Forest fire smoke is a growing public health concern as more intense and frequent fires are expected under climate change. Remote sensing is a promising tool for exposure assessment, but its utility for health studies is limited because most products measure pollutants in the total column of the atmosphere, and not the surface concentrations most relevant to population health. Information about the vertical distribution of smoke is vital for addressing this limitation. The CALIPSO satellite can provide such information but it cannot cover all smoke events due to its narrow ground track. In this study, we developed a random forests model to predict the minimum height of the smoke layer observed by CALIPSO at high temporal and spatial resolution, using information about fire activity in the vicinity, geographic location, and meteorological conditions. These pieces of information are typically available in near-real-time, ensuring that the resulting model can be easily operationalized. A total of 15,617 CALIPSO data blocks were identified as impacted by smoke within the province of British Columbia, Canada from 2006 to 2015, and 52.1% had smoke within the boundary layer, where the population might be exposed. The final model explained 82.1% of the variance in the observations with a root mean squared error of 560 m. The most important variables in the model were wind patterns, the month of smoke observation, and fire intensity within 500 km. Predictions from this model can be 1) directly applied to smoke detection from the existing remote sensing products to provide another dimension of information; 2) incorporated into statistical smoke models with inputs from remote sensing products; or 3) used to inform estimates of vertical dispersion in deterministic smoke models. These potential applications are expected to improve the assessment of ground-level population exposure to forest fire smoke.