Performance of segmentation (Watershed and Meanshift) and image transformation (MNF-Laplacian filter) methods for extracting complex coastlines from Pleiades images: the case of the Kerkena archipelago, Tunisia

Luc Simirore Diatta^1*Katia Schörle^2*Luc Lapierre^3*

• ¹UMR7330 Centre Européen de Recherche et d'enseignement de Géosciences de l'environnement (CEREGE), Aix En Provence, Provence-Alpes-Côte d'Azur, France
• ²CNRS UMR 7299 Centre Camille Jullian, Aix-en-Provence, France
• ³SFPT - Société Française de Photogrammétrie et Télédétection, Paris, France

With the ongoing surge in global coastal development, understanding shoreline dynamics has become a critical issue, given the inherent vulnerability of coastal fringes to significant mobility.Developing tools to support the sustainable management and future planning of these areas requires a robust comprehension of their dynamic behavior. Monitoring shoreline changes through coastline extraction using remote sensing is vital for quantifying the diachronic evolution of shorelines. However, the accuracy of coastline extraction methods can be hindered by various factors, including the quality of geospatial data, the characteristics of the study area, and the adequacy of pre-processing techniques applied. This study evaluates the performance of different coastline extraction methods based on the segmentation (Watershed and Meanshift) and transformation and discrimination (MNF-Laplacian filter) of very high spatial resolution Pleiades images resampled to 0.5 m. This evaluation of the performance of the automatic extraction methods was carried out by comparison with manually Digitized coastlines across different types of coastlines, a methodology that could be applied to other study areas with similar characteristics.The analysis is based on the mean distances and mean differences of the statistics obtained from DSAS (Digital Shoreline Analysis System) with Shoreline Change Envelope (SCE), Net Shoreline Movement (NSM) and End Point Rate (EPR) indices, which quantify the variations in the reference line detected by each method as well as the diachronic changes in the shoreline over 10 years (2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020)(2021)(2022). The results show that the extraction method based on the WaterShed algorithm is the most accurate compared with coastlines obtained by manual extraction. It enables the shoreline to be detected perfectly on developed coasts and sandy coasts. On cliffs, the MeanShift and MNF (Minimal Noise Fraction)-Laplacian filter algorithms perform better. Detecting the coastline on cliffs is complex, due to the shadow of the cliffs caused by the sensor's acquisition angle, and the over-segmentation of the images. The method based on the MNF-Laplacian filter combination performed best, with 98.8% of coastline extracted. Taking into account the coastline extracted by the best-performing method for each type of coastline, we could determine an average retreat of the shoreline of -0.33 m/year over 10 years.