Shape-based unmixing for vegetation mapping

Spectral mixture analyses (SMA) is often used as a tool to map complex/mixed (semi-)natural ecosystems. Yet, the performance of SMA, which traditionally uses the amplitude-based RMSE as the objective function, is often hampered by the high spectral similarity among co-occurring plant species. Experiments, based on ray-tracing simulations, in situ measured reflectance data and AVIRIS imagery demonstrated the added value of implementing shape-based error metrics in the unmixing of forests and orchards. The approach allowed to highlight the subtle spectral differences among co-occurring plant species resulting in an overall improvement of species specific mapping (i. e. decrease in MSE ≈ 40%).