Geoscience Reference
In-Depth Information
processing
. The strongest appeal of CNNs is their suitability for machine learning (i.e. compu-
tational adaptivity). Machine learning in CNNs consists of adjusting the connection weights to
improve the overall performance of a model. This is a very simple and pleasing formulation of
the learning problem. The speed of computation is another key attraction. In traditional single
processor Von Neumann computers, the speed of the machine is limited by the propagation delay
of the transistors.
However, with their intrinsic parallel distributed processing structure, CNNs can perform com-
putations at a much higher rate, especially when implemented on a parallel digital computer or,
ultimately, when implemented in customised hardware such as dedicated neurocomputing chips.
The rapid speed at which these tools can work enables them to become ideal candidates for use
in real-time applications, involving, for example, pattern recognition within data-rich geographic
information system
(GIS) and remote sensing environments. It is also clear that the ever-increasing
availability of parallel hardware and virtual parallel machines, coupled with the spatial data explo-
sion, will enhance the attractiveness of CNNs (or other parallel tools) for GC. The non-linear nature
of CNNs also enables them to perform function approximation and pattern classification operations
that are well beyond the reach of optimal linear techniques. These tools therefore offer greater rep-
resentational flexibilities and total freedom from linear model design. CNNs are also considered to
be semi- or non-parametric devices that require little or no assumptions to be made about the form
of underlying population distributions - in strong contrast to conventional statistical models. One
other important feature is the robust behaviour of CNNs when faced with incomplete, noisy and
fuzzy information. Noise, in this instance, refers to the probabilistic introduction of errors into data.
This is an important aspect of most real-world applications and neural networks can be especially
good at handling troublesome data in a reasonable manner.
CNNs have massive potential and have been applied with much success in numerous diverse
areas of geographical data analysis and environmental modelling to solve problems of various kinds
(see Fischer 1998). These are described in more detail in the following.
13.2.1 P
attern
c
laSSification
The task of pattern classification is to assign an input pattern represented by a feature vector to
one of several pre-specified class groups. Well-known applications would include spectral pattern
recognition where pixel-by-pixel information obtained from multispectral images is utilised for the
classification of pixels (image resolution cells) into given
a priori
land cover categories. However, as
the complexities of the data grow (e.g. more spectral bands from satellite scanners, higher levels of
greyscaling or finer pixel resolutions), together with the increasing trend for additional information
from alternative sources to be incorporated (e.g. digital terrain models), then so too does our need for
more powerful pattern classification tools. There is now a considerable literature on the use of CNNs
for pattern classification, particularly in relation to remote sensing, for example, Civco (1993), Foody
(1995), Miller et al. (1995), Gopal and Fischer (1997), Fischer and Staufer (1999), Tapiador and
Casanova (2003), Tatem et al. (2003), Brown et al. (2007), Wang and Xing (2008) and Tahir (2012).
13.2.2 c
luStering
/c
ategoriSation
In clustering operations (also known as unsupervised pattern classification), there are no pre-spec-
ified, or known, class group labels attached to the training data. A clustering algorithm is used to
explore the data and to determine inherent similarities that exist between the various patterns that
make up the data set. Each item is then identified as belonging to a cluster of similar patterns. Well-
known clustering applications would include data mining, data compression and exploratory spatial
data analysis. Clustering has also been used to divide river flow data into different event types or
hydrograph behaviours using a self-organising map, where an event is taken to mean a short section
of a hydrograph. Feedforward neural networks were then used to develop different models on each
