REGIONAL MAPPING OF STONE WALLS IN NORTHEASTERN U.S USING DEEP LEARNING

Stone walls are unique anthropogenic landforms that are the result of glacial history and historic land use practices in northeastern USA. Mapping stone walls across this entire region will allow for a detailed understanding how the timing and magnitude historic land use practices varied around the region. Stone walls are highly visible in LiDAR data, particularly DEM derivatives such as hillshades and slope maps, and widely available LiDAR datasets around the region have enabled users to start mapping stone walls. To date, however, efforts have largely been manual. As on-screen manual digitization is labor-intensive to cover the entire region, deep learning with LiDAR derivatives has been explored to automate the mapping process of stone walls in part of Connecticut, showing high accuracy (F1: over 80%) for the one town studied in detailed. In thinking about how to scale this up to the whole region, it is important to realize that the quality of LiDAR data varies depending on states and campaigns: Connecticut (QL 2), Rhode Island (QL 3), Massachusetts (QL 1 ~ QL 3), Vermont (QL 2), and New Hampshire (QL 1 ~ QL 3), which can lead to a large variation in model performance. To handle this issue, this study maps stone walls at a regional scale (ie., the entire extents of CT, RI, MA, VT, and NH) using a transfer learning technique. For this technique, we conducted fine-tuning of the pre-trained CT stone wall U-Net model for different states and LiDAR campaigns. Unlike training a U-Net model from scratch, transfer learning with the fine-tuning method can target each LiDAR campaign and require a small amount of tuning data, which is an efficient and target-oriented method for completing the automated mapping across the region. We successfully generated stone wall maps for each state studied and results are consistent other methods. Initial interpretation of the results highlight the dependence of stone wall on the presence of glacial till, which is an important first order control that we would expect regional mapping to show.