Classification of salt marsh vegetation using edaphic and remote sensing-derived variables

Salt marshes are well known for their striking macrophyte zonation patterns.  Although many variables affect species distribution, soil salinity and waterlogging have been shown to be two of the most important edaphic parameters.  These variables are largely determined by the frequency and duration of tidal flooding, which is dependent on topographical variations.  Light detection and ranging (LIDAR) can be used to generate digital elevation models (DEMs) from which elevation and landscape metrics can be derived with GIS, as an alternative to the collection of edaphic data in the field.  The primary objective of this study was to classify four marsh vegetation classes (tall Spartina alterniflora, medium S. alterniflora/short S. alterniflora, marsh meadow and Borrichia frutescens/J.roemerianus) based on edaphic and remote sensing-derived variables in order to determine which combination of variables best describe plant distributions in a Southeastern salt marsh.  Although multivariate statistical techniques such as linear discriminant analysis (LDA) are commonly used to classify and predict plant distributions based on edaphic and/or remote sensing-derived metrics, nonparametric classification and regression trees (CART) is being used increasingly as an alternative as it may better capture nonlinear and collinear relationships in environmental data sets.  Our second objective was to compare the performance of LDA and CART for the classification of marsh vegetation.  Models based on the edaphic variables soil organic matter content, water content, salinity and redox, attained accuracies of 0.63 and 0.72, with LDA and CART respectively.  When the remote sensing variables DEM elevation, slope, distance to mean high water line and distance to upland area were used, classification accuracies improved to 0.78 for LDA and 0.79 for CART.  The greatest accuracies (0.82 for LDA and 0.83 for CART) were attained by combining soil organic matter content with the four remote sensing metrics in the combination models.  Our results suggest that remote sensing-derived metrics can capture edaphic gradients effectively, which makes them especially suited to landscape level analyses of salt marsh plant habitats.  Although the two classification techniques had similar overall accuracies, we recommend a workflow wherein CART is used for variable reduction and selection prior to training and subsequent prediction of new observations by LDA.