Assessing Synthetic Aperture Radar (SAR)-Derived Temporal Patterns and Digital Terrain Data for Palustrine Wetland Mapping

Abstract

Palustrine wetland systems are important ecosystems and provide numerous ecosystems services to support society. Unfortunately, they remain under constant threat of devastation due to land use practices and global climate change, which underscores the need to identify, map, and monitor these landscape features. This study explores harmonic coefficients and seasonal median values derived from Sentinel-1 synthetic aperture radar (SAR) data, as well as digital elevation model (DEM)-derived terrain variables, to predict palustrine wetland locations in the Vermont counties of Bennington, Chittenden, and Essex. Support vector machine (SVM) and random forest (RF) machine learning models were used with various combinations of the three datasets: terrain, SAR seasonal medians, and SAR harmonic time series coefficients. For Bennington County, using the harmonic and terrain data with a RF model yielded the most accurate results, with an overall accuracy of 76%. The terrain data alone and RF model produced the highest overall accuracy in Chittenden County with an accuracy of 85%. In Essex County any combination of the three datasets and the RF model yielded the highest overall accuracy of 81%. Generally, this study documented better performance using the RF algorithm in comparison to SVM. Terrain variables were generally important for differentiating wetlands from uplands and waterbodies. However, Sentinel-1 data, represented as harmonic regression coefficients and seasonal medians, provided limited predictive power. Although Sentinel-1 SAR data were of limited value in the explored case studies, findings may not extrapolate to other SAR datasets using different polarizations, wavelengths, and/or spatial resolutions.