Comparison of Methods for Mathematical Processing of Long-Term Field Experiment Data Using Soil Ammonia Nitrogen as an Example

Kalichkin, V. K., Usenko, V. I., Fedorov, D. S., Maksimovich, K. Yu., Malygin, A. E., Garkusha, A. A.
Russian Agricultural Sciences · Springer Nature · 2026-04-01 · DOI: 10.3103/s1068367426700370

Abstract

Abstract The paper reports the study which aimed at comparing the information value of the data on the effects of agronomic factors on ammonia nitrogen content in the arable soil layer under long-term exposure in different agrocenoses using mathematical statistics and machine learning methods. Data obtained in a long-term field experiment carried out in the Altai Federal Scientific Centre for Agricultural Biotechnology located in the forest-steppe area of the Altai Ob River region was processed. Multivariate analysis of variance (ANOVA), multiple regression and correlation (Pearson’s correlation coefficient) analyses, random forest and extreme gradient boosting (XGBoost) algorithms, and a leave-one-out (LOO) cross-validation strategy were used. Shapley additive explanation method (SHAP) was used to explain the outputs of machine learning models and quantitatively assess the contribution of factors. ANOVA revealed that three factors (sampling depth, fertilizers, and crop rotation type) had statistically significant effect on soil ammonia nitrogen content, while multiple regression analysis and correlation analysis both revealed only one factor, namely, sampling depth, with a statistically significant effect on the studied index. Highest in the experiment ammonia nitrogen content was found in the 20–30 cm layer with a combination of fertilizer use and shallow subsurface tillage, or with no tillage and plant protection agents use. The lowest was found in the 0–10 cm layer with deep subsurface tillage with no fertilizers. In addition to confirming the dominating role of sampling depth and the secondary role of plant protection use SHAP allowed to quantitatively assess the described relationships between the factors: sampling depth |SHAP| = 0.156–0.195 > fertilizers |SHAP| = 0.064–0.096 > tillage system |SHAP| = 0.014–0.054 > crop protection products |SHAP| = 0.008–0.023. Machine learning methods with interpretation of results using SHAP should be considered a preferable approach to processing the experimental data on changes in the ammonia nitrogen content in the arable soil layer under the effect of agronomic factors in different agrocenoses.

Read at publisher →

Similar articles