LEVERAGING DEEP LEARNING APPROACHES FOR ENHANCED OUTCROP MAPPING AND ANALYSIS IN DRONE CAPTURED DATASETS (Invited Presentation)

High-resolution datasets captured with small uncrewed aerial systems (sUAS or ‘drones’) have assisted with geologic fieldwork by overcoming outcrop inaccessibility and providing new perspectives for interpretation. Transforming collections of drone images into detailed image mosaics or 3D digital outcrop models (DOMs) enables geologists to make precise, in situ, quantitative measurements over large areas. However, this remains an arduous task prone to inconsistencies and incompleteness – particularly as the size of the outcrop increases.

Recent advancements in artificial intelligence, particularly deep learning (DL), offer new possibilities for simplifying and enhancing these workflows through consistency of interpretations and efficiency performing repetitive tasks. DL involves training a computer to recognize patterns in data, similar to patterns a human would recognize. While DL has been successfully applied to many fields, its use in geology (and drone-based datasets) is only beginning to be explored.

To demonstrate the potential of DL as a tool in outcrop-scale mapping, we introduce an adaptable step-by-step workflow for training and applying DL models using sUAS datasets. We present three DL models designed for common geologic mapping tasks: 1) detecting distinct features (concretions), 2) segmenting sandstone beds, and 3) delineating geologic surfaces. We apply each model to different outcrops and geologic settings, each achieving a tested accuracy > 80%. To emphasize the versatility of DL, we apply our trained models to perform similar mapping tasks in entirely different datasets and environments, including mosaic images from the Mars rovers.

We discuss best practices for collecting, processing, and analyzing sUAS data, highlighting the benefits and challenges of using DL to streamline geologic mapping and increase accuracy. We present future outlook for sharing and developing DL models in 2D and 3D geologic mapping. Our goal is to provide a functional introduction of these advanced tools and workflows so they are accessible to all geologists, whether their data is collected by drones or from other sources.