The automated retrieval of supraglacial lake depth and extent from ICESat-2 photon clouds leveraging DBSCAN clustering

Abstract

Meltwater feedbacks are a core component to the Greenland ice sheet surface mass balance but the contributions are hard to quantify, which makes predictions very uncertain. Effort has been taken to quantify the total volume of supraglacial meltwater with optical remote sensing. However, the requirement for in-situ measurements and tuning for each hydrological feature keeps the method from scaling to an automated ice-sheet wide solution. The ICESat-2 altimeter photon retrieval product provides high resolution (0.7m footprint) height estimates over the Greenland ice sheet at regular intervals (90 day revisit time). The Advanced Topographic Laser Altimeter System (ATLAS) onboard ICESat-2 returns both photons reflected from the water surface and the lake bottom which allows for depth retrieval. In this thesis a method is proposed to classify the ATL03 photon returns as surface and bottom photons leveraging an iterative Density-Based Spatial Clustering of Applications with Noise (DBSCAN) approach. The photon classifications yield estimates for the surface and lake bottom heights across the complete ICESat-2 track and a lake classification based on the slope of the surface and the estimated depth. Extrapolation of the along-track depths across the Greenland ice sheet, using Sentinel-2 optical imagery, is investigated. For extrapolation a generalized form for an empirical, physical and machine learning technique are considered. The lake classification, as compared to Sentinel-2 lake detection has an accuracy of 97.8% and provides consistent and smooth depth estimates over the complete ICESat-2 track. Extrapolation of the depth estimates using a random forest regressor shows an out-of-sample R2 of 0.7 with an RMSE of 1.54 across the complete validation set. The results suggest that DBSCAN clustering of ICESat-2 altimeter data can yield consistent and accurate depth estimates that can be used in a generalized method to estimate lake depths across the entire Greenland ice sheet with a low error margin and small computational overhead.