Characterisation of Wetland Types Using Sentinel-1 Radar Time Series Data

Abstract

During the last decades, wetlands have been determined as one of the most valuable ecosystems on Earth. Despite their importance for both humanity and nature, they are also one of the most rapidly degrading land cover types. In order to implement and evaluate effective policy for wetland preservation, large-scale monitoring and characterisation of wetlands is needed. The use of satellite-based remote sensing techniques has proven its use for this purpose. However, wetland mapping and characterisation by using remote sensing is challenging.

The recently launched Sentinel-1 satellites acquire radar images with a relatively high spatial and temporal resolution, using C-band dual-polarimetric (VV/VH) sensors. This new data-rich information source provides a unique opportunity for more accurate wetland monitoring from space. In this research, the temporally dense Sentinel-1 radar time series data was applied for wetland characterisation and its use was assessed. The combination of Sentinel-1 and Sentinel-2 data was also applied to additionally assess the use of Sentinel-1 when combined with optical satellite data.

In order to assess the use of Sentinel-1 data for wetland characterisation, four different machine learning classifications were done in the St. Lucia wetlands in South Africa, based on a classification scheme with three levels of wetland characterisation: (1) general wetland delineation, (2) the classification of wetland vegetation types and (3) the classification of surface water dynamics. The sole use of Sentinel-1 and the combined use of Sentinel-1 and Sentinel-2 were both applied. As the C-band radar system aboard Sentinel-1 was expected to have limited capabilities in mapping high-vegetated wetlands, an additional set of classifications was done excluding these high-vegetated wetlands. Each classification was done in a Monte Carlo simulation of 100 Random Forests in order to obtain reliable results.

It was found that Sentinel-1 radar data is useful for mapping low- to medium-vegetated wetlands. However, it is incapable of distinguishing high-vegetated wetlands from upland forests. The combined use of Sentinel-1 and Sentinel-2 delivered significant accuracy improvements compared with the sole use of Sentinel-1. The value of Sentinel-2 was especially observed for general wetland delineation.