Non-invasive site characterization techniques have the potential to rapidly evaluate large subsurface volumes to guide subsequent invasive geotechnical site investigation. Among these methods, seismic full waveform inversion (FWI) stands out for its potential to recover detailed two-dimensional (2D) images of the subsurface. However, FWI’s need for substantial computational resources and sensitivity to the initial starting model has limited its utilization as a general-purpose geotechnical site characterization tool. Addressing this, prior studies have shown data-driven methods can predict 2D subsurface structures composed of soil over rock. In the present study, we aim to generalize these findings to all near-surface conditions. We propose a novel model generation framework that utilizes techniques from geostatistics to generate complex 2D subsurface models. The generated models include dipping soil and rock layers, soil lenses, boulders, and underground utilities; none of which have been considered previously. We use our model generation framework to simulate 100,000 2D subsurface models. We simulate field data acquisition along these 100,000 synthetic models, by numerically solving the elastic wave equation using an impulse source at the model’s center surrounded by 24 receivers (12 on either side). The data-driven predictive model, trained on 90% of the simulated data, achieved a mean absolute percent error on the testing set of 19%. Furthermore, these predictions are made within fractions of a second circumventing the computational and starting-model-related challenges associated with traditional 2D FWI. These results demonstrate that data-driven methods can predict complex images of the subsurface to enable rapid subsurface imaging for geotechnical applications.