Nowadays, access and manipulation of satellite imagery timeseries over large geographical scale is greatly by cloud-based datacubes platforms (Google Earth Engine, Digital Earth Australia). Satellite derived shorelines (SDS) are produced mainly using water indices, CoastSat or Convolutional Neural Networks (CNN). For robust change analysis, SDS are topographically corrected by estimating the water level at the time of satellite overpass. However, very few researchers have evaluated the accuracy of SDS against independent in-situ groundtruth shorelines due to the difficulty to obtain synchronous three-dimensional datasets of the monitored sites.
We use 47 surveyor-grade drone-derived digital surface models of 15 locations across Victoria acquired within a few hours from Sentinel-2 overpass to assess the 5 most common water indices, CoastSat and a new end-to-end CNN solution. We also tested the effects of using different wave data quality (nearshore buoys, offshore altimeters and hindcast model) and run-up models to estimate water levels. We found that the best methods are spatially and temporally extremely variable, ranging from 2 to sub-pixel (<10m) errors. Interestingly, we found that an hypothetical ‘mixed’ SDS composed by the pixel-wise best method is potentially able to have sub meter accuracy in some locations, hinting that a lot more can be done to improve space-derived shorelines extraction.