Skip to content
DSML.KZ
Publications

Talk · International Conference on Electronics, Computer and Computation · April 10, 2026

AgroExpert KG: A Hybrid Machine Learning and Rule-Based Decision Support System for Crop Yield Prediction in Kyrgyzstan

Akylai Topoeva, Tattybubu Arap kyzy, Timur Turatali, Ruslan Isaev, Emilbek Zholdoshbekov

Ruslan Isaev
Open via DOI

Abstract

This paper presents the AgroExpert KG decision support system, developed to address the important problem of data scarcity in precision agriculture for small farmers in Kyrgyzstan. The system combines the Random Forest regressor, trained on synthetic data, with transparent, rule-based agronomic recommendations. A new method for creating synthetic data is presented, which transforms limited expert standards into 5,000 training samples by including penalty functions that take into account nutrient deficiencies, pH imbalances, and insufficient precipitation. The model achieves a MAE of$0.51 \mathrm{t} / \text{ha}$and$\mathrm{R}^{2}$of 0.993 on synthetic test data, demonstrating successful approximation of the expert-defined agronomic response functions. Robustness evaluations confirm stability: noise$(\mathbf{~} \mathbf{1 0 - 2 0 \%})$increases MAE by$\mathbf{i} 10 \%$, out-of-range testing (extreme rainfall) shows graceful degradation, variance across seeds is low ($\mathbf{s t d}=\mathbf{0. 0 5} \mathbf{~ t} / \mathbf{h a}$), and sensitivity to penalty coefficients ($\alpha_{i} \pm \mathbf{2 0 \%}$) affects predictions by$\boldsymbol{;} \mathbf{1 0 \%}$. Uncertainty intervals via$\mathbf{R F}$variance average$\mathbf{0. 3}$t/ha. The integrated Decision Support System (DSS) provides operational alerts for nutrient management, while SHAP (SHapley Additive exPlanations) ensures interpretability of the model. The Streamlit web application provides farmers with personalized yield forecasts and recommendations with a user-friendly interface. The analysis of the importance of traits determines the type of crop and phosphorus level as the most influential predictors. The proposed approach represents a practical solution for improving the efficiency of fertilizer use and crop selection decisions in seven regions of Kyrgyzstan, creating the basis for sustainable agricultural intensification in regions with limited access to data. Note: Results are based on synthetic data; future validation on real fields is planned.

Smart Agriculture and AIRemote Sensing in AgricultureClimate change impacts on agriculturecrop yield predictiondecision support systemsynthetic data generation

More by Ruslan Isaev