Arctic ice-wedge landscape mapping by CNN using a fusion of Radarsat constellation Mission and ArcticDEM

In Canada's Arctic tundra region, permafrost is continuous, and the landscape is rich in patterned features. Polygonal terrain, which includes both high- and low-centered features and their wet trenches below, is considered to be high-latitude wetlands in the continuous permafrost region. These prominent hydrological features retain and transport water within widespread ice-wedge networks and govern many ecosystem dynamics. Due to the meter-scale spatial gradients of these processes, mapping of polygonal wetland networks necessitates high-resolution imagery. To date, most studies have used optical imagery for this task; however, these sensors are affected by cloud cover and polar darkness, limiting image availability and repeatability. Thus, our overall objective was to evaluate high-resolution hybrid compact polarimetric (HCP) imagery from the recently launched Radarsat Constellation Mission (RCM), in fusion with ArcticDEM topographic data, for Arctic landscape mapping with a focus on polygonal wetlands. RCM's 5 m Stripmap beam mode, which has yet to be studied for such a task, represents an innovative HCP synthetic aperture radar (SAR) data source since it allows for polarimetric decomposition methods, despite being a dual-pol system. Within this study, we present a seveninput channel Convolutional Neural Network (CNN) model for the classification of ice-wedge dominated landscapes. A range of model hyperparameters as well as the effect of SAR speckle filtering on classification accuracy, have been examined. The optimized CNN achieved a high classification accuracy (0.931 mean Intersection Over Union; mIOU) for three semantic classes representative of the study area, namely polygonal wetlands, open water, and uplands. These results were superior in comparison to a benchmark machine learning (ML) Random Forest (RF) algorithm, thus demonstrating the proposed CNN's potential for regional-scale permafrost feature mapping. Notably, the optimal CNN architecture used unfiltered SAR data as input, underscoring the importance of spatial resolution when classifying polygonal terrain with deep learning (DL). These findings have important implications regarding the design, tuning, and sensitivity of CNN algorithms, and on the efficacy of HCP SAR, for mapping Arctic regions.