Hyperspectral shape-based unmixing to improve intra- and interclass variability for forest and agro-ecosystem monitoring

The monitoring of forests and agro-ecosystems often requires the use of a spectral mixture model to provide detailed information on spatial and temporal variations in vegetation cover. Two key issues in the mapping of vegetation cover in complex ecosystems are the high spectral similarity (i.e. low interclass variability) between and the high spectral variability among different vegetation species (i.e. high intraclass variability), as they impede the performance of the Root Mean Square Error (RMSE) criterion, traditionally used in Spectral Mixture Analysis (SMA) to optimise the fit between modelled and measured mixed signal. Shape-based objective functions have been proposed as an alternative. Experiments, based on ray-tracing simulations, indeed demonstrated the added value of implementing shape-based error metrics in unmixing of vegetation in (i) reducing the effects of intra-class endmember variability (ΔfRMSE ≈ 0.21 vs ΔfSAM ≈ 0.10) and (ii) highlighting the subtle spectral differences among similar endmembers (ΔfRMSE ≈ 0.61 vs ΔfSAM ≈ 0.43). Shape-based unmixing as such has the potential to become an alternative to the traditional but CPU intensive MESMA approach. Simulated data results presented in this study show a significant increase in fraction estimate accuracy for shape-based unmixing over MESMA (ΔfMESMA,RMSE ≈ 0.15 vs ΔfsSMA,SAM ≈ 0.10) and subsequently confirmed using three different real hyperspectral data sets.