Reducing Dimensionality in Multitemporal MODIS Data Using Principal Component Analysis for Land Cover Mapping

• Activity Number: 182
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Hunter Glanz*+ and Luis E. Carvalho and Mark Friedl and Damien Sulla-Menashe
• Companies: Boston University and Boston University and Boston University and Boston University
• Keywords: expectation maximization ; principal component analysis ; classification ; land cover ; MODIS

Abstract

High dimensionality presents significant challenges for algorithms that use remote sensing to classify land cover and detect land cover changes. To address this, principal components analysis has been widely used for several decades to reduce dimensionality. Time series images are increasingly being used to monitor and map land cover, which significantly increases the dimensionality problem because each spectral band is replicated in time. Traditionally each of these replicated observations is considered a different feature in the principal components analysis. In this talk we present an approach that models the data using a covariance structure that is partitioned into spectral and temporal pieces. We estimate the parameters of this model, including this covariance, using EM and perform PCA on the spectral piece of the covariance in order to preserve temporal variation and behavior. The goal of this method is to reduce the spectral dimensionality in a way that simultaneously captures temporal variance properties of different land cover classes. We successfully test our approach using a data set composed of MODIS data from land cover training sites in North America.