Domain-Adaptive Acoustic Insect Classification for Artificial Intelligence Agriculture Grain Silo Pest Detection Under Shipping-Container Reverberation: A Computer Cross-Corridor Transfer Evaluation

Acoustic pest detection in grain silos is attractive because insects emit feeding and locomotion transients that microphones can capture without disturbing the bulk. The obstacle is that the same insect sounds arrive at the classifier through different room acoustics: a concrete farm silo, a steel shipping container and a warehouse corridor each impose a distinct impulse response, so a classifier trained in one enclosure loses accuracy in another. the present analysis treats that loss as a domain-shift problem and asks whether the shift can be characterised and bounded before field deployment. A linear time-invariant model of the enclosure is combined with a source model for the insect transient, and the resulting cross-corridor divergence is expressed as a function of reverberation time, direct-to-reverberant ratio and the spectral tilt of the source. The analysis yields a transfer bound that separates the part of the error attributable to the room from the part attributable to the classifier, and it identifies the conditions under which unsupervised domain adaptation is expected to recover most of the in-domain accuracy. The bound is then used to specify a minimal calibration protocol for a new enclosure. The contribution is a formal, testable account of cross-corridor transfer for acoustic insect classification, not a report of field measurements; the framework is intended to make subsequent empirical evaluation comparable across enclosures.