Spatially-explicit distribution maps of tree species are increasingly valuable to forest managers and researchers, particularly in light of the effects of climate change and invasive pests on forest resources. Yet, current regional and national forest cover maps provide only coarse classifications (e.g. deciduous, evergreen, or mixed) with minimal field validation. Advanced remote sensing techniques, such as spectral unmixing and object-based image analysis (OBIA), offer a novel approach to mapping species distributions. Building on the spectral depth of multi-temporal imagery, spectral unmixing allows us quantify percent basal area on a per-species basis, outperforming traditional pixel-based classifiers. This is particularly useful in northern forests where species composition is often highly heterogeneous. OBIA allows us to refine basal area maps with a series of ancillary environmental data layers on an object, rather than pixel, basis. Because spectral unmixing relies on calibration with a robust set of “pure” spectral signatures, percent basal area mapping across the region was limited to ten common species/genera including: Abies balsamea, Acer rubrum, Acer saccharum, Betula spp., Fagus grandifolia, Picea rubens, Pinus strobus, Populus spp., Quercus spp., and Tsuga canadensis. In addition to identifying pure stands of these species, which were relatively rare across the landscape, the OBIA rule set classified six common forest assemblages. Here, we present a comparison of this integrated unmixing and OBIA method for mapping of forest composition to other large-scale forest mapping efforts (National Land Cover Database and LANDFIRE) for the region covered by Landsat Path 14, Row 29 (northern New York and Vermont).
Independent validation with 67 inventory plots located across Vermont, covering a range of species compositions, indicate that this integrated classification was 59% accurate in classifying down to species-type level. Based on the same validation data set, LANDFIRE was able to classify to species-type level with 33% accuracy. Because no species-level comparison was possible with the coarse (evergreen, deciduous, mixed) classifications of the NLCD product, both the integrated and LANDFIRE classifications were fuzzed to NLCD categories. Accuracy for this coarse forest type was highest for the integrated product 72%, followed by 57% for the NLCD product, and 39% for the LANDFIRE classification. These results indicate that higher quality, more detailed species mapping is possible using this novel integrated approach. Efforts are currently underway to repeat this classification for the broader Northern Forest region.