Abstract:
Fertiliser applications in vegetable crops are one of the main input costs of
production. Thematic soil maps have been widely used for decades to
characterise soil nutrients and, therefore, apply variable rate fertilisers.
However, traditional variable rate methods used in soil sampling are time-
consuming, costly and not accurate. Thus, they fail in providing a true estimate
of the nutrients soil needs. To obtain better crop response to inputs, a rapid,
non-destructive, timely and cost-effective soil analysis are needed to enable
site-specific fertiliser applications. Proximal soil sensing with visible and near
infrared (vis-NIR) spectroscopy is a promising tool to assist in variable rate
applications. This thesis aims to develop reliable calibration models for a
previously developed on-line visible (vis) and near infrared (NIR) spectroscopy
sensor (Mouazen, 2006), for the prediction of soil properties in vegetable crop
fields for a better N fertiliser management. Experiments were established in
crops of cauliflower (Brassica oleracea) during 2013 season (two fields) and
2014 season (three fields), in UK. A mobile, fibre-type, vis–NIR
spectrophotometer (AgroSpec, Tec5 Technology for Spectroscopy, Germany)
with a measurement range of 305-2200 nm was used to measure soil spectra in
diffuse reflectance mode, measuring up to ~1500 points per ha. Four different
calibration sets were tested to establish the most accurate calibration model for
moisture content (MC), soil organic carbon (OC), pH and total nitrogen (TN),
using partial least squares (PLS) regression analysis selected according to
different spectral library size and geographical scale: Scenario 1 (SC1 (local)),
Scenario 2 (SC2 (regional)), Scenario 3 (SC3 (national)), Scenario 4 (SC4
(continental)). The best results in cross-validation were obtained for MC with
SC2 (R2[R squared] = 0.89; RPD > 2.5), followed by SC4 (R2[R squared] = 0.88; RPD = 2.91-3.31, in
2013 and 2014, respectively); and SC1 and SC4 worked very well for MC on-
line prediction (R2[R squared] > 0.90 and RPD > 2.5). SC3 and SC4 both provided the best
performance for OC and TN in cross-validation, whereas no clear trend was
observed for on-line prediction. Poor model performance was obtained for pH in
on-line predictions (R2[R squared] < 0.30 and RPD < 0.9). Although the calibration models
using the on-line vis-NIR sensor provided good and detailed information of the
soil nutrients analysed, future research will be needed to estimate these
properties more accurately, with the aim to develop reliable vis-NIR calibration
models for the on-line measurement in vegetable crop fields.