Embodiments described herein use stochastic inversion (460) in lower dimensions to form an initial model (458) that is to be used in higher-dimensional gradient-based inversion (466). For example, an initial model may be formed from 1.5-D stochastic inversions, which is then processed (464) to form a 3-D model. Stochastic inversions reduce or avoid local minima and may provide an initial result that is near the global minimum.