A machine-learning model to predict uncertainty in permafrost thaw-induced land cover transition

This study addresses the effects  of  future  climate-induced permafrost thaw  on  the distribution of land cover in the discontinuous permafrost zones of Northwest Territories (NWT) of Canada. The rapid transition from  a landscape dominated by peat plateaus to one dominated by connected wetlands (fens) and isolated wetlands (bogs) is intricately connected to permafrost thaw. To be able to predict and estimate the potential long-term evolution of these three dominant land covers, we developed a machine learning-based time series land cover change model (TSLCM). The TSLCM is trained on a set of spatio-temporal variables as driving factors of change including: the estimated summertime land surface temperature anomaly (LST), the distance to land cover interfaces, time intervals between observations, time-accumulated land surface temperature, and classified land cover maps from 1970-2008. The TSLCM is used to capture  spatial patterns of change, replicate historical land cover change, and generate reasonable estimates of future land cover evolution over time. The output of TSLCM model is the spatial distribution of fen, bogs, and peat plateaus consistent with a default 50\%\ threshold applied on the predicted probability maps. 
We here use the TSLCM to simulate land cover change under multiple plausible futures scenarios by using the most recent set of climate model projections. The simulation of the TSLCM under different scenarios helps us to:

    1: visualize the spatial pattern of change
    2: calculate the pace of evolution over time and compare results between climate scenarios
    3:  explore the sensitivity of the model to driving factors of change

 
In addition to examining uncertainty due to climate uncertainty, a probabilistic approach is used to sample the threshold value to generate a range of land cover realizations.