Object-oriented and pixel-based image classification using Landsat multispectral and Hyperion hyperspectral imagery in boreal conditions

Abstract

Current environmental trends dictate a need for new methods, initiatives, and technologies that provide reliable, up-to-date forest information. Canada, which is home to ten percent of the Earth's forests, has made national and international commitments to better monitor the sustainable development of its forest ecosystems. In Ontario, the Ministry of Natural Resources monitors its natural resources through the Ontario Land Cover Database (OLCD). The OLCD is a large area land classification that uses Landsat multispectral imagery with a traditional pixel-based classifier. The goal of this thesis is to explore new ways to improve upon large area land classifications such as the OLCD. This thesis evaluates two alternative approaches: (1) it compares Landsat-5 TM multispectral imagery to Hyperion hyperspectral imagery, and (2) it compares a traditional pixel-based classifier to eCognition's object-oriented image classifier. Eight boreal cover classes were used consisting of water, wetland (aggregated marsh, fen and bog), black spruce, jack pine, mixedwood, dense deciduous, sparse deciduous and clearcuts.