Wetland Mapping in Large Spatial Scales via Machine Learning Approaches
Abstract
Wetlands are natural infrastructures playing important hydrological, biogeochemical, and biological functions. However, existing data on potential wetland extents are either outdated or imprecise. There is a growing need for efficient and accurate wetland mapping approaches applicable over large regions. To this end, we present a Machine Learning (ML)-based approach focusing on the 314,000 sq. km U.S. Gulf Coast region and involving 18 open access datasets that are representative of geomorphic, hydroclimatic, hydrologic, soil, and vegetation characteristics of wetlands. A combined wetland dataset aggregating National Wetland Inventory (NWI), National Land Cover Dataset (NLCD), and European Space Agency (ESA) was used for training and verification of the Light Gradient Boosting Machine algorithm. Our ML model was found to be 92% and 86% accurate respectively during training and verification. The outcome was a novel dataset identifying 97,500 sq. km of potential wetland extents across the U.S. Gulf Coast. Additional statistical analysis to understand feature importance suggested that our ML model captures the governing factors of wetland dynamics well, indicating its applicability over the continental U.S.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H36C..08G