Bayesian Networks for spatio-temporal integrated catchment assessment

ABSTRACT

In this thesis, a methodology for integrated catchment water resources assessment using

Bayesian Networks was developed. A custom made software application that combines

Bayesian Networks with GIS was used to facilitate data pre-processing and spatial modelling.

Dynamic Bayesian Networks were implemented in the software for time-series modelling.

The structures of three Bayesian Network models were created automatically using a Hybrid

Genetic Algorithm (HGA) which was implemented in a custom developed software product.

The creation of the networks was done in a one step process with the discretisation of the

continuous datasets. The discretisation was done using an equal binning method and the three

networks resulted from variations in the number of intervals defined for the bins. The three

networks were scored using the error rate, the logarithmic metric, the Brier score and the

spherical score. From this evaluation, the states of the continuous variables were finalised and

the optimum Bayesian Network model (the one with the most favourable scores) emerged. The

model was then populated with the data collated for the Great Kei catchment in the Eastern

Cape Province in South Africa. The results were used to explore scenarios on the likely impacts of variations of some query

variables over other variables in the network. This was performed through sensitivity analyses,

scenario analyses and ―what if‖ analyses. The findings from the model conform to existing

knowledge on the study area which illustrated that Bayesian Networks can be successfully

applied in integrated catchment assessment. The use of Bayesian Networks for spatial

prediction was successfully proven with an example on the effects of surface water EC in one

catchment on other neighbouring catchments. This information can be used in assessing the

likely impacts of changes in surface water quality on connected catchments.

Lastly, the capability of Dynamic Bayesian Networks for temporal prediction was

demonstrated. Dynamic Bayesian Networks were tested for predicting monthly rainfall and

temperature and the results compared to that obtained from the static Bayesian Network. The

results showed that Dynamic Bayesian Networks provided better predictions mainly because of

the ability to incorporate evidence from the preceding months.

The major finding is that there is need for adequate data at the required scale. This was evident

from the fact that some well-known relationships from theory could not be established using

the automatic structure mining method used. The importance of selecting the appropriate

discretisation technique was also highlighted by the different patterns obtained with variations

in the discretisation levels. In the absence of the required data, expert knowledge should be

collected and used to inform the modification of the relationships obtained using automatic

methods and for the infilling of gaps in data.