We present a method for inferring surface morphology of ice from scanning electron microscope images. We first develop a novel functional form for the backscattered electron intensity as a function of ice facet orientation; this form is parameterized using smooth ice facets of known orientation. Three‐dimensional representations of rough surfaces are retrieved at approximately micrometer resolution using Gauss‐Newton inversion within a Bayesian framework. Statistical analysis of the resulting data sets permits characterization of ice surface roughness with a much higher statistical confidence than previously possible. A survey of results in the range −39°C to −29°C shows that characteristics of the roughness (e.g., Weibull parameters) are sensitive not only to the degree of roughening but also to the symmetry of the roughening. These results suggest that roughening characteristics obtained by remote sensing and in situ measurements of atmospheric ice clouds can potentially provide more facet‐specific information than has previously been appreciated. Key Points We discover a novel functional form for expressing backscattered electron intensity as a function of ice facet orientation Gauss‐Newton/Bayesian inversion robustly and flexibly yields three‐dimensional mesoscale morphology Surface roughness statistics are found to be sensitive not only to the degree of roughening but also to its symmetry Plain Language Summary The authors have found a way to extract three‐dimensional information about the surfaces of tiny ice crystals grown under a high‐powered scanning electron microscope. They found that these surfaces contain some unexpected features, with long, deep valleys in some instances, and rounded hills in others. The authors believe that these features may help understand how rough ice surfaces in real cirrus clouds affect Earth's climate.