DRAMM-Pest: Direct Multi-Horizon Forecasting of Rice Leaf Roller Activity from Light-Trap and Field-Survey Records

Multi-horizon forecasts of rice leaf roller activity support advance surveillance planning. Routine light-trap and field-survey records differ in measurement meaning and availability, complicating their complementary use across future monitoring periods. We developed DRAMM-Pest, a mask-aware dual-measurement model in which available field histories condition temporal states derived from light-trap history before horizon-specific aggregation. The model jointly predicts positive-catch probability and light-trap rate over six successive monitoring periods, with observed field-survey targets providing auxiliary supervision during training. On a public 24-year rice leaf roller dataset covering 11 sites, chronological testing on 2020–2023 yielded a light-trap-rate RMSLE of 0.837±0.029 and a Brier score of 0.084±0.004. The corresponding absolute reductions relative to XGBoost were 0.25023 and 0.00859, respectively. Removing all field information or retaining only its binary availability indicator increased mean light-trap-rate MAE by 3.12 and 2.74, respectively. The findings support using complementary monitoring records as historical context and training targets within a horizon-specific forecasting framework for pest-surveillance planning.