Discriminating treed and non-treed wetlands in boreal ecosystems using time series Sentinel-1 data

Wetlands are recognized for their importance to a range of ecosystem goods and services; however, detailed information on wetland presence, type, extent, and persistence is challenging to attain over large areas and/or long time periods due to the spatial complexity and temporal dynamism of wetlands. In this study we explored the potential for within-year time series of C-band Synthetic Aperture Radar (SAR) observations from the free and open Sentinel-1 data archive to improve discrimination of treed and non-treed wetlands and non-wetlands in a boreal forest environment. Through a set of 3843 classification experiments for the year 2017, we tested the influence of three factors on classification accuracy: (i) input features (two backscatter coefficients in VV and VH polarization (σVV and σVH) and four quantitative measures derived from the Stokes vector); (ii) the temporal form of features (i.e. using all within-year observations versus generalized measures such as monthly/seasonal means or annualized statistics); and (iii) missing observations in Sentinel-1 time series due to varying observation availability across space. Among the tested features, we found the greatest utility in σVV and σVH. Directly using all within-year observations yielded higher accuracy than using generalized temporal forms. Moreover, the temporal form of the features had a greater impact on classification accuracy than the features themselves. The highest overall accuracy (0.860 ± 0.002) was achieved using σVV and σVH from all within-year observations. The majority of class confusion occurred between treed wetlands and non-wetlands. We found no significant reduction in the overall accuracy by simulated missing observations in time series when using all within-year observations. With the increasing availability of free and open data from the Sentinel-1 archive, new opportunities are emerging to readily integrate within-year time series into large-area land cover mapping, particularly if analysis-ready SAR data products further reduce preprocessing requirements for end users.

Abstract licence: CC BY-NC

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.