Mapping canopy cover in African dry forests from combined use of Sentinel-1 and Sentinel-2 data: application to Tanzania for year 2018

The monitoring of tropical forests has benefited from the increased availability of high-resolution earth observation data. However, the seasonality and openness of the canopy of dry tropical forests remains a challenge for optical sensors. The availability of time series of remote sensing images at 10-meters is changing this paradigm. In the context of REDD+ national reporting requirements, we investigate a methodology that is reproducible and adaptable in order to ensure user appropriation. The overall methodology consists of three main steps: (i) the generation of Sentinel-1 (S1) and Sentinel-2 (S2) layers, (ii) the collection of an ad-hoc training/validation dataset and (iii) the classification of the satellite data. Three different classification workflows are compared in terms of their capability to capture the canopy cover of forests in East Africa. The method is tested at scale, over Tanzania. Two big data computing platforms are combined to exploit the important volume of satellite data available over a yearly period. The study also explores the accuracy of two products derived from these mapping approaches: i) binary tree cover/no tree cover (TC/NTC) map, and ii) map of canopy cover classes. We demonstrate the potential of the combination of S1 SAR and S2 optical sensors to derive an accurate map of forest cover in East Africa at a spatial resolution of 10 meters for the year 2018. Our approach uses the high temporal resolution of S2 that allows to produce bimonthly cloud-free composites that reflect the seasonality of the vegetation. An overall accuracy (OA) of about 95% is reached for the TC/NTC map. When mapping different categories of forest and canopy cover, we obtain an OA over 85% both with a per pixel accuracy approach and when considering the neighbouring pixels in the classification training. The potential of S1 and S2 data for single trees discrimination is also assessed. The reference dataset (training and validation), the three best maps and the codes to produce the S1 and S2 composites on Google Earth Engine are shared here