Sub-Pixel Waterline Extraction: Characterising Accuracy and Sensitivity to Indices and Spectra

Abstract

Accurately mapping the boundary between land and water (the 'waterline') is critical for tracking change in vulnerable coastal zones, and managing increasingly threatened water resources. Previous studies have largely relied on mapping waterlines at the pixel scale, or employed computationally intensive sub-pixel waterline extraction methods that are impractical to implement at scale. There is a pressing need for operational methods for extracting information from freely available medium resolution satellite imagery at spatial scales relevant to coastal and environmental management. In this study, we present a comprehensive evaluation of a promising method for mapping waterlines at sub-pixel accuracy from satellite remote sensing data. By combining a synthetic landscape approach with high resolution WorldView-2 satellite imagery, it was possible to rapidly assess the performance of the method across multiple coastal environments with contrasting spectral characteristics (sandy beaches, artificial shorelines, rocky shorelines, wetland vegetation and tidal mudflats), and under a range of water indices (Normalised Difference Water Index, Modified Normalised Difference Water Index, and the Automated Water Extraction Index) and thresholding approaches (optimal, zero and automated Otsu's method). The sub-pixel extraction method shows a strong ability to reproduce both absolute waterline positions and relative shape at a resolution that far exceeds that of traditional whole-pixel methods, particularly in environments without extreme contrast between the water and land (e.g., accuracies of up to 1.50-3.28 m at 30 m Landsat resolution using optimal water index thresholds). We discuss key challenges and limitations associated with selecting appropriate water indices and thresholds for sub-pixel waterline extraction, and suggest future directions for improving the accuracy and reliability of extracted waterlines. The sub-pixel waterline extraction method has a low computational overhead and is made available as an open-source tool, making it suitable for operational continental-scale or full time-depth analyses aimed at accurately mapping and monitoring dynamic waterlines through time and space.

Beach Slopes From Satellite‐Derived Shorelines

“…(b) Modeled tide levels associated with the satellite‐derived shorelines (black line). The gray shaded area indicates the overall tidal fluctuations, noting that the vertical bias here is due to the sun‐synchronous orbit of Landsat satellites (refer to Bishop‐Taylor, Sagar, Lymburner, Alam, & Sixsmith, 2019). (c) Ensemble of tidally corrected time series of shoreline change using slope values ranging from 0.01 (red) to 0.2 (green).…”

Mapping and Monitoring the Multi-Decadal Dynamics of Australia’s Open Waterbodies Using Landsat

“…This is because water management decisions are typically based on waterbodies (dams, lakes, river reaches, refugial pools) rather than pixels. Waterbody mapping with satellite imagery has almost exclusively been done on a pixel-by-pixel [12,13,18,20,21,23,24] (or sub-pixel [25][26][27][28] basis), with only a few studies delineating waterbodies as vector objects [29][30][31][32]. The delineation of waterbodies provides an object-based analysis, which characterises the dynamics of the whole waterbody, not its individual pixels.…”

“…Since DEA Waterbodies uses an automatic workflow utilising Landsat data to objectively map waterbodies, we have chosen to leave the final waterbody line work pixelated to transparently identify which Landsat pixels have been mapped inside (and outside) of each waterbody (Figure 12). Future iterations of DEA Waterbodies could make use of sub-pixel methods [25][26][27][28] or higher resolution imagery to better capture the true waterbody shape.…”

Estimating inundation of small waterbodies with sub‐pixel analysis of Landsat imagery: long‐term trends in surface water area and evaluation of common drought indices

“…Many methods are used to measure hydrologic variation in waterbodies smaller than a single Landsat pixel (i.e. sub‐pixel methods), including regression trees (Rover et al, 2010; Huang et al, 2014; Jin et al, 2017), discrete particle swarm optimization (Li et al, 2015), Dynamic Surface Water Extent (DeVries et al, 2017; Jones, 2019), and spectral unmixing methods (Halabisky et al, 2016; Liu et al, 2017; Bishop‐Taylor et al, 2019; Hong et al, 2019). Linear spectral mixture analysis, a spectral unmixing technique, estimates the relative abundance of individual components (spectral endmembers) on a pixel‐by‐pixel basis based on their unique spectral characteristics (Hu et al, 1999; Heinz and Chang, 2001), so fractional abundances can be converted to surface areas (Halabisky et al, 2016).…”