A Novel Algorithm for Ice-Water Discrimination in Large Lakes Using ICESat-2 Altimetry and Data Driven Machine Learning

Abstract

Large freshwater lakes are critical for human life, ecosystem functioning, and the global carbon cycle. However, consistent high-resolution methods to characterize ice over large lakes remain limited. Here we develop an algorithm to progress ice observations over inland bodies of water by improving surface classifications using data derived from ICESat-2, Landsat 8/9 and other operational products. This algorithm implements a hierarchical approach composed of remote sensing products and data driven machine learning. In this study we show that although the current classification method used in ICESat-2 Inland Surface Water Height (ATL13) is prone to overgeneralization and misclassification, our proposed algorithm, which integrates novel classification methods and data-driven machine learning, enhances surface classification accuracy. We tested this algorithm on a wide breadth of data, spanning four ice seasons in the Laurentian Great Lakes. In our algorithm, we developed two prediction methods that outperformed the current classification method in place for ATL13 by 26.46% and 20.37% and is scalable to other inland surface waters because of the global coverage of the necessary parameters for surface classification. Improved surface classification allows for inland surface bodies of water to be observed with greater detail, particularly using ICESat-2 data, and enables the production of improved data sets of ice concentration and thickness. Improved ice information on Earth's largest lakes will have cascading effects on not only public safety and operational efficiency, but also the monitoring of anthropogenic changes in these bodies of water.