Enhanced forest cover mapping using spectral unmixing and object-based classification of multi-temporal Landsat imagery

Abstract Spatially-explicit tree species distribution maps are increasingly valuable to forest managers and researchers in light of the effects of climate change and invasive pests on forest resources. Traditional forest classifications are limited to broad classes of forest types with variable accuracy. Advanced remote sensing techniques, such as spectral unmixing and object-based image analysis, offer novel forest mapping approaches by quantifying proportional species composition at the pixel level and utilizing ancillary environmental data for forest classifications. This is particularly useful in the Northeastern region of the United States where species composition is often mixed. Here we employed a hierarchical forest mapping approach using spectral unmixing of multi-temporal Landsat imagery to quantify percent basal area for ten common tree species/genera across northern New York and Vermont. Basal area maps were then refined using an object-based ruleset to produce a thematic forest classification. Validation with 50 field inventory plots covering a range of species compositions indicated that the quality of percent basal area mapping largely reflected the number of “pure” (> 80% BA) endmember plots available for calibration, with more common species mapped at a higher accuracy (i.e. Acer saccharum , adj. r 2  = 0.44, compared to Populus sp ., adj. r 2  = 0.24). The resulting thematic forest classification mapped 15 forest classes (nine species/genus level and six common species assemblages) with overall accuracy = 42%, KHAT = 33%, fuzzy accuracy = 86% at the pixel level, and 38%, KHAT = 29%, fuzzy accuracy = 84% at the object level. Using the validation plots to compare existing forest classification products, this hierarchical approach provided more class detail (11 represented classes) and higher accuracy than the National Forest Type Map (six represented classes, overall accuracy 18%, fuzzy accuracy 70%), LANDFIRE (five represented classes, overall accuracy 28%, fuzzy accuracy 80%) and National Land Cover Database (three represented classes, overall accuracy = 56%). These results show that more detailed and accurate forest mapping is possible using a combination of multi-temporal imagery, spectral unmixing, and rule-based classification techniques. Improved large-scale forest mapping has important implications for natural resource management and other modeling applications.