Reconstructing and identifying historic land use in northeastern United States using anthropogenic landforms and deep learning

Abstract

The northeastern U.S. experienced land use/land cover (LULC) changes over the last 300 years, including widespread deforestation during the 17th and 19th centuries, followed by reforestation in the 20th century. Reconstructing the extent of historic land use is crucial to understanding longer term impacts of anthropogenic activity. However, LULC datasets around the globe are typically only available for the last 50 years. This study aims to reconstruct the spatial distribution of cleared land with area estimation at two time periods based on stone walls using light detection and ranging (lidar), historic aerial photography, and deep learning. The two time periods are (1) the peak of agricultural activity period (mid-19th century) and (2) early-20th century. We utilized deep learning methods for (1) mapping the distribution of stone walls from lidar derivatives and 1934 aerial photography and (2) mapping historic land use from pre-satellite era aerial photography. The results of area estimation demonstrate strong consistency with the area indicated in the census record for both time periods. Notably, the incorporation of stone walls from 1934 orthomosaic plays a key role in significantly increasing the agreement of cleared area in the mid-19th century, as it compensated for the absence of walls that have been removed. By adding new historical time intervals of cleared land with our method, it allows for investigating the duration of land used for agricultural activities and the spatial pattern of LULC transition over the last 300 years.