Spatio-Temporal Modeling of Malaria Occurrence Probability Using Multi-Sensor Remote Sensing and Machine Learning
Background: Reliable mapping of malaria occurrence requires validation across environmentally distinct years and careful separation of model-derived probability from transmission intensity. This study developed a multi-sensor remote sensing and machine-learning framework for mapping annual malaria case occurrence in southeastern Iran. Methods: Annual median composites of Soil Water Index (SWI), Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST) were generated from Sentinel-1, Sentinel-2 and Landsat 8/9 at 30m resolution. After quality control, 82, 168, 775 and 483 malaria case locations were retained for the Solar Hijri years 1400–1403 (2021–2025). Five spatially balanced pseudo-absence realizations were generated annually at a 1:1 ratio with a 300m exclusion buffer. Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) were evaluated through complete four-fold leave-one-year-out validation with nested tuning. Performance, bootstrap confidence intervals, uncertainty, Getis–Ord Gi* clusters and SHAP (SHapley Additive exPlanations) were assessed. Results: Across 20 outer evaluations, RF achieved ROC-AUC 0.685±0.045, PR-AUC 0.693±0.040, balanced accuracy 0.624±0.026 and Brier score 0.230±0.015. XGBoost achieved ROC-AUC 0.687±0.044, PR-AUC 0.691±0.043, balanced accuracy 0.625±0.033 and Brier score 0.231±0.015. Year-specific ROC-AUC ranged from 0.624 to 0.726. LST was the dominant predictor (48.27%), followed by NDVI (37.04%) and SWI (14.69%). Only 3.46–6.10% of pixels exceeded probability 0.70. Mean uncertainty was low (0.030–0.032) but spatially heterogeneous. Conclusion: Both algorithms showed moderate, comparable performance with year-dependent variation. Maps provide relative probabilities for prioritizing field surveillance, not transmission estimates.