A HYBRID SARIMA-DEEP LEARNING APPROACH FOR LOCUST SWARM RISK CLASSIFICATION
MEETCON - EUROPE 4th INTERNATIONAL CONGRESS ON SCIENTIFIC RESEARCH, Madrid, İspanya, 26 - 30 Ekim 2026, cilt.1, sa.1, ss.112-118, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 1
- Basıldığı Şehir: Madrid
- Basıldığı Ülke: İspanya
- Sayfa Sayıları: ss.112-118
- Eskişehir Osmangazi Üniversitesi Adresli: Evet
Özet
Locust swarm outbreaks are among the most significant agricultural pest events threatening
global food security. This study aims to develop a comprehensive methodology for locust
swarm risk classification through a hybrid approach combining SARIMA-based residual
analysis with deep learning and ensemble learning models. A dataset covering 314 monthly
observations for the ġanlıurfa region between 2000-2026 was utilized. In the first stage, time
series residuals were calculated using a SARIMA(2,1,2)(1,1,1)₁₂ model, and binary risk labels
(Low/High) were generated through IQR-based outlier detection. In the second stage,
comprehensive feature engineering was performed, encompassing features including lagged
observations, rolling statistics, meteorological variables (temperature, precipitation, soil
moisture, wind speed), seasonal cycles, and ecological interaction terms. In the third stage,
eleven different classifiers were comparatively evaluated: LSTM and 1D-CNN deep learning
models; XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, and Gradient Boosting
ensemble models; and SVM and Robust Logistic Regression (L1 and L2 penalized)
traditional models. SARIMA residuals exhibited the highest feature importance across all
models in the consensus analysis conducted with six different methods. Among
environmental features, soil moisture and precipitation categories emerged as the most critical
determinants, while wind variables contributed significantly to class separation power.
Among ensemble learning models, XGBoost demonstrated the best overall performance with
the highest accuracy and balanced F1-score. LSTM, LightGBM, and CatBoost models also
achieved similarly high accuracy values, exhibiting strong classification performance.
Random Forest stood out with the highest accuracy value in capturing outbreak events. The
staged integration of environmental features yielded substantial accuracy improvements,
particularly in tree-based ensemble models. Among feature categories, SARIMA,
observation-based variables, and soil moisture ranked in the top three. The hybrid SARIMA
deep learning and ensemble learning approach, when enriched with environmental features,
demonstrates sufficient accuracy performance for operational early warning systems. The
critical role of soil moisture and precipitation aligns with locust breeding ecology. Future
studies aim to further enhance model accuracy through the integration of satellite-based data
such as NDVI.