Automated quantification of depressional water storage in prairie pothole landscapes using synthetic aperture radar and random forest classification

Thumbnail Image
Wong, Calvin
Major Professor
Crumpton, William G
Lu, Crystal (Chaoqun)
Gelder, Brian K
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Is Version Of
Ecology, Evolution, and Organismal Biology
Quantifying the water storage dynamics of perennial and intermittent waterbodies is crucial to understanding landscape-scale hydrologic and nutrient cycling processes, and for managing wildlife habitat, in wetland-dominated landscapes such as the North American Prairie Pothole Region (PPR). Wetlands in the PPR serve as important water and nutrient sinks and as breeding grounds for North American waterfowl, promoting biological diversity. However, a significant portion of PPR wetlands have been lost, primarily due to drainage to facilitate agriculture. Potential further wetland losses are on the horizon, as water levels in intact wetlands react to shifts in evapotranspiration rates and temperature trends due to climate change. The extent to which wetlands are inundated at a given time can be quantified with publicly available remote sensing products derived from optical sensors such as the Dynamic Surface Water Extent (DSWE). However, many of these products are of a coarse resolution and can be obscured by cloud cover, especially after heavy rain events, thus limiting their utility for tracking surface water dynamics, particularly at finer temporal scales. The objective of this study was to develop and test a methodology that leverages freely available Synthetic Aperture Radar (SAR) datasets (which are finer resolution than DSWE and can penetrate cloud cover) and random forest classification to detect standing water in perennial waterbodies over large geographic regions. The methodology was tested on three watersheds within the PPR featuring varying degrees of artificial drainage and wetland loss: the relatively unaltered and undrained lower Edmore Coulee subbasin (LECS) in North Dakota, the moderately altered and drained upper Mustinka subbasin (UMS) in Minnesota, and the extensively drained and altered South Skunk River subbasin (SSRS) in Iowa. Our classification method had good overall accuracy with an 86% accuracy over the SSRS, 80% over the LECS and 79% in the UMS. However, when compared to DSWE, our classification algorithm underestimated the areal extent of inundation in all regions. Improvements are needed before utilizing this method to analyze depressional inundation dynamics, particularly over farmed and extensively drained regions. We believe there is significant benefit to develop this method further to allow for inexpensive, fast detection of water over large areas and temporal scales, thus potentially providing a deeper understanding of inundation trends in North American wetlands.