Mapping the state of land cover and wetlands in the Northwest Territories, Canada

The Northwest Territories (NWT) is a large jurisdiction (>1.3 million km²) in Canada featuring vast areas of forests, expansive organic wetlands (peatlands), and open tundra. Within the mainland of 1.15 Mkm² (excluding the Arctic Archipelago), the diversity of terrain, climate, biotic factors, and sub-surface (e.g., permafrost) conditions give rise to heterogeneous landscapes at both small and large scales.  Land cover maps provide the basis for understanding how different types of vegetation and wetlands are distributed across landscapes, providing a foundation to subsequently derive information concerning climate change impacts and carbon/methane modelling. Through the joint Multisource Vegetation Inventory (MVI) project (Castilla et al. 2022), the Government of the Northwest Territories (GNWT) in partnership with the Canadian Forest Service of Natural Resources Canada (CFS) has developed land cover maps with improved classification accuracy of key vegetation types (forests, wetlands). The goal of this work is to produce and validate an updated land cover map of the entire NWT mainland including 10 forest classes as broad forest cover types (coniferous, broadleaf, mixedwood, treed wetland) combined with density classes (dense, open, sparse), as well as 10 non-treed classes. This presentation provides an overview of the major components in the land cover map development, with specific focus on improved (organic) wetland mapping. This initiative is based on a new network of land cover reference data consisting of thousands of points (n=24,865), forming the densest ever compilation of forest, wetland, and non-treed reference information available across NWT. The land cover map is produced from a random forest (RF) classification procedure using above reference land cover points and 30-m resolution rasters of predictive variables derived from satellite imagery and environmental datasets. Satellite imagery composites include cloud-free multispectral Sentinel-2 time-series of six spectral bands and six spectral indices which were temporally composited for each pixel over the 2020 to 2022 time period as i) seasonal summer (July-August) and winter (February-March) medians and ii) inter-annual statistics from full time-series including six percentiles (p5, p20, p40, p60 p80, p95) and two temporal variability measures (range, st. dev.). In addition, a single ca. 2020 PALSAR-2 L-band dual-polarized (HH, HV) summer composite was created, along with 26 terrain-derived data layers, 24 climatic layers, and three long-term spectral change metrics. The RF classification procedure included a spatially balanced split of the reference data into calibration and independent validation observations, a recursive feature elimination algorithm to iteratively remove the least important predictor variables, as well as a hyper-parameter optimization routine to further improve predictive performance. The resulting NWT land cover map product improves upon national mapping results (71 % vs 47 % overall accuracy all classes, improvements up to 20% in upland-wetland separation) and shows potential to provide an invaluable operational map of baseline forest and wetland information required to serve forest, wetland, wildfire, and wildlife management applications.