This paper proposes an approach to optimal planning under uncertainty with limited sensory information for an autonomous unmanned vehicle (UXV). We consider a surveillance application in which identification of environmental targets is impacted by controllable features commanded by the UXV as well as ambient features with dynamics that are uncontrollable. To manage computational complexity, mission states are defined by abstract or discretized features that are maximally influential based on Shannon information content. A receding horizon optimization method is applied to find optimal actions given uncertain and potentially erroneous sensor readings. Ambient feature transition probabilities are learned from empirical data then integrated with controllable features that evolve as a function of UXV actions.