Mapping the world's coast: a global 100-m coastal typology derived from satellite data using deep learning

Abstract. We present a globally consistent, high-resolution (100 m) coastal typology dataset derived from satellite imagery and elevation data using deep learning—the first application of its kind in coastal science. Using a supervised multi-task convolutional neural network, we classified nearly 10 million coastal transects (one million km of coast) into four coastal attributes along the cross-shore profile: (1) sediment type, (2) coastal type, (3) presence or absence of built environment, and (4) presence or absence of human-made coastal defenses. The model, trained on about 1800 globally distributed samples, achieves strong predictive performance with F1 scores ranging from 0.67 to 0.83. Results show that the global coastal sediment distribution consists of approximately 40 % sandy, gravel, or shingle; 21 % muddy; 13 % rocky; and 27 % with no sediment. Considering the coastal type, 33 % of coasts are cliffed, 22 % are sediment plains, 15 % are wetlands, and 3 % are dune systems (i.e. 26,000 km). Combining sandy, gravel, shingle, and muddy sediments, we estimate that 61 % of the global coastline consists of soft sediments that are potentially easily erodible. Among sandy, gravel, or shingle coasts specifically, 20 % are cliff-backed and 16.5 % are located on built-up coasts. This global dataset, available in a cloud-optimized format at https://doi.org/10.5281/zenodo.15599096, provides a robust foundation for coastal change analysis and erosion assessment, and enables new opportunities for broad-scale vulnerability mapping and adaptation planning in the face of accelerating sea-level rise.