Modeling stream baseflow nitrate concentration in an agricultural watershed using neural network and bootstrap method

Abstract

The evaluation of water resource vulnerability to nonpoint source pollution in the presence of uncertainty remains a crucial concern. To enhance the accuracy of the assessment and pinpoint the key factors that affect watershed water quality, the integration of an artificial neural network (ANN) into the evaluation process is imperative. The research involved the collection of streams baseflow samples from thirty-eight sites between 1994 and 1999 in the Tomorrow-Waupaca River Watershed in central Wisconsin, USA, with a focus on examining the relationship between nitrate concentrations and a range of environmental and land use variables extracted from the watershed GIS database. This study utilized ANN methodology, combined with a bootstrap technique that employed a random resampling approach from a single input dataset, to simulate monthly stream baseflow nitrate concentrations. The effectiveness and predictive ability of the ANN model were assessed by comparing it to conventional multivariate regression methods. Through the use of ANN, more precise outcomes can be achieved while taking into account the uncertainty associated with the analysis. The findings demonstrated that the ANN outperformed both the multivariate linear regression and nonlinear quadratic response surface models in explaining the variance of stream nitrate and in external prediction consistency. This study also highlighted several key variables, such as the areal percentage of agricultural land and grassland, stream order, and the average slope of the groundwater flow path, that significantly impacted the stream baseflow nitrate concentrations in this watershed that was dominated by dairy farming and groundwater. Of these variables, the percentage of agricultural land emerged as the most significant factor.