Rapid estimation of minimum depth-to-bedrock from lidar leveraging deep-learning-derived surficial material maps

Abstract

Previously glaciated landscapes often share similar surficial characteristics, including large areas of exposed bedrock, blankets of till deposits, and alluvium-floored valleys. These materials play significant roles in geologic and hydrologic resources, geohazards, and landscape evolution; however, the vast extents of many previously glaciated landscapes have rendered comprehensive, detailed field mapping difficult. While recent advances in remote sensing have facilitated mapping of surficial materials and landforms, manual map creation has remained a time-intensive task.

The development of convolutional neural networks (CNNs) for image classification has provided a new opportunity for rapid characterization of digital elevation models, thus enabling efficient mapping of surficial materials and landforms. We have developed a methodology that leverages existing geologic maps and high-resolution (1–3 m) lidar data to train a U-Net CNN to classify alluvium and exposed bedrock in previously glaciated regions. Coupled with U.S. Geological Survey-developed geomorphometry tools capable of approximating stream incision depths, these classifications can be used to estimate the minimum thicknesses of stream-proximal hillslope sediments in areas where streams have undergone minimal incision into bedrock.

We validate this approach in the context of the Neversink River watershed, a subbasin of the Delaware River Basin and significant water source for New York City. Evaluation of deep learning model performance demonstrates substantial agreement with manually drawn maps of alluvium and exposed bedrock. Validation of the minimum sediment thickness map using borehole data and passive seismic measurements shows the greatest performance for shallow materials and decreased performance in deep sediments, as well as in areas where bedrock exposures were too small to be resolved by lidar. To resolve these issues and create more accurate surficial maps, we are training new CNNs with additional geologic data and exploring advanced approaches for estimating depths of stream incision.