AI-assisted interpretation of changes in riparian woodland from archival aerial imagery using Meta's segment anything model

Abstract

An implementation of Meta's 2023 foundation artificial intelligence model, Segment Anything (SAM) is tested and used to assist in mapping changes in the extent of riparian woodland using publicly available archival aerial imagery along three gravel bed, meandering, river reaches in rural settings in the UK. Using visual prompts in interactive mode, this newly applied approach is shown to deliver substantial time savings over manual digitisation techniques and, for the type of imagery and the small-scale deployed, potentially greater accuracy. When applied to high-resolution (25 cm) aerial imagery SAM appears to be a practical and useful method for examining vegetation and landform change in a manner that has previously only been feasible through detailed field studies. The extent of riparian wood increased by 37–46% between 1999 and 2022 along all three reaches with extension occurring in three main situations: lateral expansion of existing woodland patches along stable or near stable banks; localised bankside establishment of trees transplanted under flood conditions; and progressive colonisation of point bars that developed through channel migration. Considering these factors, important conditions for the establishment, survival and expansion of riparian wood are discussed and likely differences in species distribution according to the geomorphic context are highlighted.