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

The study addresses the ongoing challenge of long-lead prediction of El Niño–Southern Oscillation (ENSO), particularly after 2000 when central Pacific events became more frequent and conventional predictors weakened. Accurate ENSO forecasts are vital due to ENSO’s broad impacts on global climate, agriculture, ecosystems, health, and society. While dynamical and traditional statistical models have improved, and recent deep learning (DL) approaches have extended effective forecast lead times, these methods still suffer from amplitude underestimation at long leads and the spring predictability barrier (SPB). Sea surface salinity (SSS) affects ocean stratification, barrier layer formation, and heat redistribution, influencing ENSO onset and evolution. The purpose of this study is to test whether incorporating SSS with SST into a DL framework can enhance ENSO forecasting skill, extend lead times, and mitigate SPB, and to clarify the time-dependent roles and spatial sources of SST and SSS information for prediction.

Prior work shows both dynamical and statistical ENSO prediction models have achieved skill, with improvements from better initialization, parameterizations, and methodologies. Recent DL models (e.g., Ham et al. 2019) demonstrated superior skill using SST, ocean heat content (OHC), and winds, reaching effective leads of ~17–21 months, but still underestimated event amplitudes for long leads and remained sensitive to SPB, with markedly shorter skill when initialized in spring. Studies in dynamical frameworks have suggested SSS and the salinity-driven barrier layer influence ENSO by modulating vertical stratification, mixed-layer depth, entrainment, and Kelvin wave dynamics. Assimilation studies of in situ and satellite-based SSS (e.g., Aquarius/SMAP) improved ENSO forecasts by months and helped alleviate SPB. However, the added value of SSS within AI/DL-based models had not been fully explored. This work builds on those findings by explicitly integrating SSS with SST in a DL architecture and employing interpretability tools to quantify their relative, time-dependent contributions.

The authors develop a deep learning architecture named Spatio-Temporal Pyramid Network (STPNet) to forecast the Niño3.4 index up to 24 months ahead. Inputs comprise global monthly anomalies of sea surface temperature (SSTA) and sea surface salinity (SSSA) for the current and previous 23 months (24 months of inputs), covering 55°S–60°N and 0–355°E at 5°×5° resolution. Data processing: All datasets are bilinearly interpolated to 5°×5°. For CMIP5/6 training data, climatologies are computed over each model’s full time span; for test datasets, 2004–2018 is used as baseline to compute SSTA and SSSA. The Niño3.4 index target is the 5-month median-filtered average SSTA over 170°–120°W, 5°S–5°N. The model input tensor is [24×2, 24, 72] representing time, variables, latitude, and longitude; the output is a 24-element vector of monthly Niño3.4 predictions for the next 24 months. Training data: CMIP5/6 historical simulations (1861–2100) provide SSTA/SSSA anomalies, yielding 1,095,273 training samples. Testing uses three sources to challenge generalization: SODA plus Argo (2000–2021), SODA3.4.2 (2000–2015), and IAP (2000–2021). Model architecture: STPNet includes (1) a multi-scale spatial pyramid that downsamples inputs to 5°, 10°, 20°, and 40° to capture multi-scale spatial features; (2) spatial feature extraction via stacked 2D convolutions (3×3 kernels; 128, 256, 32 channels) with Tanh activations; (3) a temporal feature extraction block using a Temporal Convolutional Network (TCN) with causal and dilated convolutions to model temporal dependencies at each grid point. Inputs are reshaped to [batch, lat×lon, month] for TCN processing and reshaped back. (4) Feature fusion with ResNet-style residual connections and upsampling to merge features across scales, followed by global average pooling and a fully connected layer to predict the Niño3.4 sequence. Training: Mean Squared Error loss, Rectified Adam optimizer, learning rate 3×10⁻⁵, batch size 32, Pytorch 1.9/CUDA 11.7 on 4×NVIDIA Tesla V100 GPUs. To mitigate overfitting and stochastic variability, the model is trained five times and ensemble-averaged predictions are used. Evaluation: Forecast skill is assessed for 2002–2021 using model–observation correlation and RMSE as a function of lead time and initialization month, with focus on effective forecast length (correlation ≥0.50) and SPB sensitivity. Sensitivity and interpretability: The authors perform input-masking experiments (zeroing SSTA or SSSA) to assess variable importance and SPB impacts, as well as region-masking experiments to quantify contributions from the Pacific, Indian, and Atlantic basins and from tropical vs. extratropical bands. A gradient-based saliency method weights backpropagated input gradients by observed ENSO amplitudes to visualize spatiotemporal sources of predictability from SSTA and SSSA across leads and regions.

