Epsilon Archive for Student Projects

Abstract

Ley is the most widespread crop in Sweden and is the base of ruminant diets. Improving ley quality
is necessary for increasing the proportion of forage in ruminant diets, offering both environmental
and economic benefits. Harvest timing is critical to achieving the desired ley quality, but a decision
support system for optimizing all harvests is currently lacking. Remote sensing is a promising tool,
and the Arable Mark 3 (AM3), a low-cost commercial field weather station including a spectrometer
with 21 bands, has the potential to estimate ley quality directly in-field. The purpose of this study
was to evaluate the potential of AM3 as a decision support tool to assist farmers in determining the
optimal time for harvesting ley. The AM3 was evaluated using two devices: one stationary, which
collected weather and plant data throughout the season, and one mobile unit, which served as the
primary data collector for analysis. The stationary AM3 demonstrated the versatility of the AM3 as
weather station. The mobile AM3 measured spectral reflectance from the ley, followed by cutting
the ley for laboratory analysis of nutritive values. A total of 43 ley samples were analysed across
two regions in Sweden during the 2024 growing season. Multivariable regression models, including
multiple linear regression (MLR), partial least squares (PLS and caretPLS), and support vector
machine (SVM), were applied to find relationships between spectral data and laboratory analyses of
crude protein (CP), metabolizable energy (ME), neutral detergent fibre (NDF), organic matter
digestibility (OMD), and dry matter (DM). PLS, caretPLS and SVM used internal tenfold crossvalidation
to prevent overfitting. MLR lacked regulations, likely resulting in overfitting. Using data
from narrow bands alone gave the lowest R2 (0.59-0.71), while combining narrow and wide bands
increased R2 (0.59-0.87), mainly for CP and NDF. The results showed that PLS performed best for
CP and NDF. caretPLS and SVM performed for ME, OMD, and DM, with narrow bands alone and
narrow and wide bands together. Adding additional spectral data from the AM3 such as normalized
difference vegetation index and chlorophyll index as auxiliary predictors improved R2 (0.75-0.82)
for ME, OMD and DM, with risk of overfitting. The study highlights the potential of the AM3 as a
practical decision support tool for timing all ley harvests but much larger diverse datasets are
necessary to build reliable predictive models. Involving farmers in decision support system
development will ensure applicability of the AM3.