Multi-season unmixing of vegetation class fractions across diverse Californian ecoregions using simulated spaceborne imaging spectroscopy data.


Spaceborne imaging spectrometers are expected to facilitate regional-scale vegetation analyses with multi-season hyperspectral imagery. However, we still lack a better understanding on both whether multi-season hyperspectral approaches are favorable over single-season approaches, as well as on the benefits of hyperspectral compared to multispectral data. Our study investigates the potential of multi-season unmixing of simulated Environmental Mapping and Analysis Program (EnMAP) data for vegetation class fraction mapping across diverse natural and semi-natural ecoregions in California, USA. We utilized spring, summer and fall 2013 simulated EnMAP imagery derived from Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) data covering study sites in the San Francisco Bay Area, Lake Tahoe and Santa Barbara. Regression-based unmixing with synthetic training datasets from spectral libraries was implemented for mapping needleleaf tree, broadleaf tree, shrub, herbaceous and non-vegetation fractions, and independent reference data was used for validation. Multi-season unmixing of simulated EnMAP had average Mean Absolute Errors (MAE) over all classes of 8.7% for the Bay Area, 8.5% for Lake Tahoe and 9.6% for Santa Barbara. However, larger errors in the low and high end of the fraction range remained, particularly in open-canopy woodlands and xeric shrub-dominated regions. Single-season unmixing of simulated EnMAP revealed large seasonal and regional variations within individual vegetation classes. In most cases, the best performing single-season unmixing had similar errors as the multi-season unmixing, i.e., ∆MAEs within ±1.0%. This points to the advantage of the multi-season integration strategy for more robust and generalized mapping independent from season and study site. Relative to EnMAP analyses, multi-season unmixing of multispectral Landsat composites for the same seasons yielded increases in average MAEs of +1.7%, +2.3% and +1.4% for the three study sites. This indicates that the higher spectral resolution of simulated EnMAP provides more relevant discriminative information when comparing contemporary image pairs. Unmixing of seasonal spectral-temporal metrics (STMs) from all available Landsat images for an entire year took advantage of the full temporal detail provided by these ongoing missions. We found Landsat STMs to effectively map vegetation class fractions, with average MAEs of 9.9%, 10.0% and 9.7% for the three study sites. Still, improvements particularly for mapping fractions of the woody vegetation classes through multi-season unmixing of simulated EnMAP point to the benefit of high spectral resolution data, and we assume that a comparable higher temporal resolution of hyperspectral satellites will further positively influence results. Overall, we conclude that multi-season unmixing of spaceborne imaging spectroscopy data holds great potential for advancing vegetation class fraction mapping across natural and semi-natural ecosystems.

Remote Sensing of Environment