LiDAR‐based semi‐automated mapping of drumlins and mega‐scale glacial lineations of the Green Bay Lobe, Wisconsin, USA: Ice sheet beds as glaciotribological systems

Abstract

A machine learning methodology for processing and visualizing high-resolution LiDAR digital data is used to map drumlins and mega-scale glacial lineations (MSGLs) on the bed of the Late Wisconsin Green Bay Lobe in Wisconsin, USA, which exhibited surge-like behaviour during deglaciation. Previous work has shown that streamlined bedforms are the product of erosional streamlining of pre-existing sediment. Analysis of bedform height and elongation ratio using curvature-based relief separation (CBRS) and K-means clustering of 32,003 bedforms reveals a continuum of six morphotypes ranging from drumlins, through "channeled" more elongated multi-crested drumlins, to MSGLs. Further statistical analysis shows that morphotypes cluster into six types of streamlined surfaces (S1-S6) recording progressive elimination of an antecedent overridden topography to produce a smoother bed. Initial, relatively high-relief drumlinized surfaces (S1, S2) occur around the slowerflowing lateral flanks of the lobe, where a pre-existing hummocky morainal topography was only partially modified by subglacial erosion. More streamlined surfaces (S3, S4) dominated by multi-crested more elongate drumlins of reduced relief amplitude are transitional to flow sets of MSGL-dominated surfaces (S5, S6) indicative of much faster-flowing ice streaming along the lobe's axis. Estimates of basal drag based on roughness calculations for each surface type identify a 61% reduction in frictional retardation from poorly streamlined surfaces S1 and S2 to MSGL-dominated surfaces S5 and S6, with a step-like reduction between drumlins and channeled drumlins (S3, S4) possibly recording the rapid onset of fast flow. Subglacial streamlining is argued to be accomplished by a thin (<1 m) "third layer" of deforming subglacial debris between ice and its bed which functioned as an erodent layer. A thin (<3 m) till cap, formed by aggradation of deforming debris, rests unconformably on heterogeneous core sediments. Streamlined subglacial surfaces are comparable to the "functional" surfaces resulting from erosion by a "third layer" of wear debris in engineering tribological systems, and also by gouge on faults. Pleistocene ice sheets expanded over pre-existing landsystems, pointing to the broader relevance of the methodology and findings reported here.

Automatic identification of streamlined subglacial bedforms using machine learning: an open-source Python approach

“…In this overlapping dataset, we identify the average elongation of the features as 7.83 and the average orientation of streamlined subglacial bedform long axis as 41.44 cardinal degrees. These results provide additional evidence that builds upon previous work in this region that have identified streaming conditions (Colgan and Mickelson 1997), basal drag flux (Zoet et al 2021), and bed roughness (Eyles et al 2022) of the larger Green Bay Lobe. With minimal manual effort, we have robustly derived general ice flow direction and streamlined subglacial bedform elongation of this subregion of Green Bay Lobe flow with a larger number of streamlined subglacial bedforms and the associated quantitative data.…”

“…The history of streamlined subglacial bedform identification in the field of glacial geomorphology has resulted in an incredible number of glacial feature datasets that support the ability to interpret entire ice sheet histories from bedform assemblages, build understanding of processes that control bedform formation and evolution, contribute to identifying glaciation style, identify areas of interest for field campaigns, and provide constraints to ice sheet modeling (Chandler et al 2018). Streamlined subglacial bedforms have been mapped using a wide range of techniques (e.g., Principato et al 2016, Norris et al 2017, Spagnolo et al 2017, Wang et al 2017, Saha et al 2011, Sookhan et al 2021, Eyles et al 2022, McKenzie et al 2022. While all of these methods contribute necessary knowledge to the field, each dataset is subject to varied levels of manual input and scalability has proven difficult in an era where there is unprecedented data availability.…”

“…Previous work on mapping streamlined subglacial bedforms across deglaciated landscapes has been conducted manually (e.g., Principato et al 2016, Norris et al 2017, using Fourier spectra data (Spagnolo et al 2017, e.g.,), through object-oriented automatic identification e.g., (Saha et al 2011, e.g.,), contour-tree mapping (Wang et al 2017), and by identifying slope variations in the landscape at various scales (e.g., McKenzie et al 2022, Sookhan et al 2021, Eyles et al 2022. Each of these mapping techniques has associated error, where the most inconsistent error arguably stems from manual identification.…”