Sediment Yield Assessment for Land Management Using Hydrological Models and Bayesian Network

Ani Sofini

March 2006

Abstract

The Nil catchment is a small and hilly area in the central Belgium. As the main usage of the land is for agriculture, the main water quality problems are related to sediments, pesticides, and nutrients. The last 2 parameters can be present in the water as sediment-bound pollutants and since the soil in the Nil catchment are very sensitive to erosion, the pollution is very related to sediment transport what is the subject of this study. Soil erosion depends very much on the agricultural management. Best management practice may lead to large reductions in sediment transport, such as the use of Vegetative Filter Strips (VFS). VFS is an area of vegetation used to reduce sediments or other pollutants from surface runoff. In order to find out the best method for application of VFS, the SWAT (Soil and Water Assessment Tool) model is used. Two scenarios are compared: applying 5 meters VFS along the main stream and 2 meters VFS in all field (HRUs: Hydrological Response Units). The results show clearly that the second scenario is much better with minimum reduction about 30% for whole catchment.

Finding a method of reduction is not enough. It has to be communicated to land and water managers or farmers who play an important role in the field-scale. The SWAT model is too complicated for this. Therefore, a simpler model using Bayesian Network is developed. Many parameters used by SWAT in calculating the sediments are simplified to a smaller number of parameters: land use, soil type, agricultural management, and geography. The relationship between those parameters and the sediment yield in the SWAT model results then used as cases by the Bayesian Network to be learned. Moreover, to increase the belief of the network, the uncertainty analysis which uses Monte Carlo – Latin Hypercube simulations, is used. The simulations include 100 combinations of parameters that are sensitive to the sediment predictions.

Although SWAT and Bayesian Network can be used for sediment modelling, there are some limitations. In the SWAT model, main difficulties in application lay on the calibration and the need to review the default values that are generated in the interface. In the Bayesian Network, the results provided by SWAT model can only directly be used for building a field-scale model.

Keywords: Sediment, Vegetative Filter Strips, SWAT, Bayesian Network, Monte Carlo – Latin Hypercube

Back to the list of MSc abstracts