A Deployed Vision–Language Decision-Support Platform for Tomato Harvest: Identity-Gated Grading, Uncertainty-Propagating Yield Estimation and Weather-Conditioned Advisories

Abstract

This work presents a deployed vision–language decision-support platform for tomato harvest that connects image capture, identity verification, ripeness grading, harvest guidance, yield estimation, damage assessment, weather-conditioned loss projection, multilingual reporting, and persistent provenance within a single auditable workflow. The system uses a six-stage tomato ripeness taxonomy and an identity-gated inference architecture designed to prevent non-tomato objects from entering downstream grading and yield calculations. It propagates uncertainty through a multi-layer yield estimation pipeline, distinguishes measured, derived, provisional, unavailable, and unverifiable values, supports markerless and scale-referenced estimation, and accepts weighed-harvest observations for later calibration. A deterministic advisory layer evaluates weather forecasts from multiple providers and derives reference evapotranspiration when sufficient inputs are available. The paper reports deployment measurements including controlled identity-gating experiments, repeated-run reliability analysis, detection-collapse characterization, latency, localization and report-rendering validation, and an 851-case automated test suite. The study reports 93.1% inter-run ripeness agreement (Cohen’s κ = 0.902), a derived 97.9% accuracy ceiling under the stated assumptions, and a 6.7% detection-collapse rate. It explicitly distinguishes engineering evidence from biological validation: no stage-labelled ground-truth dataset or weighed-harvest calibration set was available for the reported study, so absolute ripeness accuracy and yield error are not claimed. The accompanying software repository contains the implementation and associated research artifacts used to support reproducibility of the described system.