This thesis presents an approach of classifying multiple targets of interest in minimum time with satisfactory confidence by an imaging sensor on an underwater robot. The overall goal is achieved by sequentially solving a single target classification problem and a global target ordering problem. First, a multi-view single-target classification algorithm is developed based on the POMDP framework, which incorporates a deep convolutional neural network and a support vector machine as the observation model. The classification algorithm allows the underwater robot to adaptively select its next configuration state near the target of interest in order to maximize the increase of classification confidence. Next, a traveling salesman algorithm is used to generate the global target visiting order. Simulation results of an unmanned underwater vehicle equipped with a side-scan sonar validate the effectiveness of the proposed algorithm and demonstrates the ability to find significantly shorter path for multi-view based multi-target classification.