Applying shape selection methods to Landsat time series for mapping forest disturbance history and cause

Abstract:

In the US, nationwide processing of historic Landsat data has recently been completed to provide a comprehensive annual, wall-to-wall analysis of US disturbance history and cause 1982-2011. Here, we present new methodology used in this effort that involves fitting nonparametric shape-restricted regression splines to Landsat spectral trajectories which are sensitive to both forest structure (eg., Band 5) and leaf area (eg., NDVI ). The functions, developed in R and automated for nationwide processing, deliver a series of parameters for each pixel and band that identify the best (of seven possible) shapes of the spectral pattern, year(s) of inflection, magnitude, and pre- and post- inflection rates of growth or recovery. Parameters from model fits are then fed into random forests models along with static bioclimatic and vegetation data layers to predict causal agent through time. Results indicate shape parameters are the driving predictors in diverse Landsat scenes across the US, and are able to capture subtle patterns produced by low magnitude or slow change such as insect and disease. We discuss the interaction between shape parameters in the empirical models, and illustrate how the intermediate product of the fitted trajectories themselves can be used as ancillary data in model-assisted estimates of forest attributes through time.