USING HIGH-RESOLUTION LIDAR AND DEEP LEARNING MODELS TO GENERATE MINIMUM THICKNESS MAPS OF SURFICIAL SEDIMENTS

The areal distribution and thicknesses of sediments overlying bedrock have significant implications for groundwater storage. Characterizing surficial materials often requires extensive fieldwork to map coverage, as well as studies of outcrops and/or cores to estimate depth to bedrock. For large areas, generating informed estimates of surficial cover thickness may be a prohibitively expensive or time-consuming process. Deep learning tools in ArcGIS Pro provide a means to rapidly classify and map distinctive areas of surficial deposits as well as uncovered areas (i.e., bedrock exposures) based on their geomorphic textures as expressed in roughness and slope rasters derived from lidar elevation models. When combined with local measurements of stream incision depths obtained from lidar elevation models, these coverage maps can be interpolated to create minimum sediment thickness maps in areas that have experienced stream incision through surficial deposits but negligible incision into bedrock.

We used this methodology to generate a minimum sediment thickness map for the Neversink River watershed in New York, a tributary of the Delaware River and major source of water for New York City. Using the USGS DEM Geomorphology Toolbox, we estimated depth to bedrock for incised tributaries of the Neversink River and removed spurious datapoints near roads. Combining these estimates with the boundaries of deep learning-derived uncovered bedrock shapefiles corresponding to zero sediment thickness, we applied an inverse distance weighted interpolation to estimate minimum sediment thicknesses at 30 m resolution throughout the watershed. Our automated approach yielded minimum sediment thickness estimates (> 0-6 m) that generally agree with depth-to-bedrock estimates from local passive seismic survey data (~1.5-6 m). At present, we are working to further validate and expand this methodology to generate maps of minimum sediment thickness across broader areas. Here we present our methodology for generating specialized composite rasters that highlight the geometric signatures of bedrock and surficial cover, training deep learning models that rapidly classify these features over large areas, and generating sediment thickness maps from high-resolution lidar.