Bayesian Land Cover Classification for MODIS Data

• Activity Number: 86
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Luis Eduardo Carvalho*+ and Surajit Ray and Hunter Glanz
• Companies: Boston University and Boston University and Boston University
• Keywords: spatio-temporal prior ; gibbs sampling ; centroid estimator

Abstract

We discuss a Bayesian inferential approach to land cover classification, according to the IGBP categorization, of a Moderate Resolution Imaging Spectroradiometer (MODIS) dataset collected from a number of validated sites. The dataset comprehends time series with 46 observations per year and per site from 2001 to 2009 of an enhanced vegetation index (EVI) computed from MODIS spectral observations. In order to capture spatial and temporal co-dependencies, we adopt a Potts model prior on the latent IGBP classes. We further model a hierarchical Gaussian likelihood structure that represents expected within-year EVI intensities and aims to explain EVIs for specific site, time of the year, and year through between-year and between-site variances. We develop an efficient Gibbs sampler that allows us to conduct inference on IGBP classifications by obtaining centroid estimates and testing for changes, in time, of land cover classes. Finally, we propose directions for future work involving improved MCMC schemes and an explicit change point model.