El Niño, the warm phase of the Southern Oscillation, alters weather worldwide: it triggers droughts in Australia and floods in Peru, and it affects fisheries and crops. Accurate forecasts months in advance give time to prepare. Classical climate models (GCMs) poorly resolve small-scale processes, and regional models are limited by one-way nesting. Machine learning offers a path to more flexible analysis of large oceanographic datasets, but existing models often suffer from insufficient validation on real data and do not always capture extreme events.
The scientists compiled sea surface temperature (SST) data from NOAA ERSST v5 and ocean heat content (OHC) data from the Oceanographic Data Center of the Chinese Academy of Sciences, covering January 2000 through September 2023. The region was limited to the Niño 3.4 area (5°S–5°N, 120°–170°W). Data were converted into heat maps for the convolutional neural network (CNN) and into normalized numerical series for the long short-term memory (LSTM). The ConvLSTM-XT architecture consists of two ConvLSTM blocks, where matrix multiplications are replaced by 2D convolutions, followed by a fully connected output layer with ReLU activation and dropout of 0.3. The model was trained for 50 epochs using the Adam optimizer with a learning rate of 0.001 and batch size 32, with mean squared error as the loss function. Forecasts were evaluated using the ONI index (three-month average of SST anomalies), compared against a threshold of 0.5°C.
The model was tested on 52 months from November 2018 to September 2023. Forecast accuracy was 90.57% for configurations where at least one of the five quarters used real observations and the rest used predictions. When all five quarters were forecast without observational input, accuracy dropped to 83.02%. Qualitative analysis of heat maps from March to September 2023 showed that the model captured the transition from cold to warm anomalies characteristic of El Niño development, although it did not always exceed the 0.5°C threshold across all five periods.
The results confirm that hybrid spatiotemporal neural networks can compete with traditional models in El Niño forecasting. High accuracy with partial observational input indicates practical value for operational monitoring. The drop in accuracy for fully autonomous forecasts points to the need for additional input variables or architectural improvements.
Future research could expand the input data to include wind stress, sea level anomalies, ocean currents, and bathymetry. This would give the model a more complete picture of ocean–atmosphere interactions. Architectural modifications are also possible, such as adding attention mechanisms or using transformers to better capture long-term dependencies.
The developed approach could be adapted for El Niño early warning systems, enabling governments and humanitarian organizations to plan response measures in advance. The method is also applicable to other regions and climate indices.
The next step is to validate the model on independent data outside the training period and to test it in real time with monthly forecast updates. It is also worth investigating how adding atmospheric variables affects long-term forecast accuracy.
In the field of climate anomaly forecasting, predicting extreme events such as very strong El Niños remains an unsolved problem. The model has not yet been tested on rare but catastrophic scenarios. Another issue is how to use forecasts in real decision-making systems: the link between model accuracy and the effectiveness of humanitarian response is poorly understood.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors developed a model called ConvLSTM-XT, which combines convolutional layers for processing heat maps and LSTM for time series, and applied it to forecast sea surface temperature anomalies in the Niño 3.4 region. The model was trained on NOAA ERSST v5 data and ocean heat content, and its forecast accuracy over 52 test months reached 90.57% when partially using real observations, and 83.02% when fully autonomous forecasting for five quarters ahead.
Where its strength is
A strong point of the work is demonstrating that a hybrid spatiotemporal architecture can capture transitional processes, such as the development of El Niño, even with limited reliance on observations, which paves the way for practical early warning.
Where it could go next
Worth trying
Why this may matter
If the accuracy of autonomous forecasting can be improved, it would enable timely warnings about climate anomalies in regions where observations arrive with a delay, and reduce the economic damage from El Niño.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.57, against a usual 0.64. It looks like few have drilled nearby.
The 1982–1983 El Niño nearly wiped out the Peruvian anchovy population, which supported the world's largest fishery. Catches plummeted from 12 million tons to less than 1 million tons.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', climate models play a key role in predicting extreme weather events, and scientists struggle to make their forecasts heard by policymakers. Developing more accurate El Niño models echoes this theme: the earlier and more precise the warning, the better the chances for adaptation.