Kurmet Baibussenov , Arman Rustembayev *, Aigul Bekbayeva , Zhasulan Amanbay, Askar Zhubatkanov and Arsen Dzhumagulov
Saken Seifullin Kazakh Agrotechnical University, Astana, Republic of Kazakhstan
*Corresponding author: rustem.arman88@gmail.com
Agroclimatic variability influences harmful non-gregarious locust populations across Kazakhstan, but forecasting performance may differ among ecological zones. This study developed a machine learning–GIS framework using 7,192 phytosanitary monitoring records collected in 25 agroclimatic zones during 2003–2024. We integrated monitoring data with meteorological and remote-sensing predictors, and classified zones as high, moderate, or low risk based on mean absolute population density (Pabs). We evaluated Gradient Boosting, Support Vector Regression (SVR), Random Forest, SARIMA, and Linear Regression using a model-development period of 2003–2018 and an independent test period of 2019–2024. In the high-risk category, Gradient Boosting was selected because it achieved the lowest RMSE (2.636 individuals m⁻²) and the highest R2 (0.005), although SVR produced the lowest MAE (1.307 individuals m⁻²). SVR was selected for the moderate-risk category (MAE = 1.135; RMSE = 3.039 individuals m⁻²) and the low-risk category (MAE = 1.212; RMSE = 1.940 individuals m⁻²), although R2 remained slightly negative in both categories. Exploratory screening associated higher population densities with land-surface temperatures of 30–35°C, precipitation of 20–50 mm, normalized soil-moisture values of 0.60–0.75, and SPI values of 0–0.5. Treat these ranges as dataset-specific warning indicators. The framework can help phytosanitary agencies prioritize field surveys and direct verification toward higher-risk zones, while farmers can use zone-level alerts to guide scouting. Because forecast skill was limited and formal prediction intervals were unavailable, the outputs should support, rather than replace, field-based phytosanitary management decisions.