120-19 From supercomputer to desktop: training and deploying open-source deep learning software for mapping piles at mines and quarries

Abstract:

High-resolution lidar data are available for nearly the entire continental United States. These datasets are useful for mapping surficial geology, geomorphology, hydrology, and anthropogenic features, especially in areas covered by dense foliage. Manually outlining features in lidar has been conventionally done for mapping limited areas; however, it does not upscale to the regional extents for which lidar data are now available. Through the Appalachian Manganese Oxide Research Effort (AMORE), we developed deep learning models that were trained on mine site polygons and lidar data, and that produce high-resolution maps of surficial features over large extents. Producing maps from trained models is straightforward on office workstations; however, training models is a resource-intensive process and may take several days per model on these machines.

We overcame this computational barrier by using large-language-model-assisted coding to translate our training workflows into open-source Python scripts that run on the USGS Tallgrass Supercomputer. These scripts more efficiently use computer resources, perform grid searches across training parameters for up to 100 models per session, and enable remote training on Tallgrass. These advances improved model tuning and performance, while reducing training time by over an order of magnitude and freeing up office workstation resources.

Here, we present a model trained on Tallgrass to map piles at quarries and mines. It was trained on 274 manually drawn piles in the Winchester, Virginia 30- x 60-minute quadrangle. Piles had a median area of 1507 m² and documented commodity types included iron, manganese, crushed stone, limestone, and sand & gravel. Because the model is trained on lidar data, it maps all piles regardless of commodity. We ran the model on the Sunbury, Pennsylvania 30- x 60-minute quadrangle, which features numerous coal mines, and the Big Levels, Virginia 7.5-minute quadrangle, which features several manganese mines. The model correctly identified numerous piles yet suffered from false positives in areas with different physiographies from the original training zone. Subsequent retraining included the false positive areas as background imagery, which improved performance as the mapping footprint expanded. This iterative approach of mapping beyond a model's training area, identifying false positives, retraining with a larger footprint, and rerunning models highlights the benefit of high-performance computing in the deep learning mapping space.

Geological Society of America Abstracts with Programs. Vol. 58, No. 5, 2026

doi: 10.1130/abs/2026AM-17480

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.