Mapping 10 years of forest cover regeneration in Massachusetts: A comparison of pixel-based and object-based classification



Spring 2005



Trevor Gareth Jones, Clark University tgjones@clarku.edu



Massachusetts, forest cover change mapping, pixel-based image analysis, image segmentation, object based classification



The Massachusetts landscape has continually undergone natural and anthropogenic change since the arrival of European settlers. The long-term history of statewide spatio-temporal forest cover dynamics is responsible for creating a unique landscape. Forest regeneration is occurring across the state, primarily in former agricultural areas. Current timber harvest occurs in non-industrial private forests (NIPFs) and in peri-urban areas. Remote sensing data and techniques provide an effective means to assess land cover change over time. This paper presents the results of a comparison between forest cover change mapping using pixel-based image analysis and classification using image segmentation and subsequent object based image classification, employing Landsat TM and ETM imagery. The study area is the North Quabbin region of Massachusetts, between 1989 and 1999. Traditional forest cover change mapping utilizes pixel-based analysis of imagery. This relies on classifications of individual pixels carrying integrated spectral signals treated as individual samples, clustered in a multi-dimensional feature space. Imagery and research demands are not always met using pixel-based classification. Using image segmentation and object based classification explores image segments as spatially contiguous and spectrally homogenous groups of pixels. Homogeneous groups of pixels combine the spectral behavior of geographic features with spatial characteristics. Comparison results show that using image segmentation and object based image classification for forest cover change mapping of Landsat TM and ETM imagery is more effective than using traditional pixel-based image analysis. With multi-resolution imagery, object-based classification utilizing spectral and geometrical properties, as well as object relationships, is more appropriate and accurate in assessing forest cover change.