Contrastive multimodal alignment of Sentinel-2 and community radio transcripts for pest-outbreak early warning: a climate-analytics evaluation of false-alarm burden on extension services

Pest-outbreak early warning in smallholder systems depends on the timely fusion of satellite vegetation signals with farmer-reported observations, yet the two streams are misaligned in vocabulary, cadence and reliability. The framework is formalised as a contrastive objective over paired satellite tiles and transcribed radio segments, with a calibration layer that converts embedding similarity into outbreak probabilities. Because no paired operational corpus is available, the evaluation is analytical: a propagation model translates classifier operating characteristics into extension-worker visit loads, and a threshold-selection rule is derived that bounds expected false alarms per season. False-alarm burden is governed less by classifier accuracy than by the base rate of outbreaks and the cost asymmetry between missed detections and wasted visits. Under plausible smallholder base rates, a system with high recall still generates visit loads that exceed typical extension capacity unless the decision threshold is raised, which in turn suppresses recall.