Towards accurate mapping of forest in tropical landscapes: A comparison of datasets on how forest transition matters

Tropical forests represent half of the Earth's remaining forest area, but they are shrinking at high rates, which poses a threat to their multiple ecosystem services. As a response, international environmental agreements and related programs require information about tropical forested landscapes. Despite the increasing quantity and quality of remote sensing-based data, the effective monitoring of forests in the tropics still faces operational challenges: (a) applicability at local levels, with lack of reference or cloud-free information; (b) overcoming geographical, ecological, or biophysical variability; (c): stratification, distinguishing forest categories related to functionality and disturbance history. We conducted an extensive ground verification campaign through 36 landscapes in 9 regions of Zambia, Ecuador and Philippines, which constitute a gradient of pantropical deforestation contexts or forest transitions. We collected over 16,000 ground control points and digitized over 18,000 ha with details on land use and forest disturbance history. We trained a random forest algorithm and generated high-resolution (30 m) binary forest maps covering ~15 Mha, building on 39 optical (Landsat-8), radar (Sentinel-1) and elevation bands, indices and textures. We validated the quality of the outputs across the studied deforestation gradient and compared them to (a): 3 national land cover maps used for international reporting, (b): 4 global forest datasets (Global Forest Change, Copernicus Land Cover, JAXA and TanDEM-X Forest/Non-Forest). Our method generated highly accurate (92%) forest maps for the studied regions when compared to the global datasets, which generally overestimated forest cover. We achieved accuracies similar to the national maps, following a standardized method for all countries. The difficulties in delineating forest increased in more advanced stages of deforestation, with recurring struggles to distinguish non-forest tree-based systems (e.g. perennials, palms, or agroforestry), shrublands and grasslands. Regrowth forests were repeatedly misclassified across contexts, countries and datasets, in contrast to reference or degraded forests. Our results highlight the importance of in situ verification as accompanying method to establish efficient forest monitoring systems, especially in areas with higher rates of forest cover change and in tropical regions of advanced deforestation or early reforestation stages. These are precisely the areas where current REDD+ or Forest Landscape Restoration initiatives take place.