Quantifying Urban Vegetation Type Fractions Using Signal Unmixing with Airborne Spectroscopic Imagery

Airborne spectroscopic imagery could facilitate city-scale analyses of urban vegetation form and function. However, doing so is complicated by small-scale urban surface heterogeneity as well as vegetation spectral similarity. In 2014, NASA's HyspIRI Preparatory Campaign collected visible near infrared/shortwave infrared imagery over Santa Barbara, CA, using the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) (18 m resolution, 224 bands, 0.35 - 2.5 μm). We used endmembers (representative spectra) extracted from this imagery and Multiple Endmember Spectral Mixture Analysis (MESMA) to estimate sub-pixel fractions of impervious, pervious, and vegetated surfaces. We then used signal unmixing, in which per-pixel MESMA fractions and endmember spectral signatures are normalized to remove the non-vegetation signal and recover a "pure" vegetation spectrum. This spectrum was then classified as either turfgrass or tree. Fractions were validated using 1 m resolution aerial imagery. MESMA fractions of vegetation (r2 = 0.95), impervious (r2 = 0.92), and pervious (r2 = 0.90) surfaces were highly correlated with validation data based on manually delineated polygons. Classification of vegetation into turfgrass and tree classes following signal unmixing yielded accuracies of r2 = 0.85 for tree fractions and r2 = 0.74 for turfgrass fractions, indicating the possibility for refined inputs for future urban ecological studies. The signal unmixing/classification method was also computationally efficient, requiring MESMA to evaluate 30% fewer mixtures compared to directly estimating fractions of turfgrass and tree. Overall, these results show MESMA/signal unmixing to be a promising method for analyzing mixed vegetation spectra in urban imagery, with further potential for extracting important biophysical information from unmixed signals using vegetation indices.