Abstract
Multispectral satellite imagery enables routine surveying of the surf zone by discretising the land–sea interface at known water levels, supporting estimates of coastal recession and accretion. The daily revisit of PlanetScope provides near-continuous shoreline observations, increasing temporal resolution relative to traditional tasking. Many existing approaches delineate shorelines by applying a single spectral index threshold, typically NDWI, and contouring the resulting binary mask. We present an alternative, fully probabilistic method. An ensemble of multilayer perceptrons (MLPs) is trained to predict, for each pixel, the probability of “water” versus “land.” The shoreline is then extracted as an isoprobability contour, eliminating the need for a global threshold and allowing spatial variability in sensor response, illumination (e.g., shadows), and local geomorphology to be accommodated. Applied to PlanetScope imagery at Seaford, UK, and evaluated against a height contour referenced to the instantaneous water level, the method achieves a root-mean-square error of ≈7m for scenes with<50% cloud cover. These results indicate that probabilistic pixel wise classification, coupled with high-cadence PlanetScope acquisitions, offers robust shoreline detection suitable for high-frequency coastal monitoring.