Hyperspectral remote sensing of vegetation - a transect approach

2017-02-28T00:29:20Z (GMT) by Amiri, Reza
Human-induced global environmental changes are increasingly occurring at larger scales. Terrestrial vegetation is largely affected by such anthropologic land transformations. As a result, the ability to monitor the status of terrestrial vegetation is essential for understanding and managing these changes. The rich spectral information contained in hyperspectral data provides a promising source of information for earth observation of global change. However, the analytical methods for the retrieval of vegetation bioindicators from hyperspectral data are suggested to lack spatial transferability. This is important because spatial transferability is the underlying assumption in employing these methods at large scales. Therefore, to apply these analytical approaches confidently, study of their spatial transferability is required. Thus, the aim of this thesis is to assess the robustness of currently dominant empirical methods in the context of a sub-continental environmental gradient. In the first part of the study, the performance of commonly used spectral vegetation indices for the retrieval of leaf biochemical constituents was systematically assessed along a strong rainfall gradient in savannas of northern Australia. The results demonstrated that in cross-site situations the performance of the estimation of the foliar biochemical properties was dependent on the biochemical constituent. For example, estimation of leaf nitrogen content was largely consistent at the sampling sites while leaf chlorophyll and carotenoid contents were affected by fluctuations along the gradient. Furthermore, the study of the performance of the indices in a cross-species situation revealed that except for carotenoid content the narrowband predictors were species specific. These findings indicate that the observed inconsistency of the vegetation indices at the scale of this study is likely to affect the applications that utilise the prediction of leaf biochemical properties provided by these indices. The second part of the study assessed the robustness of partial least square regression (PLSR) multivariate technique for the retrieval of leaf biochemical properties along the NATT. The results showed that PLSR provided more consistent predictions across the sites along the gradient. This provided evidence that multivariate methods may be a better alternative in large scale estimations of biochemical constituents. Additionally, the spatial transferability of the partial least square regression technique was assessed and compared to the vegetation indices. It was demonstrated that no method was able to produce solutions transferable to the whole transect. The final part of the study incorporated the large scale transferability as an objective in a multiobjective optimisation framework to design transferable hyperspectral predictors of foliar biochemical properties. The method introduced improvements in the vegetation indices based estimations by finding an optimal waveband demonstrating both stability and performance in the predictions along the NATT. In summary, findings from this work contribute to the understanding of the reliability of the currently dominant information retrieval methods from narrowband hyperspectral reflectance data. The multiobjective optimisation method implemented in this work is of added benefit by providing a framework for addressing the issue of transferability.