Deep Learning Image Segmentation of Sandy Beaches in Australia

Full description

This collection provides mapped sandy beach polygons for the Australian coastline generated using image segmentation methods. It extends the regional-scale sandy beach segmentation framework developed by Yong et al. (2024) from southeastern Australia to a continental scale.

The collection comprises beach polygon datasets generated using a U-Net model, Segment Anything Model 3 (SAM3), and a distance-transform-based union of the two approaches. Historical beach polygon predictions generated using the U-Net model are also provided for the Collaroy–Narrabeen (NSW) and Inverloch (VIC) coastlines to demonstrate the application of the framework to archived imagery. The trained U-Net models used to generate the beach polygon datasets are also included to support reproducibility and further application of the framework.

A final cleaned dataset of Australian beach polygons was produced through manual inspection and correction of the UNetAus model outputs. This process included removing larger backshore aeolian dunes, seawalls, docks, and other non-beach features.

This collection supplements the publication: Domain-Specialised Deep Learning Segmentation Outperformed Generalised Models for National-Scale Sandy Beach Mapping in Australia (Yong et al., in prep)
Lineage: Australian coastal imagery sourced from Esri World Imagery was processed using image segmentation workflows to derive sandy beach polygons. A U-Net model trained using Microsoft Bing Maps aerial imagery and OpenStreetMap-derived beach labels was applied to the imagery. Historical imagery for the Collaroy–Narrabeen (NSW) and Inverloch (VIC) coastlines was sourced from Esri World Imagery Wayback, georeferenced and processed using the same workflow.