| dc.description.abstract | The growing global food demand of the 21st century relies increasingly on the input of fertilization
for high agricultural production quota (Ehrlich et al., 1993; Sattari et al., 2016). An example is
the fertilized grasslands from West-European high-intensive dairy farming systems, which serve
as a source for qualitative forage products for ruminants (Andersen et al., 2007). Extensive use
of fertilization has, however, a harmful effect on the grassland ecosystem (soil, air and water
quality) and the environment (greenhouse gasses). The unbalanced use of fertilization also has
negative effects on the individual farmer since a farm is dependent on the long-lasting quality of
the grassland ecosystem (Schellberg et al., 1999).
Precision farming which uses hyperspectral UAV-remote sensing methods as a tool could provide
a solution for a balanced use of fertilization for grasslands (Hoving et al., 2015). Hyperspectral
UAV-based remote sensing is a well-tested method in rangeland research (Rango et al., 2009;
Mccollum et al., 2017). However, its use for the estimation of grasslands of high-intensive dairy
farming systems is still in its infancy (Capolupo et al., 2015).
This research compares the prediction qualities of two statistical regression models for the estimation
of structural and biochemical grass traits relevant for grass forage goals for three different
moments in the grass growing season (May 2017 – October 2017), including an integrated dataset
consisting of all three moments. This research compares prediction quality of linear regression
models using narrow-band vegetation indices and Partial Least Square Regression models using
the hyperspectral dataset acquired with the HYMSY-system from Wageningen University &
Research mounted on a UAV. The research set-up exists in an experimental field with different
fertilization applications. Validation data is acquired from field measurements and weather statistics.
This research recommends the use of PLSR models for the estimation of structural and biochemical
grass traits, especially when using an integrated dataset of different acquisition moments.
Linear models using GNDVI, MTCI or TCARI/OSAVI as vegetation index have a strong
performance for structural grass traits for single moments in the growing season. The results also
show that grass has different trait-compositions for the selected moments in the growing season,
which influence the prediction qualities of the regression models accordingly. Future research
should, therefore, incorporate multiple moments in the grass growing season: This research has
indications that especially the moment after the start of the growing season (June) has a deviating
trait-pattern. | |
