Electronic Nose Flags Artificially Ripened Bananas at 97.03% Accuracy

By Olivia Hart

A five-channel metal-oxide gas sensor array coupled to XGBoost machine learning separated naturally ripened bananas from ethephon-treated fruit with 97.03% test accuracy and 94.91% 10-fold cross-validation accuracy.

The trial covered 20 Sagar Kola bananas, the BARI-1 hybrid cultivar grown widely across South Asia, and ran for 14 days. Researchers split the fruit into two groups of 10, immersing one group for five minutes in an aqueous ethephon solution of 1.2 g/L and leaving the other to ripen naturally.

Inside a custom dual-compartment metal chamber, a National Instruments myRIO data acquisition card controlled sampling and recorded voltages from five Figaro TGS sensors — TGS 2602, 2603, 2610, 2612 and 2620 — through a repeating cycle: 20 minutes of volatile accumulation, five minutes of sensing, five minutes of fresh-air purge.

Automated vacuum pumps moved the released gases from the sample compartment to the sensing unit. The fruit remained intact.

One caveat governs those figures. The source states that it reports an unedited accepted paper awaiting final editing, and that "the paper should not be regarded as conclusive or treated as established information."

The paper, "Identification of artificially ripened banana by chemical agent using gas sensor array," appeared in Scientific Reports under Jim, A.A.J., et al. (2026), DOI 10.1038/s41598-026-76940-9.

What does an electronic nose actually measure?

Bananas release volatile organic compounds (VOCs, airborne chemicals that carry odor) as they ripen, and this array reads that mixture without touching or cutting the fruit. Each Figaro device uses a solid-state metal-oxide semiconductor, a heated ceramic whose electrical resistance responds to gas.

When VOCs reach the heated sensing surface, reversible oxidation-reduction reactions alter the material's resistance. The acquisition card logs those shifts as voltages, producing a digital fingerprint of each gas mixture; combining channels lets the classifier sort samples that overlap on any single sensor.

The study generated 1,009 time-series data points spanning unripe fruit through deterioration, then tested four machine-learning models: logistic regression, support vector machines (SVMs, margin-based classifiers), decision trees, and extreme gradient boosting (XGBoost, a boosted tree ensemble). Recursive feature elimination (RFE) and Shapley additive explanations (SHAP) ranked channel usefulness and flagged redundancy.

Which classifier separated the treated fruit?

Nonlinear models beat the linear one on the binary task. XGBoost, using all five channels, reached 97.03% test accuracy and 94.91% cross-validation accuracy; the decision tree reached 96.70%, while logistic regression managed only 59.08% against overlapping response patterns.

SHAP and RFE identified TGS 2603 and TGS 2620 as the most informative sensors for detecting artificial ripening. Dropping TGS 2610 left XGBoost's test accuracy at 97.03% and lifted cross-validation accuracy to 95.05%, indicating a four-sensor build could cut complexity without sacrificing accuracy.

Can it also date the fruit?

The same array also assigned samples to 28 classes defined by their exact ripening days. With all five sensors, the SVM posted 99.95% test accuracy and 99.92% cross-validation accuracy; the decision tree reached 99.94%, and XGBoost 98.68%. Removing TGS 2612 left the SVM's test accuracy at 99.95%.

These are measured study results, not manufacturer specifications, and the source's caution applies to both tasks. The trial covered one cultivar and one chemical agent, so the numbers describe this protocol rather than field performance.

Where would such a unit sit?

Against gas chromatography, mass spectrometry, optical imaging, and wet-chemical assays, the array trades analytical detail for portability, non-destructive screening, and lower instrument cost. That makes it a candidate for packing facilities, wholesale hubs, and agricultural inspection checkpoints, where handheld or conveyor-mounted units could flag suspect batches before distribution.

The authors also point to storage. Monitoring VOCs in the ambient air of cold rooms and transport containers could track ripening and estimate shelf life, supporting storage management and reducing post-harvest losses. A miniaturized, battery-powered unit with onboard processing would extend the method to food-safety inspectors and agricultural distributors.

The authors state: "Future work should evaluate additional banana cultivars, other fruit varieties, and different ripening agents, and examine sensor performance under varying conditions."

The remaining hurdle is procedural. The source lists gas chromatography among the traditional detection methods, while this trial tested ethephon alone and left calcium carbide unexamined. The adoption question is concrete: will buyers accept a 97.03% screening classifier as compliance evidence, and which reference method and error rate will they require before an array clears a consignment?