Schetselaar, Ernst 
  
IMAGE CLASSIFICATION FROM LANDSAT TM, AIRBORNE 
MAGNETICS AND DEM DATA FOR MAPPING 
PALEOPROTEROZOIC BEDROCK UNITS, BAFFIN ISLAND, 
NUNAVUT, CANADA 
Ernst Schetselaar! and Eric de Kemp? 
l. International Institute for Aerospace Survey and Earth Sciences, 
Hengelosestraat 99, 7500 AA Enschede, The Netherlands, 
email: schetselaar G itc.nl 
2. Geological Survey of Canada, 601 Booth street, Ottawa, Canada, K1H OE8 
email: edekemp Gnrcan.gc.ca 
Working Group: TC III-5 
KEY WORDS: Bedrock Mapping, Geology, Classification, Remote Sensing, Nunavut 
ABSTRACT 
A novel approach in image classification to bedrock mapping is applied to a 1:100 000 scale map area 
of the Meta Incognita Peninsula Baffin Island, Nunavut, Canada to support mapping of the 
Paleoproterozoic basement and cover boundary. The classification algorithm was trained by input from 
ground traversed bedrock geology, and applied to 7 TM bands, airborne magnetics and di gital elevation 
data (CDED). We report the results of this investigation, with an emphasis on the comparison of 
interpretive bedrock mapping of the off-traverse areas. Focusing the training and the classification on 
the most exposed areas, by masking out overburden, vegetation and hydrographic features allows more 
specific targeting of bedrock exposures in the field. This approach supports remote geological mapping 
in similar geologic terrains were expense, access and inclement weather place severe constraints on the 
efficiency of traditional surveying. This research is a development contribution to enhance the ongoing 
natural resources management of Nunavut through partnership with the Geological Survey of Canada 
and its various geological mapping programs. 
1 INTRODUCTION 
A pilot study was undertaken to asses the usefulness of applying remotely sensed data from standard 
sensor platforms, such as airborne total field magnetic data and satellite based multispectral data (TM), 
as an aid in regional bedrock mapping in arctic terrains. The approach used was similar to previous 
Work (An and Chung, 1994; Schetselaar et. al. 2000), which employs actual ground truth data of the 
bedrock geology from traverses across the study area to define the training set. The classification was 
conducted on areas dominated by bedrock exposures masked out from densely vegetated areas on the 
basis of a green vegetation index. The main objective of this preliminary investigation was to determine 
if regionally significant bedrock features such as ‘basement ° and ‘cover’ could be predicted by 
combining various data types that individually would be less successful. The approach, if successful, 
could be employed before, during and after a detailed geological ground survey, as a tool to target 
specific tectonostratigraphic domains or areas of conflicting predictability, which require detailed 
analysis. Typically bedrock classification research has employed random spatial sampling of existing 
geological maps to act as the training set. As off-traverse areas are interpreted by geologists, and 
depending on methods used to extend lithologic units, the final geological map can have variable 
confidence. We assume that the highest confidence samples to be used for training are on the traverse 
lines. Provided the traverses reflect a reasonable sampling of the range and abundance of lithologies 
they should provide the most reliable data set to control the image classification. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 1325