Super-resolution generative adversarial network for Landsat shoreline extraction: Training strategy effects on spectral consistency and accuracy

Despite advances in shoreline extraction techniques, most satellite-derived shoreline (SDS) approaches rely on the native spatial resolution of Landsat imagery (30 m), which constrains shoreline delineation accuracy. Although Landsat-8/9 support resolution enhancement through their 15-m panchromatic band, this study instead examines a learning-based alternative that is transferable to sensors lacking a panchromatic channel. Using Landsat-8/9 as a test case, we evaluate deep learning-based super-resolution for shoreline extraction along the east coast of Gangwon Province, South Korea. A super-resolution generative adversarial network was trained using two strategies: a cross-sensor strategy with Landsat and Sentinel-2 pairs, and a synthetic strategy using downsampled Sentinel-2 imagery. Shorelines were extracted using CoastSat, a pixel classification-based method, and validated against unmanned aerial vehicle reference shorelines at two validation sites along the study coast. The spectral consistency of super-resolved imagery proved critical for reliable shoreline extraction. The cross-sensor strategy achieved higher peak signal-to-noise ratio but introduced high-frequency variability that degraded extraction performance. Meanwhile, the synthetic strategy produced more consistent outputs, reducing the mean absolute error and root mean square error by 28% and 30% at one site and by 54% and 54% at the other; block-bootstrap confidence intervals and paired tests confirmed these gains are robust. These findings suggest that super-resolution can improve SDS accuracy when spectral consistency is preserved, extending resolution enhancement to sensors that lack a panchromatic band.