Deep neural network model to predict N2O emission change by biochar amendment in upland agricultural soils

The N₂O emission change by biochar addition in soils showed inconsistent trends depending on biochar types, soil properties, environmental conditions, and soil management practices. Especially in non-flooded upland agricultural soils, due to the complexity of N₂O emission processes, which include nitrification, nitrifier-denitrification, and denitrification, there are still many gaps in the mechanistic understanding of biochar effects. In order to maximize climate change mitigating effect of biochar, the biochar application guidelines that consider N₂O emission change need to be offered to farmers. However, the current lack of knowledge makes it challenging to create mechanistic models, and new approaches are needed. Machine learning techniques can be a solution because we can find the relationship between input and output variables without explicit mechanistic understanding and mathematical description. We aimed at developing a deep neural network (DNN) model to predict the N₂O emission change from upland agricultural soils by biochar application. Among all the papers published between Jan 2007 ~ Jul 2019 collected from Web of Science Core Collection, 65 papers were chosen which report changes in N₂O emissions by biochar addition in upland agricultural soils. Eleven variables, which have been reported as important factors influencing N₂O emission, were selected as input parameters. These include 5 soil properties (Total carbon and nitrogen content, sand and clay content and pH), 3 biochar properties (Feedstock type, pyrolysis temperature and biochar application rate), and 3 agricultural practices (Fertilizer type, number of fertilization and N application rate). The output parameter is the ratio of the cumulative N₂O emission of biochar treatment and control. Using 85% of the compiled dataset (training set), the DNN model was trained to predict the changes in N₂O emission by biochar addition. The rest of the dataset (validation set) was used to validate the DNN model. As a result, the DNN model predicted the decreasing and increasing patterns of biochar driven N₂O emission change in 84% of the validation data. This preliminary result could be a basis for developing practical biochar use guidelines. Further studies will be conducted to improve the prediction accuracy of the DNN model by combining principal component analysis.