Ground-truthing a Predictive Model for Locating Fossil Vertebrate Localities in the Eocene of Wyoming

Robert Anemone, Brett Nachman, and Charles W. Emerson
Western Michigan University

Remote sensing and other tools and methods from the geographic information sciences have the potential to revolutionize how paleontological fieldwork is performed. We trained a neural network to analyze remotely sensed imagery and multiple GIS data and analytical layers in order to recognize different land cover classes, including productive localities, in Eocene deposits of the Great Divide Basin of southwestern Wyoming. Post hoc accuracy testing indicated that the model was able to recognize the spectral signatures of productive localities and four other land cover classes with a high degree of accuracy (84% correctly classified). Our predictive model was constrained by geology (limited to outcrops mapped as Wasatch formation), by topography (minimum required slope was 5%), and was limited to pixels which resembled known localities at the 98% probability level. Ground truthing of the model took place during the summer 2012 field season. The first area we surveyed yielded characteristic Eocene fossils (turtle, fish, crocodile, gastropod, bivalve) in deposits whose lithology (paper shales and stromatolites) suggested a lacustrine origin. We confirmed that these deposits have been mistakenly mapped as Wasatch formation and should instead be attributed to the Green River formation. In the second area, the lithology was clearly fluvial and the rocks were typical of the Wasatch formation (sandstones and mudstones). The area indicated by our model as having a high priority of being fossiliferous was in fact a large outcrop of heavily eroded sandstone that yielded typical Eocene terrestrial mammals, including Hyracotherium and Hyopsodus.

AAPG Search and Discovery Article #90169©2013 AAPG Rocky Mountain Section 62^nd Annual Meeting, Salt Lake City, Utah, September 22-24, 2013