GSA Annual Meeting in Seattle, Washington, USA - 2017

Paper No. 327-8
Presentation Time: 3:20 PM

REGIONAL-SCALE PROBABILISTIC PERMAFROST MAPPING OF NORTHERN ALBERTA, CANADA, USING GRASS GIS AND SCIKIT-LEARN


PAWLEY, Steven and UTTING, Daniel, Alberta Geological Survey, 402 Twin Atria Building, 4999-98 Avenue, Edmonton, AB T6B 2X3, Canada, Steven.Pawley@aer.ca

Approximately 210,000 km2 of Northern Alberta exists within the zone of sporadic permafrost, with 26% of the peatlands estimated to contain components of perennially frozen ground. Understanding the distribution of sporadic permafrost is important because ongoing thawing affects ground stability, ecology and surface hydrology. Further, the sensitivity of permafrost to disturbance by infrastructure development makes it an important consideration for land use and reclamation planning. However, to date there has been limited mapping of sporadic permafrost in Alberta at a scale sufficient for these purposes, with previous airphoto-based interpretations providing only a small-scale delineation of the forest-covered permafrost terrain. This presentation will focus on the development of a methodology that integrates Landsat best pixel mosaics, climatic data, and terrain information from a regional 15 m LiDAR DEM to classify permafrost terrain consisting of forest-covered palsa bogs and larger peat plateaus using a suite of machine learning techniques. The software r.learn.ml, which is freely available as a GRASS GIS add on, was developed for the purpose of providing a linkage between popular open-source GIS with the Python machine learning library, scikit-learn. Evaluation of the modelled results revealed high classification accuracies when compared to ground-truth data (>85%). These results also demonstrate that near-surface permafrost is significantly more extensive than previously identified in northern Alberta, even at relatively low latitudes (~56.5°).