|
2022, № 5 (53) |
|
SOIL SCIENCE
|
|
|
|
|
|
|
|
Kizimova T.A., Riksen V.S., Shpak V.A., Maksimovich K.Yu., Galimov R.R. |
|
|
Using Machine Learning Techniques to Predict Nitrate Nitrogen in SoilUsing Machine Learning Techniques to Predict Nitrate Nitrogen in Soil |
|
|
This paper presents multinomial logistic regression and neural network models that allow predicting and thereby quickly determining the content of nitrate nitrogen in the 0-40 cm soil layer before sowing. To train the models, we used the data of a long-term multifactorial field experiment of the Siberian Research Institute of Agriculture and Chemistry of the Siberian Federal Scientific Research Center of the Russian Academy of Sciences for 2009-2018, established in 1981 on leached chernozem in the OS "Elitnaya" - a branch of the Siberian Federal Scientific Research Center of the Russian Academy of Sciences. Taking into account the features of the statistical sample (observation data and analyses), the main predictors of the models that affect the content of nitrate nitrogen in the soil (target indicator) are determined, they are presented qualitative (predecessor, tillage) and quantitative (weather conditions and the content of productive moisture before sowing in the layer 0-100 cm) by factors with appropriate gradations. The models showed a fairly high reliability when verified on empirical data and can be used as a tool for forecasting. The quality of the developed multinomial logistic regression model was assessed using the coefficient of determination, which was 78% according to the Nagelkerke measure, and 72% according to the Coxsei Snell measure. To determine the predictive ability of the neural network, an ROC analysis was carried out, which showed that the area under the ROC curve for each category of the target indicator was close to 1, which indicates a high predictive power of this method. A comparative assessment of the predictive capabilities of the trained models was carried out. The overall proportion of correct predictions for multinomial logistic regression is 80.6%, in the neural network model 89.5%.
|
|