The use of machine learning algorithms to design a generalized simplified denitrification model

Abstract

Abstract. We propose to use machine learning (ML) algorithms to design a simplified denitrification model. Boosted regression trees (BRT) and artificial neural networks (ANN) were used to analyse the relationships and the relative influences of different input variables towards total denitrification, and an ANN was designed as a simplified model to simulate total nitrogen emissions from the denitrification process. To calibrate the BRT and ANN models and test this method, we used a database obtained collating datasets from the literature. We used bootstrapping to compute confidence intervals for the calibration and validation process. Both ML algorithms clearly outperformed a commonly used simplified model of nitrogen emissions, NEMIS, which is based on denitrification potential, temperature, soil water content and nitrate concentration. The ML models used soil organic matter % in place of a denitrification potential and pH as a fifth input variable. The BRT analysis reaffirms the importance of temperature, soil water content and nitrate concentration. Generalization, although limited to the data space of the database used to build the ML models, could be improved if pH is used to differentiate between soil types. Further improvements in model performance and generalization could be achieved by adding more data.

Understanding nitrogen transfer dynamics in a small agricultural catchment: Comparison of a distributed (TNT2) and a semi distributed (SWAT) modeling approaches

“…using an event based approach like in DNDC (Li et al, 1992)) or use more generalized simplified approaches (e.g. as in Oehler et al (2010) for the denitrification model). Spatially distributed measurements of mineralization dynamics in soil as well as denitrification would help to evaluate the realism of the different modeling approaches.…”

“…However, further work is needed to assess denitrification rates in such context. In the meantime, more generalized denitrification models could be used, notably based on soil organic matter content (Oehler et al, 2010), as it may be a strong limiting factor on this site.…”

Evaluation of uncertainties in N2O and NO fluxes from agricultural soil using a hierarchical Bayesian model

“…However, the HB model strongly depends on not only the dataset but also the model itself. If the aim of the study is to comprehend the responses and their functions, other statistical methods such as neural networks [ Oehler et al , 2010] are more effective than the HB method or the other inverse calibration methods. Because the HB method is based on deduction, “model uncertainty” is the most significant issue.…”

A Low-Complexity Machine Learning Nitrate Loss Predictive Model–Towards Proactive Farm Management in a Networked Catchment

“…Although this research effort reduced the number of parameters by a half, the study used 8000 training samples based on annual values, which were obtained from different sites, to achieve the optimal results. Similarly, another work further simplified the input parameters and training size requirement and used neural networks to simulate total N emissions from the de-nitrification process (N 2 O) [44]. The training set consisted of only 536 records based on input variables such as water filled pore space, nitrate concentration, soil denitrifying potential, organic matter, soil pH, bulk density and soil depth.…”