An investigation of two date unsupervised classification in the context of a national program for Landsat based forest change mapping

A number of change mapping methods are being evaluated as part of a national program for monitoring Canada´s forest cover, change and biomass (Earth Observation for Sustainable Development-EOSD). One approach being evaluated involves a simple two date multispectral unsupervised classification approach to map forest change at a national scale with Landsat data. Key factors are the types and consistency of change classes that emerge from the clustering algorithms and issues related to labelling these into desired change classes. The technique was investigated using two sites. Prince George in central British Columbia, Canada used 1990 and 1999 imagery. The second site of Petawawa in eastern Ontario, has fifteen images from a time period from 1984 to 2001 that were normalized and then analyzed in various image pairs. Utilizing the six visible and infrared Landsat bands from each of two dates in a K-means hyperclustering, classes of conifer, hardwood and mixedwood that had not changed were separated. Before any amalgamation, many clusters were associated with change, especially changes in vegetation on old cuts and open fields. Post-clearing ground vegetation density and type had a moderate influence on the clustering process. Pre-clearing forest type was only a factor in general terms, for example whether it was hardwood, softwood or mixedwood. Clearcuts often had several clusters associated with them, but were well detected when clusters were amalgamated. Partial cuts were only partly detected. The length of time after the cut was an important consideration in their detection. They were detectable for only 2 or 3 years after the cutting. The above approach is expected to be one of several tools used in the EOSD initiative.