Coastline Automatic Extraction From Sentinel-2 Images: Different Methodological Approaches In Comparison

Remote sensing techniques are useful for acquiring coastline position because they allow data to be collected in dangerous or inaccessible areas without disturbance, offer frequent updates, and reduce costs and time compared to land-based collection. Data acquired by satellite, aircraft, or drones provide a broad and detailed view, facilitating the monitoring of coastal changes over time. Automating the process of coastline detection in satellite images dramatically reduces processing time compared to manual methods, enabling the analysis of large areas over long periods. Automation facilitates the management and organization of large amounts of data and helps speed up the analysis of time series of images to study phenomena such as erosion or sediment accumulation. The literature describes various methodological approaches for the automated extraction of coastlines from multispectral optical images. However, a common thread is generally found: the use of unsupervised classification algorithms such as K-means or Iterative Self-organizing Data Analysis Techniques (ISODATA). These are applied either directly to the images or after processing them. This paper compares three different methodological approaches for automatically extracting coastlines from Sentinel-2 multispectral images. The study area includes the coastline around the Gulf of Gaeta (Italy). All three methods analyzed use ISODATA to distinguish water from non-water pixels in the scene. Unsupervised classification is performed starting from: the synthetic image obtained via Normalised Difference Water Index (NDWI), the first component provided by the Principal Component Analysis (PCA) application and the three multispectral images selected based on the Optimal Index Factor (OIF) index that certifies the maximum level of decorrelation between the bands. The results are compared taking into account the level of accuracy achieved in the three cases analyzed. Test sites are used to construct the confusion matrix and calculate thematic accuracy indices. A comparison is also made between the position of the line manually vectorized through photointerpretation of the scene true-color Red Green Blue (RGB) composition, and the position of the line obtained as result of each methodological approach. Both tests prove a higher level of accuracy for the PCA-based method.