Assistant Professor University of Texas at Arlington
Wetlands are natural infrastructures, playing crucial hydrological, biogeochemical, and biological functions. However, currently available data on potential wetland extents is either obsolete or inaccurate. In this presentation, we will demonstrate how 17 different open-source geospatial datasets, each representing the geomorphic, hydroclimatic, hydrologic, soil, and vegetation properties of the landscape, can be harnessed to develop a machine learning (ML)-based approach for wetland mapping on large scales and at high spatial resolutions. Focusing on the 314,000 sq. km. U.S. Gulf Coast region as the study area, we integrated the National Wetland Inventory (NWI), National Land Cover Dataset (NLCD), and European Space Agency (ESA) wetland datasets to train and verify the Light Gradient Boosting Machine algorithm. During training and verification, our ML model was found to be 92% and 86% accurate, respectively. The outcome was a novel dataset that identified 97,500 sq. km of potential wetland extents along the Gulf Coast of the United States. Our ML model captures the controlling drivers of a wetland’s existence very well, according to the additional statistical analysis performed to determine feature importance, showing the applicability of our approach for wetland mapping over the continental U.S.
