The Conditional Nonlinear Optimal Perturbation (CNOP) method works essentially for conventional numerical models; however, it is not fully applicable to the commonly used Deep Learning forecasting Models (DLMs), which typically input multiple time slices without deterministic dependencies. In this study, CNOP for Deep Learning forecasting model (CNOP-DL) is proposed as an extension of the CNOP in the time dimension. This method is useful for targeted observations as it indicates not only where but also when to deploy additional observations. The CNOP-DL is calculated for a forecast case of Sea Surface Temperature in the South China Sea with a DLM. The CNOP-DL identifies a sensitive area northwest of Palawan Island at the last input time. Sensitivity experiments demonstrate that the sensitive area identified by the CNOP-DL is effective not only for the CNOP-DL itself, but also for random perturbations. Therefore, this approach holds potential for guiding practical field campaigns. Notably, forecast errors are more sensitive to time than to location in the sensitive area. It highlights the crucial role of identifying the time of the sensitive area in targeted observations, corroborating the usefulness of extending the CNOP in the time dimension.

deep learning forecasting model,conditional nonlinear optimal perturbation,targeted observation,sensitive area