Predicting national-scale tile drainage discharge in Denmark using machine learning algorithms

Abstract

Study RegionDenmark. Study FocusArtificial drainage systems can significantly improve water management in agricultural fields. Nonetheless, they transport contaminants originating from fertilizers and pesticides, threatening aquatic ecosystems. Determining the quantity of drainage discharge is an important factor for implementing constructed wetlands and other mitigation techniques. In Denmark tile drainage systems are present in more than 50 % of the total agricultural area, and the main objectives of this study were to i) estimate the annual tile drainage discharge using machine-learning algorithms and ii) to assess the importance of predictor variables for the models. Data from 53 drainage stations and 25 predictor variables including precipitation, percolation and geographic variables such as clay contents, and elevation were used. Four machine learning models were used to predict annual drainage discharge. New Hydrological Insights for the RegionRandom Forest and Cubist models demonstrated the best performance and the results highlighted the importance of cross validation. Predictor variable importance analysis showed that after precipitation/percolation, elevation, and clay content had the largest effect on tile drainage discharge. This work opens up for a better understanding of the effects of topographical and geological characteristics on tile drainage discharge, proving that machine-learning algorithms could be utilized as strong predictive models with a low complexity in the specific concept. The developed models could be used for mapping tile drainage discharge in geographic space.