Multimodal Sensor-Fusion Inference for Artificial Intelligence Agriculture Pest Detection Under Missing Modality Streams: A Computer Evidential-Calibration Reliability Bound

Multimodal pest-detection pipelines in precision agriculture fuse optical, thermal and acoustic streams, yet field deployments routinely lose one or more modalities to power failure, occlusion, dust or radio dropout. The problem is framed as inference over a Dempster–Shafer belief structure whose focal masses are produced by modality-specific classifiers and then combined under a calibration map that must remain valid when a stream is absent. A formal definition of the missing-modality fusion operator is given, followed by a theorem establishing a lower bound on the reliability of the fused decision as a function of per-modality calibration error, the number of surviving streams and the mass assigned to the universal hypothesis. The bound is then specialised to the two-modality and three-modality cases and compared with the naive renormalisation rule that practitioners commonly apply. A replication-based evaluation of the before-after-control-impact estimator is used to characterise the operating characteristics of the design under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64); power was estimated across 2000 design cells. Renormalisation inflates confidence precisely when the discarded stream carried the discriminating evidence, whereas the evidential bound degrades gracefully and predicts the cluster count at which a five-percentage-point effect reaches eighty per cent power (54 clusters). The contribution is a computable reliability certificate that a fusion pipeline can emit alongside each pest decision, together with the design parameters under which that certificate remains informative.