Geoscience Reference
In-Depth Information
•
Efficient operational capabilities for dealing with large, ultra-large and massive spatial
data sets, together with the associated prospect of obtaining better results through being
able to process finer-resolution data or to perform real-time geographical analysis
•
Built-in dynamic capabilities for adapting connection weights with changes in the sur-
rounding environment (dynamic learning)
•
Good generalisation (out-of-sample performance) capabilities that work in a specific and,
in general terms, quite satisfying manner
•
Potential improvements in the quality of results associated with a reduction in the number
of rigid assumptions and computational shortcuts that are otherwise introduced using con-
ventional methodologies and techniques
13.8.2 a
PPlication
d
oMainS
CNN models, in particular hidden-layered feedforward networks, together with their wide range
of different recognised learning techniques are now able to provide geographers with novel,
elegant and extremely valuable classes of mathematical tools - all based on sound theoretical
concepts - for geographical data analysis and modelling. Moreover, such tools are not intended
to be substitutes for traditional methods, but should instead be viewed as being non-linear exten-
sions to conventional statistical methods such as regression models, spatial interaction models,
linear discriminant functions, pattern recognition techniques and time series prediction tools
(Pao 1989; White 1989; Fischer and Gopal 1994; Fischer et al. 1994; Fischer 2013, 2014). Much
work has to date been done in what are now seen as being the two major domains wherein these
tools are most applicable:
•
As universal function approximators in areas such as spatial regression, spatial interaction,
spatial site selection and space-time series modelling
•
As pattern recognisers and classifiers, which function as intelligent aids and allow the user
to sift through copious amounts of digital data in a fast and efficient manner, to implement
multidimensional data reduction based on otherwise unknown properties and, where
possible, to find patterns of interest in data-rich environments, for example, census small
area statistics and high-resolution remote sensing images
Feedforward CNN networks, within a geographical analysis and modelling context, are often
implemented for complex function approximation purposes. A simple three-stage process has
therefore been proposed for the application of such tools and an illustration of this method is
provided in Fischer and Gopal (1994):
1. Identification of a candidate model from a range of multilayered feedforward CNN options
and specific types of non-linear PE (e.g. perceptron or RBF)
2. Estimation of network parameters for the selected CNN model and optimisation of model
complexities with respect to a given training set (using regularisation, network pruning or
cross validation)
3. Appropriate testing and evaluation of the final CNN model in terms of its generalisation
capabilities (out-of-sample performance)
13.8.3 e
xaMPleS
In the following paragraphs, two different geographical examples are provided in order to give a
general impression of what is and is not possible with regard to CNN usage.
In the first example, Fischer and Gopal (1994) used a one-hidden-layer backpropagation
network, with sigmoidal PEs, to model interregional telecommunication traffic in Austria.
