GEOSCAN Search Results: Fastlink


TitleMultispectral permafrost terrain classification, Rankin Inlet, Nunavut
LicencePlease note the adoption of the Open Government Licence - Canada supersedes any previous licences.
AuthorOldenborger, G AORCID logo; Faucher, B; LeBlanc, A -MORCID logo
SourceGeological Survey of Canada, Open File 8824, 2021, 38 pages, Open Access logo Open Access
PublisherNatural Resources Canada
Documentopen file
Mediaon-line; digital
RelatedNRCan photo(s) in this publication
RelatedThis publication is related to Performance analysis of RapidEye multispectral land-cover mapping for the western coast of Hudson Bay, Nunavut
File formatpdf
NTS55K/15; 55K/16
AreaKivalliq; Hudson Bay; Rankin Inlet
Lat/Long WENS -92.5081 -92.1158 62.9042 62.8572
Subjectsgeophysics; surficial geology/geomorphology; hydrogeology; Nature and Environment; Science and Technology; permafrost; ground ice; periglacial features; ice-wedge polygons; patterned ground; thermokarst; frost action; remote sensing; satellite imagery; mapping techniques; textures; vegetation; sediments; gravels; beach deposits; glacial deposits; eskers; tills; wetlands; bedrock geology; surface waters; drainage; topography; WorldView-2; Classification; Hydrology
Illustrationstables; geoscientific sketch maps; satellite images; photographs; bar graphs
ProgramClimate Change Geoscience, Permafrost
Released2021 09 21
This Open File reports on permafrost terrain classification using multispectral WorldView-2 satellite imagery over Rankin Inlet, Nunavut. A suite of images was processed to yield a single corrected multispectral mosaic image for a 1360 km2 area inland of the Hamlet of Rankin Inlet where permafrost studies are ongoing. Terrain classes relevant to permafrost conditions and thaw sensitivity were defined using existing on-the-ground knowledge of vegetation, surficial geology, hydrology, ground temperature, and ground ice occurrence for the region. At locations for two separate study sites (15 km2 and 7 km2), a number of reference areas were established and classified using visual interpretation of the imagery in combination with ground truth information from the sites. Given the reference classifications, permafrost terrain mapping was performed using maximum likelihood classification of the multispectral data alone (MS), and in conjunction with the derivative measure of texture (T), and the independent variable of topography transformed to topographic position index (TPI). Classification performance was assessed using true positive rate (TPR) and positive predictive value (PPV), along with detailed analysis of the confusion matrix. Classification results were validated by visual examination of the class maps and imagery, and by qualitative comparison to surficial geology. The full MS+T+TPI feature set provides the best overall classification with prediction accuracy for the reference areas of approximately 85% (TPR and PPV) for both study sites. However, significant misclassification persists as indicated by the full confusion matrix. In some cases, misclassification occurs between classes with similar spectral and topographic characteristics, but also similar permafrost conditions, such that misclassification is of limited consequence. In other cases, misclassification occurs between classes with similar spectral and topographic characteristics, but distinct thaw sensitivity, and the potential for misclassification must be carefully considered.
Summary(Plain Language Summary, not published)
Classification of satellite imagery provides a means of remotely predicting characteristics of the Earth's surface. This report presents the idea of permafrost terrain classification for which terrain types are defined based on permafrost conditions, whereby each terrain type has different thaw sensitivity (the response of the ground to climate warming). Representative areas for the terrain types were identified on multi-spectral WorldView-2 satellite imagery at two study sites inland of the Hamlet of Rankin Inlet, Nunavut. These areas were used to establish signatures for the terrain types, and machine learning was used to extend the classification over the satellite images. Accuracy is approximately 85% for both study sites. However, misclassification can occur both between terrain types with similar permafrost conditions, and between terrain types with distinct thaw sensitivity. The latter is of greater consequence if the classification is used to infer permafrost degradation potential.

Date modified: