Global hybrid forest mask: Synergy of remote sensing, crowd sourcing and statistics

Many global and regional forest cover products have recently become available. The most advanced and comprehensive of these include the global land cover datasets (GLC2000, MODIS, GLOBCOVER), MODIS Vegetation Continuous Fields (VCF), LANDSAT based (e.g. Sexton et al., 2013) and radar based (e.g. Saatchi et al., 2010; Baccini et al., 2012; Santoro et al., 2012) products. However, they often contradict each other and are typically inconsistent with forest statistics. In particular, global land cover datasets contradict each other in many areas, have limited information about forest density and are not consistent with forest statistics. VCF most likely provides the most comprehensive information about forest density with a spatial resolution of 230m during 2000-2010. However when observing VCF dynamics for individual pixels, one can see variation that cannot be explained by forest cover dynamics, but instead by unstable pixel geometry and clouds. Landsat based products also suffer from cloud cover and cannot recognize sparse forest with canopy closure of 30% or less. Space-based radar is free from cloud, but still cannot reliably delineate areas as forest/non forest (Santoro, 2012). We compare all of the above mentioned remote sensing products with a sample of high resolution imagery provided by Google Earth. We have applied the crowd sourcing platform Geo-Wiki (Fritz et al., 2010, 2012) to collect 22K training points where the percentage of forest cover was estimated for a 1km pixel size. We applied the method of geographically weighted regression to calculate the map of probability of forest cover and the map of forest share. This involved the use of the Geo-Wiki training points in combination with the land cover products, MODIS VCF and LANDSAT. The synergy of remote sensing, statistics and crowd sourcing approaches was investigated to better understand the spatial distribution of forests. Both calibrated (using FAO FRA statistics) and non-calibrated (.best guess.) forest cover datasets were obtained based on this method. We compared the Geo-Wiki training points with initial datasets and the final hybrid forest cover (Table). 8505 (out of 22K) Geo-Wiki points were classified as forest. The hybrid product shows a very close approximation in terms of the amount of forest pixels while other datasets vary from -28% to +62%. The hybrid dataset shows the highest overall agreement (when 22K points are compared) -89%. Globcover and VCF have the highest forest agreement (8505 forest points compared) because of overestimates of forest area. The highest correlation (R2=0.76) is obtained with the hybrid dataset when comparing the forest share per pixel. The hybrid forest cover is currently under validation and available for visualization at http://biomass.geo-wiki.org.