Multinomial simulation of land cover evolution in discontinuous permafrost zones of Northwest Territories.

Land cover evolution is one of the major responses of nature to global warming. Each terrestrial ecosystem is susceptible to specific driving forces of change originating from climate warming. In the lowland discontinuous permafrost regions of the Taiga Plains, permafrost thaw is the main cause of land cover change. Investigations of the discontinuous permafrost regions demonstrate that the distribution of the dominant land types (permafrost plateaus, fens, and isolated bogs) and their hydrological properties has been greatly affected by thawing.Here, we developed a machine learning-based model to estimate the evolution of the main hydrologically-important land covers in the discontinuous permafrost region. The multinomial time series land cover model (TSLCM) is able to simulate historical landcover transitions, simulate spatial patterns of change, and replicate the long-term evolution of landcover at the Scotty Creek Research Station (SCRS), NWT, and similar landscapes. Our input data includes a set of spatio-temporal variables: the estimated land surface temperature (LST), the distance to landcover interfaces, time horizon (from 0 to 38 years), time-accumulated land surface temperature, and classified landcover maps from 1970-2008. The process of implementing TSLCM starts with training a generative model by employing a random forest (RF) and a Multiple Linear Regression (MLR) method; it improves the performance of the TSLM in extrapolating time series change by boosting the initial data-set. Then, We used a MLR and an extreme gradient boosting (XGBoost) model to simulate landcover change.Compared to the MLR method, the RF method showed better results in replicating historical land cover change; our main concern was that the model did not perform well in generating plausible future scenarios. To overcome this problem, we added new data instances to the initial data-set by combining the predicted land cover change and the observational dataset. The final outputs of MRL and XGBoost trained on boosted data-set confirm that Ensemble Learning (EL) models are weak learners in forecasting time series change and capturing the correlation between spatial and temporal variables. The predicted time series land cover maps to the year 2100 suggest that permafrost plateaus are rapidly transforming to fens.