Configurable Multi-Stage Vision Pipeline for Crop Disease and Pest Diagnosis
Naga Ganesh, Naga Ganesh, Chandrashekar M S, Chandrashekar M S, Lakshmi Pedapudi, Lakshmi Pedapudi, Aakash Singh, Vineet Singh
cs.CV · cs.CL · cs.LG
Quick verdict
Production-data study rebuilding crop diagnosis as three configurable stages; DaViT-Base tops Plantix and a fine-tuned VLM, but panel-made labels cap the evidence.
Summary
This systems study analyses about 1.16 million photographs submitted to Digital Green's FarmerChat advisory service across Ethiopia, India, Kenya and Nigeria, and proposes a three-stage diagnosis pipeline — quality gate (M0), crop detection (M1), disease/pest detection (M2) — with independently tunable thresholds, replacing the closed GPT-4o-plus-Plantix production path. A MobileNetV3-small gate matches the paid GPT-4o gate's decisions at 86.9% F1 in 12 ms. On a shared 10,335-image test set, fine-tuned DaViT-Base reaches 95.41% crop accuracy and 73.26% diagnosis accuracy, against Plantix at 91.46%/42.61% and a fine-tuned Qwen3-VL-4B at 94.66%/33.04%; the specialist route is modelled at $100/month versus $924 for the GPU route. Key limitation: the reference labels come from an eight-model council, not expert annotation, and GPT-5.4 both helped vote those labels and was scored against them, so scores measure agreement with model consensus rather than ground-truth accuracy — a caveat the authors themselves stress. No confidence intervals or significance tests, unmeasured false-rejection rate, and mixed speed measurement protocols leave the ranking conclusions provisional pending expert relabelling.
Review complete
Assessment at a glance
4
risk flags identified
● high 2
● medium 2
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The production funnel fails materially: 46.8% of judged images were rejected, 27.0% of sent images got no crop name, and 35.8% of labelled problems are pests misfiled under disease (§2, Tables 3/6).
- A MobileNetV3-small quality gate reproduces the GPT-4o gate's decisions at 86.91% F1 with 12 ms median latency, replacing the slowest and paid production stage (§5.1, Table 19).
- On one common 10,335-image test set, fine-tuned DaViT-Base achieves the best crop accuracy (95.41%) and diagnosis accuracy (73.26%), exceeding Plantix (91.46%/42.61%) and the fine-tuned Qwen3-VL-4B (94.66%/33.04%) (§5.2, Table 21).
- Splitting disease (crop-masked) from pest (unmasked) detection in a shared four-head backbone improves robustness to wrong crop predictions (§3.4, §5.6, Table 24).
- The specialist CPU route is both more accurate and roughly nine times cheaper to host at measured demand ($100 vs $924 monthly), so the GPU route's case rests on its follow-up-question and reasoning abilities (§5.3, §6.1, Tables 22/26).
Strengths
Rare production-scale failure analysis
§2 quantifies the live funnel over 1,163,658 images from four countries: gate rejected 46.8% of the 821,787 it judged, 27.0% of sent images got no crop name, and 35.8% of labelled problems are pests (Table 6) — concrete motivation for the redesign, plus per-country breakdowns (Table 7: India 45.0% vs Kenya 18.3% rejection).
Unusually candid evaluation discipline
The paper repeatedly undermines its own headline: GPT-5.4 'helped write the reference that its own diagnosis column is scored against' (§5.2), Route A's disease accuracy is 'label-confounded' (§3.5.3), M0 is judged against GPT-4o's own unstable labels (§3.2.1), and the conclusion states no diagnosis number is final until production answers leave training. Rare self-skeptical framing.
Practical, honest M0 replacement economics
Reasoned M2 split with a falsifiable design choice
Matched-comparison methodology
Weaknesses
Reference labels are model consensus, not expert truth
Scorer-competitor circularity
No statistical inference anywhere
The motivating harm is never measured
Speed and cost comparisons are apples-to-oranges
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is an evaluation-and-design study built on one of the largest production corpora in agricultural AI: 1,163,658 farmer photographs submitted to Digital Green's FarmerChat across Ethiopia, India, Kenya and Nigeria (§2, Table 3). The paper's first contribution is diagnostic: the closed production path (a GPT-4o ten-check quality gate at $0.0106–0.0132 and 6.2–8.2 s per image, then a Plantix call) rejects 46.8% of judged images, leaves 27.0% of sent images without a crop name, and files pests under "disease" 35.8% of the time (Tables 2, 3, 6). Per-country analysis (Table 7) sharpens this: India's rejection share is 45.0% against Kenya's 18.3%. The second contribution is a redesigned, configurable three-stage pipeline — M0 quality, M1 crop, M2 disease/pest — instantiated in two routes: Route A, a fine-tuned Qwen3-VL-4B answering everything in one call with conditional follow-up photograph requests; Route B, orchestrated specialists (MobileNetV3-small gate; DaViT-Base or YOLO26x-cls four-head classifier with crop-masked disease and unmasked pest heads).
The headline claims are: MobileNetV3-small matches the paid gate at 86.91% F1 in 12 ms (Table 19); on a common 10,335-image test set, DaViT-Base leads all seven systems at 95.41% crop accuracy and 73.26% diagnosis accuracy versus Plantix at 91.46%/42.61% and the Route A fine-tune at 94.66%/33.04% (Table 21); and the CPU route costs $100/month modelled versus $924 for GPU (Table 22).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does the proposed system beat Plantix for crop disease diagnosis?
On the paper's own 10,335-image shared test set, yes: the fine-tuned DaViT-Base model scores 95.41% crop accuracy and 73.26% diagnosis accuracy versus Plantix at 91.46% and 42.61%. However, the reference labels were produced by a panel of AI models rather than human experts, so the numbers measure agreement with that panel consensus; the authors themselves state the diagnosis numbers are not final until expert validation.
How were the labels created — did humans annotate the images?
Mostly no. Eight AI models independently labelled each image; five of them voted on the final label, and when the vote failed, the stored production answer from Plantix was used as the reference. Human review covered only a small subset. Plantix, the off-the-shelf Qwen3-VL, and Gemma-3 were excluded from voting so that systems under test did not write their own reference labels.
What is the quality gate and how much does it save?
The quality gate decides whether a farmer's photo is usable before diagnosis. Production uses one GPT-4o call costing roughly $0.0106–0.0132 and taking 6.2–8.2 seconds per image. The proposed MobileNetV3-small gate matches its decisions at 86.9% F1 in 12 milliseconds on CPU, removing the per-image fee — though it is trained to imitate GPT-4o, so it may inherit that gate's over-rejection behaviour.
Why split disease and pest detection into separate heads?
Because crop identity helps pathogen identification (the same disease looks different on different plants, so the candidate list can be narrowed) but hurts pest identification (a caterpillar looks similar on any plant, and a wrong crop call would eliminate the right insect). The paper's four-head model therefore filters only the disease head by crop, leaving the pest head to choose from all 92 classes.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–12 (pipeline diagrams, example photographs, coverage curves) could not be analysed from the extracted text; judgments rest on tables and prose only.
- Minor numeric inconsistencies noted (e.g., 27.2 min per DaViT run in text vs Table 25's 54.4 min summed over two seeds); model names reflect the paper's 2026 context and were taken as printed.
- The separate four-head label manifest (Table 24's 16,273-image test) could not be cross-checked against the 10,335-row scoring set; the paper asserts the two experiments are never numerically compared.
- Diagnosis metrics exclude healthy, unspecified, and Plantix-referenced rows, so headline accuracies are not comparable to the funnel percentages in §2.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.