Source: arXiv

Paper

Environmental sensor readings in two crop disease image datasets identify the session in which each image was taken — Sungwoo Kang (Department of Electrical and Computer Engineering, Korea University; arXiv 2610.06369, Oct 2026).

What it claims

Ag-AI papers often report that fusing leaf photos with IoT soil and greenhouse readings beats image-only classifiers. Kang audits that claim on two Korean benchmarks — the Crop Disease Diagnosis (CDD) challenge (LG AI Research / Dacon) and the government AI Hub subtropical pest-and-disease release.

The issue is session structure: one farm or greenhouse on one day is a session, and datasets attach one sensor panel or environment series to many images from that visit. If disease classes differ between sessions, models can cheat by memorizing which reading or timestamp maps to which label — not by reading the plant’s environment.

Kang’s diagnostic reads no images. A granularity probe counts shared sensor readings; a timestamp baseline pits soil-chemistry panels against date-time-only random forests. On AI Hub, 13,150 distinct panels span 274,000 rows; on CDD, 780 series cover 5,767 images. 91.9% of CDD test images duplicate training sensor values. Across seven AI Hub crops, timestamp-only classifiers match or beat soil panels; on CDD’s published split, timestamps alone reach the macro-F1 of a published fusion model — without any leaf pixels.

Fusion gains on random splits, Kang argues, cannot be separated from session leakage. The fix: session-held-out splits and datetime baselines beside sensor scores.

The breakdown

This is an audit paper, not a new architecture — a pre-training check before claiming “sensors help.”

When one pH snapshot labels dozens of photos from the same visit, the reading encodes where and when, not necessarily what the leaf experienced. If date-time predicts disease as well as a nine-variable soil panel, the multimodal lift may be shortcut learning. On CDD, sensor-only macro-F1 falls from 0.885 on a random split to 0.653 when whole environment series are held out — much of the headline score comes from recognizing series seen in training. Random image splits that mix farm-day sessions across train and test repeat a mistake spatial vision benchmarks already warn about.

Why readers outside the lab should care

Korea ag-tech teams citing AI Hub or Dacon fusion numbers in pitch decks need this as due diligence. Overseas builders importing Korean crop-disease data should demand held-out farms/dates and datetime baselines — session leakage does not show up in leaderboard screenshots. Procurement reviewers funding smart-farm pilots can ask for session-disjoint eval norms before greenlighting field trials.

What builders and Korea-touching teams should watch

• Do count distinct sensor readings versus images before training; if readings ≪ images, session shortcuts are likely.
• Don’t report fusion gains on random splits without a session-held-out split and a timestamp-only baseline.
• Expect timestamp classifiers to match sensor panels on AI Hub’s seven-crop comparison — treat that as a red flag.
• Re-check demos showing +2–3 pp macro-F1 from “adding environment”; ask whether test images share training sessions.
• Hold out whole photo dates or farms when classes span multiple sessions.

Context

Read this as a Korean-dataset reality check on ag-AI fusion hype, not proof that field sensors are useless. Korelay’s frame: pairing leaves with IoT readings sounds grounded — but when one reading labels a whole farm visit, models may memorize which day you showed up, not diagnose the crop.

Source

Primary: arXiv:2610.06369 (abstract and framing cited; open the OA PDF for methods and full results). Do not republish the PDF.