Evaluating Reliability-Gated Multimodal Fusion for Winter Wheat Increment Forecasting

Agricultural sensors are informative at different times and drop out for different reasons: a UAV flight is canceled, a weather window is empty, a ground visit is skipped. A fusion rule that cannot see those conditions must average over them. This paper tests a reliability-conditioned mixture of experts for winter wheat leaf area index (LAI) and biomass increments: modality weights may depend on availability, input typicality and expert uncertainty, optionally together with agronomic context. The scientific contribution is not a new sensor or a new increment target. It is a comparative test of that reliability hypothesis against availability-only gating, nonnegative stacking and nested expert-dropout, under nested field-grouped cross-validation, systematic expert withdrawal, and concatenated walk-forward. Typicality and uncertainty do not earn a consistent field-level role. Availability plus agronomic context is the simpler alternative among the gates. On fully observed ΔLAI, a nonnegative stacker and UAV patch statistics beat the full gate; the gap is consistent with a high-capacity gate over-fitting a small out-of-fold expert table rather than with unused reliability information. Reconstruction y^t+1=yt+Δy^ uses the measured current state, so the task is incremental change given yt. Concatenated walk-forward R2 is negative for every method because campaign increment means are non-stationary; field-out scores are therefore not a temporal forecasting skill. Consecutive elementary sampling unit (ESU) visits are kept when their 20 m windows overlap (n=105 LAI, n=141 biomass). Paired tests are exploratory and unadjusted.