Fine-scale mapping of Oncomelania hupensis habitats in eastern China using multi-season Sentinel-2 imagery and semi-supervised deep learning.

Accurate mapping of Oncomelania hupensis, the intermediate host of Schistosoma japonicum, is challenging due to sparse field data and dynamic seasonal flooding. We developed SnailMatch, a semi‑supervised deep learning framework combining multi‑seasonal Sentinel‑2 imagery to delineate snail habitats in Guichi, China. The model used 229 field-confirmed snail presence sites, 229 systematically sampled absence sites, and 5,759 unlabeled locations. Twelve cloud-free Sentinel-2 scenes covering all seasons of 2023 are processed through an attention-based Multi-Seasonal Fusion module that captures subtle seasonal and ecological patterns relevant to snail habitat suitability. In parallel, a Random Forest model trained on seven environmental covariates provided complementary ecological information. A Bayesian‑optimised ensemble produced habitat maps at a resolution of 100 m. Spatial block cross-validation (6-fold) demonstrated the ensemble achieved superior performance (accuracy=92.44%, AUC=96.31%, F1=91.67%) compared to individual models. High - probability habitats (>70 %) were concentrated in flooded vegetation, agricultural lands, and dense bushland, capturing 73.80 % of presence sites while covering 57.56 % of the study area. Post-hoc environmental factor analysis and shapley analysis indicated that seasonal inundation and vegetation density increased suitability, whereas high land-surface temperature combined with strong nighttime illumination-reflecting urban heat island conditions-reduced suitability. Model transferability was evaluated at the Anhui-Jiangxi border, where the ensemble achieved similarly strong performance (accuracy=91.94%, AUC=96.25%, F1=91.88%), demonstrating robustness across distinct landscapes. This cost-effective, interpretable approach supports targeted molluscicide application and vegetation management, aligning with WHO's 2030 elimination targets and demonstrating the utility of multi-seasonal remote sensing and explainable AI in environmental health surveillance under climate and land-use changes.