High Resolution Forest Maps from Interferometric TanDEM-X and Multitemporal Sentinel-1 SAR Data

Abstract

In this study, a workflow for a semi-automated forest/non-forest detection is proposed that is based on multitemporal Sentinel-1 ground range detected (GRD) C-band backscatter and TanDEM-X Coregistered Single look Slant range Complex (CoSSC) X-band imagery and an unsupervised random forest classification approach. Therefore, numerous features that refer to frequency, polarisation, and texture were extracted from SAR data of different seasons. The aim was to develop a processing scheme that is feasible for semi-automated forest mapping and monitoring from SAR data at high spatial resolution and on annual scale. It was tested for seven study sites in Germany and Canada which represent different biomes and forest types. Results were validated against field observations and existing forest maps. The best performance for the German study sites was achieved with multitemporal Sentinel-1 backscatter data from the onset of the growing season with small incidence angle and VH polarisation, together with extracted textural features and TanDEM-X data. Producer’s accuracies for the forest class of the different study sites ranged from 88.4 to 98.0%. User’s accuracies ranged from 85.5 to 87.0%. Using Sentinel-1 data covering the whole growing season at a 12 day repetition rate, ascending and descending orbits and VV and VH polarisations led to comparable results. Limited data availability for the Canadian study sites resulted in on average to less reliable results than at the German sites with a higher range of producer’s (62.4–98.8%) and user’s accuracies (46.2–90.2%).