Prediction of nitrate nitrogen content in soil using machine learning

Abstract

The possibilities and feasibility of using the Bayesian network of trust and logistic regression to predict the content of nitrate nitrogen in the 0-40 cm soil layer before sowing have been investigated. Data from long-term multifactor field experience at the Siberian Research Institute of Farming and Agricultural Chemization of SFSCA RAS for 2013-2018 were used to train the models. The experiment was established on leached chernozem in the central forest-steppe subzone in 1981 in the Novosibirsk region. Considering the characteristics of the statistical sample (observation and analysis data), the main predictors of the models affecting nitrate nitrogen content in soil were identified. The Bayesian trust network is constructed as an acyclic graph, in which the main (basic) nodes and their relationships are denoted. Network nodes are represented by qualitative and quantitative plot parameters (soil subtype, forecrop, tillage, weather conditions) with corresponding gradations (events). The network assigns a posteriori probability of events for the target node (nitrate-nitrogen content in the 0-40 cm soil layer) as a result of experts completing the conditional probability table, taking into account the analysis of empirical data. Two scenarios were analyzed to test the sustainability of the network and satisfactory results were obtained. The result of the logistic regression is the coefficients characterizing the closeness of the relationship between the dependent variable and the predictors. The coefficient of determination of the logistic regression is 0.7. This indicates that the quality of the model can be considered acceptable for forecasting. A comparative assessment of the predictive capabilities of the trained models is given. The overall proportion of correct predictions for the Bayesian confidence network is 84%, for logistic regression it is 87%.