A Comparison of Different Remotely Sensed Data for Classifying Bedrock Types in Canada’s Arctic: Application of the Robust Classification Method and Random Forests
The Geological Survey of Canada, under the Remote Predictive Mapping project of the Geo-mapping for Energy and Minerals program, Natural Resources Canada, has the mandate to produce up-to-date geoscience maps of Canada’s territory north of latitude 60°. Over the past three decades, the increased ava...
Published in: | Geoscience Canada |
---|---|
Main Authors: | , , , |
Format: | Text |
Language: | English |
Published: |
The Geological Association of Canada
2014
|
Subjects: | |
Online Access: | https://doi.org/10.12789/geocanj.2014.41.062 http://id.erudit.org/iderudit/1062261ar |
Summary: | The Geological Survey of Canada, under the Remote Predictive Mapping project of the Geo-mapping for Energy and Minerals program, Natural Resources Canada, has the mandate to produce up-to-date geoscience maps of Canada’s territory north of latitude 60°. Over the past three decades, the increased availability of space-borne sensors imaging the Earth’s surface using increasingly higher spatial and spectral resolutions has allowed geologic remote sensing to evolve from being primarily a qualitative discipline to a quantitative discipline based on the computer analysis of digital images. Classification of remotely sensed data is a well-known and common image processing application that has been used since the early 1970s, concomitant with the launch of the first Landsat (ERTS) earth observational satellite. In this study, supervised classification is employed using a new algorithm known as the Robust Classification Method (RCM), as well as a Random Forest (RF) classifier, to a variety of remotely sensed data including Landsat-7, Landsat-8, Spot-5, Aster and airborne magnetic imagery, producing predictions (classifications) of bedrock lithology and Quaternary cover in central Victoria Island, Northwest Territories. The different data types are compared and contrasted to evaluate how well they classify various lithotypes and surficial materials; these evaluations are validated by confusion analysis (confusion matrices) as well as by comparing the generalized classifications with the newly produced geology map of the study area. In addition, three new Multiple Classification System (MCS) methods are proposed that leverage the best characteristics of all remotely sensed data used for classification. Both RCM (using the maximum likelihood classification algorithm, or MLC) and RF provide good classification results; however, RF provides the highest classification accuracy because it uses all 43 of the raw and derived bands from all remotely sensed data. The MCS classifications, based on the generalized training dataset, ... |
---|