Water Surface Extraction From Single-Date Sentinel-2 Imagery Using Object-Based Random Forest Classification

Abstract

Inland water bodies are among the most important natural resources on Earth due to their ecological, economic, and social importance, supporting life, agriculture, biodiversity, and disaster management. Monitoring and mapping inland water bodies from remote sensing imagery has become critical for the sustainability of ecosystems. Traditional water index methods require spatially variable threshold values to obtain accurate results, which represent a limiting factor in their applicability. On the other hand, machine learning techniques achieve highly successful results in detecting water surfaces from remote sensing imagery. In this study, an approach is presented to extract inland water bodies from Sentinel-2 imagery using object-based machine learning classification. The approach was implemented on a study area located in the Lakes Region of Türkiye using a single date Sentinel-2 imagery. The image segmentation process necessary for object-based classification was carried out using the Simple Linear Iterative Clustering (SLIC) superpixel segmentation algorithm. Training samples were selected automatically with the help of the Global Surface Water (GSW) dataset. Image classification was performed in R Studio using the random forest (RF) algorithm. In addition to 10m and 20m bands of Sentinel-2 data, the Normalized Difference Water Index (NDWI) was calculated and used as an additional band in classification. A comprehensive evaluation based on validation samples revealed overall accuracy higher than 98.4% and Kappa value higher than 95.7 %. The achieved results suggest that the presented approach is promissing in water body mapping from Sentinel-2 imagery with very high accuracy. Keywords: Inland Water Bodies Extraction, Sentinel-2, Machine Learning, Random Forest, R Studio, Automatic Training Data, SLIC, Segmentation, Object Based Image Analysis, Classification.