Download - Institute of Photogrammetry and Remote Sensing

Abstract
High resolution multi- and hyperspectral remote sensing data form the basis of important spatially distributed
biogeophysical products in support of crop management at a local to regional scale. Due to the
infrequent and irregular data availability from single high resolution sensors, multi-sensor approaches are
required to fill the gaps in data sequences and to guarantee data takes at critical time steps during the
seasonal life-time of a crop. Prerequisite for such combined use is the consistency of single sensor products
and the compatibility between the products of different origin. Developing robust algorithms for the
retrieval of vegetation variables from remote sensing data is often hindered by the underdetermined nature
of the problem, which is caused by the inferior number of independent earth observation dimensions
compared to the large number of canopy elements causing variations in the radiometric signal. The main
objective of this thesis is to address this issue in the context of developing an automated, image-based
approach which is applicable to a wide variety of airborne and spaceborne high resolution sensors.
The proposed approach, entitled CRASh, provides a concurrent retrieval of the key agricultural variables
leaf area index (LAI), leaf chlorophyll content (Cab), leaf water content (Cw), and leaf dry matter
content (Cdm), based on the combined inversion of the leaf optical model PROSPECT (Jacquemoud and
Baret, 1990; Fourty et al., 1996) and the 1-D turbid medium canopy structure model SAILh (Verhoef,
1984; Verhoef, 1985). The inversion of these radiative transfer models allows for the exploitation of the
entire information content contained in the data and facilitates optimizing for illumination/observation
geometry and adapting to site specific phenology, background reflectance, and atmospheric conditions.
In CRASh, model inversion is based on a lookup table (LUT) approach, which, apart from being computationally
fast, offers maximum flexibility to changing input parametrization.
The presented approach entirely relies on the spectral image content and the information products
provided by automated preprocessing of the data. Under this assumption, the lack of a priori knowledge
on land cover causes the inversion process to be strongly underdetermined and ill-posed, indicating that
the system has multiple solutions. Two novel regularization techniques are proposed in this respect: the
incorporation of an automated spectral land cover classifier, and the assimilation of preliminary, a priori
estimates of the output vegetation variables, calculated in place using predictive regression equations.
Moreover, a new method, based on the local spatial neighborhood of the pixel under inversion, is suggested
for dealing with systematic small scale attribution inconsistencies resulting from the land cover
classification.
The suggested land cover classification facilitates a more explicit characterization of spectral uncertainties
and appeared very effective in reducing the uncertainties related to the individual inversion
results, since the LUTs used for model inversion could be optimized for each land cover type. Moreover,
the LUTs that are calculated separately for every land cover class and illumination/observation geometry
allow for the generation of semi-empirical predictive equations optimized for each specific situation.
These equations are based on the regression between spectral vegetation indices (VIs) calculated from
the reflectance spectra contained in the LUT, and the variables used to simulate this canopy reflectance
in the forward mode. The predictive equations obtained in this way are subsequently used to calculate a
priori estimates of each variable and for every vegetation pixel in the image. The prior estimates, and the
assessment of covariance between the several variables based on these estimates, play a decisive role in
the stabilization of the model inversion and the reduction of ambiguous results between variables invoking
complemental spectral behavior.