Comparison of dynamic feature map models for environmental monitoring

The self organizing feature map (SOM) model developed by Kohonen has the ability to generate a mapping from a possibly high-dimensional feature space onto a discrete two-dimensional space of activation in a lattice of neuron-like processing elements. Although this model has proved efficient in a variety of applications, it suffers from some limitations such as the lack of adaptation of the map architecture. This problem is particularly acute in online monitoring applications, in which the input distribution varies as a function of time, and the training set cannot be assumed to be representative of all possible system states. In this paper, three topology feature map models with architecture adaptation are applied to water quality data and compared to the basic SOM model with fixed architecture using five objective performance criteria. The BungySOM model is shown to outperform the other constructive algorithms and to carry out faster adaptation to new inputs as compared to the SOM model