Random Forests vs. Markov Random Fields for Land-Cover Classification

• Activity Number: 254
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Jason Stover*+
• Companies: Georgia College
• Keywords: iterated conditional modes ; Potts model ; kernel density estimates ; remote sensing ; MAP estimation

Abstract

We present two approaches to classifying land cover: Random forests and Markov random fields. The data used to train each classifier consist of ground observations which appear on LANDSAT and ASTER satellite images taken over two years over central Georgia, U.S.A. Each type of classifier is then used to classify the remaining pixels from a central image, using reflectivities from it and thirty-seven others, along with georectified land-cover surveys from the U.S. Multi-Resolution Land Characteristics Consortium and Georgia Land Use Trends. Though the random forest classifier has a much lower misclassification rate, the sampling of the training data may make the random forest unable to detect important patterns that the Markov random field classifier can detect. This shortcoming of the random forest is not apparent from the estimated misclassification rate alone.