BRENNEN

Industries / Food, farming and forestry / Food and beverage processing / Rice milling

Sorting and grading

Sorting removes discolored grains, seeds and stones from milled rice by machine, and grading measures what is left, such as whole and broken grains, chalky grains and moisture, so that the rice can be priced.

6

Milestones achieved, from 1947 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1947

    A machine that sees color

    In May 1947 the Balint brothers, owners of Gunson Seeds in Britain, set up a firm to replace the hand-picking of seeds. Its first machine, the G1 or Gunson's Sortex Electronic Separator, looked at each item and pushed aside discolored ones, at 10 kg an hour; Bühler, which now owns Sortex, says its sorters now handle 10,000 kg an hour.

    world-grain.com

    Credit

    Dr. Okolicsanyi and Herbert Fraenkel, for the Balint brothers' firm (now Bühler Sortex)

  2. 1979

    Sorters come to rice mills

    In late-1970s Japan, weeds spread in lightly tended paddies, and their seeds and pest-stained grains got into the harvest. Satake began work on an optical sorter for rice mills in 1978 and released it in 1979: a 10-channel machine, each channel a lane of falling grains watched by a light sensor, that grew to 80 channels by 1986. After Japan's poor 1994 harvest forced imports, exporters abroad bought sorters to meet Japanese buyers' standards, by the same Satake account.

    LessonA European sorter tried in Japanese rice mills in the early 1960s did not catch on, because it cost a lot and buyers did not yet demand rice free of discolored grains.

    gfmt.blogspot.com

    Credit

    Satake Corporation

  3. 1995

    One world standard for rice

    The Codex Alimentarius Commission adopted its Standard for Rice, CXS 198-1995, for husked, milled and parboiled rice sold for eating. It sets a maximum moisture of 15% by mass and defines grading terms by appearance: a head rice kernel is at least three-quarters of the average length of a whole kernel, and a chalky kernel has at least three-quarters of its surface "opaque and floury". It names ISO 7301 as the method for counting defects.

    fao.org

    Credit

    Codex Alimentarius Commission (FAO and WHO)

  4. 2004

    DNA tests catch fake basmati

    Basmati is a group of aromatic long-grain varieties from the Himalayan foothills of India and Pakistan, sold at a premium, and its grains are hard to tell from cheaper long-grain rice by eye. In a survey of nearly 300 packs, reported in July 2004, the UK Food Standards Agency found that nearly one in five held more than 20% non-basmati rice, and in one in ten the share reached 60%. Its guidance of February 2004 limited the name to 11 Indian and five Pakistani varieties, and laboratories offered DNA screening to detect mixtures.

    LessonGrain shape cannot prove a variety; a premium name needs a genetic test.

    foodnavigator.com

    Credit

    UK Food Standards Agency; Reading Scientific Services Ltd (DNA screening)

  5. 2024

    Machine graders at the depot

    The Food Corporation of India (FCI), which buys rice from mills for India's public food stocks, installed automatic grain analyzers that use artificial intelligence (AI) at all its rice acceptance depots in Punjab, to cut officials' room for discretion. On 15 February 2024 it told staff to check rice by hand again whenever it came close to the limits for damaged, discolored and sprouted grains and moisture, after evaluations showed the machines' learning needed samples from more locations in the state. Punjab's rice millers' association objected.

    LessonA grader trained on samples from elsewhere needed local data, and borderline lots went back to people.

    tribuneindia.com

    Credit

    Food Corporation of India

  6. 2026

    US approves an image grader

    On 27 April 2026 the US Federal Grain Inspection Service (FGIS) approved Cgrain's Value analyzers for official grading of whole kernels in medium-grain rough rice from California and nearby areas. The instrument feeds kernels one at a time past a camera and mirrors, and a neural network, a model trained on labeled examples, judges each kernel. FGIS began the evaluation in 2023 to replace the Foss GrainCheck 2312 and found the results equivalent to its approved methods; the machine's sorting option is not approved.

    ams.usda.gov

    Credit

    Cgrain AB and TAS Scientific; FGIS

Watch for

  1. Phone grading for African millers

    UNIDO and the Africa Rice Center (AfricaRice) ran two public contests on Zindi. In the first, from 24 December 2025 to 1 February 2026, teams trained models to predict 15 quality measures from photos of paddy, brown and white rice taken under controlled conditions, including counts of broken, chalky, red, yellow and green grains, grain length and width, and color, scored against values from "industry-standard rice analysis equipment". In the second, from 18 February to 1 March 2026, teams built phone apps that run the winning models without a network connection and warn when a photo is too dark or blurred.

