El Niño–Southern Oscillation (ENSO) is considered the main source of interannual weather variability. It shifts flood timing in Africa and alters the discharge of the Ganges, Amazon, and Congo rivers. However, forecasting ENSO itself years ahead remains difficult, and so does forecasting its hydrological consequences. One reason is the arbitrariness of rectangular regions used to compute ENSO indices. There have been hints that ENSO is part of a larger system of interconnected ocean temperature oscillations, but how to extract this information and transfer it to rivers remained unclear. This work stands at the periphery: there is almost no one nearby; the closest neighbor in the archive is a study on predicting El Niño a year and a half ahead, but it does not address rivers. Other neighboring works use neural networks to analyze water, superalloys, and solar neutrinos, but they do not link the ocean to hydrology.
Monthly sea surface temperature (SST) data from 1950 to 2005 were taken from 32 CMIP5 models and three reanalyses (Hadley-OI, COBE, ERSSTv5). The region covers the low latitudes of the Pacific and Indian Oceans from 37.5°N to 42.5°S and from 50.5°E to 0.5°W, a rectangle of 80 by 300 cells. A convolutional neural network with four convolutional and three fully connected layers was trained to predict the three-month moving average discharge of the Amazon (Obidos station) and Congo (Kinshasa station). For comparison, an ensemble of six machine learning models was used, fed only with the Niño 3.4 index, the average SST anomaly in a small rectangle (5°S–5°N, 170°W–120°W). Then saliency maps were constructed, which are derivatives of the network output with respect to input pixels, showing which ocean regions matter most for the forecast. Additionally, climate networks were built: Pearson correlations between SST time series at all point pairs were computed, and edges were marked when exceeding a threshold, separating local connections from remote teleconnections.
The neural network that saw the full SST field predicted discharge for both rivers better than the ensemble of models based on the Niño 3.4 index, and better than the historical mean. For the Amazon, the root mean square error (RMSE) dropped to 0.294, compared to 0.461 for the best index-based method and 0.287 for the historical mean. For the Congo, the gain was more modest: RMSE of 0.476 versus 0.711 for the best index-based method and 0.462 for the historical mean. Saliency maps showed that when using model data, the model relies on the tropical Pacific and Indian Oceans, not just the canonical ENSO region. For reanalyses, the maps were blurred, but the linear and nonlinear information about discharge in the global SST was confirmed. Climate networks revealed strong teleconnections between tropical regions of the Pacific, Indian, and even Atlantic Oceans in model data, whereas reanalyses had noticeably fewer long-range connections. This is consistent with the known fact that Earth system models tend to exhibit stronger coupling between components than observations.
This work differs from neighbors by combining explainable deep learning with climate networks and transferring ENSO predictions to river discharge. Neighboring studies either predict the ENSO index itself or apply neural networks to other tasks (water clarity, superalloys, neutrinos), but they do not extract interpretable spatial signals for hydrology. Here it is shown for the first time that information beyond the canonical ENSO region improves discharge forecasts, and that model data contain stronger teleconnections than reanalyses. This raises the question of whether data and machine learning can quantitatively assess and reduce the gap between models and observations.
If the approach is applied to more rivers and longer time series, a global map of discharge sensitivity to different ocean regions could be built. This would help understand which teleconnections are truly important for regional hydrology and how they change under different climate scenarios. Another direction is to use saliency maps to improve Earth system models themselves, pointing to excessive coupling or missing processes.
The method could improve seasonal and interannual forecasts of river discharge for water resource management, agriculture, and hydropower in tropical regions. It also provides a tool for checking climate models for realistic teleconnections.
Test the approach on other major rivers (Ganges, Mekong, Nile) and on longer time series, including data after 2005. Compare saliency maps across different CMIP5 and CMIP6 models to understand which models better reproduce observed connections.
In hydrology, predicting discharge on interannual timescales remains an unsolved problem: physical models often fail to capture nonlinear teleconnections, while statistical methods are limited by feature selection. This work shows that deep learning with explanation can extract these connections automatically, but transfer from model data to observations is still unreliable due to differences in correlation structure.
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.