The predictability study of oceanic deep learning models: Taking Kuroshio intrusion into South China Sea as an example

Many previous studies have delved into the predictability of atmosphere and ocean in numerical models, which are crucial for guiding and improving predictions. Currently, deep learning prediction models have developed rapidly, yet their predictability remains largely unexplored. This study endeavors to probe the predictability of deep learning models by focusing on the Kuroshio intrusion (KI) into the South China Sea, utilizing the Conditional Nonlinear Optimal Perturbation (CNOP) approach. We first construct a deep learning model for the KI prediction based on the Unet, which can well predict the KI with a lead time of 14 days. By integrating this model with a nonlinear optimization algorithm, we calculate two types of CNOPs: one with a positive sea surface height anomaly (SSHA) error, denoted as CNOP1, and another with a negative error, labeled as CNOP2. These CNOP errors can grow quickly and exert significant effects on the KI prediction: CNOP1 tends to yield an anticyclonic SSHA error, which aligns with the loop path of the Kuroshio, thereby amplifying the intrusion, while CNOP2 has an almost opposite effect. Furthermore, the sensitive area is identified by the spatial structure of the CNOP error, which is mainly located around Luzon strait. Reducing the input data errors in the CNOP sensitive area will more remarkably improve the KI prediction with a relative improvement rate surpassing 20%, compared to the sensitive area identified by occlusion method and other artificially determined areas. Such findings have the potential to elevate the KI prediction skills of deep learning models.