Applying Statistical Analysis and Machine Learning to Improve the Ice Sensing Algorithm

Abstract

The detection of sea ice is a major problem faced by Argo floats operating in polar regions. In these areas, the presence of sea ice threatens to damage or destroy floats in the event of an impact at the surface. While methods have been proposed and implemented to combat this danger, the most successful of which is the Ice Sensing Algorithm (ISA), further work is necessary to fully mitigate the risks, particularly in the Arctic. In this analysis, past CTD profiles from the Arctic are compiled and matched with sea ice data to examine the performance of the ISA and recommend potential changes and new methods to further improve its accuracy. This is accomplished by fitting the data to statistical and machine learning models to predict the presence of ice and analyzing the results. Results show that both modifications to current methods and the inclusion of new variables may increase the predictive power of the ISA. Specifically, the analysis shows that the use of point measurements (as opposed to a metric over a pressure range) at the shallowest allowable depth provides the best performance. The additional inclusion of practical salinity and time of year as predictive variables also increases the performance of the algorithm. Results and statistics on the performance of the algorithm are provided and analyzed in various regions.