The interpretation of geophysical survey results to answer hydrologic, engineering, and geologic questions is critical to diverse problems for management of water, energy, and mineral resources. Although geophysical images provide valuable qualitative insight into subsurface architecture and conditions, translating geophysical images into quantitative information (e.g., saturation, concentration, and hydraulic properties) often involves substantial nonuniqueness and uncertainty owing to the limited resolution of geophysical imaging and uncertainty in petrophysical relations. We have developed a machine-learning approach to address these challenges in the context of a field-based investigation to map zones where a hydrocarbon plume was discharging to surface water at the National Crude Oil Spill Fate and Natural Attenuation Research Site in Bemidji, Minnesota, USA. The two-step approach combines multiple types of geophysical and direct information and effectively bypasses inversion and its associated assumptions. Integrating multifrequency electromagnetic induction, ground-penetrating radar, and fluid-sampling data, we first identify discharge zones and second estimate specific conductance versus depth. Compared with conventional inversion results, the machine-learning results (1) directly address the study objectives (delineating the discharge zones); (2) better extract depth-dependent information from the data, for which sensitivity diminishes rapidly with depth; and (3) quantify the uncertainty of the predictions (i.e., discharge versus nondischarge zones), rather than the uncertainty of the geophysical estimates (i.e., the standard error of estimation for the logarithm of electrical conductivity).