Land use cover change (LUCC) is a major driver of global environmental and socio-economic transformations, with implications for carbon emissions, biodiversity, and sustainable development. However, robust historical analyses have often been limited by a lack of high-quality, spatially detailed baseline data. This study addresses this gap by being the first to apply deep learning-based image segmentation techniques to extract a comprehensive set of land use cover (LUC) information from historical topographic maps at a fine spatial resolution and at a cross-country scale. Specifically, we utilized topographic maps (1:50000) created by the U.S. Army Map Service during the Vietnam War (1963–1973) to create detailed historical LUC maps of Laos and Vietnam. We compared multiple model architectures on the manually labeled training data, with the UNet++ achieving the best performance. The resulting maps, produced at 4 m and 30 m resolutions, include 10 LUC classes and achieved high overall accuracies of 98.8% for Laos and 98.6% for Vietnam on separate test sets. Analysis of the maps revealed forest cover losses of 18.2% in Laos and 25.0% in southern Vietnam (below 17° N) by 1990 and a 36.8% reduction of mangrove forests in Vietnam by 1996. Transitions from forest to shrubland dominated in northern parts of Laos and Vietnam while transitions to cropland were most prevalent in Savannakhet province of Laos and the Southeast region of Vietnam. The maps provide a novel baseline for assessing post-war LUCC dynamics in Southeast Asia allowing for spatially explicit analyses of conflict and policy impacts at the time. Furthermore, the study demonstrates a transferable methodology that makes archival map collections accessible for large-scale historical global change research.

Read original article

Read original article