<![CDATA[Image-based rice disease screening requires assessment of unfamiliar conditions, degraded photographs, and the evidence used for advice. This study evaluates a modular image–text workflow with separately trained visual and retrieval components. A rotation- and reflection-aware audit of 5,932 rice leaf images identified 343 similarity groups for leakage-reduced training, validation, and testing. Seven lightweight classifiers and a 24,724-parameter CNN were evaluated alongside open-set rejection, image-quality triage, simulated recapture, and external testing on 10,333 exact-unique Paddy Doctor images. Color-texture logistic regression achieved 78.41% grouped-test accuracy and the CNN achieved 81.03%; their accuracies on 4,232 external images from the four known classes fell to 13.33% and 12.55%, respectively. External unknown-detection AUROCs were 0.512 for Mahalanobis scoring and 0.486 for CNN nonconfidence. Across 7,745 scoped CROP records, reranked retrieval recovered the paired answer within five candidates for 934/1,168 test queries (79.97%). Evidence-assisted benchmark accuracy was 65.21% on 365 questions and 41.34% on 179 without an exact question match in the retrieval collection. Four advisory gates enforced disease, location, treatment, and numeric-rate rules, with a trade-off between answer delivery and reference-answer retention. Nepal-context retrieval remained preliminary: the reranker succeeded on 7/8 queries (87.50%; Wilson 95% CI, 52.91–97.76%). Classification, triage, and advisory retrieval provide separate levels of evidence. The findings support further investigation of the workflow, while the external failures preclude claims of reliable field deployment.]]>
Copyrights © 2025
