Iceberg Stability Investigations Using Machine Learning for Alaska and Greenland

Abstract

Iceberg-capsize tsunamigenesis is a coastal hazard not often discussed or observed due to the lack of frequency of which this process occurs. Greenlandic communities are proximal to coasts and are at risk of tsunamis caused by unprompted iceberg-capsize. We use time-lapse imagery from our own sites in Greenland (QORQ, UMNQ, NIAQ, INKF) accompanied with imagery from LeConte Glacier, Alaska field campaigns to identify and measure unstable (capsizing, fragmenting, disintegrating) icebergs and compare them to stable icebergs. While there are studies that investigate the mechanics of capsize, there are little to none which offer applicable results to implement in real-world scenarios for hazard mitigation. We use ground imagery to observe icebergs and investigate the probability of an iceberg being unstable based on its geometry. We also use air temperature collected from meteorological data from respective sites to compare with unstable event occurrences. We find that it is not likely there are implications that air temperature has an influence on iceberg stability for both Alaska and Greenland. For Alaska, we find that logistic regression models may be able to differentiate between stable and unstable icebergs. Model scores for Alaska data and respective sampling methods: no sampling (85.07%), RMU (65.10%), SMOTE (66.14%). For Greenland, we speculate the data set may be too small and generalizes most icebergs as stable and the logistic regression seemingly has more difficulty with classification: no sampling (79.16%), RMU (57.14%), SMOTE (80.10%).