Significant strides have been made in understanding El Niño-Southern Oscillation (ENSO) dynamics, yet its long-lead prediction remains challenging, especially for the El Niño events after 2000. Sea surface salinity (SSS) is known to affect ENSO development and intensity by influencing ocean stratification and heat redistribution and therefore, when combined with sea surface temperature (SST) data, can potentially enhance ENSO forecast skill. In this study, we develop a deep learning (DL) model that incorporates a multiscale-pyramid structure and spatiotemporal feature extraction blocks, and the model successfully extends effective ENSO forecast lead time to 24 months for 2000–2021 with reduced effect of the spring predictability barrier (SPB). Interpretable methods are then applied to reveal the time-dependent roles of SST and SSS in ENSO forecast. More specifically, SST is critical for short-medium lead forecasts (<1 year), while SSS is important for medium-long lead forecasts (>6 months). Furthermore, we track global SST and SSS spatiotemporal shifts related to subsequent ENSO development, highlighting the importance of ocean inter-basin and tropics-extratropics interactions. With increasing availability of satellite SSS observations, our findings unveil unprecedented potential for advancing ENSO long-lead forecast skills.

在理解厄尔尼诺-南方涛动 (ENSO) 动力学方面已经取得了重大进展，但其长期预测仍然具有挑战性，特别是对于 2000 年之后的厄尔尼诺事件。众所周知，海面盐度 (SSS) 通过以下方式影响 ENSO 的发展和强度：影响海洋分层和热量重新分布，因此，与海面温度 (SST) 数据相结合，可以潜在地提高 ENSO 的预报技能。在本研究中，我们开发了一种深度学习（DL）模型，该模型结合了多尺度金字塔结构和时空特征提取模块，该模型成功地将 2000 年至 2021 年 ENSO 有效预报提前期延长至 24 个月，同时减少了春季可预测性的影响屏障（SPB）。然后应用可解释的方法来揭示 SST 和 SSS 在 ENSO 预报中的时间依赖性作用。更具体地说，SST 对于中短期提前预测（<1 年）至关重要，而 SSS 对于中长期提前预测（>6 个月）也很重要。此外，我们跟踪与随后 ENSO 发展相关的全球海温和海温时空变化，突出了海洋盆地间和热带-温带相互作用的重要性。随着卫星 SSS 观测的可用性不断增加，我们的研究结果揭示了提高 ENSO 长期预报技能的前所未有的潜力。