Assessing Predictability of Marine Heatwaves With Random Forests

Abstract

Marine heatwaves (MHWs) have increased in frequency and duration over the last century and are expected to intensify in the future. Such events have become an increasing threat for marine ecosystems and subsequently the economies and populations that rely on them. Here we apply random forests to assess skill in forecasting MHWs onset and severity at multiple prediction lead times. Random forests models are trained on a range of atmospheric and oceanic conditions to identify precursors of MHWs. The best performing random forest model accurately captures (76%) MHW presence/absence in the northeast Pacific and is capable of forecasting realistic extreme sea surface temperature patterns at weekly lead times. However, the total accuracy drops to 38% when forecasting MHW severity. Machine learning algorithms affirm further exploration as forecasting tools and have the potential to accelerate our predictive ability and preparedness against upcoming extreme climate changes.