The Second Generation Canadian Wetland Inventory Map at 10 Meters Resolution Using Google Earth Engine
Masoud Mahdianpari, Brian Brisco, Jean Granger, Fariba Mohammadimanesh, Bahram Salehi, Sarah Banks, Saeid Homayouni, Laura Bourgeau‐Chavez, Qihao Weng
Abstract
Recently, there has been a significant increase in efforts to better inventory and manage important ecosystems across Canada using advanced remote sensing techniques. In this study, we improved the method and results of our first-generation Canadian wetland inventory map at 10-m resolution. Iin order to increase wetland classification accuracy, the main contributions of this new study are adding more training data to the classification process and training Random Forest (RF) models on the Google Earth Engine (GEE) platform within the boundaries of ecozones rather than provinces. A considerable effort has been devoted to data collection, preparation, standardization of datasets for each ecozone. The data cleaning reveals a data gap in several Northern ecozones. Accordingly, high-resolution optical data, from Worldview-2 and Pleiades, were acquired to delineate wetland training data based on visual interpretation in those regions. By using this well-distributed training data, this second generation wetland inventory map represents an improvement of 7% compared to the first generation map. Accuracy varied from 76% to 91% in different ecozones depending on available resources. Furthermore, the results of RF variable importance, which was carried out for each ecozone, demonstrate that |SVV |2|SVH |2 and NDVI extracted from Sentinel-1 and Sentinel-2 data, respectively, were the most important features for wetland mapping.