Speaker
Description
Gas transport within permeable waste rock may result in the exhalation of hazardous gases, such as radon-rich or oxygen-deficient air, driven by seasonal stack-effect airflow. While the airflow regime can be identified by differential pressure monitoring or ground surface temperature analysis, characterizing the spatial heterogeneity of emissions across an entire landform remains a significant challenge. High-resolution infrared (TIR) imagery captured via UAV during pre-dawn windows provides a vital visual proxy for identifying localized exhalation hotspots. However, IR imagery acquisition is frequently restricted by specific atmospheric requirements, airspace regulations, and the logistical effort required for nocturnal field operations. To address these limitations, this study introduces a predictive segmentation framework that utilizes deep learning to localize radon hotspots using only openly available geospatial data. By training a convolutional neural network on an input stack comprising high-resolution RGB orthophotos and Digital Elevation Model (DEM) derivatives, including slope and Topographic Position Index (TPI), we aim to predict surface venting without the need for repeated, labour-intensive UAV thermography. The preprocessing pipeline utilizes localized feature scaling and a specialized loss function to optimize the detection of sparse exhalation masks. This scalable, data-driven approach streamlines the assessment of exhalation risks by focusing the in-situ monitoring to high-risk regions of interest.