Examining hyperspectral imaging in recycling, waste management

If you’ve shopped for optical sorters in the last decade, “hyperspectral imaging” likely is a term you have encountered in some suppliers’ literature and on their websites.

Hyperspectral imaging promises to provide material recovery facility (MRF) operators and other recyclers with a fuller picture of the materials on their conveyor belts, enabling the efficient, flexible and high-accuracy sorting of recyclables.

How hyperspectral imaging works

Hyperspectral cameras measure and analyze the spectrum of light reflected by or transmitted through materials using near-infrared (NIR) spectroscopy. This analysis provides the chemical composition of the material, enabling the optical sorter to differentiate among the materials reliably.

The software paired with these cameras helps to provide flexibility in sorting as new algorithms can be introduced as they become available to address new or modified packaging in the recycling stream.

Hyperspectral cameras are an alternative to the multispectral cameras commonly used in optical sorters, and they provide a richer dataset of spectral information for each pixel in the image, improving sorting accuracy.

Tim Barry, market manager, waste, at Steinert US Inc. in Walton, Kentucky, explains that three types of vision sensing technology can be used in optical sorters: traditional (or multispectral) NIR, hyperspectral NIR and color, or RGB, cameras (which can be used with artificial intelligence, or AI, to make sorting decisions).

“Your color camera is going to give you data like the human eye sees,” he says, which includes the object’s shape and color. “And that’s about the extent of the data you get from RGB or a color camera.”

Standard multispectral NIR cameras can result in a 14-to-18-point line graph.

“You connect those dots with a line and it gives you a wave. That wave is significant to proving what type of polymer it is,” Barry explains as an example.

Hyperspectral imaging, however, can result in a 256-point wave, he says of Steinert’s technology, which was introduced a decade ago.

“Now, we can take those waves made of 256 points and start comparing and contrasting all the little intermediate peaks and valleys to see if there are trends or indicators,” Barry says. “[T]he intermediate peaks and valleys with all the extra data can tell a clamshell package from a standard PET [polyethylene terephthalate] bottle. The intermediate data can tell you a Grade A wood versus B-grade wood. It starts to get into some of the finer detail, and that finer detail is more and more of what the industry is asking for these days.”

Dirk Balthasar, Ph.D., vice president and head of core R&D at Tomra Recycling, which is based in Germany, says that company began using hyperspectral imaging in 2002.

Tomra employs a whisk-broom hyperspectral imaging system, which has two core elements: a point spectrometer to capture high-fidelity spectral data from a single sample point and a scanning element to sweep that point across the conveyor belt in a perpendicular motion.

“As the conveyor moves, successive lines stack up into a 2D hyperspectral image: Each pixel holds a full spectrum,” Balthasar says. “Most systems illuminate the entire scan line with high-power halogen bulb arrays, which consume a lot of energy and generate excessive heat. At Tomra, we instead use the same scanning element to project a narrow ‘flying beam’ of light precisely onto the detection spot, introduced in 2012. By illuminating only the point being measured, we achieve much higher energy efficiency.”

He says Tomra’s flying-beam approach also offers continuous self-calibration using scanning optics.

What hyperspectral imaging can do

Machinex introduced hyperspectral technology to its optical sorters in 2012.

David Marcouiller, executive vice president of sales engineering, at Machinex, based in Plessisville, Quebec, says using multispectral technology, vendors predetermine the specific wavelengths needed to identify a PET bottle, for example. But with hyperspectral technology, the entire spectrum of an object is captured. This enables Machinex to add “triggers” to target specific wavelengths during sorting.

“Hyperspectral imaging is giving you a much more precise view … instead of acting on predetermined wavelengths,” Marcouiller says.

This specificity enables more detailed sorting and offers flexibility.

In the case of C&D wood, for example, he says hyperspectral imaging can identify chemical additives in the wood by a change in the spectral curve, triggering the air ejection nozzles.

Bulk Handling Systems (BHS), based in Eugene, Oregon, introduced hyperspectral imaging to its NRT optical sorters in 2023. BHS Chief Technology Officer Thomas Brooks, who is based in Nashville, Tennessee, says hyperspectral imaging allows for hypersegmentation of the full spectrum of an object.

