Abstract
<title>Abstract</title> <p>Anthropogenic landform classification is critical for hazard mitigation, heritage conservation, and ecological restoration in complex geological terrains like the Appalachian region. Within these areas, extensive surface mining has drastically altered local topography, creating a complex mosaic of natural and 'synthetic' landscapes. Specifically, operations predating the Surface Mining Control and Reclamation Act (SMCRA) of 1977 lacked strict backfilling requirements, leaving enduring anthropogenic footprints of exposed highwalls and mine benches. Today, these geomorphological remains are recognized as an integral part of the region's ‘mining heritage’, serving as permanent, tangible records of human interaction with the environment. However, accurate mapping of these post-mining landscapes is challenging due to dense forest canopy regrowth and their morphological similarities to the rugged natural terrain. Consequently, existing historic maps and databases remain incomplete for these surface mines. To address this, an automated deep learning (DL) framework was developed in this study to detect and delineate relict surface mining features. High spatial resolution light detection and ranging (LiDAR)-derived land surface parameters (LSPs) were utilized to train fully convolutional networks across southern West Virginia. A comparative analysis of four U-Net architectures (Base U-Net, ResU-Net, DiU-Net, and DiResU-Net) coupled with systematic hyperparameter optimization was conducted to determine the optimal configuration. Results indicate that the optimized Base U-Net bypassed the ‘complexity trap’ of deeper networks, achieving the highest performance with an F1-score of 0.452 and an overall accuracy of 81.4%. Conversely, more complex models struggled with over-segmentation and noise. While the generated inference maps can isolate anthropogenic landforms, future improvements- such as hybrid data fusion with optical imagery, transfer learning across diverse geologic settings, and topology-aware training- may further refine segmentation accuracy, thereby advancing post-mining landscape inventories.</p>