Crop Contamination Forecasting Based on Machine-Learning Approaches

• Published in: Russian agricultural sciences 2022, Vol.48 (2), p.115-122
• Main Authors: Kalichkin, V. K., Alsova, O. K., Maksimovich, K. Yu, Vasilyeva, N. V.
• Format: Article
• Language: English
• Subjects: Agricultural pollution; Agriculture; Air temperature; Algorithms; Automation; Biomedical and Life Sciences; Cause-effect relationships; Contamination; Crop production; Crops; Data mining; Decision analysis; Decision support systems; Decision trees; Digitalization; Electrification; Fallowing; Field tests; Forecasting; Life Sciences; Machine learning; Mathematical models; Mechanization; Meteorological parameters; Model accuracy; Nonparametric statistics; Normal distribution; Parameter identification; Plant Biochemistry; Polls & surveys; Qualitative analysis; Root-mean-square errors; Soil contamination; Statistical analysis; Statistical methods; Steppes; Weather stations
• ISSN: 1068-3674; 1934-8037
• DOI: 10.3103/S1068367422020069

The outcomes of the survey to assess the effects of qualitative factors and meteorological parameters on crop contamination are provided. A forecasting model for crop contamination has been built based on the decision tree analysis with the limited sample coverage. The data collected in the long-term field experiments conducted in the forest steppe area in Novosibirsk oblast and the information on meteorological parameters for 1996–2018 from the Novosibirsk weather station are used in the survey. Different types of data-mining techniques, including the nonparametric probability and statistics approaches, visualization methods, and the decision tree analysis, are used for the survey’s task solution. The choice of the survey methods is caused by the specific structure and the distribution pattern of the initial data, such as disparity between the model and the normal distribution law, a comparatively little sampling coverage, qualitative and quantitative predictors, and compound nonlinear correlations between crop contamination rates and meteorological parameters. The qualitative factors determining the crop contamination level have been identified. They include background chemicalization, crop after fallowing, soil-treatment systems, and meteorological parameters (average 10-day-period air temperatures and precipitation amount in the period from the third 10-day period in April to the end of May). Their contribution and statistical significance are assessed. The crop-contamination forecasting model and the weed-spread logic rules relative to the effects of management practices and agro-meteorological conditions have been developed based on the CART building algorithm for decision tree constructing. The model accuracy is assessed based on the Mean Absolute Error (MAE) = 3.75, the Root Mean Square Error (RMSE) = 5.7, and the Coefficient of Determination/R Squared (denoted R2) = 0.8. The set of logic rules describes the structure of cause-effect relationships, which may be subsequently used to create a decision support system in crop production.