High-Resolution Remote Sensing and Deep Learning for Assessing Tree Regeneration on Boreal Disturbances

Abstract

In western Canada, the leading cause of forest fragmentation is seismic lines—corridors cleared for hydrocarbon exploration. These linear features have been linked to habitat degradation and predation-driven declines of boreal woodland caribou populations, prompting efforts to restore tree cover on seismic lines within caribou habitat. Prior to restoration, existing tree regeneration is assessed using aerial surveys and field visits to triage lines for treatment. However, these approaches are costly and labour-intensive and could be improved through automated methods that integrate high-resolution remote sensing and deep learning for individual tree detection and crown delineation (ITDCD) and species classification. Despite growing interest, the data sources and seedling conditions (i.e., species and size) under which these approaches are effective remain insufficiently explored. This thesis explores the use of a YOLO (You Only Look Once) object detection model and three remote sensing data sources—standard RGB (red–green–blue) imagery, multispectral imagery (RGB–red edge–near infrared), and LiDAR (light detection and ranging)—for ITDCD and species classification of black spruce (Picea mariana) and tamarack (Larix laricina) on regenerating seismic lines in the Canadian boreal forest. I found that standard RGB imagery (1 cm ground sample distance; GSD) consistently outperformed multispectral imagery (5 cm GSD), with F1 score improvements of up to 17.8 percentage points (pp). This was observed for both species and was most pronounced for smaller seedlings. At equivalent GSD, the inclusion of red edge and near-infrared bands yielded modest F1 improvements (up to 4.0 pp), although these effects were species- and size-dependent. Low-level fusion of optical imagery with LiDAR-derived surface models did not consistently improve performance relative to optical-only models (0.0–1.2 pp F1), with variable effects across seedling species and sizes. These findings indicate that sensor selection and data acquisition parameters should prioritize high spatial resolution for assessing tree regeneration on seismic lines, with additional spectral bands as a secondary consideration. Under the strategies tested, LiDAR provided limited additional benefit and may not justify the added operational complexity.