Hyperspectral remote sensing of biochemical and biophysical parameters: the derivate red-edge "double-peak feature", a nuisance or an opportunity?

Improved quantification and monitoring of biophysical and biochemical attributes is required to predict the response of ecosystems to climate change and acquire deeper understanding of the carbon cycle. Remote sensing is widely viewed as a time- and cost-efficient way to proceed with large-scale monitoring of vegetation parameters. For over thirty years, use has been made of broadband sensors such as Landsat TM/ETM+. The advent of hyperspectral remote sensing or imaging spectrometry enlarged the number of available bands within the visible, near-infrared (NIR) and shortwave infrared (SWIR). Hyperspectrally detectable variables associated with 1eaf chlorophyll content, phenological state and vegetation stress such as the spectral shift of the red-edge (670-780 nm) slope and its inflection point termed the red-edge position (REP), are not accessible with broadband sensors. State of the art indices and analytical techniques applied for broad-band remote sensing are not always suitable for information extraction from high dimensional hyperspectral data. This study aimed to develop new hyperspectral indices and propose innovative ways for empirically estimating biochemical and biophysical parameters from hyperspectral data.The red edge position is estimated using the first derivative of the spectral curve. Existing curve fitting approaches localise the REP while assuming a derivative curve with a single peak. The proposed linear extrapolation method localises the red edge position while explicitly considering two peaks in the derivative curve. The major contribution of this study is that the linear extrapolation method allows optimised estimates of 1eaf chlorophyll or nitrogen content while minimising the confounding effects of background and the structure of leaves and canopy. By minimising these canopy effects, the linear extrapolation may be useful for detecting early physiological stresses associated with changes in 1eaf chlorophyll/nitrogen levels. The linear extrapolation method also shows high potential for discriminating tree and shrub species at both the leafand canopy scales. Lastly, it could be used as a more stable predictor for monitoring green grass biomass in the Majella National park, Italy compared with two-band vegetation indices. The method is simple to implement, but sensitive to spectral noise. Spectral smoothing is recommended when noise is a problem.The study also highlights the utility of partial least squares (PLS) regression based on airborne hyperspectral imagery (HyMap) for estimating grass biomass and beech (Fagus sylvatica L.) forest mean diameter-at-breast height (DBH) in the Majella National Park, Italy. PLS regression produced lower prediction errors for grass biomass and beech forest mean DBH compared with univariate regression involving vegetation indices such as NDVI. NDVI may be simple to implement but could be lacking in terms of exploiting the information content inherent in several narrow bands.In a nutshell, this study makes a contribution in the domain of information extraction from hyperspectral data for estimating vegetation parameters such as 1eaf chlorophyll/nitrogen concentration, grass biomass and forest structural parameter using empirical models. Other studies are focused on developing physically based methods given the lack of robustness and portability of empirical models for varying environmental conditions. However, empirical models that are less sensitive to environmental conditions such as models based on the linear extrapolation REP could be used to support the development of physically based models, particularly to estimate the values of the model parameters, or to refine the underlying concepts on which the model is constructed. The future of hyperspectral remote sensing could hinge or enhancing the link between empirical and physically based approaches.