are well known to be strongly influenced by the shallow subsurface conditions at the recording site. We make use of this fact to determine optimum side models from SH-wave microearthquake seismograms using a non-linear waveform inversion technique based on simulated annealing. Site models are optimized in terms of thicknesses, S-wave velocities, and densities of the uppermost layers (<100 m) for 1D plane layer models. In addition, the source parameters in terms of Brune model source time functions are optimized simultaneously. The synthetic seismograms are calculated using a modification of Haskell's plane SH-wave algorithm. Using aftershock recordings of the Loma Prieta earthquake, a 1D plane layer model has been obtained for a selected site in Central California which has been occupied during the aftershock experiment. For this dataset, cross-correlation coefficients between optimized synthetics and observed records reach values above 0.8. The site model at the selected station for varying sources is fairly stable, especially for the uppermost layers. We observe systematic variations of layer parameters with back-azimuth which would be in agreement with slightly dipping layer interfaces. ¿ American Geophysical Union 1995