Targeted observation is an appealing procedure for improving model predictions through the assimilation of additional collected measurements. However, studies on targeted observations in the oceanic field have been largely based on modeling efforts, and there is a need for field validating operations. Here, we report the results of a field program that is designed based on the sensitive areas identified by the Conditional Nonlinear Optimal Perturbation (CNOP) approach to improve the short-range (7 days) summer thermal structure prediction in the Yellow Sea. We found good spatial consistency in the locations of the identified sensitive areas among the hindcast and climatology runs. By introducing the technique of cycle data assimilation and the new concept of time-varying sensitive areas, we designed an observing strategy based on the identified sensitive areas, and conducted a set of Observing System Simulation Experiments prior to assessing the effectiveness of the plan on later observations. On this basis, the impact of targeted observations was investigated by a choreographed field campaign in the summer of 2019. The results of the in-field Observing System Experiments show that compared to conventional local data assimilation, conducting targeted observations in sensitive areas can double the benefits of data assimilation in thermal structure prediction. Furthermore, dynamic analysis demonstrates that the refinement of vertical thermal structures is mainly caused by the changes in the upstream horizontally advected temperature driven by the Yellow Sea Cold Water Mass circulation. This study highlights the effectiveness of targeted observations on reducing the forecast uncertainty in the ocean.