Northeastern Section - 57th Annual Meeting - 2022

Paper No. 15-5
Presentation Time: 3:00 PM

MAPPING SURFICIAL MATERIALS IN THE DELAWARE RIVER BASIN USING A DEEP LEARNING MODEL APPLIED TO HIGH-RESOLUTION LIDAR DATA


ODOM III, William, Florence Bascom Geoscience Center, U.S. Geological Survey, 12201 Sunrise Valley Drive, Mail Stop 926A, Reston, VA 20192, BURKE, Caitlin E., U.S. Geological Survey, Florence Bascom Geoscience Center, 12201 Sunrise Valley Drive, MS 926A, Reston, VA 20192 and DOCTOR, Daniel, U.S. Geological Survey, Florence Bascom Geoscience Center, 12201 Sunrise Valley Drive, Reston, VA 20192

The Delaware River Basin (DRB) covers over 35,000 km2 and provides water to millions of people in Delaware, New Jersey, New York, and Pennsylvania. Originating in the Catskill Mountains and flowing southward to Delaware Bay, the Delaware River crosses a variety of formerly glaciated and unglaciated physiographic areas. These areas feature a diverse assemblage of Pleistocene landforms and sediments for which no basin-wide surficial maps exist.

The advent of high-resolution (1-3 m) lidar and user-friendly deep learning tools has recently enabled more rapid, precise classification of surficial materials and landforms than previously possible. We applied these tools to the DRB to generate high-resolution preliminary surficial maps. We trained a resnet-18 deep learning model using a topographic position index-slope composite image and existing geologic/land use datasets from New Jersey, New York, and Pennsylvania (total shapefile area = 11,921 km2). The model was trained to identify 15 distinctive textures, ranging from alluvium and bedrock to urban zones and water bodies.

Preliminary results indicate that overall model performance is particularly effective for mapping areas of exposed bedrock (>90% accuracy) and urban development (>70% accuracy). More nuanced surficial classes, including boulder-dominated colluvium and Quaternary alluvium, were identified with >60% accuracy. Our model may be improved by additional, targeted training data and larger training image tiles. While no deep learning model – regardless of data resolution – can replace field investigations, this approach shows promise for rapidly generating useful preliminary surficial maps over large areas.