Predicting tile drainage discharge using machine learning algorithms

Abstract. Drainage systems can significantly improve the water management in agricultural fields. However, they may transport contaminants originating from fertilizers and pesticides and threaten ecosystems. Determining the quantity of drainage water is an important factor for constructed wetlands and other drainage mitigation techniques. This study was carried out in Denmark where tile drainage systems are implemented in more than half of the agricultural fields. The first aim of the study was to predict the annual discharge of tile drainage systems using machine-learning methods, which have been highly popular in recent years. The second objective was to assess the importance of the parameters and their impact on the predictions. Data from 53 drainage stations distributed in different regions of Denmark were collected and used for the analysis. The covariates contained 35 parameters including the calculated percolation and geographic variables such as drainage probability, clay content in different depth intervals, and elevation, all extracted from existing national maps. Random Forest and Cubist were selected as predictive models. Both models were trained on the dataset and used to predict yearly drainage discharge. Results highlighted the importance of the cross-validation methods and indicated that both Random Forest and Cubist can perform as predictive models with a low complexity and good correlation between predicted and observed discharge. Covariate importance analysis showed that among all of the used predictors, the percolation and elevation have the largest effect on the prediction of tile drainage discharge. This work opens up for a better understanding of the dynamics of tile drainage discharge and proves that machine-learning techniques can perform as predictive models in this specific concept. The developed models can be used in regard to a national mapping of expected tile drain discharge.