https://ui.adsabs.harvard.edu/abs/2020AGUFMH194.0002A/abstract

Akbarpour, S., Craig, J. R. (2020) Modelling Evolution of Discontinuous Permafrost Landscapes and Hydrology, American Geophysical Union, Fall Meeting 2020, abstract #H194-0002

The rapid evolution of the lowland discontinuous permafrost region of the Scotty Creek Research Station (SCRS) located in the Northwest Territories of Canada is caused by permafrost thaw and other effects of climate warming. Permafrost thaw is affecting hydrological and ecological characteristics of the SCRS, and it is changing the borders and distribution of the three main landform types in this region (peat plateaus, bogs, and fens). Land-cover change models help us to analyze the causes and effects of evolution in the SCRS. To understand the relationships between the driving factors of permafrost thaw and transformation of the land-covers in the SCRS, we applied statistical and machine learning methods informed by a set of spatial independent variables affecting the probability of conversion from one type of land cover to others. The selection of the independent variables of models was dependent upon the availability of data, the strength of relationships between the probability of change in each land-cover and the known drivers of lateral permafrost thaw Logistic regression and random forest models are devised by using environmental and proximity factors as driving forces of change (e.g., the estimated summertime Land surface temperature (LST), Euclidean distance, cost distance to land cover interfaces, Normalized difference vegetation index (NDVI), soil moisture). Accuracy of the implemented methods will be compared to evaluate their performance for modeling the land-cover change in the SCRS. Based on the probability maps generated from the logistic and random forest models, the rate of change from peat plateaus to fen and bog is higher than the other transitions, as it is consistent with permafrost thaw impacts. It is also found that rate of transition from peat plateaus to fens and bogs is higher when the LST of each specific location is higher than the mean of the whole domain. Our results illustrate that the random forest model outperformed the logistic regression model for modelling most land-cover changes due to its higher accuracy. Logistic regression is more accurate for the specific land-cover transition. The result also reveals that the primary driver of land-cover change is proximity to thaw edges, especially for the transition of peat plateaus to fen and bog.