Temporal seagrass mapping using machine learning and field-validated satellite imagery to inform restoration and management

Seagrasses are vital coastal ecosystems, but their growth and ecological dynamics in response to seasonal and interannual changes are still poorly understood, limiting our ability to detect their responses to climate and human-driven impacts. Remote sensing is increasingly used to assess seagrass presence and extent to support management and restoration. However, most remote sensing approaches provide only a static snapshot of seagrass distribution making it difficult to identify seasonal and interannual trends in seagrass dynamics. In this study we mapped seagrass extent at a restoration site in Western Port, Victoria, using Sentinel-2 imagery calibrated with drone-derived ground control data, and cloud-based processing platform Google Earth Engine (GEE). A Random Forest classifier was trained to identify seagrass from satellite imagery achieving an overall accuracy of 87.0%. This model was then back-cast through the catalogue of available Sentinel-2 imagery to characterise seasonal dynamics, ephemerality, and trajectories of seagrass meadows in the region. Although seagrass extent remained relatively stable at the regional scale, the focal restoration site showed high variability, with patterns of seasonal dieback and regrowth. Developing semi-automated seagrass mapping pipelines can therefore offer vital temporally explicit insights into intra- and inter- annual seagrass dynamics, providing essential context for conservation and restoration outcomes.