Industry · 1 Nov 2026

How AI cameras detect trash in cane trucks

AI cameras learn to tell cane from tops, leaves and other material in images, and research systems convert what they see on a conveyor or spread sample into an estimated trash share. Estimating the trash percentage of a whole loaded truck from a gate camera is much harder, so a camera can support sampling and cannot replace it.

Why trash matters

Trash, or extraneous matter, is everything delivered with the cane that is not millable stalk: tops, leaves, roots and soil. The mill pays by weight, so trash can cost the mill in price and in lost recovery. Deductions to cover it cost the farmer, which makes them a source of dispute.

The numbers vary by source and harvest method. A 2001 survey of Indian factories reported extraneous matter of 3 to 10 percent in cane supplies. A study at three Haryana mills in 2019 found that mixed produce from mechanical harvesters had 0.71 units lower sugar recovery than manually harvested cane.

Deductions are contested. In October 2026, Union Minister Pralhad Joshi asked Karnataka to act on complaints of deductions of about eight percent on mechanically harvested cane. A camera would be attractive if it gave a fair and consistent figure, so it is worth asking how close the technology is.

How the technology works

The third step is the hard one. A camera sees area and a scale measures mass. Light leaves take up a lot of the picture and little of the weight, while wet soil does the opposite.

A camera takes images of the cane under controlled lighting.
A model trained on labelled images marks each region as cane, top, leaf or other material.
Software converts those regions into an estimated mass, using assumptions about density or shape.
The estimate is compared with a reference: the same sample separated by hand and weighed.

What research shows so far

Note where these cameras sit. These results come from cane that is spread out, mixed in view or on a conveyor, where the material is visible. The 2026 paper states the problem directly: dense stacking cuts machine vision accuracy.

Most of this work covers machine-harvested, chopped cane. Results may not transfer to whole-stalk loads on Indian tractors and trucks.

Work	What was imaged	Reported result	Limit
South China Agricultural University, 2023	Samples from machine-harvested loads, spread by hand on a belt	6.5% average relative error on impurity ratio, 70 samples	One variety, one mill; soil and sand not covered
IEEE Access, 2024	Mixtures of cane segments, tops and leaves, with a binocular camera	4.58% total impurity rate error on 830 test images	Abstract read only; no full truck load result reported there
Agriculture, 2026	Piled cane spread out first by a conveyor and baffle device	94.90% average detection accuracy	Needs a device to undo stacking; we read the abstract only
QUT and Tully Mill, Australia	Billets on the receival conveyor, from an overhead camera	We found no published accuracy figure	Project listed to run to February 2027

Why a truck at the gate is harder

The camera sees only the top and sides of the load.
Trash can sit low in the load, out of sight.
Lighting, dust, rain and night shifts all change the image.
Vehicle types and load shapes vary widely at Indian gates.
Soil is hard to measure from an image. One study on shredded cane instead tested near-infrared sensing, which is another route worth watching.

What our own trial taught us

We trialled a camera system at a real gate. It classified whether a truck was carrying cane or not. It did not measure trash percentage, which is a far harder task. Even so, the trial gave four useful lessons.

Fixed objects get mistaken for vehicles unless the system compares each frame against the empty scene.
Detection has to be trained on the site's own footage. Web photos are not enough.
A head-on view hides the load, so camera position decides what can be known.
Accuracy has to be checked against manual records. A demo video proves little.

How to judge a camera system

Ask whether it classifies loads or estimates a trash percentage
Ask what the reference method was, and how many samples were used
Run it beside your existing sampling for a period, across varieties and shifts
Test it at night, in rain and with each vehicle type
Keep the image and the estimate for each load, so disputes can be reviewed
Do not link any deduction to a camera figure until it has been validated at your mill
Treat it as a flag that tells you which loads to sample

Frequently asked questions

Can a camera measure trash percentage in a loaded truck?

We found no published proof of that at the accuracy needed for payment. The research with stronger results images spread samples or cane on a conveyor.

Can a camera replace trash sampling?

No. Current research still checks its estimates against manual separation and weighing. An Australian project description also says manual sampling is the only current way to measure these values, and that is the problem it aims to solve.

Where is the best place to put a camera?

It depends on the question. Overhead or side views show more of a load than a head-on view, and a conveyor view shows the most cane. Test on your own site.

Is a camera useful if it cannot measure trash?

It may be. It can record each load, count vehicles and flag unusual loads for sampling, provided that it has been trained and checked on your own gate.

Sources

Talk to SugarOS