Visualisation of Hidden Neuron Behaviour in a Neural Network Rainfall-Runoff Model

Abstract

This chapter applies graphical and statistical methods to visualise hidden neuron behaviour in a trained neural network rainfall-runoff model developed for the River Ouse catchment in northern England. The methods employed include plotting individual partial network outputs against observed river levels; carrying out correlation analyses to assess relationships among partial network outputs, surface flow and base flow; examining the correlations between the raw hidden neuron outputs, input variables, surface flow and base flow; plotting individual raw hidden neuron outputs ranked by river levels; and regressing raw hidden neuron outputs against river levels. The results show that the hidden neurons do show specialisation. Of the five hidden neurons in the trained neural network model, two appear to be modelling base flow, one appears to be modelling surface flow, while the remaining two may be modelling interflow or quick sub-surface processes. All the methods served to provide confirmation of some or all of these findings. The study shows that a careful examination of a trained neural network can shed some light on the sub-processes captured in its architecture during training.