Chapter 2
Ecological Applications of Unsupervised Artificial
Neural Networks
J.L. Giraudel . S. Lek
2.1 Introduction
For ecological data, cluster analysis (CA) is basically a classification technique for
sorting sampie units into groups based upon their resemblance. Several algorithms
permit similar work, but due to the heuristic nature of the methods, it is impossible
to choose the 'best' one (Van Tongeren 1995). Some conventional methods need
prior knowledge of the number or of the size of the clusters. On the other hand,
particular shapes of the clusters may lead to erroneous conclusions. The results are
commonly displayed on dendrograms becoming hard to interpret for huge
datasets.
Inspired by the structure and the mechanism of the human brain, the Artificial
Neural Networks (ANNs) should be a convenient alternative tool to traditional
statistical methods. ANNs have already been successfully used in ecology (Lek
and Guegan 1999). Whereas, the backpropagation algorithm in a supervised
learning way is mostly used in the ecological applications (Lek et al. 2000), only a
few works use unsupervised learning and more specially the Self-Organizing
Maps (SOM) algorithm. Nevertheless, SOM has been demonstrated in
patternizing ecological communities (Chon et al. 1996) and has been applied to
the analysis of community data (Foody 1999) or to model microsatellite data
(Giraudei et al. 2000).
The Kohonen SOM is the most famous neural network with an unsupervised
learning rule; it performs a topology-preserving projection of the data space onto a
regular two-dimensional space and can be used to visualize clusters effectively.
The main aim of this paper is to demonstrate a practical methodology for the
best use of SOMs for community classification. The presentation will be made
using a well-known dataset: upland forests in Wisconsin, USA (Peet and Loucks
1977). Firstly, the methodology leading to a good learning process will be
specified. It will be shown that the SOM is a good tool to interpret the
classification with abundance data or abiotic variables. Then, computing the
unified-matrix, an effective way to obtain clusters will be explained and some
indices to evaluate the quality of the map will be given.
Ecological Applications of Unsupervised Artificial
Neural Networks
J.L. Giraudel . S. Lek
2.1 Introduction
For ecological data, cluster analysis (CA) is basically a classification technique for
sorting sampie units into groups based upon their resemblance. Several algorithms
permit similar work, but due to the heuristic nature of the methods, it is impossible
to choose the 'best' one (Van Tongeren 1995). Some conventional methods need
prior knowledge of the number or of the size of the clusters. On the other hand,
particular shapes of the clusters may lead to erroneous conclusions. The results are
commonly displayed on dendrograms becoming hard to interpret for huge
datasets.
Inspired by the structure and the mechanism of the human brain, the Artificial
Neural Networks (ANNs) should be a convenient alternative tool to traditional
statistical methods. ANNs have already been successfully used in ecology (Lek
and Guegan 1999). Whereas, the backpropagation algorithm in a supervised
learning way is mostly used in the ecological applications (Lek et al. 2000), only a
few works use unsupervised learning and more specially the Self-Organizing
Maps (SOM) algorithm. Nevertheless, SOM has been demonstrated in
patternizing ecological communities (Chon et al. 1996) and has been applied to
the analysis of community data (Foody 1999) or to model microsatellite data
(Giraudei et al. 2000).
The Kohonen SOM is the most famous neural network with an unsupervised
learning rule; it performs a topology-preserving projection of the data space onto a
regular two-dimensional space and can be used to visualize clusters effectively.
The main aim of this paper is to demonstrate a practical methodology for the
best use of SOMs for community classification. The presentation will be made
using a well-known dataset: upland forests in Wisconsin, USA (Peet and Loucks
1977). Firstly, the methodology leading to a good learning process will be
specified. It will be shown that the SOM is a good tool to interpret the
classification with abundance data or abiotic variables. Then, computing the
unified-matrix, an effective way to obtain clusters will be explained and some
indices to evaluate the quality of the map will be given.