We introduce a mathematical framework for the inverse medium problem arising commonly in geotechnical site characterization and geophysical probing applications, when stress waves are used to probe the material composition of the interrogated medium. Specifically, we attempt to recover the spatial distribution of Lame's parameters ( and μ) of an elastic semi-infinite arbitrarily heterogeneous medium, using surface measurements of the medium's response to prescribed dynamic excitations. The focus is on characterizing near-surface deposits, and to this end, we develop a method that is implemented directly in the time-domain, is driven by the full waveform response collected at receivers on the surface, while the domain of interest is truncated using Perfectly-Matched-Layers (PMLs) to limit the originally semi-infinite extent of the physical domain. There are two key issues associated with the problem at hand: (a) the forward problem, namely the numerical simulation of the wave motion in the domain of interest; and (b) the framework and strategies for tackling the inverse problem. To address the forward problem, it is necessary that the domain of interest be truncated, and the resulting finite domain be forced to mimic the physics of the original problem: to this end, we introduce unsplit-field PMLs, and develop and implement two new formulations, one fully-mixed and one hybrid (mixed coupled with a non-mixed approach) that model wave motion within the, now PML-truncated, domain. To address the inverse problem, we adopt a partial-differential-equation-constrained optimization framework that results in the usual triplet of an initial-and-boundary-value forward problem, a final-and-boundary-value adjoint problem, and a time-independent boundary-value control problem. This triplet of boundary-value-problems is used to guide the optimizer to the target profile of the spatially distributed Lame parameters. Given the multiplicity of solutions, we assist the optimizer, by deploying regularization schemes, continuation schemes (regularization factor and source-frequency content), as well as a physics-driven simple procedure to bias the search directions. We report numerical examples attesting to the quality, stability, and efficiency of the forward wave modeling. We also report moderate success with numerical experiments targeting inversion of both smooth and sharp profiles in two dimensions. / text