“You’re just breaking that up into a lot of little, tiny slices all the way across,” he says. “That’s really powerful from a couple of different perspectives.”

The more granular information hyperspectral imaging provides, he adds, can be important when sorting packaging that is comprised of a base material with various labels and barriers added.

John Green, the president of Green Machine, headquartered in Whitney Point, New York, says his company began using hyperspectral imaging in 2008. The technology combines NIR spectroscopy with high-resolution imaging to create a 3D image for identifying materials with multiple planes by chemical composition.

Green adds that while AI can be used to sort some items more efficiently, such as food-grade plastic packaging from nonfood-grade plastic packaging, “there’s a lot of sortation which only the hyperspectral or full spectral NIR analysis can identify” when the chemical composition of an item is required.

“In other words, you get two items that look the same, what makes them different? That’s where the hyperspectral comes in,” he says.

“Without that 3D capability, you’re going to have more mishits from our standpoint in that you’re looking for flat plains to pull information. But, if you’ve got a shielded plane, which is at a 30-degree angle to the camera, that’s going to reflect the light differently than another flat plane, where you’re at 90 degrees to the camera.”

While hyperspectral imaging can stand on its own, it also can be combined with color cameras and AI, Steinert’s Barry says.

“If you just need your good PETs or your good films or things like that, then we can get away with just the hyperspectral imaging for NIR,” he says. “If you need to start getting AI involved, though, that’s when we pair our hyperspectral imaging with a color camera. This combination brings together two different ways of ‘seeing,’ like using two senses at once. The hyperspectral sensor provides the material’s chemical fingerprint, while the RGB camera adds visual cues, such as color and shape. Through sensor fusion, we can merge these data streams and provide the AI with a much more complete basis for decision-making, enabling more precise, application-specific sorting.”

The software tie-in

Green Machine pairs its hyperspectral imaging technology with what Green says is advanced AI software to collect and make sense of the captured data and positively or negatively classify material.

“The camera literally draws a picture of that item as it is crossing over an air jet bar where you have individually actuated air jets that are on half-inch centers which fire in the outline of that product and eject it off the belt without touching any other product on the belt,” he says.

Green adds that the software package “makes sense of the reflective light” and improves the classification of objects.

Brooks says one of the benefits of hyperspectral imaging is that “all your processing is typically done in software, so now it’s very adaptable.

“What that does is it allows really quick and adaptable changes to what somebody wants to do,” he says. “It really sets you up where you’re semi future-proof. It also then gives you the ability for some of these more difficult sorts that you’re looking at,” whether that be dirty material or multilayer packaging.

“Hyperspectral is a pretty big hammer to solve that problem.”

Marcouiller says the software is the brains behind hyperspectral sorting, and the data generated by these machines are essential to a MRF’s efficient functioning.

“That data from what’s happening in the MRF, it’s very important for the end users, for the operators, to fine-tune their machines [and] to give reports to municipalities. Data is more and more at the center of operations,” he says.

In addition to its adaptability, Brooks says another positive associated with hyperspectral imaging can include lifetime calibration, as in the case of NRT.

Green says the depth of the range offered by hyperspectral imaging allows for finer tuning of the products sorted.

“We can pick every individual grade of plastics [No. 1-7], we can pick maple from oak if we cared to,” he says. “The important distinction when sorting wood is our ability to identify pressure-treated, arsenic-laden wood from clean wood. We can pick different grades of paper, whether it’s newsprint, whether it’s office paper, whether it’s cardboard.”

But to sort food-grade PET thermoforms from PET bottles, Green Machine adds an AI-trained camera to its Green Eye optical sorters that use the company’s patented hyperspectral technology.

“This is a perfect case where spectral analysis is insufficient in identifying shapes required to classify them,” Green continues.

“I don’t think we’ve even touched the surface of the capabilities of what this technology will end up showing us,” Brooks adds of hyperspectral imaging.

He mentions chemical recycling. “Some of these companies that are really concerned about very specific chemical compositions…, hyperspectral will be one of the major technologies to be able to discern all those different things.”

The author is editorial director of the Recycling Today Media Group and can be reached at dtoto@gie.net.