Wetland mapping with multi-temporal sentinel-1 & -2 imagery (2017 – 2020) and LiDAR data in the grassland natural region of alberta

Abstract

ABSTARCT In the Grassland Natural Region (GNR) of southern Alberta, wetlands are relatively small-sized disconnected prairie pothole marshes, swamps, and shallow open water habitats often surrounded by grasslands, parkland forests, agricultural lands, and urban areas. These wetlands are susceptible to climatic variability, resulting in temporally and spatially dynamic habitats that are difficult to map accurately. This study hypothesizes that seasonal synthetic aperture radar (SAR) and optical imagery will capture temporal variations of wetlands in the spring/summer and fall months of 2017, 2018, 2019, and 2020. We propose that these data combined with topographic variability offered by LiDAR-derived topographic wetness index (TWI) shall result in the accurate delineation of the wetlands. Using a combination of open-access government databases, we generated ground and training data to develop the classification models and perform accuracy assessments. The wetland map products’ overall accuracy results ranged from 63.2% to 75.7%. The pixel-based random forest (RF) classified dataset (Dataset 5 – multi-temporal (2017–2020) S1 SAR (VH) and S2 optical (B8 and B11) bands fused with TWI) had the highest overall accuracy (75.6%). The RF result significantly outperformed similar CART (Classification and Regression Trees) and SVM (Support Vector Machine) classifications, which had overall accuracies of 67.4% and 63.2%, respectively. In addition, the RF optimal wetland product had the best combination of F-score values for wetland and upland classes: 0.61 (marsh), 0.82 (open water), 0.75 (swamp), and 0.8 (uplands). Overall, the methodology adopted in this study is promising for mapping the spatial distribution of wetland habitats across the seasonally dynamic GNR of Alberta.