This paper investigates the use of random forests and spatial random forests (RFsp) for the classification of coastal dune areas along 41km of Lake Michigan’s shoreline using a lidar- derived DEM. Terrain variables across a range of spatial neighborhood scales are utilized, and for two different cell resolutions. Distance is explicitly incorporated into the RFsp models through the calculation of buffer distances around small numbers (6-13) of gridded points in the study area. While classification accuracy is high generally, RFsp produced much more accurate results. At the fine scale, topographic variables and their neighborhood ranges were not predictive of dune areas, perhaps because large (> 0.1 hectare) neighborhoods were not tested at that scale. At the coarse scale these variables were much more important. The use of small numbers of gridded (non-sample) points to improve spatial prediction warrants further investigation.
Machine learning for cross-scale geomorphometric classification of landforms: a day at the beach