Industries / Food, farming and forestry / Tobacco
Leaf processing
Turning cured leaf into graded, cleaned, dried and packed raw material, and preparing it as filler for factories: inspection and grading, sale, threshing the leaf from its stems, redrying, aging, and cutting and treating the shreds.
6
Milestones achieved, from 1730 to 2005.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1730
Inspection before export
Virginia's General Assembly, urged by Lieutenant Governor Sir William Gooch, passed the act popularly known as the Tobacco Inspection Act. From 1 August 1731 all tobacco exported from the colony had to pass through a public warehouse, where inspectors opened every hogshead (a large barrel), weighed and stamped leaf they judged sound and free of trash, sand and dirt, and had "bad, unsound, or unmerchantable tobacco" burned in their presence. Inspectors issued notes for accepted tobacco that the act declared current in all tobacco payments, so the notes served as money.
Credit
Virginia General Assembly; William Gooch
1858
Selling leaf in open piles
At Neal's Warehouse in Danville, Virginia, the "Danville System" of selling tobacco began. Before, tobacco had been sold by a sample drawn from a packed hogshead; under the new system it was auctioned in open, loose piles so buyers could examine the whole lot. Virginia's 1969 state marker says the system was then "in general use".
Credit
Neal's Warehouse, Danville
1935
Federal grades at auction
The US Tobacco Inspection Act of 23 August 1935 let the Secretary of Agriculture designate tobacco auction markets, but only where two-thirds of growers voting in a referendum agreed. Once a market was designated, no tobacco could be offered for sale there until a federal inspector had inspected and certified it against official grade standards.
Credit
US Congress; US Department of Agriculture
1974
Puffing cut leaf with CO2
Roger de la Burde and Patrick Aument of Philip Morris filed a patent in 1974 for soaking cut tobacco in liquid carbon dioxide and then heating it, so the shreds expand and the same weight fills more cigarette; in one of their examples, corrected filling power rose from 38 to 71 cubic centimeters per 10 g. Philip Morris's internal product reports, analyzed by K. Michael Cummings and colleagues in 2023, show expanded tobacco at 3% of Marlboro's blend in 1975 and 12% by 1990, while reconstituted sheet, a paper made from stems and scraps, rose from 16% in 1965 to 22%, and the tobacco in each Marlboro fell from 0.88 g in 1960 to 0.75 g in 1990.
LessonProcessing changed the product as much as the leaf did: by 1990 about a third of Marlboro's filler was expanded or reconstituted tobacco, by the reports' figures.
Credit
Philip Morris (Roger de la Burde, Patrick Aument)
2004
Quotas and auctions end
The US Fair and Equitable Tobacco Reform Act of 2004 ended the federal quota and price-support program for tobacco, in a buyout of $10.14 billion paid for by assessments on manufacturers, $9.6 billion of it going to growers and quota holders over ten years. At a 2005 Virginia Tech conference, USDA's Agricultural Marketing Service reported that mandatory grading had ended, inspection was now voluntary, and its tobacco staff had fallen from 169 to 46; the flue-cured growers' cooperative went from 800,000 members to 3,800, and a University of Kentucky economist estimated that half of Kentucky's growers quit in 2005. Buyers now set prices in contracts rather than at public auction.
Credit
US Congress; US Department of Agriculture
2005
Zimbabwe turns to contracts
Zimbabwe introduced contract farming for tobacco in 2005: a leaf buyer such as British American Tobacco supplies seed, fertilizer and other inputs and buys the farmer's whole crop. By 2023, 95% of Zimbabwe's crop was grown under contract, by Xinhua's figure as reported by Dialogue Earth, and only the 5% of farmers who finance themselves sell on the auction floors.
Credit
Government of Zimbabwe; Tobacco Industry and Marketing Board; leaf merchants
Watch for
AI grading at Indian auctions
The Tobacco Board of India has partnered with Constems AI to grade flue-cured tobacco bales from photographs at a few auction platforms in Andhra Pradesh. A vision model scores color, texture and visible leaf condition and sends a grade to the classifier's phone. The Board says classification is the only manual step left in its electronic auctions.
Achieved whena party outside both publishes how often the system's grades match the Board's classifiers across a full season's bales.
Claimed by
Tobacco Board of India and Constems AI, August 2026
Open
01
A grader agreement audit
A blind regrading test that measures how often trained human classifiers agree on the grade of the same leaf, published with the photos, so that grading machines can be judged against a known human range.
The briefCredit · unclaimed02
From studio to buying station
A public test set of phone photos of cured leaf taken under buying-station light, graded by classifiers, to measure how much accuracy grading models trained on studio photos lose in the field.
The briefCredit · unclaimed
Open 01 · the brief
A grader agreement audit
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A blind regrading test that measures how often trained human classifiers agree on the grade of the same leaf, published with the photos, so that grading machines can be judged against a known human range.
Why it is still open
The grade sets the farmer's price, and grading is still done by eye; the Tobacco Board of India calls it the only manual operation in its auctions. Grading systems are scored against labels from one source: Altria's 2018 hyperspectral "AutoGrader" was compared with its own company grader, and a 2024 Tanzanian dataset of 49,779 leaf images in 22 grades was labeled by domain experts without a published measure of how often experts agree. A search found no published study of grader-to-grader agreement.
Built with what exists
A phone camera and a light box; a spreadsheet or a free labeling tool such as Label Studio; Fleiss' kappa, a standard measure of agreement among several raters; the Tanzanian dataset and open models in PyTorch to test machine graders against the human range.
First step
Draft the protocol and ask the Tanzanian dataset's authors, Nguleni and colleagues, whose dataset is hosted by the Nelson Mandela African Institution of Science and Technology, whether their partner classifiers would regrade a sample.
Done looks like
At least 300 hands (bundles of leaf) each graded blind by at least three classifiers, 50 of them regraded a week later by the same people, and a published agreement score per grade group, followed by a test of whether a machine grader falls inside or outside that range.
Open 02 · the brief
From studio to buying station
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of phone photos of cured leaf taken under buying-station light, graded by classifiers, to measure how much accuracy grading models trained on studio photos lose in the field.
Why it is still open
The largest open dataset, from Tanzania, was shot with a Canon 5D Mark III and a 100 mm macro lens; India's field deployment works from bale photos on a phone but has published no accuracy figures. A 2019 Philippine study photographed air-cured leaf in a grading room (Marzan and Ruiz), and several hobby grading projects on GitHub, including one for Malawi, train on photos, but no published test was found of how a model trained in one setting does in another.
Built with what exists
Two or three mid-range phones; the Tanzanian studio dataset for training; open image models in PyTorch (for example EfficientNet or ConvNeXt); a color reference card in each photo.
First step
Train a baseline model on the Tanzanian dataset this month, then photograph 100 graded leaves under ordinary indoor light with the same grade labels and report the drop.
Done looks like
At least 2,000 graded phone photos released openly, and a table of accuracy for at least three models trained on the studio set, before and after fine-tuning on 10% of the phone photos.