T46 The Past, Present, and Future of the Digital Data Revolution and Its Effect on the Geosciences (Posters)

Sunday, 15 October 2023
8:00 AM - 5:30 PM
Hall B

AUTOMATED MAPPING OF HISTORIC APPALACHIAN MANGANESE OXIDE MINES USING CONVOLUTIONAL NEURAL NETWORKS APPLIED TO HIGH-RESOLUTION LIDAR DERIVATIVES

Paper number: T46-15
Program: Technical Sessions
Time: 8:00 AM – 5:30 PM

Authors

Abstract

The eastern United States hosts hundreds of defunct manganese oxide mining operations that have lain untouched for decades. These inactive sites are often identifiable by distinctive features such as prospect pits, linear trenches, large excavations, and waste piles. Recent geochemical investigations of manganese oxide mine waste piles have shown these materials to be enriched in numerous critical elements and most rare earth elements relative to the average continental crust. However, the precise locations and sizes of these mines and their waste piles are poorly known in most instances, rendering detailed analysis of their resource and hazard potential difficult.

We are exploring the use of deep learning mapping techniques to rapidly locate and precisely map the distinctive topographic signatures of surficial mining operations. Focusing on documented locations of Virginian manganese oxide mines, we created a training dataset with >1700 polygons that outline prospect pits (n = 1149), linear trenches (n = 587), and large excavations (n = 34). These polygons were coupled with high-resolution (one-meter) lidar derivatives to generate training images, of which 10% were reserved for validation during training. Multiple models were trained using default settings of various convolutional neural networks (DeepLabV3, Faster R-CNN, Mask R-CNN, RetinaNet, and Single Shot Detector) to gauge relative model performance. Of the trained models, only DeepLabV3 successfully mapped significant portions of manganese oxide mines. Preliminary experimentation with different training data tile sizes for DeepLabV3 models shows a scale-based impact on performance; models using default tile sizes of 256x256 pixels excel at mapping small features such as prospects and trenches, while models using larger 1024x1024 pixel tiles more accurately map large excavations. These results demonstrate that while AI-based tools show promise for mapping and characterizing abandoned mines, careful consideration of mapping targets is required when generating models.

Geological Society of America Abstracts with Programs. Vol. 55, No. 6

doi: 10.1130/abs/2023AM-391336

© Copyright 2023 The Geological Society of America (GSA), all rights reserved. Permission is hereby granted to the author(s) of this abstract to reproduce and distribute it freely, for noncommercial purposes. Permission is hereby granted to any individual scientist to download a single copy of this electronic file and reproduce up to 20 paper copies for noncommercial purposes advancing science and education, including classroom use, providing all reproductions include the complete content shown here, including the author information. All other forms of reproduction and/or transmittal are prohibited without written permission from GSA Copyright Permissions.