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-۱, Sentinel-۲ and Landsat ۸/۹ at ۳۰m resolution. After quality control, ۸۲, ۱۶۸, ۷۷۵ and ۴۸۳ malaria case locations were retained for the Solar Hijri years ۱۴۰۰–۱۴۰۳ (۲۰۲۱–۲۰۲۵). Five spatially balanced pseudo-absence realizations were generated annually at a ۱:۱ ratio with a ۳۰۰m 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 ۲۰ outer evaluations, RF achieved ROC-AUC ۰.۶۸۵±۰.۰۴۵, PR-AUC ۰.۶۹۳±۰.۰۴۰, balanced accuracy ۰.۶۲۴±۰.۰۲۶ and Brier score ۰.۲۳۰±۰.۰۱۵.
XGBoost achieved ROC-AUC ۰.۶۸۷±۰.۰۴۴, PR-AUC ۰.۶۹۱±۰.۰۴۳, balanced accuracy ۰.۶۲۵±۰.۰۳۳ and Brier score ۰.۲۳۱±۰.۰۱۵. Year-specific ROC-AUC ranged from ۰.۶۲۴ to ۰.۷۲۶. LST was the dominant predictor (۴۸.۲۷%), followed by NDVI (۳۷.۰۴%) and SWI (۱۴.۶۹%). Only ۳.۴۶–۶.۱۰% of pixels exceeded probability ۰.۷۰. Mean uncertainty was low (۰.۰۳۰–۰.۰۳۲) 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.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-۱, Sentinel-۲ and Landsat ۸/۹ at ۳۰m resolution. After quality control, ۸۲, ۱۶۸, ۷۷۵ and ۴۸۳ malaria case locations were retained for the Solar Hijri years ۱۴۰۰–۱۴۰۳ (۲۰۲۱–۲۰۲۵). Five spatially balanced pseudo-absence realizations were generated annually at a ۱:۱ ratio with a ۳۰۰m 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 ۲۰ outer evaluations, RF achieved ROC-AUC ۰.۶۸۵±۰.۰۴۵, PR-AUC ۰.۶۹۳±۰.۰۴۰, balanced accuracy ۰.۶۲۴±۰.۰۲۶ and Brier score ۰.۲۳۰±۰.۰۱۵.
XGBoost achieved ROC-AUC ۰.۶۸۷±۰.۰۴۴, PR-AUC ۰.۶۹۱±۰.۰۴۳, balanced accuracy ۰.۶۲۵±۰.۰۳۳ and Brier score ۰.۲۳۱±۰.۰۱۵. Year-specific ROC-AUC ranged from ۰.۶۲۴ to ۰.۷۲۶. LST was the dominant predictor (۴۸.۲۷%), followed by NDVI (۳۷.۰۴%) and SWI (۱۴.۶۹%). Only ۳.۴۶–۶.۱۰% of pixels exceeded probability ۰.۷۰. Mean uncertainty was low (۰.۰۳۰–۰.۰۳۲) 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.