Neural Network

• Setting Analysis Parameters
• Generate Random Training Points
• Generate Neural Network Input Files
• Run Neural Network Module
• Read Results from Neural Network Module

Note: Neural Network Analysis is performed in a Visual C++ program called "DataXplore" and "GeoXplore".

RBFLN stands for radial basis functional link network. This method uses a training dataset of "deposit" and "non-deposits" combined. This is a type of supervised classification.

The fuzzy clustering method finds clusters of unique conditions in the evidential themes and does not use training data (unsupervised).

The Probabilistic Neural Network (PNN) uses training sites of "deposits" and "Not deposits" combined and allows for a fuzzy membership attribute of these training sites. This is a type of supervised classification.

The PNN method is applied to the same files as the RBFLN. The Fuzzy clustering can be trained on the class.dta or train.dta tables. Training is limited to 300 exemplars; so if the class.dta table has more than 300 unique conditions, use the train.dta for training..

A fuzzy membership attribute between [0,1] can be added to the training points for use with the RBFLN and PNN methods. This fuzzy membership provides a way of ranking the training sites. For example a model of large deposits could give the Deposits Training sites a fuzzy membership based on the size of the deposits. These might go from 0.5 for a small deposit to 1 for a large deposit. Then the "Not Deposits" training sites could all have memberships of zero or some other appropriate value.

The evidence rasters used to create the unique conditions raster (SDMUC#) are recorded in the Description box of the General tab in the raster Properties window.

Symbolization of the results from GeoXplore have been a problem. It has been found that symbolization works best using the quantile method with a larger number of classes. Fifteen to 20 classes is often a good starting place.

Setting Analysis Parameters

1. Select 'Set Analysis Parameters...' from the ArcSDM3 menu.
2. It is required that a "non-deposit" training point shapefile theme exists before this process can be completed. One way to create a "non-deposits' training set, is to use the 'Generate Random Training Points' function.
3. Check the RBFLN (Supervised) and/or Fuzzy Clustering (Unsupervised) options listed under 'Neural Network Analysis'.
4. Select:
   • a study area grid theme
   • unit area
   • "deposit" training point theme (Supervised option)
   • "non-deposit" training point theme (Spervised option)
   • default value for defining missing data
5. Click OK.