Machine Learning Modeling of Climate Variability Impact on River Runoff

Abstract

The hypothesis of this study was one of existence of spatially organized links between the time series of river runoff and climate variability indices, describing the oscillations in the atmosphere–ocean system: ENSO (El Niño–Southern Oscillation), PDO (Pacific Decadal Oscillation), AMO (Atlantic Multidecadal Oscillation), and NAO (North Atlantic Oscillation). The global river flow reconstructions (ERA-20-CM-R) for 18 study areas on six continents and climate variability indices for the period 1901–2010 were used. The split-sample approach was applied, with the period 1901–2000 used for training and 2001–2010 used for testing. The quality measures used in this paper were mean absolute error, dynamic time warping, and top extreme events error. We demonstrated that a machine learning approach (convolution neural network, CNN) trained on climate variability indices can model the river runoff better than the long-term monthly mean baseline, both in univariate (per-cell) and multivariate (multi-cell, regionalized) settings. We compared the models to the baseline in the form of heatmaps and presented results of ablation experiments (test time ablation, i.e., jackknifing, and training time ablation), which suggested that ENSO is the primary determinant among the considered indices.