Efficiency analysis of machine learning models for simulating nitrate movement in soils from Illinois State

Agricultural application of nitrogen fertilizers such as urea is needed because of the lack of available nitrogen forms (NO3^- and NH4⁺) that plants absorb. Although crop yield and nitrogen application are highly correlated, over applying this nutrient threatens the environment causing groundwater acidification and water eutrophication. Thus, characterizing NO3^- movement in the soil is crucial to evaluate potential impacts to the environment of intensive nitrogen application, as well as to assist in the adoption of practical tools that aim to reduce environmental contamination and optimize the nitrogen use. The hypothesis of this work is that data-driven models can be a simple-to-use though powerful tool for characterizing NO3-movement in the soil. Therefore, the objective of this study was to compare different machine learning methodologies (Random Forest, Decision Tree, Neural Network) with the traditional numerical modeling (Hydrus-1D) for predicting nitrate contamination classes in soils from Illinois State. First, breakthrough curves were adjusted, and transport parameters were estimated with STANMOD for 10 soil types from Illinois. Then, simulations of nitrate movement in a 30-year range using HYDRUS-1D were done. Partial date-interval was used as training dataset of the Machine Learning methodologies for the nitrate classes and the full simulated dataset was compared with the machine learning classifications. It was concluded that machine learning methodologies, especially artificial neural network, performed well predicting the nitrate contamination classes and can be used as a tool for improving best management practices and decision-making process.