Agriculture Reference
In-Depth Information
dependent data (X variables) express common information, as usually happens in spectral
data. The required number of PCs is typically smaller than that in a PCR calibration model
for similar model performance (Wold et al. 1983; Esbensen et al. 2002; Nicolaï et al. 2007).
2.2.4 Artificial neural networks
Artificial neural networks (ANNs) are based on their ability to “learn” during a training
process where they are presented with inputs and a set of expected outputs. The basic
structure of an ANN consists of three layers of “nodes” or “neurons”: an input layer (i.e.
spectral data or PCs), a hidden layer (which can consist of one of more nodes), and an
output layer (which combines the outputs of the hidden layer into a single output). The
node is a fundamental processing unit; each node has a series of weighted inputs, either
from an external source or the output from other nodes. The inputs to the node are
analogous to synapses, and the weights correspond to the strength of the synaptic
connection. The sum of the weighted inputs is transformed with linear or nonlinear transfer
functions, a popular nonlinear transformation function being the sigmoid function. The
learning (or training) is an iterative process in which the resultant output is compared to the
provided expected output, and an algorithm adjusts the weights accordingly. This method
was first tested in the field of spectroscopy on simulated data (Long et al. 1990). It was
proven to provide the best results in modeling soil constituents in a recent large-scale study
that included over 1,100 soil samples (Rossel & Behrens 2010) .
3. Heavy metals
3.1 Introduction
Heavy metals are released into the environment—the air (e.g.
during combustion, extraction
and processing), surface water (via
runoff and release from storage and transport) and soil
(and hence into the groundwater and crops). Although the adverse health effects of heavy
metals have long been known, exposure to heavy metals continues and is even increasing in
some areas, especially in less developed countries. The main threats to human health from
heavy metals are associated with exposure to Pb, Cd, Hg and As. Heavy metal exposure
may cause headaches, irritability, abdominal pain, kidney damage, skeletal damage, acute
pulmonary effects, cardiovascular disease, chronic renal failure, cancer, lung damage,
neurological and psychological symptoms, nervous system damage and much more (Järup
2003). Heavy metals as soil contaminants typically occur in low concentrations (10,000 mg
kg
-1
dry soil). Inorganic material in general does not exhibit characteristic absorption
features in the VNIR-SWIR wavelength region. It is therefore considered impossible to
directly detect inorganic soil contaminants in general, and heavy metals in particular
(Winkelmann 2005). The conventional method of estimating the spatial distribution of heavy
metals is by raster sampling and a time-consuming laboratory analysis, followed by
geostatistical interpolation (Kemper & Sommer 2002). While pure metals do not absorb in
the VNIR-SWIR region, they may be detected indirectly via their complexing with organic
matter (OM), association with moieties such as hydroxides, sulfides, carbonates, or oxides
which are detectable, or adsorption to clays that absorb light in this wavelength range
(Malley & Williams 1997). To explain possible variations in spectral signals due to heavy
metals bound to minerals, it is necessary to consider the binding reaction of the metal onto
the mineral surface. This approach starts from the premise that the spectrally assigned
position of minerals can change with chemical composition and surface activity (Ben-Dor et
