Mapping stone walls in Northeastern USA using deep learning and LiDAR data

Abstract

ABSTRACT Stone walls are widespread and iconic landforms found throughout forested terrain in the Northeastern USA that were built during the 17th to early 20th centuries to delineate property boundaries and the edges of agricultural fields and pastures. As linear, or broadly curved, features that are typically >0.5 m high, 1–2 m wide, and >4–8 m long, stone walls are highly visible in LiDAR data, and mapping them is of broad interest to the cultural heritage sector as well as to researchers specifically focused on historic landscape reconstruction. However, existing mapping attempts have commonly relied on field surveys and manual digitization, which is time-consuming, especially when trying to complete mapping at broader scales. In response to this limitation, this study: (1) presents a novel framework to automate stone wall mapping using Deep Convolutional Neural Networks (DCNN) models (U-Net and ResUnet) and high-resolution airborne LiDAR, (2) evaluates model performance in two test sites against field verified stone walls, (3) investigates the factors that can influence model performance in terms of the quality of LiDAR data (e.g. ground point spacing), and (4) suggests post-processing for town-level mapping of stone walls (~120 km2). Both models performed well with respect to the Matthews Correlation Coefficient (MCC) score. U-Net scenario 3 achieved an MCC score of 0.87 at test site 1, while ResUnet scenario 3 (S3) had an MCC score of 0.80 at test site 2. In town-level test site 3, ResUnet S3 achieved the best F1 score of 82% after post-processing. This study demonstrates the potential of automated mapping of anthropogenic features using our models.