Future high autonomy missions such as Endurance A on the moon require autonomous targeting capabilities during long, pre-planned drives where humans are not always in the loop. Previous missions were limited to fixed campaigns developed manually by machine learning engineers interfacing with scientists to program valuation models for automatically identified targets. We present a novel machine teaching user interface, ReRank, that allows scientists to directly program valuation models by demonstration through interactive targeting and visualization. ReRank places scientists in an immersive 2D and 3D environment of a solar scene and leverages direct manipulation to automatically learn preferences when scientists click on targets of interest. We visualize the preferences for learned features of each target and apply Ranking SVM to learn a model of science value that can be applied in new scenes to automatically rank unseen targets. We envision applying ReRank to planetary missions such as Endurance A that are likely to change on touchdown and where human attention is divided between multiple spacecraft. Our approach will broadly enable more flexible and rapid campaign updates that increase science return by empowering scientists to be the programmer.