Assessing forest/non-forest separability using Sentinel-1 C-band SAR

Abstract

Synthetic Aperture Radar data has a unique potential for continuous forest mapping as it is not affected by cloud cover. While longer wavelengths such as L-band are commonly used for forest applications, in this paper we assess the aptitude of C-band Sentinel-1 data for this purpose, for which there is much interest due to its high temporal resolution (5 days) and `free, full and open' data policy. We tested its ability to distinguish forest from non-forest in six study sites, located in Alaska, Colombia, Finland, Florida, Indonesia, and the UK. Using the time series for a full year significantly increases the classification accuracy compared to a single scene (a mean of 84.6% compared to 76.8% across the study sites). Our results show that we can further improve the mean accuracy to 86.4% when only considering the annual mean and standard deviation of VV and VH backscatter. In this case, separation accuracies of up to 93% (in Finland) are possible, though in the worst case (Indonesia) the highest possible accuracy using these variables was 82%. The best overall performance was observed when using a Support Vector Machine classifier, outperforming random forest, k-Nearest-Neighbors and Quadratic Discriminant Analysis. We further show that the small information content we found in the phase data is an artifact of terrain slope orientation and has a negligible impact on classifier performance. We thus conclude that for the purposes of forest mapping the smaller file size and easier to process GRD data is sufficient, with little benefit to downloading the SLC data. Possible uses of the phase data in this context relate to its temporal coherence, which was not tested in this study.