River level forecasting and visualisation: the integration of neural networks and GIS

Abstract

This project was designed to combine pre-existing technologies and techniques to provide a customised interface to aid in the decision making process of river monitoring practices, it is a pilot study built using a variety of components any of which are upgradeable due the open source platform of Visual Basic upon which it is deployed.

A basic neural network was employed to generate a water level estimate from rainfall data supplied by NIWA. River level estimations were not statistically significant and the model did not perform as expected.

The Hukahuka river catchment is very small which initially was an attraction for the use of the river but its size meant its reactions were very rapid and in many cases could potentially exacerbate the stochastic elements involved with river levels.

Further analysis of the data revealed the presence of a large baseflow component which resulted in the river flowing year round despite only 43 days of rainfall during the test data year of 1995. Because of the large baseflow in the river the link between rainfall and river levels was minor which affected accuracies of the estimations from the neural networks. This combined with the inability of the neural network to consider historical rainfall events and flow decay curves accounted for the problems encountered with the application of a basic neural network.

The creation of a lightweight viewer application was successful as a means to rapidly display the data generated by the neural network. ESRI’s map objects provided a quick and simple method to assemble a GIS capable stand-alone application. However real-time data inking between the neural network and the viewer was not achieved due to time and technical considerations

Ultimately the project demonstrates that the technology has potential but careful selection of the river site is required to ensure that this approach is suitable for the river.