Remote sensing of forested wetlands: application of multitemporal and multispectral satellite imagery to deterimne plant community composition and structure in southeastern USA

A hierarchical classification of forested wetland communities was developed for the lower Roanoke River floodplain of northeastern North Carolina, USA, through the use of multitemporal and multispectral satellite digital data. Landsat Thematic Mapper (TM) images from different seasons (March-April, May-June, July-August) throughout a single year were used to exploit the phenological variability of forest communities for generating a landcover classification of ecologically important vegetation types within the floodplain. A hierarchical classification scheme was developed that relied upon customized spectral 'feature sets' of Landsat TM bands and their transformations to generate the classified images for each level of the forest community classification scheme. The objective was to enhance the discrimination of the community types at subsequent levels of the hierarchical classification scheme through different spectral inputs from the assembled satellite time series in conjunction with detailed floristic information collected though in-situ methods. As such, general landcover classes were iteratively reclassified into more detailed classes at correspondingly 'deeper' levels or nodes in the hierarchy. Vegetation classes included 21 forest communities and several other ecologically important classes in the study area. The integration of detailed field data permitted spatially-explicit and highly descriptive definitions of the forest types occurring within the floodplain. Additional field data were used to validate the compositional and structural characteristics of the mapped plant communities described by the classification scheme. Use of fuzzy set theory in the accuracy assessment provided details on the magnitude and direction of errors in the classification, and permitted ecological interpretation of those errors. The application of fuzzy set concepts to the mapping of bottomland forest communities is significant because these forests typically exhibit substantial variation in species composition and support diverse canopy dominants. Unlike the discrete classification assessments that are traditionally employed, fuzzy sets report accuracy according to the degree of correctness of a mapped class. By this method, the natural variability of the forest communities can be reported relative to a continuous scale ranging from full membership, to partial membership, to zero membership. Using the most stringent rules for class membership, the classification was 92.1% accurate, but was 96.6% accurate when fuzzy (transitional) relationships between forest types were considered. Diagnostic statistics indicated the magnitude of classification correctness and the degree of confusion and/or ambiguity for classes at various levels of the classification scheme. Assessing classification accuracy through a continuous scale of membership simulated the natural variability and transitional nature of the forested wetland communities within the study area.