AUTOMATED MAPPING OF STONE WALLS IN NORTHEASTERN USA USING DEEP LEARNING

Stone walls are ubiquitous features in the forested landscape of northeastern USA that mark 17^th to early 20^th century property boundaries and agricultural practices. Mapping and documenting stone walls, therefore, has great potential for studying and analyzing the spatial extent and variation of these past human impacts. Stone walls can be mapped on the ground during field surveys or with remote sensing data such as 1m LiDAR. However, field mapping can only cover small tracts of land at a time, and although LiDAR is widely available, on-screen manual digitization is time-consuming, especially when trying to complete mapping over vast areas of land. To overcome these limitations, deep learning has great potential to automate the mapping process, and the U-Net convolutional neural network model has shown promise incorporating image classification techniques and remote sensing data such as LiDAR. This study has two goals: 1) apply the deep learning, U-Net model to automate mapping of stone walls and compare model performance in various terrain conditions, and 2) compare the results of the U-Net model against field verified walls to address accuracy and efficiency. To achieve these goals, we built a ground truth dataset using field mapping, the LiDAR hillshade and aerial imagery, and trained U-Net model with 1 m LiDAR derivatives (hillshade and slope). Model accuracy was calculated based on F1 score. Stone walls in smooth terrain were identified with the best accuracy (F1 score: 89%).