ACCURACY EVALUATION OF COASTLINE EXTRACTION METHODS IN REMOTE SENSING: A SMART PROCEDURE FOR SENTINEL-2 IMAGES

Abstract. Different algorithms are available in literature to extract coastline from remotely sensed images and different approaches can be adopted to evaluate the result accuracy. In every case, a reference coastline is suitable to compare alternative solutions: usually, the visual photointerpretation on the RGB composition of the considered imagery and the manually vectorization of the coastline allow an accurate term of comparison, but they are laborious and time consuming. This article aims to demonstrate that a smart procedure is possible using a LiDAR-generated Digital Elevation Model (Lg-DEM) as a useful source from which to rapidly extract the reference coastline. The experiments are carried out on Sentinel-2 imagery, using six indices: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), Enhanced Vegetation Index (EVI), Red-Green Ratio (RGR) and NIR-Red Ratio (NRR). The unsupervised classification algorithm named K-Means transforms each index resulting product in two clusters, i.e. water and no-water, while the automatic vectorization allows to detect the coastline as separation between land and sea. The coastline from Lg-DEM and the manually achieved one using photointerpretation are both assumed as references for testing result accuracy. In every case, the performance analysis of the six indices products induces similar results, confirming the combination of NDWI and K-Means as the most performing approach. The tests demonstrate that, when Lg-DEM and satellite images concern the same area in the same period or in absence of variations, the coastline extracted from Lg-DEM is useful as reference to compare various methods.