Near-Real-Time Prediction of Desert Locust Presence Risk in Ethiopia Using Satellite-Derived Environmental Data and Machine Learning

Abstract

Desert locusts (Schistocerca gregaria) pose a long-standing challenge to crop production and food security in Ethiopia. Current warning systems have several limitations. These include relatively coarse spatial resolutions, rule-based inferences, and delayed reports. These limitations restrict their operability and effectiveness. This study developed a near-real-time, spatially validated framework for predicting locust presence in Ethiopia using multisource satellite data. The model integrates MODIS NDVI, CHIRPS rainfall, and ERA5 temperature datasets covering 2015–2024, combined with 18,742 quality-controlled locust occurrence records from the FAO DLIS and the DLCO-EA. All data were harmonized to a spatial resolution and a 16-day temporal interval. A total of 32 environmental predictors were derived, including lagged variables, anomalies, and seasonal indicators. Four models (LR, RF, XGBoost, and LightGBM) were tested using spatial block cross-validation. LightGBM achieved the best spatial performance (AUC = 0.740), and a stacked ensemble (LightGBM + XGBoost + calibrated logistic regression) achieved a final test set AUC of 0.780 (95% CI: 0.765–0.795) with a conservative temporal validation AUC of 0.85 under a strict two-year gap split.