This paper describes a novel method for identifying multiple targets with multiple robots in a partially known environment. Two main issues are addressed. The first relates to the use of motion planning algorithms to determine whether robots can reach ``good'' positions that offer the most informative measurements. The second concerns the use of predictive sensing to decide where sensor measurements should be taken. The problem is formulated similar to a next-best-view problem with differential constraints on the robots' motion, with additional layers of complexity due to visual occlusions as well as navigational obstacles. We propose a new distributed sensing strategy that exploits the structure of image manifolds to predict the utility of the measurements at a given position. This information is encoded in a cost map that guides a motion planning algorithm. Coordination among robots is achieved by incorporating additional information in each robot's cost map. A range of simulations indicates that our approach outperforms current approaches and demonstrates the advantages of predictive sensing and accounting for reachability constraints.