Potential of Sentinel-1 time series for deforestation and forest degradation mapping in temperate and tropical forests

We want to present results of the Sentinel4REDD project. The overall aim of the Sentinel4REDD-Project is the development of new remote sensing based methods using Sentinel-1 and Sentinel-2 data to support UNFCC (United Nations Framework Convention on Climate Change) REDD+ MRV (Measurement, Reporting and Verification) Systems. We show results on the example of two testsites located in Mexico, one is over deciduous and evergreen forests in Hidalgo and the other is over tropical dry forests in Yucatan. For a signature analysis, we first selected forested, deforested and degraded areas based on visual interpretation of multi-temporal very high resolution (1 m) optical imagery. We developed a new speckle filter which retains the spatial resolution by using the temporal domain. We plotted filtered Sentinel-1 time series for the three reference classes and were able to determine the time frame of deforestation and forest degradation. The initial analyses showed promising results regarding the separation of forest and forest-change classes with filtered Sentinel-1 data in contrast to original SAR backscatter images. To generate REDD+ products (forest/non-forest, deforestation, forest degradation, reforestation maps), we apply two approaches using dense time series from Synthetic Aperture Radar (SAR) and optical sensors. The first method is based on multi-temporal metrics, e.g., mean, standard deviation and different percentiles of the SAR backscatter and surface reflectances for different time periods. The second is based on different seasonalities of land cover classes. For this, the time series of every pixel is decomposed into subsignals with differing temporal frequencies. From these subsignals different statistics are calculated. These statistics can then be used to derive forest/non-forest maps for different time periods to then obtain resulting deforestation and reforestation maps or to directly gain deforestation and degradation maps via the application of machine learning techniques. Furthermore, we present preliminary deforestation maps for study sites in Mexico and South Africa based on Bayesian probability approach and filtered Sentinel-1 backscatter time series