Event Runoff and Sediment-Yield Neural Network Models for Assessment and Design of Management Practices for Small Agricultural Watersheds

Abstract

This study presents the development of novel artificial neural networks (ANN) models for assessment of best management practices (BMPs) for controlling runoff and sediment yield from small agricultural watersheds. The ANN models integrate complex nonlinear effects of key climatic, topographic, drainage, and management characteristics and can evaluate BMP effectiveness without presumptions about their physical mechanisms or performance. Thirty-two ANN models were developed and tested. Penalty-related criteria and statistical model performance evaluation parameters were used to select the two models (one for runoff, one for sediment yield) with an optimum number of input parameters and hidden nodes. Event based monitoring data (n=248) at the outlet of seven watersheds (1.4–30.2 ha) in southwestern Wisconsin were used to train, validate, and test the models. Statistical parameters (e.g., R2=0.76–0.93) suggested that the ANN models performed well. Sensitivity analysis for the BMP parameters showed that the runoff model was heavily influenced by length of grassed waterway and channel density; the sediment-yield model was mainly affected by upland crop type and tillage.