Authors: Ji Won Suh*, Department of Geography, University of Connecticut, William Ouimet, Department of Geosciences, Department of Geography, University of Connecticut, Jonathan Leonard, Department of Geography, University of Connecticut
Topics: Remote Sensing
Keywords: Stone wall mapping, Deep learning, UNet, ResUNet, Northeastern USA
Session Type: Virtual Paper
Start / End Time: 3:05 PM / 4:20 PM
Room: Virtual 43
Presentation File: No File Uploaded
Stone walls are ubiquitous features in the forested landscape of northeastern USA that mark 18th-to-early 20th century property boundaries and agricultural activities. Identifying and mapping stone walls throughout this region, therefore, allows for a detailed understanding of human impacts and land-use/land cover prior to the mid-to-late 20th century, when this information can be studied using aerial and satellite imagery. The most widely used method that has been applied for mapping stone walls is manual digitization using 1m LiDAR data. However, this method is time-consuming, especially when trying to complete mapping over vast areas of land. In response to this limitation, deep learning has great potential to automate the mapping process. UNet models have shown promise using image classification techniques. This study aims to apply UNet and ResUNet to map stone walls and compare model performance depending on various terrain conditions (rugged vs smooth) and modern land cover (forest, farm fields, developed/suburban). To achieve this goal, we first built ground truth data using aerial imagery, google street view, and field mapping. Secondly, we created training data using 1m LiDAR derivatives; hillshade and slope maps. From here, we trained two models and compared their performance using F1 score, which resulted in the UNet model having the best performance of the two. In addition, stone walls in smooth terrain show the highest detection accuracy (F1 score: 79%). This approach shows value regarding speed and accuracy in understanding the spatial distribution of past land-use activities and agricultural legacy with airborne LiDAR and supplementary materials.