Using TimeSync and a New Statistical Method to Validate a Forest Disturbance Model
Abstract:VeRDET, a new automated method for remotely sensing forest disturbance through time using Landsat imagery, provides two different types of data for each pixel on the landscape: the year in which disturbance or regeneration events occur (vertex) and the severity of this event (slope of the line between vertices). We developed a new method for simultaneously evaluating the fit of this data to data derived from an expert interpreter.
The study focused on eastern deciduous forests in the southeastern US, with 192 pixels selected for error evaluation over 25,000 sq km in TN, NC, VA and KY. Land cover change for those pixels from 1984-2011 was determined by an expert using TimeSync, a tool that allows a user to select start and end dates of disturbance, regenerating and stable periods by examining Landsat time series and other data sources (Google earth, aerial detection surveys, localized studies and data, and their own expert knowledge). For each evaluation pixel and for each of four vegetation indices (Tassel-Cap Angle, NDVI, NBR, and NDMI), the disturbance/regeneration trajectories identified by VeRDET were compared with those expert-determined trajectories by calculating the mean squared error (MSE) between the two fit lines. We then generated 1000 random trajectories for each pixel, using resampling to generate the number of vertices, assigning verteces to random years, fitting a piecewise linear regression, and calculating the MSE of these random trajectories as compared to the expert-determined trajectories. The number of random trajectories with a higher MSE than VeRDET, converted to a percentile, is used as the dissimilarity score. This approach takes into account both the agreement in position of vertices as well as the shape of the resulting trajectory, and weights subtle changes less than large ones.
Comparing the four vegetation indices in periods of disturbance, regeneration and stability, NDMI was found to best match the expert interpretation, with VeRDET outperforming 80% of the random null models 89% of the time. Segmentations using NBR and NDVI performed less well overall and Tassel-Cap Angle performed the worst.