Measure the false-rejection rate of the M0 image-quality gate

Determine how many usable farmer-submitted images the M0 image-quality gate rejects by conducting a recall-versus-threshold sweep of Gate B.

Background

The proposed pipeline uses M0 as an image-quality gate before crop and disease-or-pest diagnosis. Gate B is a lightweight MobileNetV3-based model trained to reproduce the decisions of the GPT-4o reference gate, and its threshold can be adjusted. The reported evaluation measures agreement with the reference gate, but it does not establish how many genuinely usable images are incorrectly rejected. The authors therefore identify a recall-versus-threshold analysis as necessary to quantify the trade-off between rejecting unusable photographs and retaining usable ones.