Abstract
Variability in the El Nino-Southern Oscillation has global impacts on
seasonal temperatures and rainfall. Current detection methods for
extreme phases, which occur with irregular periodicity, rely upon sea
surface temperature anomalies within a strictly defined geographic
region of the Pacific Ocean. However, under changing climate conditions
and ocean warming, these historically motivated indicators may not be
reliable into the future. In this work, we demonstrate the power of data
clustering as a robust, automatic way to detect anomalies in climate
patterns. Ocean temperature profiles from Argo floats are partitioned
into similar groups utilizing unsupervised machine learning methods. The
automatically identified groups of measurements represent spatially
coherent, large-scale water masses in the Pacific, despite no inclusion
of geospatial information in the clustering task. Further, temporal
dynamics of the clusters are strongly indicative of El Nino events, the
Pacific warming phase of the El Nino-Southern Oscillation. The
unsupervised clustering task successfully identifies changes in the
vertical structure of the temperature profiles through reassignment to a
different group, concisely capturing physical changes to the water
column during an El Nino event, such as tilting of the thermocline.
Clustering proves to be an effective tool for analysis of the
irregularly sampled (in space and time) data from ocean floats and may
serve as a novel approach for detecting future anomalies given the
freedom from thresholding decisions. Unsupervised machine learning
approaches could be particularly valuable due to their ability to
identify patterns in datasets without user-imposed expectations,
facilitating further discovery of anomaly indicators.