With the increase of computational power, recently seismic full-waveform inversion (FWI) has been used to obtain more reliable shear-wave velocity (VS) distribution in the shallow subsurface. However, successful FWI of near-surface seismic data remains challenging due to several reasons, e.g., lack of a good initial model, small seismic wavelength in soft soils rendering simulations of wavefield computationally demanding, cycle skipping problem, and attenuation of high frequencies in the soil. Additionally, for dynamic site response analyses, one needs – in addition to soil stratigraphy - other in-situ soil properties, which are usually obtained from cone penetration tests (CPTs). For instance, CPT cone-tip resistance (qc) is used to obtain undrained shear strength of saturated cohesive soils and friction angle of sands. In this research we have extended the resolution and accuracy of FWI through efficient forward modelling using spectral element method, robust inversion using the gradient descent method in combination with an optimal transport (OT) misfit function (Metivier et al., 2016), and high-frequency shear-wave data. A careful adaptation of these elements in the FWI has enabled us to stretch the resolution and accuracy of the VS field derived from surface-seismic data to that obtained from high-density (25 cm depth interval) seismic cone penetration tests (SCPTs). However, SCPT offers only 1D information at discrete field locations, whereas from FWI we achieve nearly the same resolution and accuracy, but in a continuous, laterally varying 2D or 3D manner. Once the high-resolution, accurate 2D/3D VS field is obtained by FWI, in the next step we attempt to predict reliably qc from VS. We develop new, dedicated schemes of machine learning utilizing site-/region-specific SCPT databases. The accuracy of prediction is generally high. This results in a laterally continuous CPT field. This is invaluable for estimating laterally varying soil parameters needed to forecast more accurate seismic ground accelerations.