Document: FAW-3-89-9

Historical vegetation mapping using unsupervised classification of Landsat imagery and ancillary data.

GADALLAH, FAWZIAH(ZUZU)* ¹, F.CSILLAG ¹ and R.K.BROOK ²

University of Toronto, Toronto, Ontario, Canada, M5S 3G3 ¹
University of Manitoba, Winnipeg, Manitoba, Canada, R3T 2N2 ²

Abstract:
Satellite imagery is increasingly seen as an inexpensive and consistent data source for mapping and monitoring large areas of ecological interest. It also presents an opportunity to examine changes over time, since imagery has been collected over large areas of the Earth's surface for several decades. With increasing public concern over regional-scale environmental issues, accurate vegetation maps are needed both for management decision-making and for public education. This is particularly true when management is expensive or controversial. Such is the case at La Perouse Bay, Manitoba, (58 44' N, 93 25' W), the site of a breeding colony of lesser snow geese (Anser caerulescens caerulescens). Increasing numbers of birds in recent decades have damaged the sub-arctic coastal marshes on which they feed in summer. Negative effects of this damage to the geese and other species have been documented by the Hudson Bay Project. Because the geese are an economic, aesthetic and visible species, plans for population control via increased harvest are controversial. Documentation of the extent, pattern and timing of habitat damage over large spatial and temporal scales requires reconstruction of the historic vegetation. For the La Perouse Bay area, a 1984 Landsat TM image was classified using a commercial clustering algorithm (PCIwork's IsoData). The lack of concurrent ground information and the predominance of mixed-class pixels limits the use of supervised classification techniques. Unsupervised classification, by contrast, clusters spectrally similar pixels into classes, and avoids the problems introduced by determining target classes in advance. Sixty-three classes produced this way were identified using a combination of historic and current aerial photos, current vegetation data, and a knowledge of local vegetation trajectories in time. These initial classes were then aggregated into 14 ecologically meaningful classes. Water forms the largest part of the study area (60% of a total 32936 ha). Salt marsh, the habitat type most heavily used and damaged by the geese, was found to be 4.9% of the land area (967 ha of a total 19869 ha land). Map accuracy was tested with vegetation data collected independently at a number of precisely located points. This has resulted in a detailed base map against which changes can be measured.

Keywords: Lesser snow geese, Hudson Bay, vegetation change, vegetation mapping, remote sensing, Landsat, unsupervised classification

This abstract is being presented at: 10:30 AM in session:
REMOTE SENSING