Arctic shrub expansion revealed by Landsat-derived multitemporal vegetation cover fractions in the Western Canadian Arctic

• Published in: Remote Sensing of Environment
• Main Authors: Nill, Leon, Grünberg, Inge, Ullmann, Tobias, Gessner, Matthias, Boike, Julia, Hostert, Patrick
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: 2022
• Subjects: Arctic; Mackenzie Delta; Tundra
• Online Access: https://epic.awi.de/id/eprint/57000/; https://epic.awi.de/id/eprint/57000/1/Nill_etal_2022.pdf; https://www.sciencedirect.com/science/article/pii/S0034425722003340; https://hdl.handle.net/10013/epic.edecc4c2-6452-4e09-bd9b-65819b8555e0

Warming induced shifts in tundra vegetation composition and structure, including circumpolar expansion of shrubs, modifies ecosystem structure and functioning with potentially global consequences due to feedback mechanisms between vegetation and climate. Satellite-derived vegetation indices indicate widespread greening of the surface, often associated with regional evidence of shrub expansion obtained from long-term ecological monitoring and repeated orthophotos. However, explicitly quantifying shrub expansion across large scales using satellite observations requires characterising the fine-scale mosaic of Arctic vegetation types beyond index-based approaches. Although previous studies have illustrated the potential of estimating fractional cover of various Plant Functional Types (PFTs) from satellite imagery, limited availability of reference data across space and time has constrained deriving fraction cover time series capable of detecting shrub expansion. We applied regression-based unmixing using synthetic training data to build multitemporal machine learning models in order to estimate fractional cover of shrubs and other surface components in the Mackenzie Delta Region for six time intervals between 1984 and 2020. We trained Kernel Ridge Regression (KRR) and Random Forest Regression (RFR) models using Landsat-derived spectral-temporal-metrics and synthetic training data generated from pure class spectra obtained directly from the imagery. Independent validation using very-high-resolution imagery suggested that KRR outperforms RFR, estimating shrub cover with a MAE of 10.6 and remaining surface components with MAEs between 3.0 and 11.2. Canopy-forming shrubs were well modelled across all cover densities, coniferous tree cover tended to be overestimated and differentiating between herbaceous and lichen cover was challenging. Shrub cover expanded by on average + 2.2 per decade for the entire study area and + 4.2 per decade within the low Arctic tundra, while relative changes were strongest in the northernmost ...