The orchestra plays a symphony, and you listen only to the first violin. That is roughly how river discharge forecasts have worked for decades: the entire El Niño signal is reduced to a single number calculated from a small rectangle in the Pacific Ocean. Yet the ocean is a vast ensemble, where heat flows from one region to another, and its full score contains far more information than a single note.
The authors of the new study decided to let a neural network listen to the entire orchestra. They took sea surface temperatures over 56 years across a vast area from the Indian Ocean to the mid-Pacific and trained a convolutional neural network to predict how much water would flow through the Amazon and Congo in the coming months. For comparison, other models received only the classic Niño 3.4 index, that very first violin.
The result was telling. The neural network that saw the full temperature field made smaller errors than an ensemble of models based on a single index: for the Amazon, the error dropped from 0.461 to 0.294; for the Congo, from 0.711 to 0.476. But the main point is not even accuracy, but where the network looked. Saliency maps showed that it relied not only on the canonical El Niño region but also on the Indian Ocean, as well as on teleconnections between tropical regions of the Pacific, Indian, and even Atlantic Oceans.
However, the orchestra sounds different in different halls. When the network was trained on model data, it found strong teleconnections, sometimes spanning half the globe. But in real observations, such teleconnections turned out to be noticeably weaker. This is consistent with a known fact: Earth system models tend to exhibit stronger coupling between components than observations do. Models may overstate long-range interactions, and this should be taken into account in forecasts.
Still, the main lesson is that the ocean conducts the rivers not with a single baton but with the whole orchestra. Listening only to the first violin, we miss important harmonies. A neural network trained on the full score already predicts discharge better today, and tomorrow it will help verify how realistically climate models reproduce these teleconnections. Perhaps this is how we will learn to understand how the ocean governs rivers in a changing climate.
The threshold for teleconnections in model data was set at a distance of 19000 km, which is almost half the Earth's circumference. In reanalyses, the threshold had to be lowered to 15000 km because there are fewer long-range connections.