Machine Learning Approaches for Coastline Extraction from Sentinel-2 Images: K-Means and K-Nearest Neighbour Algorithms in Comparison

Abstract

AbstractIn the last decades, machine learning techniques have increasingly spread to many application fields of remote sensing and, more recently, have also involved the extraction of the coastline from satellite images. The presence of different machine learning algorithms as well as the availability of different types of remote sensing data, make it necessary to further investigate in order to identify methodological solutions for providing accurate results. This article aims to compare two alternative and typical methodological approaches of machine learning, one unsupervised, the other supervised, represented respectively by the K-Means (KM) and K-Nearest Neighbour (KNN) algorithms. The experiments are conducted on Sentinel-2 satellite images, limited to the bands with the highest geometric resolution (10 m). The dataset includes also the image resulting from the application of the Normalized Different Water Index (NDWI), which is particularly effective for distinguishing water/non-water. The coastline obtained by manual vectorization on the Sentinel-2 RGB composition is the term of comparison for evaluating the result accuracy. The Distributed Ratio Index (DRI) is applied for this purpose. The use of training sites with the KNN method allows to obtain a more reliable classification in the presence of multiple spectral bands. On the contrary, using only the NDWI layer the KM method produces better results, demonstrating how in this case the land-sea distinction is clearer and the automatic clustering, as it is not affected by human error that accompanies the detection of the training sites, is more reliable.KeywordsCoastline extractionMachine learningK-MeansK-Nearest NeighbourSentinel-2Accuracy evaluation