The role of sea surface salinity in ENSO forecasting in the 21st century

The El Niño-Southern Oscillation (ENSO) is a climate pattern with significant global impacts on climate, agriculture, ecosystems, health, and society. Accurate and early ENSO prediction is crucial for mitigating these impacts. Current ENSO prediction models include dynamical models (ranging from intermediate to fully coupled general circulation models), traditional statistical models (like multivariate linear regression and Markov Chain-based models), and artificial intelligence (AI) models. While dynamical and statistical models are widely used, recent advancements in deep learning (DL) have shown promise in improving ENSO prediction skill. Previous DL models, such as those developed by Ham et al. (HKL19), have demonstrated improved forecasting skill compared to traditional methods by utilizing SST, ocean heat content, and sea surface winds. However, these models still suffer from limitations such as underestimating peak values in long-term forecasts and the influence of the spring predictability barrier (SPB), a phenomenon that reduces forecast skill before boreal spring. This study addresses these challenges by incorporating SSS data into a novel DL model to enhance long-lead ENSO forecasting skill, particularly for the period after 2000 when predicting El Niño events has proven more difficult due to weaker variations in warm water volume.

Existing literature demonstrates the role of sea surface salinity (SSS) in ENSO development. SSS influences ocean stratification and heat redistribution, affecting the thickness of the ocean barrier layer. A thicker barrier layer can enhance the response of the shallower mixed layer to westerly wind bursts, a key El Niño trigger, and suppress vertical mixing, facilitating El Niño onset. Studies have shown that assimilating SSS observations into dynamical models improves ENSO predictions and alleviates the SPB. However, the extent to which incorporating SSS into AI-based models can further advance ENSO forecast skill needed further investigation. Previous work using deep learning focused primarily on SST and other variables, showing some success in extending prediction lead times, but still faced challenges with the spring predictability barrier and accurate prediction of peak ENSO intensity. This paper builds upon this existing work by adding SSS data as an input, leveraging the power of deep learning to capture complex nonlinear relationships between SSS and ENSO.

This study developed a novel deep learning model called Spatio-Temporal Pyramid Network (STPNet) for ENSO prediction. STPNet utilizes global monthly anomalies of SST and SSS as input variables. The model architecture comprises three main components: a multiscale spatial feature structure (downsampling input variables to various scales to capture diverse spatiotemporal features), a spatiotemporal feature extraction block (employing Temporal Convolutional Networks (TCN) for time feature extraction and convolutional layers for spatial features), and a feature fusion block (combining spatiotemporal features of different scales using ResNet residual connections and upsampling). The model uses 24-month-long global SST and SSS anomaly fields to predict the Niño3.4 index for the next 24 months. Ocean heat content (OHC) is not explicitly used as an input; its information is implicitly included in the long-term SST data. The Coupled Model Intercomparison Project 5/6 (CMIP5/6) dataset was used for training, while the Simple Ocean Data Assimilation (SODA) reanalysis data and Argo buoy observation data served as the test set. To evaluate the model's performance, the study employed a multi-error statistical analysis technique focusing on effective prediction length and the reduction of SPB influence. The model was trained using a Rectified Adam optimizer and Mean Squared Error loss function for 130,000 iterations. To further analyze the model's behavior, the study incorporated interpretable methods such as a saliency map method and backpropagation-based visualization to reveal the relative importance of SST and SSS in ENSO prediction and identify key spatiotemporal regions involved in ENSO development.

STPNet significantly outperforms existing DL models in ENSO prediction, achieving an effective prediction duration of up to 24 months for the period after 2000. The inclusion of SSS information significantly improves the model's forecasting ability, particularly in long-lead forecasts (>1 year). The model successfully reduces the impact of the spring predictability barrier (SPB), maintaining high correlation between predicted and observed Niño3.4 indices even when forecasts are initiated in spring. Sensitivity tests reveal that SST is crucial for short-lead forecasts (<1 year), while SSS becomes more important for long-term forecasts (>1 year), particularly in the equatorial central Pacific. Interpretable methods show SST's dominant role in short-term predictions (0-13 months) and SSS's influence on medium-to-long-term predictions (beyond 14 months). The analysis of spatiotemporal patterns reveals the importance of ocean inter-basin interactions (Indian and Atlantic Oceans) and extratropical ocean influences (both hemispheres) for long-term ENSO forecasts. Masking experiments demonstrate that the inclusion of SST and SSS data from the Indian Ocean significantly improves the model's forecast skill, extending the effective forecast lead time from 12 to 24 months, while adding Atlantic Ocean information provides further improvements but to a lesser extent. Including extratropical ocean information also extends the forecast length.

This study's findings highlight the crucial role of SSS in improving long-lead ENSO forecasts using deep learning. STPNet's superior performance stems from its ability to capture complex spatiotemporal relationships between SST, SSS, and ENSO. The model's success in mitigating the SPB and accurately predicting ENSO events even after 2000, a period previously considered more challenging, demonstrates the value of incorporating SSS information. The model's interpretability enables the identification of key geographical regions and the time-dependent roles of SST and SSS in ENSO development, providing insights into the underlying physical mechanisms. The importance of inter-basin and extratropical interactions further underscores the need for holistic approaches to ENSO forecasting. This work supports the hypothesis that the complex and potentially nonlinear salinity-ENSO relationships can be effectively captured by deep learning, leading to improved forecast skill. While some limitations exist in the datasets used (CMIP5/6 models may overestimate SSS seasonal variations), the strong correlation between model outputs and independent datasets (SODA and Argo) validates the reliability of the results.

This study demonstrates the significant improvement in ENSO forecasting skill achieved by incorporating sea surface salinity (SSS) into a deep learning model. The STPNet model extends the effective forecast lead time to 24 months, reduces the impact of the spring predictability barrier, and reveals the time-dependent roles of SST and SSS in ENSO development. Future research should focus on exploring the underlying physical mechanisms through which salinity influences ENSO beyond a year and improving the accuracy of SSS datasets to further enhance model performance. Investigating the potential benefits of including subsurface temperature and salinity data could provide additional improvements in forecasting accuracy.

The study utilizes CMIP5/6 datasets for training, which have been shown to potentially overestimate the seasonal variations of SSS in tropical regions. While the model's strong performance on independent datasets suggests robustness, this remains a potential limitation. The model's performance might also be influenced by the specific architecture and hyperparameters chosen for STPNet. While efforts have been made to improve generalizability and avoid overfitting, it would be useful to perform further testing with a wider range of datasets and models. Further investigation into the physical processes explaining the longer-term effects of salinity on ENSO is also warranted.