Modelling sediment transfer in Malawi: comparing backpropagation neural network solutions against a multiple linear regression benchmark using small data sets

Abstract

The recent growth in neural network hydrological modelling has focused on the provision of river flow estimates of one kind or another. Little or no scientific research has been undertaken to assess the potential benefits for modelling sediment transfer. Some initial pathfinder experiments were therefore conducted to assess the competence of a backpropagation network to produce a combined model of sediment transfer occurring under different types of agriculture and land management conservation regimes. The results of this investigation demonstrate that a neural network solution is able to exceed the limitations of traditional multiple linear regression. The potential to create multiple solutions at different levels of generalisation and robust solutions that can be transferred to unknown catchment types is illustrated.