Description: High resolution (0.15 m (0.05 ft)) four-band orthoimagery was acquired from December 2012 to January 2013 as part of the Coastal Imagery Project for Chatham, Bryan, Liberty, McIntosh, and Glynn counties. Camden County separately acquired three-band, 0.15 m (0.5 ft) resolution orthoimagery from February to March 2012. All orthoimagery data have a horizontal accuracy of ≤ 0.76 m (2.5 ft). For this project, the orthoimagery was subsetted by county and resampled to a 4 m (13.12 ft) spatial resolution. Orthoimagery for each county was classified using the random forest classifier using the R (version 3.2.1) package randomForest. Three raster image products were used as predictor variables in random forest to produce classified maps of habitat distributions: orthoimagery reflectance (3 or 4 bands), DEM elevations, and NWI Plus class. Four orthoimagery bands (blue, green, red, near infrared (NIR)) were used in the classification of Chatham, Bryan, Liberty, McIntosh, and Glynn counties. As Camden County orthoimagery was acquired without the NIR band, three bands (blue, green, red) were used for its classification. Training data for random forest were generated by digitizing vegetation areas from field maps at 596 randomly selected ground control points. In the field at each sampling location, large-scale maps of the site and surrounding area where annotated with vegetation boundaries and notes on the distribution of marsh vegetation. Polygons were digitized from the field maps using the resampled 4 m (13.12 ft) orthophotography at a minimum scale of 1:500. All dominant tidal marsh covers were classified, including non-vegetated areas. This resulted in eight classes for which training data were generated: three height classes of Spartina alterniflora (tall, medium, short); Juncus roemerianus/Schoenoplectus spp.; Spartina cynosuroides/Schoenoplectus tabernaemontani; marsh meadow (Salicornia virginica(Sarcoconia sp.), Distichlis spicata, Batis maritima); mud; and unvegetated areas. The mud class included both mud found at higher elevations on the marsh platform and lower elevations along creek banks and exposed tidal flats. Unvegetated areas were comprised of highly reflective surfaces including shell, sand, beach, pavement, and docks. Classification accuracy was calculated based on the out-of-bag error estimate. Each county was classified independently and mosaicked together post-classification. Focal statistics (3x3 window) were used to remove the salt and pepper appearance of the classification and remove isolated pixels. The random forest classifier performed well, with a combined overall accuracy of 0.90 for an eight-class habitat map. The overall accuracy of individual county random forest classifications ranged from 0.85 (Camden) to 0.95 (Bryan).