Modelling Sediment Trapping by Non-Submerged Grass Buffer Strips Using Nonparametric Supervised Learning Technique

Abstract

Grass strips are known as one of the most effective management practices in controlling sediment loss to rivers and other surface water bodies. Some physically-based models have been previously developed to predict the amount of sediment retention in grass strips. Although physically-based models can explain the effects and interactions of various factors, they tend to be sophisticated as they require a large amount of input data. A nonparametric supervised learning statistical model was developed to predict the efficiency of grass strips in trapping sediments. Grass type and density, inflow sediment particle size distribution, slope steepness, length of strip, and the antecedent soil moisture were the five major factors on which the statistical model was built. The model was assessed by comparing with an independent dataset. Estimated bias, coefficient of model efficiency, mean absolute percentage error, Pearson product-moment correlation coefficient of the model were 1.01, 0.54, 18.1and 76% respectively. Testing the model predictions, permuting the input data, showed that inflow sediment particle size distribution, length of the buffer strip, and the antecedent soil moisture are the most important factors upon the performance of grass strips in trapping sediments. From the model outputs for a range of likely scenarios it was concluded that very long strips are needed in extreme conditions such as steep slopes, wet soil and sparse grass strips in order to trap sediments effectively.

Keywords: grass strip, model, regression tree, sediment, supervised learning