OSAIN: An Ontology-Guided Sparse Additive-Interaction Network for Interpretable Multi-Horizon Forecasting of Diamondback Moth (Plutella xylostella) Abundance

Forecasting diamondback moth (Plutella xylostella) abundance months ahead could support timely crop protection, but short monitoring records challenge both accuracy and interpretability. We examined whether ecological structure could improve multi-horizon forecasts while making their predictive evidence auditable. We developed OSAIN, a joint ontology-guided, horizon-conditioned sparse additive-interaction formulation with separate pest-memory and weather pathways. Forecasts decompose into a horizon bias, pest memory, additive weather effects and restricted ecological interactions. Using monthly larval records (larvae per 100 plants) from Huiyang and Shantou, Guangdong, the model forecasts abundance one to four months ahead from 12 months of 137 meteorological descriptors and 1 pest-history descriptor. Under chronological, target-disjoint evaluation against fifteen baselines, OSAIN achieved the lowest mean root mean squared error (RMSE) and mean absolute error across ten runs at both sites. RMSE was 2.4% lower than the strongest baseline at Huiyang, where the difference was not significant, and 7.1% lower at Shantou (Holm-adjusted p = 0.001). Individual component forecasts used approximately one-eighth of the descriptors, and fitted gate rankings agreed with occlusion-based dependence more strongly than those of untrained models. Cross-site transfer reduced accuracy in both directions, indicating the need for local calibration and prospective evaluation before operational use.