    Achieved whenan app is in use at mills and an outside party publishes how its counts compare with a laboratory grader on samples photographed outside the controlled setup.

    zindi.worldzindi.world

    Claimed by

    UNIDO and AfricaRice, February and March 2026

  2. AI paddy grading in Gujarat

    The Gujarat State Civil Supplies Corporation used AI grain analyzers at 10 paddy buying centers in four districts in the 2025–26 kharif (monsoon crop) marketing season. The machines checked foreign matter, damaged and discolored grains, immature and shrunken grains, admixture (other varieties mixed in) and moisture. Of 6,191 samples, 54.6% were graded A, 45.4% were graded Common and 341 were rejected, with rejection rates from none at Viramgam to 24% at Dholka.

    Achieved whenthe state publishes, or an outside party audits, a comparison of machine and manual grades on split samples, and the analyzers run a full season without a return to manual grading.

    newkerala.com

    Claimed by

    Gujarat State Civil Supplies Corporation, September 2026

Open

  1. 01

    Good rice in sorter rejects

    A public tally of how much sound rice color sorters throw away. Collect samples of the reject stream from working sorters, count sound and defective grains by hand under the Codex definitions, photograph each sample, and publish the counts with each sorter's make, model and settings, together with an open model that estimates the sound share from a photo.

    The brief
    Credit · unclaimed
  2. 02

    One sample, many graders

    A ring test for milled rice from Asian and African markets, in which identical portions of the same rice go to many graders. Split 30 well-mixed lots into identical sealed portions, send them to human graders, laboratory image analyzers and phone apps, and publish how far their counts of broken, chalky and discolored grains disagree.

    The brief
    Credit · unclaimed

Open 01 · the brief

Good rice in sorter rejects

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public tally of how much sound rice color sorters throw away. Collect samples of the reject stream from working sorters, count sound and defective grains by hand under the Codex definitions, photograph each sample, and publish the counts with each sorter's make, model and settings, together with an open model that estimates the sound share from a photo.

Why it is still open

A sorter trades catching defects against losing good grain, and what is said about the loss comes from the makers. Bühler says ejection spikes in a typical optical sorter "often result in the loss of good product" and gives no percentage. Kubota calls the problem carryover, good grains blown out with defective ones, and says farms took to feeding the rejects back through the machine, again with no figure. Web and Crossref searches found no independent field measurement of sound grain in rice sorter rejects.

foodprocessing.com.aukubota.com

Built with what exists

A riffle divider, a box of chutes that splits a sample into equal parts; a phone camera and a light box; open image models such as Meta's Segment Anything to outline grains and Ultralytics YOLO to classify them.

  1. First step

    Ask one mill for three 200 g samples of its sorter rejects, sort them by hand into sound and defective grains, and weigh each pile.

  2. Done looks like

    Reject samples from 20 sorters at 10 or more mills, each counted by hand twice by different people, published with the percentage of sound grain by weight, and a model whose estimates match the hand counts within 2 percentage points on samples it was not trained on.

Open 02 · the brief

One sample, many graders

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A ring test for milled rice from Asian and African markets, in which identical portions of the same rice go to many graders. Split 30 well-mixed lots into identical sealed portions, send them to human graders, laboratory image analyzers and phone apps, and publish how far their counts of broken, chalky and discolored grains disagree.

Why it is still open

The grading terms are defined by appearance, and the one comparison of many methods found is from the United States. Anna McClung and colleagues (Cereal Chemistry, 2020) graded chalk in 20 US varieties by three visual methods, used by mills, an exporter and government inspectors, and by three imaging systems; the methods differed significantly but agreed on the best and worst varieties. An earlier paper from the same group (2014) found that mills "differ greatly in how they judge rice samples". The new graders have each been checked against their own reference: FGIS approved one image analyzer in 2026, only for whole kernels in one type of rough rice; PhilRice said in April 2025 that its phone app for grain size and shape is 98% accurate, without naming the reference; the UNIDO and AfricaRice photos were taken under controlled conditions; and in Punjab in 2024 the FCI sent borderline lots back to hand checks. No comparison of AI analyzers, phone apps and human graders on the same samples of broken and discolored grain was found.

repository.lsu.eduars.usda.govams.usda.govphilrice.gov.ph

Built with what exists

A riffle divider and sealed sample bags; phone apps built on the UNIDO and AfricaRice winning models or similar; Segment Anything and YOLO for an open baseline; volunteer graders at mills, traders and university laboratories.

  1. First step

    Split one 5 kg bag of shop rice into 30 sealed portions, and have two people each grade the same three portions for broken and chalky grains, to see how far two careful people disagree.

  2. Done looks like

    30 lots, each graded by at least 8 graders or instruments, with the spread of each measure (for example, the range and standard deviation of percent broken) published with the raw counts.