Ethan P. Carr

Ice‑marginal lakes (IMLs) are increasing in number and size around the Greenland Ice Sheet and its peripheral glaciers. Yet consistent long‑term mapping of IML surface area and detection of sudden drainages, possibly as glacial lake outburst floods (GLOFs), remain unrealized. This study evaluated 10 spectral indices and 28 machine‑learning models to identify an optimal method for extracting IML surface area in Greenland. Performance was assessed using expert‑elicited pixel classifications and stratified random samples from six known GLOF‑producing IMLs. Spectral index performance varied substantially: the Water Ratio Index (WRI) and Icy Lakes Index (ILI) performed well across highly weighted metrics and produced the highest weighted rank-sum values (0.8990 and 0.8987). The widely used Normalized Difference Water Index (NDWI) performed poorly, scoring the third‑lowest weighted rank-sum value (0.7126). Machine‑learning models, developed using a clustering-to-random-forest methodology, generally outperformed spectral indices, with 10 models exceeding the best spectral index scores. The top performer, Model_1i, trained on Landsat 8’s seven surface reflectance bands and including lake ice pixels in the training data, achieved a weighted rank-sum value of 0.9327. An expert intercomparison dataset of 300 commonly interpreted pixels showed a mean agreement of 72% among experts, underscoring the need for rigorously trained and tested automated approaches. These combined results demonstrate that machine‑learning methods can provide the most consistent, scalable, and expert‑aligned approach for tracking Greenland IML areas and their potential drainages. Future work will apply the optimal model (Model_1i) to produce a Greenland‑wide, multi‑decadal database of IML dynamics that will, notably, be used for GLOF detection.