Mapping tropical disturbed forests using multi-decadal 30 m optical satellite imagery

Abstract

Tropical disturbed forests play an important role in global carbon sequestration due to their rapid post-disturbance biomass accumulation rates. However, the accurate estimation of the carbon sequestration capacity of disturbed forests is still challenging due to large uncertainties in their spatial distribution. Using Google Earth Engine (GEE), we developed a novel approach to map cumulative disturbed forest areas based on the 27-year time-series of Landsat surface reflectance imagery. This approach integrates single date features with temporal characteristics from six time-series trajectories (two Landsat shortwave infrared bands and four vegetation indices) using a random forest machine learning classification algorithm. We demonstrated the feasibility of this method to map disturbed forests in three different forest ecoregions (seasonal, moist and dry forest) in Mato Grosso, Brazil, and found that the overall mapping accuracy was high, ranging from 81.3% for moist forest to 86.1% for seasonal forest. According to our classification, dry forest ecoregion experienced the most severe disturbances with 41% of forests being disturbed by 2010, followed by seasonal forest and moist forest ecoregions. We further separated disturbed forests into degraded old-growth forests and post-deforestation regrowth forests based on an existing post-deforestation land use map (TerraClass) and found that the area of degraded old-growth forests was up to 62% larger than the extent of post-deforestation regrowth forests, with 18% of old-growth forests actually being degraded. Application of this new classification approach to other tropical areas will provide a better constraint on the spatial extent of disturbed forest areas in Tropics and ultimately towards a better understanding of their importance in the global carbon cycle.