The feasibility of using national-scale datasets for classifying wetlands in Arizona with machine learning

The advent of machine learning techniques has led to a proliferation of landscape classification products. These approaches can fill gaps in wetland inventories across the United States (U.S.) provided that large reference datasets are available to develop accurate models. In this study, we tested the feasibility of expediting the classification process by sourcing requisite training and testing data from existing national-scale land cover maps instead of customized sample sets. We created a single map of water and wetland presence by intersecting water and wetland classes from available land cover products (National Wetland Inventory, Gap Analysis Project, National Land Cover Database and Dynamic Surface Water Extent) across the U.S. state of Arizona, which has fewer wetland-specific mapping products than other parts of the U.S. We derived classified samples for four wetland classes from the combined map: open water, herbaceous wetlands, wooded wetlands and non-wetland cover. In Google Earth Engine, we developed a random forest model that combined the training data with spatial predictor variables, including vegetation greenness indices, wetness indices, seasonal index variation, topographic parameters and vegetation height metrics. Results show that the final model separates the four classes with an overall accuracy of 86.2%. The accuracy suggests that existing datasets can be effectively used to compile machine learning training samples to map wetlands in arid landscapes in the U.S. These methods hold promise for the generation of wetland inventories at more frequent intervals, which could allow more nuanced investigations of wetland change over time in response to anthropogenic and climatic drivers.