Applications of neural networks in West Antarctic meteorology and climatology

A deeper understanding of regional meteorology in the Antarctic is required for improved interpretations of the ever-growing body of ice-core-based paleoclimate records from this region. Benefits would also accrue in other areas such as operational forecasting at McMurdo and global change research. Artificial neural network (ANN) techniques offer new approaches to improving our record of surface observations and our understanding of the regional atmospheric circulation, two keys to this important problem.

Apart from a small number of mostly coastal manned stations, the only source of continuous, direct measurements of near-surface weather on the West Antarctic ice sheet is the network of automatic weather stations (AWS) begun in 1980. Unfortunately, the AWS are widely and non-uniformly distributed, record only a few basic measurements, and are subject to instrument failures that may go for months without repair due to limited opportunities for service. The records are also relatively short with many stations in place only since the late 1980’s/early 1990’s. This makes it difficult to use these highly valuable records for comprehensive climatological studies. Recent work with ANN-based methods has shown a way around some of these problems. Records for temperature and pressure are built from existing AWS observations and ANN predictions from GCM-scale upper air data. Errors for temperature prediction are approximately equal to those from a satellite-based methodology but with no exposure to problems from surface melt events or sensor changes. The significant biases seen in ECMWF surface temperatures are also absent from our predictions. We have developed complete 15-year temperature and pressure records (1979-93) for six West Antarctic AWS: Siple Station, Byrd, Lettau, Marilyn, Elaine and Ferrell.

Previously infeasible climatological studies become possible with complete records from multiple sites. Annual temperature anomalies are significantly above average in 1980 and 1988, due to an unusually warm winter and spring, respectively. Seasonal cold anomalies occur at most sites in fall (1982, 1987), winter (1986) and spring (1981). The spatial pattern of temperature anomalies and principal component analysis suggests significant differences both between the ice shelf and ice sheet sites and between the two ice sheet sites.

Self-organizing maps (SOMs), a second ANN technique, have proved useful for analysis of synoptic-scale circulation in temperate latitudes. The use of SOMs allows development of synoptic climatologies with an arbitrary number of smoothly transitioning climate states, in contrast to traditional synoptic classification techniques. Results from SOM analyses are applicable both to the ice core interpretation problem and to studies of global change. SOM-derived maps of synoptic variables such as temperature and geopotential height can be compared to ice core data to examine the relationship between the proxy and the atmosphere. Climate change can be studied by looking at how state transitions evolve over time. We anticipate new insights and improved ice-core interpretations from application of SOMs to the West Antarctic atmosphere.