Methodological advancement of optical and radar remote sensing for Arctic forest-tundra ecotone vegetation analysis

Ecosystems in the high north are vulnerable to both the direct and indirect impacts of climate change. Rising temperatures, tall shrub encroachment and advancing treelines, increasing frequency of canopy defoliation events, and more extreme weather will all influence the Arctic forest-tundra ecotone, the transitional zone between Subarctic forest and Low Arctic tundra. Efficient large-scale monitoring of these areas requires extracting information from satellite remote sensing (RS) products using machine learning (ML) methodologies. The overarching goal of this thesis has been to develop such techniques for monitoring the Arctic forest-tundra ecotone as part of the "Methodological advancement of Climate-Ecological Observatory for Arctic Tundra" (COAT Tools) project.

Several challenges need to be overcome to provide efficient ML tools applicable for satellite RS-based monitoring of this region. The frequent cloud cover limits the opportunities for imaging by multispectral optical sensors. Observations from largely weather independent synthetic aperture radar (SAR) satellites can mitigate this, especially if the full information potential of the sensor is utilised and integrated with multispectral data. The sparse and scattered distribution of vegetation classes critical for understanding this region also poses a challenge. This is aggravated by a general lack of ground reference data suitable for training ML models on satellite imagery with sufficient resolution to capture this variability.

This thesis presents methodological advancements that address these challenges. A resolution-preserving method for estimating the polarimetric SAR covariance matrices using an optical guide image is demonstrated to help differentiate live from defoliated forest canopy. Further, this method is used to generate the post-event image for a semi-supervised targeted change detection mapping of forest mortality. By applying image-to-image translation, the change detection is performed against a multispectral optical image captured before the outbreak of the defoliating pest insect that caused the disturbance. A targeted semi-supervised approach is also developed to map forest and tall shrub cover. The method exploits the similar appearance of forest and tall shrub in RS imagery to first differentiate them from all other land cover types, before a second ML model is trained to distinguish between them. The map is then created from a dataset where stacks of freely available SAR and multispectral images have been combined and utilised to perform multitemporal filtering.

The results in this thesis demonstrate that these methodological advancements can contribute to accurate and reliable large-scale monitoring of the Arctic forest-tundra ecotone based on SAR and multispectral optical satellite remote sensing.