EXPLORING DEEP LEARNING MODELS AS A TOOL TO AID IN MAPPING SURFICIAL GEOLOGY IN MULTIPLE PHYSIOGRAPHIC REGIONS OF NEW YORK

Deep learning, a type of machine learning that uses training data to self-learn and perform tasks such as pixel classification, is being gradually explored as an approach to surficial mapping to aid more time- and labor-intensive conventional means. In New York, glaciation resulted in the deposition of an array of surficial deposits across multiple physiographic regions. Key to the development of deep learning models is the availability of highly detailed surficial geologic maps in each of these regions for use in training the models. We trained two deep learning models and explored their potential to aid in mapping surficial geology in two physiographic regions in upstate New York: the high relief Appalachian Plateau and the Central Lowlands bordering Lake Ontario.

Through exploring different groupings of surficial deposits and different combinations of spatial data, the optimized models reproduced existing mapping in greater than 79% of the training areas, and with similar accuracy in validation areas within the same physiographic region. These models used only two simple inputs: 1-meter lidar data and previously mapped surficial geologic training data (maps) published by the U.S. Geological Survey and New York State Geological Survey. This simple approach is intended to make the methods and models created easily reproducible for a wide range of audiences using widely accessible datasets. This work helps provide an essential understanding of the strengths and limitations of lidar- derived surficial geologic models and elucidates how grouping surficial types, characteristics of different physiographic regions, and testing models in physiographic regions outside of their training areas effect model creation and performance. In addition, the models created may provide a useful preliminary tool for mapping surficial geology in similar physiographic regions, in addition to establishing a framework for creating similar models in different regions.