- Incorporating SSS with SST in STPNet extends effective ENSO forecast lead time to 24 months for 2002–2021 and reduces the spring predictability barrier (SPB). Correlation remains >0.50 at long leads for forecasts from any calendar month, with no rapid degradation after crossing two springs. For example, a forecast initialized in March attains correlation ≈0.83 at 21-month lead with SSTA+SSSA; when trained with SSTA only, this drops to ≈0.68. - Long-lead RMSE remains below ~0.5 °C for the SSTA+SSSA model, outperforming other DL baselines (CNN, ResCNN) and the SSTA-only STPNet. - Variable importance is time-dependent: SSTA dominates short to medium leads (<~12–13 months), whereas SSSA becomes increasingly critical for medium to long leads (>~6 months) and is dominant beyond ~14 months. When SSSA is masked, short-term forecasts retain some skill but SPB effects become more pronounced; when SSTA is masked, skill for winter ENSO events can still exceed 0.5, indicating a robust salinity–ENSO association. - Useful input memory differs by variable: SSTA contains predictive information up to ~12 months back, while SSSA contributes useful information up to ~24 months back. - Spatial origins of predictability evolve with lead: for leads <4 months, key SSTA signals reside in the eastern equatorial Pacific; between 4–12 months they expand to the Indo–western Pacific and Atlantic, indicating inter-basin interactions. Beyond ~1 year, SSS signals become dominant, with a critical region in the equatorial central Pacific near ~160°W. - Region-masking shows strong inter-basin and extra-tropical contributions: masking the Indian and Atlantic Oceans limits effective lead time to ~12 months; including the Indian Ocean substantially boosts skill, with the Atlantic adding further but smaller benefits. Inclusion of extratropical oceans (both hemispheres) extends effective lead from ~12 to ~24 months, consistent with the seasonal footprinting mechanism connecting extratropical variability to tropical ENSO evolution.

The findings demonstrate that adding sea surface salinity information to a deep learning ENSO forecasting framework significantly enhances long-lead predictability and mitigates the SPB, addressing a key limitation in existing dynamical and DL-based systems. The model’s interpretability analyses clarify that SST primarily constrains short-term evolution due to its direct correlation with the Niño3.4 index, while SSS provides longer-memory constraints through stratification, barrier layer dynamics, and pycnocline adjustments that influence ENSO onset and amplitude over multi-season timescales. The saliency and regional masking experiments reveal that inter-basin (Indian and Atlantic) and extratropical signals contribute materially to medium- and long-lead skill, supporting hypotheses of tropical–extratropical and inter-basin teleconnections modulating ENSO. By achieving robust correlation beyond two springs and maintaining low long-lead RMSE, the approach advances operational potential, particularly in the post-2000 period when ENSO predictability has been more challenging due to shifts toward central-Pacific events and weaker warm water volume precursors. The results suggest DL’s capability to capture nonlinear, nonstationary salinity influences that may be underrepresented in traditional models, providing a pathway for improved extended-range ENSO outlooks leveraging expanding satellite SSS observations.

This study introduces STPNet, a DL architecture that integrates global SSTA and SSSA over 24 months to predict Niño3.4 up to 24 months ahead. Incorporating SSS markedly enhances long-lead forecast skill and reduces SPB sensitivity compared to SST-only models and other DL baselines. Interpretability analyses reveal the time-varying roles of SST (short lead) and SSS (medium–long lead), identify key spatial source regions including the equatorial central Pacific for SSS, and highlight the importance of inter-basin and extratropical interactions. These advances underscore the value of satellite SSS in operational extended-range ENSO forecasts. Future work should: (1) investigate physical mechanisms underlying salinity’s influence beyond one year; (2) incorporate subsurface temperature and salinity to further improve skill; and (3) leverage improved SSS products to enhance training and operational performance.

- Training relies on CMIP5/6 simulations for SSS and SST anomalies; recent assessments suggest CMIP5/6 may overestimate tropical SSS seasonal variability, potentially biasing learned relationships. Nonetheless, strong correlation with SODA/Argo/IAP tests supports robustness. - The model does not explicitly use subsurface temperature or salinity, which are known to be important for ENSO dynamics; adding these may further improve skill. - Despite ensemble training (five runs) to reduce overfitting, DL models remain sensitive to data characteristics and training choices; code availability is on request rather than fully open, limiting external reproducibility. - Evaluation emphasizes 2002–2021; performance in earlier periods or under climate change scenarios requires further assessment.