From percent tree cover to categorical forest cover and change: propagating the uncertainty in detecting forest disturbance

Friday, 19 December 2014
Joseph O Sexton1,2, Praveen Noojipady3, Anupam Anand4, Xiao-Peng Song2, Sean McMahon5, Chengquan Huang6, Min Feng2, Saurabh Channan2 and John R Townshend7, (1)University of Maryland, College Park, MD, United States, (2)Global Land Cover Facility, Univeristy of Maryland, College Park, MD, United States, (3)National Wildlife Federation Reston, Reston, VA, United States, (4)Global Land Cover Facility, College Park, MD, United States, (5)SERC, Edgewater, MD, United States, (6)University of Maryland College Park, Geographical Sciences, College Park, MD, United States, (7)University of Maryalnd- College Park, College Park, MD, United States
Several remotely sensed datasets are now available to detect and analyze forest disturbances. Each of these regional and global datasets is estimated with varying degrees of systematic and unsystematic error (i.e., uncertainty), and so rigorous analysis of forest disturbance requires detected changes to be accompanied by estimates of error. Currently, errors are assessed predominantly by external, post hoc validation—i.e., comparison of detected changes to independent observations that are assumed to be true. However, validation yields only broadly aggregated, regional summaries of error but does not describe its spatial and temporal variation. Alternatively, modeling the propagation of error through the processes of estimation, classification, and change-detection provides an internal means to infer classification and change-detection error at a scale equivalent to that of cover and change—i.e., at pixel resolution and global extent. Using sequential observations from the first global, Landsat-based dataset of fractional tree cover, we derive and demonstrate a rigorous, probabilistic method for propagating the uncertainty of fractional tree-cover estimates to categorical forest-cover classification and change detection. Modeling tree cover as a probability distribution in each pixel represents estimation error locally and enables its transmission into inferences such as change detection. Propagation thus enables mapping and visualization of the spatial distribution of uncertainty resulting from estimation of tree cover and from selection of the threshold of tree-cover to define “forest”. The approach described here provides a useful, per-pixel description of forest classification and change-detection certainty and can accommodate any definition of “forest” based on tree cover—an especially important consideration for monitoring disturbance, given the variety of institutional definitions of forest cover based on remotely sensible characteristics.