Date of Completion


Embargo Period



William Ouimet, Chandi Witharana, Andrew Jolly-Ballantine

Field of Study



Master of Arts

Open Access

Open Access


Advanced machine learning combined with widespread, publicly available, airborne light detection and ranging (LiDAR) data has great potential to automate the extraction and classification of landforms and 17th-early 20th century land use features preserved under forest canopy throughout the northeast US landscape. Previous studies have shown that stone walls, house foundations and relict charcoal hearths (RCHs) stand out clearly in derivative LiDAR digital elevation model (DEM) products such as slope and hillshade maps, but to date, mapping has been mainly carried out by on screen digitization. In this study, a deep learning convolutional neural network (CNN) algorithm was employed to extract relict charcoal hearth features from LiDAR data. With the application of CNN algorithms on LiDAR based slope maps and edge detection rasters, the network was successfully able to identify locations likely to be RCHs. The model results were further refined using object-based segmentation and image analysis methods, and compared with a dataset of manual digitized RCHs in 5 test area that cover a range of landscape types (steep terrain, rough terrain, lower gradient valleys and wetlands, developed, and forested). The CNN approach offers value in speed and ease and performed best when applied to forested hillslopes >10 degrees and areas lacking an overprint of mid to late 20th century development. Overall, the results of this study offers a unique insight into mapping past land use activity anywhere LiDAR data exists and the landscape experienced similar land use change during the 17th-20th centuries.

MajorĀ Advisor

William Ouimet