Hyper-temporal Water Body Dynamics Mapping using Sentinel-1 Time Series Clustering

13 Nov 2018, 14:20
20m
Soil and Hydrology Soil & Hydrology Session

Speaker

Mikhail Urbazaev (Friedrich-Schiller-University Jena)

Description

This study aims at providing water body dynamics maps at highest temporal resolution for better wetland characterization from space and as such is a mapping component of H-2020 project Satellite-based Wetland Observation Service (SWOS) (Grant No 642088). By depending on freely available data and demanding a high acquisition frequency across a large number of test sites from the North of Sweden to Central Africa, the Sentinel-1 SAR mission was chosen as the primary source of satellite images. Data from this mission acquired in Ground Range Detected (GRD) Interferometric Wide Swath (IW) mode VV polarization was selected as suited best for the mapping task due to storage volume (as compared to SLC data) and acquisition frequency (as compared to VH and HH polarization).
While SAR offers great possibilities for mapping water bodies from space, due to its insensitivity to clouds and independence from sunlight, difficulties arise from varying sea states. A roughened sea surface, returning more radiation to the sensor than a smooth surface, might be indistinguishable from land in the image. This effect of reduced contrast between land and water is particularly prominent in images from short wavelength SAR systems like X-Band and C-Band. Thus, common thresholding approaches like Otsu and Kittler-Illingworth, which assume a bimodal distribution of gray values in the image for distinguishing between dark water pixels and brighter land pixels, are likely to be inaccurate. However, by incorporating e.g. interferometric and polarimetric techniques in combination with change detection approaches, a high accuracy for flood detection can be achieved regardless of wavelength. Yet, many of these methodologies are impossible to use on large spatial scales and high temporal frequency due to data scarcity and/or cost.
The choice of data, a single polarization C-Band VV setup, is not the most favorable for per-image water mapping due to the explained sensitivity to sea state, precluding land-water thresholding as described above for a large number of acquisitions. Thus, it was decided to perform the classification purely in the temporal domain making use not only of image intensity but also its variability over time. In this context a bright pixel resulting from high sea state can be seen as an outlier in an otherwise series of dark backscatter pixel acquisitions and thus be discarded from being classified as land unless it is succeeded by other bright pixels indicating a change from water to land. This combination of pixel value and time series neighbor relationships is transformed into a synthetic data space and subsequently clustered using a simple K-means approach for all pixels. After restoring the original time series with the clustering result a classification of land and water is achieved for each pixel in time.
Though being clustered/classified completely independent from each other, the pixels from the same acquisition form a low-noise and accurate classification map even at highest sea states. This approach has proven highly successful under different conditions in the observed wetlands while being conceptually simple and completely automated.

Primary author

John Truckenbrodt (Friedrich-Schiller-University Jena, Institute of Geography, Department for Earth Observation)

Co-authors

Mr Victor Onyango Odipo (Friedrich-Schiller-University Jena, Institute of Geography, Department for Earth Observation) Mr Jan Bongard (Friedrich-Schiller-University Jena, Department for Earth Observation) Mr Stefan Werner (Friedrich-Schiller-University Jena, Institute of Geography, Department for Earth Observation) Dr Marcel Urban (Friedrich-Schiller-University Jena, Institute of Geography, Department for Earth Observation) Prof. Christiane Schmullius (Friedrich-Schiller-University Jena, Institute of Geography, Department for Earth Observation) Mrs Kathrin Weise (Jena-Optronik GmbH)

Presentation materials

There are no materials yet.