Geometric Accuracy Comparison of Shorelines Derived from Multitemporal Sentinel-2A Imagery with Various Image Spectral Transformations (Case Study: Marine Deposition Coasts in Bantul Regency, Yogyakarta, Indonesia)

Abstract

Shoreline extraction from Landsat imagery is limited to 30m spatial resolution. The presence of Sentinel-2A imagery opens up opportunities for shoreline extraction for larger-scale mapping; thus, it is necessary to investigate the capability and limitations of various spectral transformations in mapping multitemporal shoreline positions. This study presents a methodology for processing Sentinel-2A imagery and methods for semi-automatic shoreline extraction. Experiments were carried out on marine deposition coasts as the coastal physical typology developing on most Indonesian coasts. The 20m Sentinel-2A imagery data acquired in 2015 until 2020 were compared with several spectral transformation methods: Automated Water Extraction Index Non-Shadow (AWEInsh), Automated Water Extraction Index Shadow (AWEIsh), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Water Index (NDWI), Band Ratio (B3/B11), and Single Band (B11). The use of multitemporal imagery aims to test image and method consistencies in extracting shorelines accurately. The image acquisition time for each year was selected by taking into account cloud cover, land-sea contrast boundary (delimited with Otsu’s image thresholding algorithm), and data availability. The accuracy was assessed with small-format aerial photographs of 0.3m spatial resolution. The results indicated that MNDWI produced shorelines with RMSE ranging from 32 to 64 m, which meets the class-3 standard for 1: 50,000 scale maps per the Indonesian Geospatial Information Agency’s (Badan Informasi Geospasial, BIG) Regulation Number 6 of 2018. Therefore, 20m Sentinel-2A can be used as a data source in rapid shoreline extraction. Several factors contributing to RMSE that is 3 to 6 times greater than the spatial resolution of Sentinel-2A images include cloud cover, mixed pixels, foam, and tidal bias, creating data uncertainty values ranging from 28 to 45 m.